Skin cancer represents one of the most common and clinically significant malignancies worldwide, with melanoma being particularly aggressive due to its high metastatic potential and strong dependence on early diagnosis for favorable treatment outcomes.

Personalized cancer treatment planning is a complex clinical decision-making process that requires the optimal selection of therapeutic strategies while balancing tumor control, patientspecific anatomical constraints, treatment toxicity, and long-term survival outcomes. Radiotherapy remains a major treatment modality for many cancer types; however, designing high-quality radiation treatment plans is often time-consuming and highly dependent on expert clinical judgment.

Sepsis is a life-threatening clinical condition that requires rapid recognition and timely intervention, particularly in Emergency Department (ED) settings where patients often present with heterogeneous symptoms, incomplete medical histories, and rapidly changing physiological status. Delayed diagnosis of sepsis may lead to septic shock, multi-organ failure, prolonged hospitalization, and increased mortality

The rapid growth of unstructured clinical text within Electronic Health Records (EHRs), discharge summaries, medication notes, progress reports, and spontaneous drug safety reports has created a valuable but underutilized source of information for pharmacovigilance and clinical decision support. Drug Adverse Events (ADEs) represent a major challenge in healthcare systems, as they may lead to prolonged hospitalization, increased medical costs, treatment discontinuation, and serious patient safety risks. 

Medical image segmentation plays a fundamental role in computer-aided diagnosis, treatment planning, surgical navigation, radiotherapy guidance, and longitudinal disease monitoring. However, developing robust deep learning segmentation models typically requires access to large and diverse medical imaging datasets collected from multiple healthcare institutions. In real-world clinical environments, direct centralization of medical images is often restricted due to patient privacy regulations, institutional data-governance policies, ethical constraints, and concerns related to data ownership.

Cardiovascular disease (CVD) remains one of the most critical public health challenges worldwide, requiring early identification of high-risk patients to reduce adverse outcomes such as myocardial infarction, stroke, heart failure, and cardiovascular-related mortality. Traditional risk prediction tools, including rule-based scoring systems, provide valuable clinical guidance but often have limited ability to capture complex nonlinear relationships among heterogeneous patient variables stored in Electronic Health Records (EHRs). 

Breast cancer remains the most prevalent malignancy among women globally, accounting for over 2.3 million new diagnoses annually, making early and accurate diagnosis a decisive factor in reducing mortality and improving therapeutic outcomes. Machine learning-based diagnostic systems, particularly Support Vector Machines (SVMs), have demonstrated considerable promise in automated tumor classification; however, their performance is critically hindered by two persistent challenges

Brain tumors represent one of the most aggressive and life-threatening neurological malignancies, with accurate and timely classification being paramount for determining appropriate treatment strategies and improving patient prognosis

Accurate prediction of mortality rates among Intensive Care Unit (ICU) patients is a critical challenge in modern clinical medicine, as timely risk stratification enables physicians to prioritize interventions, allocate limited resources effectively, and improve patient survival outcomes. Despite significant advances in machine learning-based prognostic models, existing approaches often suffer from limited generalization, sensitivity to missing clinical data, and insufficient exploitation of temporal dependencies inherent in physiological time-series records. 

Diabetic retinopathy (DR) remains one of the leading causes of preventable blindness worldwide, with its prevalence continuing to rise in parallel with the global diabetes epidemic. Early and accurate detection is critical to preventing irreversible vision loss; however, manual screening by ophthalmologists is time-consuming, costly, and subject to inter-observer variability. This paper presents a deep learning-based framework for the automated early detection and grading of diabetic retinopathy from fundus retinal images, leveraging the complementary strengths of Convolutional Neural Networks (CNNs) and Transfer Learning.

Domain-specific research literature in renewable energy, medical sciences, and finance contains extensive specialized knowledge, including technical terminology, methodological descriptions, experimental findings, clinical indicators, financial risk factors, performance metrics, and evidence-based conclusions. However, extracting accurate and context-aware information from such literature remains challenging due to domain-specific vocabulary, semantic ambiguity, long document structures, rapidly expanding publication volumes, and the limitations of general-purpose natural language processing models in understanding specialized scientific content. This paper proposes a large language model fine-tuning framework for domain-specific knowledge extraction across renewable energy, medical, and financial research literature. The proposed framework adapts pre-trained transformer-based language models to specialized corpora using supervised fine-tuning, instruction tuning, and parameter-efficient learning strategies such as Low-Rank Adaptation and adapter-based fine-tuning. The system supports multiple knowledge extraction tasks, including named entity recognition, relation extraction, keyphrase extraction, evidence summarization, methodology classification, citation-aware retrieval, and domain-specific question answering. Large-scale corpora were constructed from peer-reviewed journal articles, conference papers, clinical studies, renewable energy forecasting reports, financial risk assessment literature, and benchmark scientific datasets. A comprehensive preprocessing pipeline was applied, including document parsing, abstract and section segmentation, metadata extraction, duplicate removal, terminology normalization, annotation refinement, and train-validation-test splitting to ensure robust evaluation. Experimental results demonstrate that the proposed fine-tuned language model achieves superior performance compared with general-purpose language models, traditional NLP pipelines, BiLSTM-CRF models, BERT-based architectures, SciBERT, BioBERT, and finance-oriented transformer baselines across multiple extraction tasks. The proposed framework records higher precision, recall, F1-score, exact-match accuracy, semantic similarity, and factual consistency, while reducing hallucinated outputs and improving the reliability of extracted information in technical contexts. Furthermore, explainability and traceability mechanisms were incorporated through attention analysis, confidence scoring, evidence-span highlighting, and citation-grounded response generation, allowing users to verify extracted knowledge against the original literature. The findings indicate that fine-tuned large language models provide a scalable, accurate, and adaptable solution for domain-specific knowledge extraction, offering practical value for researchers, clinicians, renewable energy analysts, financial experts, systematic reviewers, and intelligent literature-mining platforms operating across complex scientific domains.

Top of Form

Multi-domain regression and classification problems are characterized by heterogeneous data structures, nonlinear input-output relationships, uncertainty, noise, missing values, and domain-specific feature interactions, making it difficult for conventional statistical and machine learning models to achieve robust and generalizable performance across different application areas. Adaptive Neuro-Fuzzy Inference Systems offer a powerful hybrid modeling approach by combining the learning capability of neural networks with the linguistic interpretability of fuzzy logic; however, their effectiveness is highly dependent on the appropriate selection of membership functions, fuzzy rules, premise parameters, and consequent parameters. This paper proposes an Adaptive Neuro-Fuzzy Inference System optimized by the Cuckoo Search Algorithm for multi-domain regression and classification problems. The proposed framework employs the global search capability of Cuckoo Search to automatically optimize ANFIS structure and parameters, including membership function type, number of fuzzy rules, rule weights, learning coefficients, and parameter initialization, thereby reducing the limitations of manual configuration and improving convergence toward high-quality solutions. The framework was evaluated across diverse domains, including medical diagnosis, energy forecasting, environmental monitoring, financial risk prediction, industrial fault detection, and benchmark regression datasets. A comprehensive preprocessing pipeline was applied, including missing value imputation, outlier treatment, normalization, feature selection, categorical encoding, class imbalance handling, and train-validation-test splitting to ensure reliable model assessment. For regression tasks, the optimized ANFIS model was assessed using RMSE, MAE, MAPE, and coefficient of determination, while classification tasks were evaluated using accuracy, sensitivity, specificity, precision, recall, F1-score, and AUC-ROC. Experimental results demonstrate that the proposed CS-ANFIS framework achieves superior predictive performance compared with conventional ANFIS, standalone fuzzy inference systems, artificial neural networks, Support Vector Machine, Random Forest, Gradient Boosting, Genetic Algorithm-optimized ANFIS, and Particle Swarm Optimization-based ANFIS baselines. The optimized framework records improved prediction accuracy, lower error rates, faster convergence, and stronger generalization across heterogeneous datasets. Furthermore, the fuzzy rule base generated by the proposed model provides interpretable decision patterns that help explain the relationship between input features and predicted outcomes, supporting transparency in high-impact domains. The findings indicate that Cuckoo Search-optimized ANFIS provides an adaptive, interpretable, and computationally efficient solution for multi-domain regression and classification, offering practical value for intelligent decision-support systems that require both predictive accuracy and human-understandable reasoning.

Smart industrial systems increasingly depend on heterogeneous Internet of Things sensor streams to monitor equipment health, production stability, energy consumption, environmental conditions, and operational safety in real time. However, anomaly detection in such environments remains challenging due to the high dimensionality of sensor data, nonlinear temporal dependencies, irregular sampling rates, noise, missing values, sensor drift, and the coexistence of multiple operating regimes across industrial processes. Traditional statistical monitoring methods and conventional machine learning approaches often fail to capture complex cross-sensor interactions and long-range temporal patterns, particularly when abnormal events are rare, diverse, and difficult to label. This paper proposes an attention Transformer-based model for anomaly detection across heterogeneous IoT sensor streams in smart industrial systems. The proposed framework leverages self-attention mechanisms, temporal positional encoding, multi-head attention layers, and feed-forward representation learning to model both short-term fluctuations and long-term dependencies among multivariate sensor signals. The system integrates data from diverse industrial sensors, including vibration, temperature, pressure, humidity, current, voltage, acoustic emission, flow rate, machine speed, and environmental monitoring devices. A comprehensive preprocessing pipeline was applied, including missing value imputation, noise filtering, temporal synchronization, normalization, sliding-window segmentation, feature embedding, and imbalance-aware training to improve robustness under real-world industrial conditions. The model identifies abnormal behavior by learning normal operating patterns and detecting deviations through reconstruction error, prediction residuals, attention-weight shifts, and anomaly scoring mechanisms. Experimental results demonstrate that the proposed attention Transformer-based framework achieves superior anomaly detection performance compared with traditional control charts, Isolation Forest, One-Class SVM, Random Forest, Autoencoder, LSTM, GRU, and Temporal Convolutional Network baselines. The proposed model records higher accuracy, precision, recall, F1-score, AUC-ROC, and lower false alarm rates across multiple industrial monitoring scenarios. Furthermore, attention-based interpretability analysis was incorporated to identify the most influential sensors and time intervals contributing to detected anomalies, enabling maintenance engineers to localize potential faults and understand abnormal system behavior more effectively. The findings indicate that attention Transformer models provide a scalable, adaptive, and interpretable solution for anomaly detection in heterogeneous IoT sensor streams, offering practical value for predictive maintenance, industrial safety, energy efficiency, fault diagnosis, and intelligent decision support in smart manufacturing and Industry 4.0 environments.

Top of Form

Explainable artificial intelligence has become essential for improving transparency, trust, accountability, and regulatory acceptance of predictive models deployed in high-impact domains such as healthcare, energy, and finance. Although advanced machine learning and deep learning models often achieve strong predictive performance, their complex decision mechanisms can limit practical adoption when users require clear justification, risk awareness, and domain-specific interpretability. This paper presents a benchmark study for the comparative analysis of explainable artificial intelligence methods across healthcare, energy, and financial prediction tasks. The proposed study evaluates multiple predictive models, including Random Forest, XGBoost, LightGBM, Support Vector Machine, Convolutional Neural Networks, Long Short-Term Memory networks, and Transformer-based architectures, combined with widely used explainability techniques such as SHAP, LIME, Grad-CAM, Integrated Gradients, attention visualization, permutation importance, and counterfactual explanations. The healthcare tasks include disease diagnosis, medical image classification, and clinical risk prediction using imaging, tabular, and time-series patient data. The energy tasks focus on electricity load forecasting, renewable energy output prediction, and equipment fault detection using meteorological, sensor-based, and operational variables. The financial tasks include credit default prediction, fraud detection, market volatility forecasting, and systemic risk identification using transaction records, market indicators, borrower profiles, and macroeconomic features. A unified experimental protocol was designed, including standardized preprocessing, feature engineering, model training, hyperparameter optimization, cross-validation, and domain-specific evaluation metrics to ensure fair comparison across heterogeneous datasets. Predictive performance was assessed using accuracy, sensitivity, specificity, F1-score, AUC-ROC, RMSE, MAE, and MAPE, while explanation quality was evaluated using fidelity, stability, sparsity, consistency, computational efficiency, and domain expert alignment. Experimental results demonstrate that no single explainability technique consistently dominates across all domains and data modalities; instead, explanation effectiveness depends strongly on task type, model architecture, input representation, and stakeholder requirements. SHAP provides strong global and local interpretability for structured healthcare, energy, and financial data, Grad-CAM offers clinically meaningful visual explanations for medical imaging tasks, and attention-based methods provide useful temporal insights for forecasting and risk-monitoring applications. Furthermore, the benchmark highlights trade-offs between predictive accuracy and explanation complexity, showing that highly accurate models may require complementary explanation methods to support reliable interpretation. The findings provide practical guidance for selecting appropriate explainable artificial intelligence techniques across critical prediction tasks, supporting safer, more transparent, and more trustworthy deployment of intelligent decision-support systems in healthcare, energy management, and financial risk analysis.

Water quality monitoring, energy output forecasting, and financial volatility prediction represent three critical but heterogeneous time-dependent prediction problems that support environmental sustainability, energy system reliability, and financial risk management. Although these domains differ in data characteristics and operational objectives, they share common modeling challenges, including nonlinear dynamics, temporal dependencies, noise, missing observations, external shocks, uncertainty, and the need for accurate multi-output decision support. This paper proposes a multi-task learning architecture for the simultaneous prediction of water quality, energy output, and financial volatility indices within a unified deep learning framework. The proposed architecture employs shared representation layers to capture common temporal and nonlinear patterns across domains, while task-specific prediction branches are designed to learn domain-dependent features associated with environmental indicators, renewable energy generation, and financial market risk. The framework integrates Long Short-Term Memory networks, Gated Recurrent Units, Temporal Convolutional Networks, attention mechanisms, and Transformer-based encoders to model short-term fluctuations, long-range dependencies, and cross-task relationships among heterogeneous time-series inputs. The water quality task incorporates features such as pH, dissolved oxygen, turbidity, temperature, conductivity, nitrate concentration, biochemical oxygen demand, and chemical oxygen demand. The energy forecasting task utilizes solar irradiance, wind speed, temperature, humidity, historical power generation, load demand, and weather-related variables. The financial volatility task includes market returns, trading volume, volatility indices, moving averages, interest rates, macroeconomic indicators, and risk sentiment measures. A comprehensive preprocessing pipeline was applied, including missing value imputation, outlier treatment, normalization, temporal alignment, sliding-window construction, feature engineering, and chronological train-validation-test splitting to prevent data leakage. Experimental results demonstrate that the proposed multi-task learning framework achieves superior predictive performance compared with single-task learning models and conventional machine learning baselines across all three domains, recording lower RMSE, MAE, and MAPE values while improving prediction stability under noisy and volatile conditions. Furthermore, attention-based interpretability and task-specific feature importance analysis reveal the dominant influence of physicochemical parameters on water quality, meteorological variables on energy output, and lagged volatility and market stress indicators on financial risk prediction. The findings indicate that multi-task learning provides a scalable, adaptive, and data-efficient approach for simultaneous cross-domain forecasting, offering practical value for environmental agencies, energy operators, financial analysts, policymakers, and integrated decision-support platforms.

Top of Form

Bottom of Form

Resource-constrained edge computing environments require intelligent models that can operate under strict limitations related to memory capacity, computational power, energy consumption, communication bandwidth, and real-time response requirements. Although deep neural networks have achieved strong performance in many artificial intelligence applications, their training and deployment at the edge remain challenging due to high parameter complexity, slow convergence, large optimization search spaces, and sensitivity to hyperparameter configuration. This paper proposes a quantum-inspired optimization framework for training deep neural networks in resource-constrained edge computing environments. The proposed framework integrates quantum-inspired metaheuristic algorithms with lightweight neural architectures to enhance convergence efficiency, reduce computational overhead, and improve model performance under limited hardware resources. Quantum-inspired optimization mechanisms, including quantum-behaved Particle Swarm Optimization, quantum genetic search, quantum-inspired evolutionary operators, and probability amplitude-based solution representation, are employed to optimize critical neural network parameters and hyperparameters such as learning rate, weight initialization, pruning thresholds, dropout rates, batch size, number of hidden units, and layer configurations. The framework was evaluated using edge-oriented learning tasks, including image classification, anomaly detection, sensor-based activity recognition, and Internet of Things predictive monitoring, using datasets collected from embedded devices, smart sensors, wearable systems, and low-power edge nodes. A comprehensive preprocessing and compression pipeline was applied, including feature normalization, data augmentation, quantization-aware training, model pruning, knowledge distillation, and lightweight architecture selection to improve deployability on constrained platforms. Experimental results demonstrate that the proposed quantum-inspired optimization framework achieves superior performance compared with conventional training strategies, including stochastic gradient descent, Adam, RMSprop, classical genetic algorithms, Particle Swarm Optimization, and standard evolutionary optimization baselines. The optimized models achieve higher accuracy, faster convergence, reduced training loss, lower memory footprint, decreased inference latency, and improved energy efficiency across multiple edge computing scenarios. Furthermore, computational complexity analysis confirms that the proposed framework maintains strong scalability while reducing unnecessary parameter updates and resource-intensive search operations. The findings indicate that quantum-inspired optimization provides an effective, adaptive, and resource-aware solution for training deep neural networks at the edge, offering practical value for intelligent IoT systems, autonomous monitoring platforms, wearable healthcare devices, smart manufacturing applications, and real-time cyber-physical systems operating under limited computational resources.

Smart city infrastructure management requires continuous monitoring, intelligent decision-making, and proactive maintenance across highly interconnected urban systems, including transportation networks, energy grids, water distribution systems, waste management, public safety services, environmental monitoring, and building operations. The rapid expansion of Internet of Things technologies has enabled large-scale collection of real-time urban data from sensors, cameras, meters, vehicles, mobile devices, and cyber-physical infrastructure; however, transforming these heterogeneous, high-volume, and dynamic data streams into reliable predictive insights remains a major challenge. This paper proposes an Internet of Things-enabled predictive analytics framework using deep learning for integrated smart city infrastructure management. The proposed framework combines distributed IoT sensing, edge-cloud data processing, and advanced deep learning models, including Convolutional Neural Networks, Long Short-Term Memory networks, Gated Recurrent Units, Temporal Convolutional Networks, Graph Neural Networks, and Transformer-based architectures, to predict infrastructure conditions, detect anomalies, and support early intervention across multiple urban domains. The system analyzes multimodal data sources such as traffic flow, energy consumption, air quality, weather conditions, water pressure, equipment health indicators, public transport activity, surveillance streams, and citizen service requests. A comprehensive preprocessing pipeline was applied, including data cleaning, missing value imputation, noise filtering, temporal synchronization, sensor fusion, normalization, spatial-temporal feature extraction, and real-time stream segmentation. The proposed deep learning framework enables multi-task predictive analytics, including traffic congestion forecasting, energy demand prediction, water leakage detection, air pollution estimation, equipment failure prediction, and infrastructure risk assessment. Experimental results demonstrate that the proposed framework achieves superior predictive performance compared with traditional statistical models, rule-based monitoring systems, and standalone machine learning approaches, recording higher accuracy, precision, recall, F1-score, and lower RMSE, MAE, and MAPE across diverse smart city applications. Furthermore, the integration of edge computing reduces latency and communication overhead, while explainable AI techniques provide interpretable insights into the key sensor variables and spatial-temporal patterns influencing infrastructure risks. The findings indicate that IoT-enabled deep learning provides a scalable, adaptive, and intelligent solution for integrated smart city infrastructure management, enabling municipal authorities, urban planners, utility providers, and emergency response agencies to improve operational efficiency, reduce maintenance costs, enhance sustainability, and increase urban resilience.

Cross-domain transfer learning has become an important research direction for developing intelligent systems that can generalize across heterogeneous data environments, particularly in domains where labeled data are scarce, expensive, imbalanced, or highly specialized. Medical imaging, energy forecasting, and financial risk assessment represent three critical application areas with distinct data structures, including visual diagnostic images, temporal energy demand patterns, and multivariate financial risk indicators. Despite their differences, these domains share common modeling challenges, such as distribution shift, limited labeled samples, nonlinear feature interactions, noise, uncertainty, and the need for interpretable decision support. This paper proposes a unified deep learning framework for cross-domain transfer learning across medical imaging, energy forecasting, and financial risk assessment. The proposed framework combines convolutional neural networks, recurrent neural networks, transformer-based architectures, autoencoders, and attention mechanisms within a modular architecture capable of extracting transferable representations from image, time-series, and tabular data. Domain adaptation layers, shared feature encoders, task-specific prediction heads, and fine-tuning strategies are incorporated to enable knowledge transfer from source domains with abundant data to target domains with limited labeled observations. The framework was evaluated using representative datasets from retinal disease detection and brain tumor classification, short-term electricity load and renewable energy forecasting, and credit default and systemic financial risk prediction. A comprehensive preprocessing pipeline was applied for each domain, including image enhancement and augmentation for medical imaging, normalization and sliding-window construction for energy forecasting, and missing value treatment, feature engineering, and imbalance handling for financial risk datasets. Experimental results demonstrate that the proposed unified framework improves predictive performance compared with domain-specific standalone models and conventional transfer learning baselines across all evaluated tasks. The framework achieves higher classification accuracy, sensitivity, specificity, F1-score, AUC-ROC, and lower forecasting errors measured by RMSE, MAE, and MAPE, while requiring fewer labeled target-domain samples. Furthermore, attention visualization, Grad-CAM, SHAP analysis, and latent feature interpretation were integrated to enhance model transparency and provide domain-relevant explanations for clinical, energy, and financial decision-makers. The findings indicate that unified cross-domain transfer learning can provide a scalable, adaptive, and data-efficient solution for complex real-world prediction problems, supporting reliable decision-making in healthcare diagnostics, smart energy management, and financial risk monitoring.

Bankruptcy prediction in emerging market economies is a complex and high-stakes financial classification problem due to unstable macroeconomic conditions, limited disclosure quality, market inefficiencies, currency fluctuations, liquidity constraints, and the nonlinear interaction between firm-level financial indicators and external economic shocks. Traditional statistical models and standalone machine learning approaches often struggle to capture these complex relationships, particularly when bankruptcy cases are rare, datasets are imbalanced, and firm behavior varies significantly across sectors and economic cycles. This paper proposes a Harris Hawks Optimization-tuned Extreme Gradient Boosting framework for accurate and interpretable bankruptcy prediction in emerging market economies. The proposed framework integrates the strong nonlinear learning capability of XGBoost with the global search efficiency of Harris Hawks Optimization to automatically tune key hyperparameters, including learning rate, maximum tree depth, number of estimators, subsampling ratio, column sampling ratio, regularization parameters, minimum child weight, and gamma. The framework was evaluated using firm-level financial datasets from emerging markets, incorporating accounting ratios, profitability indicators, liquidity measures, leverage ratios, solvency metrics, cash-flow variables, firm size, market-based indicators, sectoral attributes, and macroeconomic variables such as inflation, interest rates, exchange rate volatility, GDP growth, and credit conditions. A comprehensive preprocessing pipeline was applied, including missing value imputation, outlier treatment, winsorization of extreme financial ratios, feature normalization, categorical encoding, correlation analysis, and imbalance-aware learning using SMOTE and cost-sensitive loss weighting. Experimental results demonstrate that the proposed HHO-XGBoost framework achieves superior predictive performance compared with traditional bankruptcy models, including Altman Z-score, logistic regression, discriminant analysis, Support Vector Machine, Random Forest, standalone XGBoost, PSO-XGBoost, and GWO-XGBoost baselines. The optimized model records higher accuracy, sensitivity, specificity, F1-score, AUC-ROC, and precision-recall AUC, while significantly improving the detection of financially distressed firms and reducing false negatives. Furthermore, SHAP-based explainability analysis was incorporated to identify the most influential bankruptcy risk factors, revealing the dominant role of leverage intensity, declining profitability, weak liquidity, negative operating cash flow, high debt servicing burden, reduced asset turnover, and adverse macroeconomic conditions. The proposed framework provides a robust, scalable, and transparent decision-support tool for banks, investors, auditors, regulators, and policymakers, enabling earlier identification of corporate distress, improved credit risk assessment, and stronger financial stability monitoring in emerging market environments.

Top of Form

Candlestick chart analysis is widely used in financial markets to interpret price behavior, identify trend continuation or reversal patterns, and generate trading signals; however, traditional technical analysis often depends on manually defined rules, subjective visual interpretation, and fixed pattern templates that may fail under noisy, volatile, and rapidly changing market conditions. This paper proposes a Convolutional Neural Network-based pattern recognition framework for automated technical analysis and trading signal generation using candlestick chart representations. The proposed framework transforms historical open, high, low, close, and volume data into image-based candlestick sequences, enabling deep convolutional models to learn discriminative visual patterns associated with future price movements. Advanced CNN architectures, including VGGNet, ResNet, EfficientNet, and custom lightweight CNN models, are employed to extract spatial features from candlestick formations, trend structures, support and resistance zones, volatility patterns, and price momentum signals. The framework was evaluated using historical financial data from multiple asset classes, including equities, foreign exchange pairs, commodities, and cryptocurrencies, across different time horizons and market regimes. A comprehensive preprocessing pipeline was applied, including OHLCV data normalization, sliding-window chart generation, image resizing, noise reduction, class labeling based on future returns, and chronological train-validation-test splitting to avoid look-ahead bias. To improve robustness and reduce overfitting, data augmentation, dropout regularization, batch normalization, transfer learning, and class-weighted loss functions were incorporated during model training. Experimental results demonstrate that the proposed CNN-based framework achieves superior pattern recognition and trading signal prediction performance compared with rule-based candlestick strategies, traditional technical indicators, Support Vector Machine, Random Forest, and recurrent neural network baselines. The proposed model records higher classification accuracy, precision, recall, F1-score, AUC-ROC, and directional prediction performance, while generating more reliable buy, sell, and hold signals under volatile market conditions. Furthermore, Grad-CAM visualization was integrated to highlight the most influential chart regions contributing to each prediction, enhancing interpretability by revealing how the model focuses on relevant candlestick structures and price formations. The findings indicate that CNN-based candlestick pattern recognition provides an adaptive, scalable, and interpretable approach for automated technical analysis, offering practical value for algorithmic trading systems, financial analysts, quantitative traders, and decision-support platforms operating in complex and dynamic financial markets.

Decentralized portfolio management under stochastic market conditions is a complex sequential decision-making problem characterized by uncertain asset returns, dynamic correlations, transaction costs, liquidity constraints, volatility clustering, and rapidly changing risk preferences. Traditional portfolio optimization models often depend on centralized decision structures, fixed assumptions about return distributions, and static risk-return trade-offs, which limit their ability to adapt to heterogeneous investor objectives and unstable market regimes. This paper proposes a multi-agent deep reinforcement learning framework for decentralized portfolio management in stochastic financial environments. The proposed framework models each trading agent as an autonomous decision-maker responsible for learning adaptive asset allocation, rebalancing, and risk-control policies through continuous interaction with a simulated market environment. Multiple deep reinforcement learning algorithms, including Multi-Agent Deep Deterministic Policy Gradient, Multi-Agent Proximal Policy Optimization, Independent Deep Q-Networks, and actor-critic-based architectures, are employed to coordinate decentralized trading decisions while preserving agent-level autonomy. The state space incorporates historical asset prices, log returns, trading volume, volatility indicators, moving averages, momentum signals, correlation measures, macro-financial variables, and portfolio-specific risk exposures, while the action space represents dynamic portfolio weights, buy-sell decisions, and rebalancing strategies across multiple asset classes. A comprehensive preprocessing pipeline was applied, including missing value treatment, feature normalization, rolling-window state construction, covariance estimation, transaction cost modeling, and chronological train-validation-test splitting to prevent look-ahead bias. The reward function was designed to optimize long-term risk-adjusted performance by integrating portfolio return, Sharpe ratio, downside risk, maximum drawdown, turnover penalty, and diversification constraints. Experimental results demonstrate that the proposed multi-agent framework achieves superior performance compared with traditional benchmarks, including buy-and-hold, equal-weighted allocation, mean-variance optimization, risk parity, momentum-based strategies, and single-agent reinforcement learning models. The proposed approach records higher cumulative returns, improved Sharpe and Sortino ratios, lower maximum drawdown, reduced portfolio volatility, and stronger resilience during stochastic market shocks. Furthermore, inter-agent coordination analysis reveals that decentralized agents can learn complementary investment behaviors, adaptive diversification patterns, and regime-sensitive allocation strategies under uncertain market dynamics. The findings indicate that multi-agent deep reinforcement learning provides a scalable, adaptive, and robust solution for decentralized portfolio optimization, offering practical value for quantitative asset management, automated trading systems, robo-advisory platforms, and financial institutions operating in volatile and uncertain investment environments.

Top of Form

Financial time series are highly complex, nonlinear, and nonstationary, frequently exhibiting abnormal movements, volatility clustering, structural breaks, and sudden regime transitions caused by market shocks, liquidity stress, macroeconomic announcements, geopolitical events, and investor behavioral changes. Accurate anomaly detection and regime change identification are therefore essential for risk management, algorithmic trading, portfolio protection, and early-warning financial surveillance systems. However, traditional statistical models and conventional machine learning methods often struggle to capture hidden temporal dependencies, latent market structures, and rare abnormal patterns in noisy financial data. This paper proposes a hybrid deep learning framework that combines Variational Autoencoders with Long Short-Term Memory networks for anomaly detection and regime change identification in financial time series. The proposed model leverages the Variational Autoencoder to learn compact probabilistic latent representations of normal market behavior, while the LSTM component captures sequential dependencies and temporal evolution across historical observations. The framework was evaluated using multivariate financial datasets including stock indices, exchange rates, commodities, volatility indices, trading volume, technical indicators, returns, realized volatility, moving averages, momentum indicators, and liquidity-related features. A comprehensive preprocessing pipeline was applied, including missing value imputation, normalization, log-return transformation, rolling-window segmentation, noise reduction, and chronological train-validation-test splitting to avoid look-ahead bias. Anomalies were identified based on reconstruction error, latent-space deviation, and temporal prediction residuals, while regime changes were detected by monitoring shifts in latent distributions and volatility-sensitive sequence patterns over time. Experimental results demonstrate that the proposed VAE-LSTM framework achieves superior anomaly detection performance compared with baseline models including ARIMA, GARCH, Isolation Forest, One-Class SVM, standard Autoencoder, standalone LSTM, and GRU-based approaches. The proposed framework records higher precision, recall, F1-score, AUC-ROC, and detection stability, while reducing false alarms during normal high-volatility periods. Furthermore, latent-space visualization and reconstruction-based explainability were incorporated to distinguish between transient anomalies and persistent regime transitions, enabling clearer interpretation of market stress conditions. The findings indicate that the proposed hybrid VAE-LSTM model provides an adaptive, robust, and interpretable solution for identifying abnormal financial behavior and structural regime shifts, offering practical value for financial institutions, quantitative traders, portfolio managers, and risk-monitoring systems operating in uncertain and rapidly changing market environments.

Top of Form

Digital payment platforms process massive volumes of transactions in real time, creating significant challenges for fraud detection systems due to the high velocity of data streams, evolving fraud strategies, severe class imbalance, and the need for immediate and interpretable decision-making. Traditional rule-based systems and conventional machine learning models often fail to adapt to emerging fraudulent behaviors and may generate excessive false alarms, while complex deep learning models frequently operate as black boxes, limiting their acceptance in regulated financial environments. This paper proposes an explainable artificial intelligence-driven fraud detection system for real-time transaction monitoring in digital payment platforms. The proposed framework integrates advanced machine learning and deep learning models, including XGBoost, LightGBM, Random Forest, Autoencoders, Long Short-Term Memory networks, and Graph Neural Networks, to detect suspicious transaction patterns across heterogeneous payment data. The system analyzes multiple transaction-level and user-level features, including transaction amount, frequency, merchant category, device fingerprint, geolocation, payment channel, transaction time, user behavioral history, account age, velocity indicators, and network relationships among accounts, merchants, and devices. A comprehensive preprocessing pipeline was applied, including missing value handling, categorical encoding, feature normalization, temporal aggregation, anomaly feature construction, and imbalance-aware learning using SMOTE, focal loss, and cost-sensitive classification. To support real-time deployment, the framework incorporates streaming-based inference, low-latency feature extraction, threshold optimization, and adaptive risk scoring, enabling immediate classification of transactions as legitimate, suspicious, or fraudulent. Explainability was achieved using SHAP, LIME, attention-based visualization, and rule extraction techniques to identify the most influential fraud indicators behind each prediction, allowing analysts to understand model decisions and investigate suspicious activities more effectively. Experimental results demonstrate that the proposed explainable fraud detection system achieves superior performance compared with traditional rule-based engines, logistic regression, Support Vector Machine, standalone Random Forest, and conventional neural network baselines. The proposed system records higher accuracy, precision, recall, F1-score, AUC-ROC, and precision-recall AUC, while significantly reducing false positives and improving detection of rare fraudulent transactions. Furthermore, interpretability analysis reveals that abnormal transaction velocity, unusual geolocation, high-risk merchant categories, device switching, atypical transaction amounts, and irregular account behavior are among the most important predictors of fraud. The proposed framework provides a scalable, adaptive, and transparent solution for real-time fraud monitoring, offering practical value for banks, fintech companies, payment gateways, and regulatory bodies seeking to improve transaction security, reduce financial losses, and strengthen trust in digital payment ecosystems.

Credit default risk prediction is a fundamental task in banking institutions, as inaccurate assessment of borrower risk can lead to increased non-performing loans, capital losses, regulatory pressure, and reduced financial stability. Traditional credit scoring approaches often rely on linear assumptions and manually selected variables, which limits their ability to capture complex nonlinear relationships among borrower characteristics, repayment behavior, macroeconomic conditions, and institutional lending policies. This paper proposes a swarm intelligence-enhanced gradient boosting framework for accurate and interpretable credit default risk prediction in banking institutions. The proposed framework integrates powerful gradient boosting models, including XGBoost, LightGBM, and CatBoost, with swarm intelligence optimization algorithms such as Particle Swarm Optimization, Grey Wolf Optimizer, Whale Optimization Algorithm, and Firefly Algorithm to automatically tune critical hyperparameters and improve predictive generalization. The optimization process focuses on selecting optimal learning rates, tree depths, number of estimators, regularization coefficients, subsampling ratios, feature sampling rates, and class-weight parameters, thereby reducing reliance on manual trial-and-error configuration and conventional grid search. The framework was evaluated using large-scale banking and credit datasets containing demographic information, income level, employment status, loan amount, credit history, repayment behavior, debt-to-income ratio, collateral indicators, transaction patterns, and macro-financial variables. A comprehensive preprocessing pipeline was applied, including missing value treatment, outlier handling, categorical encoding, normalization, feature engineering, correlation analysis, and imbalance-aware sampling to address the typically low proportion of default cases in real-world credit portfolios. Experimental results demonstrate that the proposed swarm intelligence-enhanced gradient boosting framework achieves superior predictive performance compared with traditional credit scoring models, standalone gradient boosting algorithms, Support Vector Machine, Random Forest, logistic regression, and neural network baselines. The optimized framework records higher accuracy, sensitivity, specificity, F1-score, AUC-ROC, and precision-recall performance, while reducing false negatives associated with high-risk borrowers. Furthermore, SHAP-based explainability analysis was incorporated to identify the most influential default risk factors, revealing the dominant contribution of repayment history, credit utilization, debt burden, income stability, loan duration, delinquency records, and macroeconomic stress indicators. The proposed framework provides a robust, scalable, and interpretable decision-support solution for banking institutions, enabling more reliable credit approval, proactive risk monitoring, improved capital allocation, and enhanced compliance with data-driven risk management practices.

Financial market crashes and systemic risk events represent highly disruptive phenomena characterized by abrupt price collapses, volatility explosions, liquidity shortages, contagion effects, investor panic, and strong interdependence across financial institutions and asset classes. Early detection of such events is essential for risk management, portfolio protection, regulatory supervision, and the design of timely intervention strategies; however, conventional econometric models and traditional machine learning methods often struggle to capture nonlinear temporal dependencies, cross-market spillovers, and hidden warning signals preceding periods of financial instability. This paper proposes an attention-based Bidirectional Long Short-Term Memory network for the early detection of financial market crashes and systemic risk events. The proposed framework combines the sequential learning capability of BiLSTM networks with an attention mechanism that dynamically assigns higher importance to the most informative time steps and market indicators associated with emerging instability. Historical financial data from multiple markets, including equity indices, bond yields, foreign exchange rates, commodities, volatility indices, trading volume, credit spreads, and macro-financial indicators, were used to construct a comprehensive multivariate time-series dataset. A robust preprocessing pipeline was applied, including missing value imputation, normalization, return transformation, volatility estimation, rolling-window segmentation, feature lagging, and chronological train-validation-test splitting to prevent look-ahead bias. The model was trained to identify early-warning patterns preceding crash periods and systemic risk episodes by learning from both forward and backward temporal dependencies in financial sequences. To address the rarity of crisis events, class-weighted loss functions, focal loss, and oversampling strategies were incorporated during training. Experimental results demonstrate that the proposed attention-based BiLSTM model achieves superior predictive performance compared with baseline models including logistic regression, Random Forest, Support Vector Machine, XGBoost, standard LSTM, GRU, and conventional BiLSTM architectures. The proposed framework records higher accuracy, sensitivity, F1-score, AUC-ROC, and precision-recall performance, while reducing false negatives in crash-event detection. Furthermore, the attention mechanism enhances interpretability by identifying critical pre-crash indicators such as rising volatility, declining liquidity, widening credit spreads, abnormal trading volume, sharp correlation increases, and persistent negative return patterns. The findings indicate that the proposed model provides an adaptive, interpretable, and data-driven early-warning system for detecting financial market crashes and systemic risk events, offering practical value for investors, financial institutions, regulators, and automated risk-monitoring platforms operating in complex and highly interconnected global financial markets.

Tunnel CCTV monitoring plays a critical role in ensuring traffic safety, emergency response, and infrastructure protection, particularly in environments where accidents can rapidly escalate due to limited visibility, confined road geometry, restricted escape routes, smoke accumulation, illumination variation, and delayed human intervention. However, conventional video surveillance systems depend heavily on manual observation or rule-based detection mechanisms, which are often insufficient for identifying unforeseen accidents under challenging real-world conditions such as low light, shadows, occlusion, camera vibration, weather-related artifacts, congestion, sudden vehicle stops, collisions, fire, smoke, and abnormal pedestrian movement. This paper proposes a deep learning-based framework for the automatic detection of unforeseen accidents in CCTV monitoring environments within tunnels. The proposed system integrates spatial feature extraction and temporal motion modeling through advanced deep learning architectures, including Convolutional Neural Networks, 3D CNNs, Long Short-Term Memory networks, ConvLSTM, and Vision Transformer-based video analysis models, to capture both appearance-based abnormalities and dynamic behavioral changes across consecutive video frames. The framework processes tunnel surveillance footage using a comprehensive preprocessing pipeline that includes frame sampling, illumination enhancement, noise reduction, motion stabilization, object detection, region-of-interest extraction, and temporal sequence construction to improve robustness under complex visual conditions. To address the rarity and diversity of accident events, the proposed approach incorporates data augmentation, anomaly-aware learning, class-balanced training, and transfer learning from large-scale video recognition datasets. The model is trained to distinguish normal traffic flow from abnormal events such as vehicle collisions, sudden stops, wrong-way driving, stalled vehicles, smoke emergence, fire incidents, falling objects, and pedestrians entering restricted tunnel areas. Experimental evaluation demonstrates that the proposed deep learning framework achieves high detection accuracy, sensitivity, specificity, F1-score, and AUC performance while maintaining low false alarm rates and real-time inference capability suitable for operational surveillance systems. Furthermore, explainable visual attention maps were integrated to highlight accident-relevant regions within video frames, supporting human operators in rapidly verifying detected incidents and improving trust in automated alerts. The proposed system offers a scalable, intelligent, and reliable decision-support solution for tunnel traffic management centers, enabling earlier accident detection, faster emergency response, reduced secondary collisions, and improved safety in complex underground transportation environments.

Volatile financial markets are characterized by rapid price fluctuations, nonlinear asset interactions, regime shifts, liquidity instability, and high exposure to systemic and unsystematic risks, making portfolio optimization and algorithmic trading highly challenging tasks. Traditional portfolio allocation models, such as mean-variance optimization and rule-based trading strategies, often rely on restrictive assumptions regarding return distributions, market stationarity, and fixed risk preferences, which limits their adaptability under dynamic market conditions. This paper proposes a deep reinforcement learning-based algorithmic trading framework for portfolio optimization in volatile financial markets. The proposed strategy formulates portfolio management as a sequential decision-making problem in which an intelligent trading agent learns optimal asset allocation policies by interacting with a simulated financial environment. Deep reinforcement learning algorithms, including Deep Q-Networks, Deep Deterministic Policy Gradient, Proximal Policy Optimization, and Soft Actor-Critic, are employed to learn adaptive trading decisions involving buy, sell, hold, and portfolio rebalancing actions across multiple risky assets. The framework utilizes historical market data from equities, exchange-traded funds, commodities, and cryptocurrencies, incorporating technical indicators, return-based features, volatility measures, momentum signals, moving averages, relative strength index, MACD, Bollinger Bands, trading volume, and risk-adjusted performance indicators. A comprehensive preprocessing pipeline was applied, including missing value handling, feature normalization, rolling-window state construction, transaction cost modeling, and chronological train-validation-test splitting to prevent data leakage. The reward function was designed to balance profitability and risk by integrating cumulative portfolio return, Sharpe ratio, maximum drawdown, volatility penalty, and transaction cost constraints, enabling the agent to optimize long-term risk-adjusted performance rather than short-term gains alone. Experimental results demonstrate that the proposed deep reinforcement learning strategy achieves superior portfolio performance compared with traditional benchmarks, including buy-and-hold, equal-weighted allocation, mean-variance optimization, momentum-based trading, and conventional machine learning-driven strategies. The proposed framework records higher cumulative returns, improved Sharpe and Sortino ratios, reduced maximum drawdown, and stronger resilience during periods of market turbulence. Furthermore, sensitivity analysis confirms that the model maintains stable performance across different rebalancing frequencies, transaction cost assumptions, and market volatility regimes. The findings indicate that deep reinforcement learning provides an adaptive, data-driven, and scalable approach for algorithmic portfolio optimization, offering practical value for institutional investors, asset managers, quantitative traders, and automated trading systems operating in uncertain and rapidly changing financial environments.

Top of Form

Foreign exchange rate forecasting is a highly complex financial prediction problem due to the nonlinear, nonstationary, and volatile behavior of currency markets, which are strongly influenced by macroeconomic indicators, interest rate differentials, geopolitical events, inflation dynamics, market sentiment, and global liquidity conditions. Conventional econometric models often struggle to capture these complex dependencies, while standalone machine learning models may suffer from limited robustness, overfitting, and sensitivity to manually selected hyperparameters. This paper proposes an optimized ensemble machine learning framework for accurate Forex exchange rate forecasting using Bayesian hyperparameter tuning. The proposed framework integrates multiple predictive learners, including Random Forest, Gradient Boosting, XGBoost, LightGBM, CatBoost, Support Vector Regression, and Extra Trees, within a unified ensemble architecture designed to improve forecasting stability and generalization across different currency pairs and market regimes. Historical Forex data for major currency pairs, such as EUR/USD, GBP/USD, USD/JPY, USD/CHF, AUD/USD, and USD/CAD, were used together with technical indicators including moving averages, exponential moving averages, Bollinger Bands, relative strength index, MACD, stochastic oscillator, average true range, momentum indicators, and lagged return features. A comprehensive preprocessing pipeline was applied, including missing value treatment, outlier detection, normalization, stationarity-aware feature engineering, rolling-window construction, and chronological train-validation-test splitting to prevent look-ahead bias. Bayesian optimization was employed to automatically tune critical model parameters, such as tree depth, learning rate, number of estimators, regularization coefficients, subsampling ratios, kernel parameters, and ensemble weights, thereby reducing the computational cost associated with exhaustive grid search while improving convergence toward optimal model configurations. Experimental results demonstrate that the proposed Bayesian-optimized ensemble framework achieves superior forecasting performance compared with traditional models, including ARIMA, GARCH, standalone SVR, Random Forest, XGBoost, and LSTM baselines, across multiple forecasting horizons. The optimized ensemble records lower RMSE, MAE, and MAPE values, while achieving higher directional accuracy and stronger robustness under volatile market conditions. Furthermore, SHAP-based explainability analysis was incorporated to identify the most influential forecasting variables, revealing that lagged exchange rates, volatility indicators, moving average crossovers, momentum features, and interest rate-related variables play a dominant role in predicting future currency movements. The proposed framework provides a reliable, adaptive, and interpretable decision-support tool for Forex traders, financial analysts, portfolio managers, and algorithmic trading systems, offering improved forecasting accuracy and practical applicability in dynamic global currency markets.

Cryptocurrency markets are characterized by extreme volatility, nonlinear price dynamics, strong temporal dependencies, and sensitivity to market sentiment, liquidity fluctuations, and macroeconomic uncertainty, making accurate multi-step ahead forecasting a highly challenging task. Traditional statistical models and conventional machine learning approaches often fail to capture long-range temporal patterns, regime shifts, and complex interactions among heterogeneous market indicators, while many deep learning models provide deterministic predictions without quantifying the uncertainty associated with future price movements. This paper proposes a Temporal Fusion Transformer (TFT)-based forecasting framework for multi-step ahead cryptocurrency price prediction with integrated uncertainty quantification. The proposed model combines recurrent sequence processing, interpretable attention mechanisms, gated residual networks, and variable selection layers to effectively learn temporal dependencies from both historical and known future inputs. The framework was evaluated using high-frequency and daily market data from major cryptocurrencies, including Bitcoin, Ethereum, and Binance Coin, incorporating multiple predictive features such as open, high, low, close, trading volume, technical indicators, volatility measures, moving averages, relative strength index, MACD, and market momentum signals. A comprehensive preprocessing pipeline was applied, including missing value handling, normalization, feature engineering, sliding-window sequence construction, and train-validation-test splitting based on chronological order to prevent data leakage. Unlike conventional point forecasting models, the proposed TFT framework produces probabilistic forecasts through quantile regression, enabling the estimation of prediction intervals and supporting risk-aware decision-making under market uncertainty. Experimental results demonstrate that the proposed model achieves superior forecasting performance compared with baseline models including ARIMA, SVR, Random Forest, LSTM, GRU, BiLSTM, and standard Transformer architectures across multiple forecasting horizons. The model records lower RMSE, MAE, and MAPE values while maintaining well-calibrated prediction intervals with strong coverage probability. Furthermore, the attention-based interpretability mechanism reveals the relative importance of historical price trends, volatility indicators, trading volume, and momentum-based features in driving future cryptocurrency price movements. The findings indicate that the proposed TFT-based framework provides an accurate, interpretable, and uncertainty-aware solution for cryptocurrency price forecasting, offering practical value for traders, portfolio managers, financial analysts, and automated decision-support systems operating in highly volatile digital asset markets.

Top of Form

Bottom of Form

High-frequency stock market prediction is a challenging financial forecasting task due to the nonlinear, noisy, and rapidly changing nature of intraday price movements. Traditional forecasting models often rely primarily on historical market indicators, such as price, volume, volatility, bid–ask spread, and technical signals, while neglecting the growing influence of investor sentiment expressed through social media platforms. In modern financial markets, social media posts, financial news discussions, online investor communities, and real-time public opinions can affect short-term trading behavior, market liquidity, volatility spikes, and price momentum. However, extracting meaningful predictive information from social media is difficult because of informal language, sarcasm, misinformation, noisy text, temporal misalignment, and the high velocity of data streams. To address these challenges, this paper proposes a hybrid deep learning and sentiment analysis framework for high-frequency stock market prediction using social media data.

The proposed framework integrates market microstructure data with sentiment-enriched textual features to improve short-term forecasting accuracy. Historical intraday stock data, including open, high, low, close prices, trading volume, returns, volatility, moving averages, order book indicators, and technical momentum signals, are combined with social media-derived features extracted from platforms such as financial forums, microblogs, and investor discussion networks. A comprehensive preprocessing pipeline is applied to clean textual data, remove spam and irrelevant posts, normalize financial slang, identify stock-related entities, handle emojis and hashtags, and align sentiment features with high-frequency market intervals. Sentiment analysis is performed using lexicon-based methods, transformer-based language models, and domain-specific financial NLP models to estimate positive, negative, neutral, and intensity-weighted investor sentiment.

The predictive architecture combines deep learning models for both numerical and textual data streams. Convolutional neural networks and transformer encoders are used to extract semantic representations from social media text, while LSTM, GRU, temporal convolutional networks, and attention-based models are employed to capture sequential dependencies in high-frequency market data. The extracted sentiment embeddings and market features are fused using a hybrid feature-level or decision-level fusion strategy to predict short-term stock price direction, return movement, volatility change, or trading signal classification. Attention mechanisms are incorporated to dynamically weight the most relevant time windows, sentiment events, and market indicators, allowing the model to focus on periods where investor sentiment has stronger predictive influence.

The proposed framework can be evaluated using high-frequency stock market datasets combined with timestamped social media streams over multiple trading periods. Model performance is assessed using accuracy, precision, recall, F1-score, AUC-ROC, mean absolute error, root mean square error, directional accuracy, Sharpe ratio, maximum drawdown, and simulated trading profitability. Comparative analysis is conducted against traditional statistical models, technical indicator-based machine learning methods, standalone sentiment models, LSTM-only models, transformer-only models, and non-hybrid forecasting baselines. Experimental evaluation is expected to demonstrate that integrating sentiment signals with deep temporal market modeling improves prediction robustness, particularly during periods of high market uncertainty, news-driven volatility, and abnormal investor attention.

The proposed system provides a scalable and adaptive framework for high-frequency financial forecasting by combining social media sentiment mining with deep learning-based market prediction. By transforming noisy public opinion streams into structured predictive signals, the framework can support algorithmic trading, risk monitoring, volatility forecasting, portfolio adjustment, and real-time decision support. Its ability to jointly model textual sentiment, temporal market behavior, and attention-driven feature interactions makes it a promising approach for next-generation intelligent financial analytics systems.

Electrodialysis Reversal desalination is an important membrane-based separation process for treating brackish water and moderately saline feed streams, particularly because it can periodically reverse electrode polarity to reduce scaling, limit fouling accumulation, and improve long-term operational stability. Despite these advantages, Electrodialysis Reversal systems involve complex electrochemical, hydraulic, and mass-transfer interactions among feed salinity, ion concentration, applied voltage, current density, flow rate, membrane selectivity, stack configuration, temperature, recovery ratio, and reversal frequency. These nonlinear and dynamic relationships make accurate process simulation and operational optimization challenging using conventional empirical equations or simplified mechanistic models. To address these limitations, this paper proposes a generative deep learning model for simulation and optimization of Electrodialysis Reversal desalination processes.

The proposed framework utilizes historical operating data, pilot-scale experiments, and process simulation records to learn the underlying behavior of Electrodialysis Reversal systems under diverse feedwater and operating conditions. Input variables include feed conductivity, total dissolved solids, ionic composition, applied voltage, current efficiency, stack current, concentrate flow rate, diluate flow rate, membrane area, spacer characteristics, temperature, pressure drop, recovery ratio, energy consumption, product water quality, and polarity reversal interval. A comprehensive preprocessing pipeline is applied to clean noisy measurements, handle missing values, normalize process variables, align time-dependent operational records, and construct performance indicators such as salt removal efficiency, specific energy consumption, water recovery, current utilization, membrane scaling tendency, and desalination cost.

The generative deep learning component is designed to simulate realistic Electrodialysis Reversal operating scenarios and support optimization under limited experimental data availability. Advanced architectures such as variational autoencoders, generative adversarial networks, conditional generative models, and diffusion-based models can be employed to generate synthetic but physically plausible process data across different salinity levels, membrane configurations, and operating strategies. These generated scenarios are then used to enrich training datasets, explore untested operating regions, and improve the robustness of predictive models for product water quality, ion removal rate, energy consumption, and membrane performance. To ensure practical reliability, the generative model incorporates process constraints related to ion mass balance, voltage limits, current density boundaries, membrane selectivity, water recovery requirements, and safe operating pressure ranges.

The optimization module integrates the generated process scenarios with data-driven predictive models and multi-objective optimization techniques to identify operating conditions that minimize energy consumption and desalination cost while maximizing salt removal, water recovery, membrane lifetime, and process stability. Decision variables include applied voltage, flow rate, reversal frequency, recovery ratio, stack operation mode, and concentrate recirculation strategy. Model performance is evaluated using simulation accuracy, synthetic data quality, prediction error, salt removal efficiency, specific energy consumption, recovery improvement, constraint violation rate, and optimization convergence behavior. Comparative analysis is conducted against conventional process simulation models, response surface methodology, standard machine learning models, and non-generative optimization approaches.

Experimental evaluation is expected to demonstrate that the proposed generative deep learning framework can accurately reproduce complex Electrodialysis Reversal process behavior while supporting efficient exploration of operating conditions that are costly or time-consuming to test experimentally. By combining synthetic scenario generation with optimization, the framework can improve process understanding, reduce experimental burden, enhance operational decision-making, and identify energy-efficient desalination strategies. Its integration into smart desalination control platforms can support adaptive operation, improved brackish water treatment efficiency, reduced scaling risk, optimized reversal scheduling, and more sustainable freshwater production.

High-pressure desalination membranes are essential components in modern desalination systems, particularly in Reverse Osmosis and nanofiltration processes, where they enable efficient salt removal and freshwater production under elevated operating pressures. However, membrane fouling remains a major operational challenge that can reduce permeate flux, increase differential pressure, decrease salt rejection efficiency, raise specific energy consumption, shorten membrane lifespan, and lead to frequent chemical cleaning or unplanned shutdowns. Early identification of fouling is therefore critical for maintaining stable plant performance, reducing operational costs, and preventing irreversible membrane damage. Conventional fouling monitoring approaches often rely on fixed thresholds, manual inspection, or delayed performance indicators, which may fail to detect subtle early-stage anomalies before significant efficiency losses occur. To address these limitations, this paper proposes a transfer learning-based anomaly detection framework for early fouling identification in high-pressure desalination membranes.

The proposed framework utilizes operational data collected from desalination monitoring systems, including feed pressure, permeate pressure, concentrate pressure, differential pressure, feed flow rate, permeate flow rate, recovery ratio, feed salinity, permeate conductivity, temperature, turbidity, pH, normalized permeate flux, salt rejection rate, and specific energy consumption. A comprehensive preprocessing pipeline is applied to handle missing sensor readings, remove abnormal noise, normalize process variables, align multi-rate time-series measurements, and construct fouling-sensitive indicators such as flux decline rate, pressure increase trend, salt passage variation, and membrane resistance index. Transfer learning is employed by adapting knowledge from pre-trained anomaly detection or time-series representation models trained on related membrane systems, pilot-scale datasets, or historical operating regimes to improve fouling detection performance in target desalination plants with limited labeled fault data.

The anomaly detection model is designed to learn normal membrane operating behavior and identify deviations that may indicate early fouling development. Deep learning architectures such as autoencoders, LSTM autoencoders, temporal convolutional networks, and transformer-based encoders can be used as feature extractors or anomaly scoring models within the transfer learning framework. The transferred model is fine-tuned using plant-specific normal operation data to capture local operating characteristics, membrane type, feedwater quality, pressure range, and seasonal variations. During real-time inference, deviations between predicted and observed membrane behavior are quantified using reconstruction error, prediction residuals, latent-space distance, and adaptive anomaly thresholds. Persistent abnormal patterns are flagged as potential early fouling events before severe performance deterioration becomes visible.

The proposed framework can be evaluated using historical records from high-pressure Reverse Osmosis or nanofiltration desalination systems, pilot-scale fouling experiments, and simulated membrane degradation scenarios. Performance is assessed using anomaly detection metrics such as precision, recall, F1-score, false alarm rate, detection delay, early warning lead time, area under the precision–recall curve, and robustness under variable feedwater conditions. Comparative analysis is conducted against conventional threshold-based monitoring, statistical process control, principal component analysis, isolation forest, one-class support vector machine, standard autoencoder, and non-transfer deep learning baselines. Experimental evaluation is expected to demonstrate that transfer learning improves early fouling detection accuracy, especially when labeled fouling data are scarce or when membrane operating conditions differ across plants.

The proposed system provides a scalable and data-efficient solution for predictive membrane maintenance in high-pressure desalination processes. By transferring learned representations from related membrane datasets and adapting them to site-specific operating conditions, the framework reduces the dependence on large labeled fouling datasets and improves generalization across different desalination plants. Its integration into plant monitoring and supervisory control systems can support early fouling alerts, optimized cleaning schedules, reduced energy consumption, extended membrane lifetime, and more reliable freshwater production.

Large-scale Reverse Osmosis desalination networks are increasingly used to meet rising freshwater demand in water-scarce regions, but their operation is strongly constrained by high energy consumption, variable electricity prices, membrane performance limitations, feedwater quality fluctuations, and the need to maintain reliable water production. Optimizing these systems requires balancing multiple conflicting objectives, particularly minimizing energy consumption and operational cost while maintaining permeate quality, production capacity, membrane safety, and hydraulic stability. Conventional single-objective optimization and rule-based operating strategies may fail to capture the complex trade-offs among pressure settings, recovery ratio, flow allocation, pump scheduling, membrane train operation, storage utilization, and time-dependent energy tariffs. To address these challenges, this paper proposes a Multi-Objective Particle Swarm Optimization framework for analyzing and optimizing the energy–cost trade-off in large-scale Reverse Osmosis desalination networks.

The proposed framework formulates desalination network operation as a constrained multi-objective optimization problem. The objective functions include minimizing specific energy consumption, total operating cost, peak electricity demand, membrane stress, brine disposal cost, and chemical usage, while maximizing freshwater production, salt rejection efficiency, system reliability, and operational flexibility. Decision variables include high-pressure pump operating points, feed flow distribution, membrane train activation, recovery ratio, pressure setpoints, energy recovery device operation, cleaning schedule, storage tank dispatch, and grid electricity purchasing strategy. Operational constraints are incorporated to ensure safe and feasible operation, including pressure limits, membrane flux boundaries, recovery limits, permeate quality requirements, pump capacity, storage capacity, brine concentration limits, and minimum water demand satisfaction.

The Particle Swarm Optimization algorithm is extended into a multi-objective structure to generate a set of Pareto-optimal solutions representing different energy–cost operating strategies. Each particle represents a candidate network operating schedule, and its position is updated based on individual experience, swarm knowledge, constraint penalties, and Pareto dominance criteria. External archive management, crowding distance, adaptive inertia weighting, and mutation operators are incorporated to maintain solution diversity and improve convergence toward the optimal Pareto front. The framework uses historical plant operation data, feedwater salinity profiles, electricity tariff structures, water demand patterns, pump efficiency curves, membrane performance models, and energy recovery system characteristics to simulate realistic large-scale desalination network behavior.

The proposed approach can be evaluated using operational records from large Reverse Osmosis plants, high-fidelity process simulation models, or benchmark desalination network scenarios. Performance is assessed using total energy consumption, specific energy consumption, operational cost, permeate production, salt rejection rate, peak load reduction, Pareto-front quality, convergence speed, constraint violation rate, and computational efficiency. Comparative analysis is conducted against single-objective optimization, rule-based scheduling, Genetic Algorithm, Differential Evolution, Grey Wolf Optimizer, and conventional Particle Swarm Optimization methods. Experimental evaluation is expected to demonstrate that the proposed multi-objective PSO framework provides superior trade-off solutions by reducing energy use and operating cost while maintaining water quality and production reliability.

The resulting Pareto-optimal solutions provide plant operators and decision-makers with flexible operating alternatives that can be selected according to real-time priorities, such as minimizing cost during peak tariff periods, maximizing production during high-demand intervals, or reducing energy intensity under constrained power availability. By explicitly modeling the energy–cost trade-off, the proposed framework supports more informed desalination network management, improves utilization of pumps and energy recovery devices, reduces unnecessary energy expenditure, and enhances long-term operational sustainability. Its integration into desalination supervisory control and energy management systems can contribute to more efficient, economical, and resilient freshwater production in large-scale Reverse Osmosis desalination networks.

Desalination plant performance is influenced by a complex interaction of operational, environmental, and water quality parameters that affect freshwater production, energy consumption, membrane integrity, thermal efficiency, fouling tendency, scaling risk, and overall process reliability. In both membrane-based and thermal desalination systems, variables such as feedwater salinity, temperature, turbidity, pH, operating pressure, flow rate, recovery ratio, pretreatment efficiency, chemical dosage, membrane age, differential pressure, top brine temperature, steam consumption, and cleaning frequency can significantly alter plant efficiency. Conventional performance assessment methods often rely on empirical correlations, fixed operating thresholds, or operator experience, which may be insufficient to capture nonlinear relationships among process variables and may provide limited insight into the root causes of performance decline. To address these limitations, this paper proposes an explainable machine learning framework for identifying key operational parameters affecting desalination plant performance.

The proposed framework utilizes historical and real-time data collected from desalination plant monitoring systems, including process sensors, water quality analyzers, energy meters, maintenance logs, chemical dosing records, and supervisory control systems. A comprehensive preprocessing pipeline is applied to handle missing values, remove abnormal readings, normalize numerical variables, align multi-rate measurements, and construct performance indicators such as specific energy consumption, permeate flux, salt rejection rate, recovery ratio, gain output ratio, production efficiency, fouling index, scaling tendency, and plant availability. Multiple machine learning models, including Random Forest, XGBoost, LightGBM, CatBoost, support vector regression, and artificial neural networks, are developed to predict desalination performance outcomes under varying operating conditions.

To enhance transparency and practical usability, explainable artificial intelligence techniques are incorporated into the modeling framework. Global interpretability methods, including feature importance analysis, SHAP values, permutation importance, and partial dependence analysis, are used to identify the most influential variables affecting plant-level performance. Local explanation methods are employed to explain individual performance deviations, enabling operators to understand why specific operating periods are associated with reduced production, increased energy consumption, lower salt rejection, or elevated fouling risk. These explanations can reveal critical operational drivers such as high feed salinity, increased differential pressure, reduced pretreatment efficiency, elevated turbidity, excessive recovery ratio, membrane aging, steam flow variation, or suboptimal chemical dosing.

The proposed framework can be evaluated using operational datasets from Reverse Osmosis, Multi-Stage Flash, Multi-Effect Distillation, or hybrid desalination plants. Model performance is assessed using regression and classification metrics, including Mean Absolute Error, Root Mean Square Error, coefficient of determination, accuracy, sensitivity, specificity, F1-score, and prediction stability. The explanatory outputs are validated through expert review, operational event analysis, and comparison with established desalination process knowledge. Experimental evaluation is expected to demonstrate that explainable machine learning models can accurately predict plant performance while providing actionable insights into the operational parameters responsible for efficiency losses and reliability issues.

By combining predictive modeling with transparent interpretation, the proposed framework provides a practical decision-support tool for desalination plant optimization. It can assist operators in prioritizing corrective actions, optimizing pressure and recovery settings, improving pretreatment control, adjusting chemical dosage, planning membrane cleaning, reducing energy consumption, and preventing performance degradation. The proposed explainable machine learning approach supports more reliable, efficient, and sustainable desalination operation by transforming complex plant data into interpretable knowledge that can guide real-time monitoring, predictive maintenance, and process optimization.

Top of Form

Bottom of Form

Thermal desalination plants are widely used for large-scale freshwater production, particularly in regions with high seawater salinity and limited natural water resources. However, scaling formation remains one of the most critical operational challenges affecting the efficiency, reliability, and lifespan of thermal desalination systems. Scale deposits caused by calcium carbonate, calcium sulfate, magnesium hydroxide, and other sparingly soluble salts can reduce heat transfer efficiency, increase thermal energy consumption, restrict brine flow, accelerate corrosion, and require frequent shutdowns for cleaning and maintenance. Effective prediction of scaling tendency and accurate chemical dosage optimization are therefore essential for maintaining stable plant operation, minimizing operational cost, and improving desalination performance. To address these challenges, this paper proposes a metaheuristic-optimized Artificial Neural Network framework for scaling prediction and chemical dosage optimization in thermal desalination systems.

The proposed framework utilizes historical and real-time plant data, including seawater temperature, feed salinity, pH, total dissolved solids, calcium concentration, magnesium concentration, sulfate concentration, bicarbonate alkalinity, brine concentration factor, top brine temperature, stage pressure, recovery ratio, flow rate, heat-transfer performance, antiscalant dosage, and cleaning history. A comprehensive preprocessing pipeline is applied to remove abnormal readings, handle missing sensor values, normalize process variables, and construct scaling-related indicators such as saturation index, concentration polarization tendency, brine supersaturation level, and heat-exchanger fouling rate. The Artificial Neural Network is designed to model the nonlinear relationships between feedwater chemistry, thermal operating conditions, chemical dosing, and scaling risk.

To improve prediction accuracy and reduce dependence on manual tuning, metaheuristic optimization algorithms are incorporated to optimize key ANN parameters, including the number of hidden neurons, learning rate, activation functions, initial weights, biases, regularization coefficients, and training configuration. Algorithms such as Particle Swarm Optimization, Grey Wolf Optimizer, Genetic Algorithm, Whale Optimization Algorithm, and Differential Evolution can be employed to identify near-optimal network structures and dosing strategies. The optimized ANN predicts scaling probability, severity level, and expected deposition risk under different operating conditions, while the dosage optimization module recommends appropriate antiscalant dosing levels that minimize chemical consumption without compromising scale control.

The proposed model can be evaluated using operational records from thermal desalination plants, pilot-scale experiments, and process simulation data. Performance is assessed using classification and regression metrics, including accuracy, sensitivity, specificity, F1-score, Mean Absolute Error, Root Mean Square Error, coefficient of determination, false alarm rate, and early warning lead time. Chemical optimization performance is evaluated through reduction in antiscalant consumption, improvement in heat-transfer efficiency, decrease in cleaning frequency, reduction in specific energy consumption, and maintenance of safe scaling indices. Comparative analysis is conducted against conventional saturation-index methods, rule-based dosing strategies, standard artificial neural networks, support vector regression, random forest, and non-optimized machine learning models.

Experimental evaluation is expected to demonstrate that the metaheuristic-optimized ANN provides more accurate scaling prediction and more efficient chemical dosage recommendations than conventional approaches. By capturing complex nonlinear interactions among seawater chemistry, temperature, brine concentration, and plant operating conditions, the proposed framework can support proactive scale management and prevent excessive or insufficient chemical dosing. Its integration into thermal desalination monitoring and control systems can improve operational stability, reduce chemical cost, enhance heat-transfer performance, extend equipment lifetime, and contribute to more energy-efficient and sustainable freshwater production.

Coastal desalination intake systems are highly sensitive to variations in seawater quality, as changes in physical, chemical, and biological parameters can directly affect pretreatment efficiency, membrane performance, energy consumption, fouling potential, and overall desalination plant reliability. Parameters such as temperature, salinity, turbidity, pH, dissolved oxygen, conductivity, chlorophyll-a concentration, total suspended solids, organic matter, and algal bloom indicators often fluctuate due to tidal cycles, seasonal changes, rainfall events, coastal currents, industrial discharge, and biological activity. These dynamic variations make accurate water quality forecasting essential for proactive plant operation, chemical dosing optimization, intake management, and early warning of harmful conditions. To address these challenges, this paper proposes a Long Short-Term Memory network for forecasting water quality parameters in coastal desalination intake systems.

The proposed framework utilizes historical time-series measurements collected from coastal intake monitoring stations, online water quality sensors, meteorological records, and oceanographic observations. Input variables include seawater temperature, salinity, turbidity, pH, conductivity, dissolved oxygen, oxidation-reduction potential, chlorophyll-a, total suspended solids, tide level, wave height, rainfall, wind speed, and previous intake water quality trends. A comprehensive preprocessing pipeline is applied to handle missing sensor readings, remove outliers, smooth noisy measurements, normalize variables, align multi-source data streams, and construct sliding-window sequences suitable for temporal forecasting. The LSTM architecture is employed to capture both short-term fluctuations and long-term dependencies in water quality behavior, enabling the model to learn recurring temporal patterns associated with tidal movement, seasonal variation, storm events, and biological activity.

The proposed model is designed to generate short-term and multi-step forecasts for key intake water quality parameters that influence desalination performance. Forecasting accuracy is evaluated using standard regression metrics such as Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, coefficient of determination, and forecasting skill score. Comparative analysis is conducted against persistence models, ARIMA, support vector regression, random forest, gradient boosting models, standard recurrent neural networks, and gated recurrent unit networks. Experimental evaluation is expected to demonstrate that the LSTM-based model provides improved forecasting accuracy by effectively modeling nonlinear temporal dependencies and delayed environmental effects in coastal water quality dynamics.

The forecasting outputs can be integrated into desalination plant decision-support systems to optimize pretreatment operation, adjust chemical dosing, schedule membrane cleaning, manage intake depth or location, and issue early warnings for high turbidity events, salinity shocks, organic loading, or algal bloom risks. By providing timely and reliable predictions of intake water quality, the proposed LSTM framework can reduce operational uncertainty, minimize membrane fouling, improve energy efficiency, enhance process stability, and support sustainable operation of coastal desalination facilities.

Reverse Osmosis desalination is widely used for freshwater production because of its high salt removal efficiency and comparatively lower energy demand than many thermal desalination processes. However, maintaining optimal performance remains challenging due to the complex interactions among feedwater salinity, operating pressure, membrane properties, recovery ratio, temperature, fouling behavior, pretreatment efficiency, and flow conditions. Two of the most important performance indicators in Reverse Osmosis systems are salt rejection rate and energy consumption, as they directly influence permeate water quality, operational cost, membrane lifespan, and overall plant sustainability. Conventional empirical models and rule-based operational strategies often have limited ability to capture nonlinear process behavior under variable feedwater and operating conditions. To address these limitations, this paper proposes an optimized Random Forest model for predicting salt rejection rate and energy consumption in Reverse Osmosis desalination systems.

The proposed framework utilizes historical and real-time operational data collected from Reverse Osmosis plants, including feedwater conductivity, total dissolved solids, temperature, pH, turbidity, feed pressure, permeate pressure, concentrate pressure, feed flow rate, permeate flow rate, recovery ratio, membrane age, cleaning history, and specific energy consumption records. A comprehensive preprocessing pipeline is applied to remove abnormal readings, handle missing sensor values, normalize process variables, and construct relevant performance indicators. Feature selection and correlation analysis are incorporated to identify the most influential variables affecting salt rejection and energy demand. The Random Forest model is then optimized using systematic hyperparameter tuning to determine the optimal number of trees, maximum tree depth, minimum samples per split, feature selection strategy, and bootstrap configuration, thereby improving prediction accuracy and generalization.

The proposed model is designed to predict both salt rejection rate and energy consumption under diverse operating scenarios, including changes in feed salinity, pressure adjustment, temperature variation, membrane fouling progression, and recovery ratio modification. Model performance is evaluated using regression metrics such as Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, coefficient of determination, and prediction stability across different operational periods. Comparative analysis is conducted against conventional regression models, support vector regression, artificial neural networks, gradient boosting models, and non-optimized Random Forest baselines. Experimental evaluation is expected to demonstrate that the optimized Random Forest model achieves accurate and robust prediction of desalination performance while maintaining strong interpretability and computational efficiency.

The proposed framework can support desalination plant operators by providing reliable performance forecasting, identifying energy-intensive operating conditions, and assisting in operational decision-making. Furthermore, feature importance analysis is incorporated to reveal the dominant factors influencing salt rejection and energy consumption, such as feed pressure, feed salinity, recovery ratio, membrane condition, temperature, and permeate flow rate. By enabling early identification of inefficient operating regimes and supporting data-driven process optimization, the optimized Random Forest model can contribute to improved water quality, reduced energy consumption, lower operational cost, and enhanced sustainability of Reverse Osmosis desalination plants.

Solar-powered seawater desalination systems offer a sustainable solution for freshwater production in water-scarce regions by combining renewable energy generation with desalination technologies. However, their operation is strongly affected by the intermittent nature of solar energy, variable seawater quality, fluctuating water demand, changing weather conditions, energy storage limitations, and the nonlinear behavior of desalination processes. Conventional rule-based and static optimization strategies often fail to adapt effectively to these dynamic operating conditions, leading to increased energy consumption, unstable water production, reduced system efficiency, and suboptimal utilization of available solar power. To address these challenges, this paper proposes a Deep Reinforcement Learning framework for dynamic operational optimization of solar-powered seawater desalination systems.

The proposed framework formulates desalination plant operation as a sequential decision-making problem, in which an intelligent agent continuously interacts with the desalination-energy environment to learn optimal control policies. The system state includes solar irradiance, photovoltaic power output, battery state of charge, seawater temperature, feed salinity, water demand, permeate flow rate, recovery ratio, membrane pressure, specific energy consumption, storage tank level, and historical operating conditions. Based on these observations, the reinforcement learning agent selects adaptive operational actions such as adjusting desalination production rate, controlling pump operation, regulating membrane pressure, managing battery charging and discharging, scheduling water storage, and allocating solar energy between direct plant operation and energy storage. The reward function is designed to maximize freshwater production and solar energy utilization while minimizing energy consumption, operating cost, membrane stress, battery degradation, water shortage, and renewable energy curtailment.

Advanced deep reinforcement learning algorithms, including Deep Q-Network, Deep Deterministic Policy Gradient, Proximal Policy Optimization, and Soft Actor–Critic, are investigated to support both discrete and continuous control strategies. A comprehensive simulation environment is developed using historical solar radiation data, seawater quality profiles, desalination process models, battery storage dynamics, and water demand patterns. The framework incorporates operational constraints related to membrane pressure limits, recovery ratio boundaries, battery capacity, pump efficiency, water storage capacity, and minimum freshwater demand. Performance is evaluated using total freshwater production, specific energy consumption, solar fraction, operating cost, renewable curtailment rate, battery cycling cost, water shortage probability, and system reliability.

Experimental evaluation is expected to demonstrate that the proposed deep reinforcement learning framework outperforms conventional rule-based control, model predictive control, heuristic optimization, and static scheduling methods by learning adaptive policies that respond effectively to rapid changes in solar availability and desalination demand. The framework can improve energy efficiency, stabilize water production, reduce dependence on grid or backup power, and enhance the long-term operational sustainability of solar-powered desalination systems. By integrating intelligent decision-making with renewable energy and desalination process control, the proposed system provides a scalable and adaptive solution for efficient freshwater production in arid, coastal, and off-grid regions.

Multi-Stage Flash (MSF) desalination remains one of the most established thermal desalination technologies for large-scale freshwater production, particularly in regions with high seawater salinity, limited freshwater resources, and access to thermal energy from power generation or industrial processes. Despite its operational robustness and ability to handle challenging feedwater conditions, MSF desalination is energy-intensive due to the substantial thermal energy required for brine heating and the electrical energy consumed by pumping, circulation, and auxiliary systems. Inefficient operating conditions, suboptimal brine recirculation, scaling formation, heat-transfer degradation, and variations in seawater temperature can significantly increase specific energy consumption and reduce plant productivity. Conventional optimization strategies often depend on simplified thermodynamic models, fixed operating rules, and manual operator experience, which may be insufficient to capture nonlinear interactions among process variables in complex multi-stage desalination plants. To address these limitations, this paper proposes a machine learning-driven energy optimization framework for MSF desalination plants using gradient boosting algorithms.

The proposed framework utilizes historical and real-time operational data collected from MSF plant monitoring systems, including top brine temperature, seawater intake temperature, brine heater outlet temperature, recycle brine flow rate, distillate production rate, stage pressure, stage temperature, flashing chamber performance, blowdown salinity, steam flow rate, cooling water flow rate, condenser performance, pump power consumption, heat-transfer coefficients, antiscalant dosing records, and cleaning history. A comprehensive preprocessing pipeline is applied to handle missing sensor readings, remove abnormal operating points, normalize numerical variables, align multi-rate process measurements, and construct energy-performance indicators such as gain output ratio, performance ratio, specific thermal energy consumption, specific electrical energy consumption, and total specific energy consumption.

Gradient boosting algorithms, including XGBoost, LightGBM, CatBoost, and Gradient Boosting Regression Trees, are employed to model the nonlinear relationship between plant operating conditions and energy efficiency outcomes. These models are trained to predict specific energy consumption, freshwater production rate, thermal efficiency, and operational performance under varying feedwater and process conditions. Feature selection, correlation analysis, Bayesian hyperparameter tuning, and cross-validation are incorporated to improve model generalization and reduce overfitting. The trained predictive models are then integrated into an optimization layer that identifies energy-efficient operating setpoints while satisfying technical constraints related to top brine temperature limits, allowable salinity, stage pressure boundaries, scaling risk, minimum production demand, steam availability, and pump operating limits.

The proposed framework can be evaluated using real-world MSF plant datasets, pilot-scale desalination records, or high-fidelity process simulation data. Model performance is assessed using Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, coefficient of determination, prediction stability, and computational efficiency. Energy optimization performance is evaluated through reduction in specific energy consumption, improvement in gain output ratio, increase in distillate production efficiency, reduction in steam usage, pump energy savings, and maintenance of safe operating conditions. Comparative analysis is conducted against conventional regression models, artificial neural networks, support vector regression, random forest, rule-based plant optimization, and thermodynamic baseline models.

Experimental evaluation is expected to demonstrate that gradient boosting-based models provide high predictive accuracy and strong interpretability for MSF energy optimization. Their ability to capture nonlinear interactions among thermal, hydraulic, and salinity-related variables allows the framework to identify operating regimes that reduce unnecessary energy consumption without compromising water production or equipment safety. To improve transparency and operator trust, explainability techniques such as SHAP, feature importance analysis, partial dependence plots, and sensitivity analysis are incorporated. These tools help identify the most influential energy drivers, including top brine temperature, steam flow rate, recycle brine flow, seawater intake temperature, condenser performance, and stage pressure distribution.

The proposed machine learning-driven optimization framework offers a practical, scalable, and data-driven solution for improving energy efficiency in MSF desalination plants. By combining gradient boosting prediction models with operational constraint-aware optimization, the system can support real-time decision-making, reduce thermal and electrical energy consumption, improve plant productivity, and lower operational costs. Its integration into supervisory control and plant energy management systems can assist operators in achieving more sustainable desalination performance, particularly in water-stressed regions where energy-efficient freshwater production is essential.

Nanofiltration (NF) membranes have become an important separation technology for water treatment, desalination pretreatment, wastewater reuse, hardness removal, organic micropollutant rejection, and selective ion separation. Permeate flux is one of the most critical performance indicators in nanofiltration systems because it directly reflects membrane productivity, hydraulic efficiency, fouling progression, and overall process stability. However, accurate prediction of permeate flux remains challenging due to the complex nonlinear interactions among operating pressure, feed concentration, temperature, pH, cross-flow velocity, membrane properties, solute characteristics, concentration polarization, scaling tendency, and fouling behavior. Conventional empirical models and manually tuned machine learning approaches often have limited ability to generalize across varying feedwater compositions and dynamic operating conditions. To address these limitations, this paper proposes a hybrid Convolutional Neural Network and Whale Optimization Algorithm framework for accurate prediction of permeate flux in nanofiltration membranes.

The proposed framework integrates the automatic feature extraction capability of Convolutional Neural Networks (CNNs) with the global optimization strength of the Whale Optimization Algorithm (WOA). The CNN model is designed to learn hidden patterns from structured membrane process data, including transmembrane pressure, feed flow rate, permeate flow rate, feed temperature, feed conductivity, total dissolved solids, solute concentration, membrane pore size, molecular weight cut-off, recovery ratio, pH, turbidity, and historical flux measurements. Input variables are transformed into one-dimensional or two-dimensional feature representations to allow convolutional filters to capture local relationships among operating parameters and temporal flux variations. The WOA component is employed to optimize critical CNN hyperparameters, including learning rate, number of convolutional filters, kernel size, number of dense neurons, dropout rate, batch size, optimizer configuration, and regularization coefficients. By reducing dependence on manual tuning, the WOA-enhanced CNN improves convergence efficiency, prediction accuracy, and robustness under diverse operating conditions.

A comprehensive preprocessing pipeline is applied to experimental and operational NF datasets, including missing value imputation, outlier removal, noise filtering, normalization, feature scaling, correlation analysis, and construction of lagged process variables. To improve model generalization, the framework considers variations in membrane material, feedwater quality, ion composition, organic matter concentration, pressure range, and fouling stage. The proposed model can be evaluated using laboratory-scale nanofiltration experiments, pilot plant monitoring data, or full-scale membrane process records. Performance is assessed using regression metrics such as Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, normalized RMSE, coefficient of determination, and computational training time. Comparative analysis is conducted against empirical flux models, support vector regression, random forest, XGBoost, multilayer perceptron, standalone CNN, standalone WOA-optimized neural networks, and conventional artificial neural network baselines.

Experimental evaluation is expected to demonstrate that the proposed CNN-WOA model achieves superior permeate flux prediction performance by effectively capturing nonlinear relationships between membrane operating conditions and productivity. The optimized CNN architecture is anticipated to provide lower prediction error, faster convergence, and improved stability compared with manually configured deep learning models. In addition, sensitivity analysis and explainability techniques such as SHAP, permutation feature importance, and response surface visualization are incorporated to identify the most influential factors affecting permeate flux, including transmembrane pressure, feed temperature, solute concentration, cross-flow velocity, recovery ratio, and fouling-related indicators. These insights can help membrane operators understand process behavior and support evidence-based operational adjustment.

The proposed hybrid CNN-WOA framework provides a scalable, accurate, and computationally efficient solution for permeate flux prediction in nanofiltration membrane systems. By combining deep learning-based pattern recognition with metaheuristic optimization, the model can support real-time process monitoring, fouling-aware operation, energy-efficient control, and predictive maintenance. Its integration into smart membrane management platforms can improve water treatment efficiency, reduce operational uncertainty, optimize cleaning schedules, and enhance the long-term reliability of nanofiltration systems.

Nanofiltration (NF) membranes have become an important separation technology for water treatment, desalination pretreatment, wastewater reuse, hardness removal, organic micropollutant rejection, and selective ion separation. Permeate flux is one of the most critical performance indicators in nanofiltration systems because it directly reflects membrane productivity, hydraulic efficiency, fouling progression, and overall process stability. However, accurate prediction of permeate flux remains challenging due to the complex nonlinear interactions among operating pressure, feed concentration, temperature, pH, cross-flow velocity, membrane properties, solute characteristics, concentration polarization, scaling tendency, and fouling behavior. Conventional empirical models and manually tuned machine learning approaches often have limited ability to generalize across varying feedwater compositions and dynamic operating conditions. To address these limitations, this paper proposes a hybrid Convolutional Neural Network and Whale Optimization Algorithm framework for accurate prediction of permeate flux in nanofiltration membranes.

The proposed framework integrates the automatic feature extraction capability of Convolutional Neural Networks (CNNs) with the global optimization strength of the Whale Optimization Algorithm (WOA). The CNN model is designed to learn hidden patterns from structured membrane process data, including transmembrane pressure, feed flow rate, permeate flow rate, feed temperature, feed conductivity, total dissolved solids, solute concentration, membrane pore size, molecular weight cut-off, recovery ratio, pH, turbidity, and historical flux measurements. Input variables are transformed into one-dimensional or two-dimensional feature representations to allow convolutional filters to capture local relationships among operating parameters and temporal flux variations. The WOA component is employed to optimize critical CNN hyperparameters, including learning rate, number of convolutional filters, kernel size, number of dense neurons, dropout rate, batch size, optimizer configuration, and regularization coefficients. By reducing dependence on manual tuning, the WOA-enhanced CNN improves convergence efficiency, prediction accuracy, and robustness under diverse operating conditions.

A comprehensive preprocessing pipeline is applied to experimental and operational NF datasets, including missing value imputation, outlier removal, noise filtering, normalization, feature scaling, correlation analysis, and construction of lagged process variables. To improve model generalization, the framework considers variations in membrane material, feedwater quality, ion composition, organic matter concentration, pressure range, and fouling stage. The proposed model can be evaluated using laboratory-scale nanofiltration experiments, pilot plant monitoring data, or full-scale membrane process records. Performance is assessed using regression metrics such as Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, normalized RMSE, coefficient of determination, and computational training time. Comparative analysis is conducted against empirical flux models, support vector regression, random forest, XGBoost, multilayer perceptron, standalone CNN, standalone WOA-optimized neural networks, and conventional artificial neural network baselines.

Experimental evaluation is expected to demonstrate that the proposed CNN-WOA model achieves superior permeate flux prediction performance by effectively capturing nonlinear relationships between membrane operating conditions and productivity. The optimized CNN architecture is anticipated to provide lower prediction error, faster convergence, and improved stability compared with manually configured deep learning models. In addition, sensitivity analysis and explainability techniques such as SHAP, permutation feature importance, and response surface visualization are incorporated to identify the most influential factors affecting permeate flux, including transmembrane pressure, feed temperature, solute concentration, cross-flow velocity, recovery ratio, and fouling-related indicators. These insights can help membrane operators understand process behavior and support evidence-based operational adjustment.

The proposed hybrid CNN-WOA framework provides a scalable, accurate, and computationally efficient solution for permeate flux prediction in nanofiltration membrane systems. By combining deep learning-based pattern recognition with metaheuristic optimization, the model can support real-time process monitoring, fouling-aware operation, energy-efficient control, and predictive maintenance. Its integration into smart membrane management platforms can improve water treatment efficiency, reduce operational uncertainty, optimize cleaning schedules, and enhance the long-term reliability of nanofiltration systems.

Membrane Distillation (MD) has emerged as a promising thermally driven separation technology for brackish water treatment, particularly because of its ability to operate at relatively low temperatures, utilize low-grade heat or renewable thermal energy, and achieve high salt rejection through a hydrophobic membrane barrier. Despite these advantages, the practical deployment of membrane distillation systems remains limited by several operational challenges, including membrane wetting, fouling, scaling, temperature polarization, concentration polarization, unstable permeate flux, variable feedwater quality, and high specific energy consumption under suboptimal operating conditions. Conventional monitoring and control strategies often rely on fixed thresholds, manual inspection, and simplified process models, which may be insufficient for real-time adaptation to dynamic brackish water characteristics and changing thermal conditions. To address these limitations, this paper proposes an Artificial Intelligence-enabled framework for real-time monitoring and control of membrane distillation systems for efficient brackish water treatment.

The proposed framework integrates process sensor data, water quality measurements, and intelligent predictive models to support continuous system supervision and adaptive operational control. Input variables include feed temperature, permeate temperature, feed flow rate, permeate flow rate, feed salinity, permeate conductivity, transmembrane temperature difference, membrane surface temperature, pressure variation, pH, turbidity, total dissolved solids, recovery ratio, permeate flux, thermal efficiency, and specific energy consumption. A comprehensive preprocessing pipeline is applied to clean sensor readings, handle missing values, remove outliers, normalize variables, synchronize multi-rate data streams, and derive operational indicators such as flux decline rate, salt rejection efficiency, wetting probability, scaling tendency, and energy performance index.

The core predictive component employs machine learning and deep learning models, including Random Forest, XGBoost, Long Short-Term Memory networks, Gated Recurrent Units, Temporal Convolutional Networks, and Transformer-based time-series architectures. These models are trained to forecast permeate flux, detect early membrane wetting, estimate fouling and scaling risk, predict permeate quality, and identify abnormal operating conditions in real time. For control optimization, the predictive outputs are integrated with intelligent control strategies such as reinforcement learning, model predictive control, fuzzy logic control, and multi-objective optimization. The control module dynamically adjusts operating parameters, including feed temperature, circulation flow rate, recovery ratio, thermal input, and cleaning or flushing intervals, to maximize water production, maintain salt rejection, reduce energy consumption, and prevent membrane deterioration.

The proposed system can be evaluated using pilot-scale membrane distillation experiments, real-time process monitoring data, and simulation-based brackish water treatment scenarios. Performance assessment includes both prediction and control metrics, such as Mean Absolute Error, Root Mean Square Error, coefficient of determination, detection accuracy, sensitivity, specificity, false alarm rate, permeate flux improvement, salt rejection stability, specific energy consumption reduction, membrane wetting prevention rate, and cleaning frequency reduction. Comparative analysis is conducted against conventional threshold-based monitoring, proportional–integral–derivative control, rule-based operation, statistical regression models, and non-adaptive machine learning baselines. Experimental evaluation is expected to demonstrate that the proposed AI-enabled monitoring and control framework improves operational stability, enhances permeate production, reduces energy consumption, and provides earlier warning of membrane performance deterioration.

To support operator trust and practical deployment, explainable artificial intelligence techniques such as SHAP, feature importance analysis, temporal contribution mapping, and sensitivity analysis are incorporated. These methods identify the most influential process variables affecting membrane distillation performance, including feed temperature, flow rate, salinity, permeate conductivity, temperature polarization, and scaling-related indicators. The resulting explanations help plant operators understand the causes of performance decline and validate recommended control actions. The proposed framework offers a scalable and intelligent solution for real-time membrane distillation management, supporting efficient brackish water treatment, predictive maintenance, energy-aware operation, and sustainable freshwater production. Its integration into smart water treatment platforms can contribute to more reliable, adaptive, and resource-efficient desalination systems, particularly in regions facing water scarcity and variable-quality brackish water resources.

Seawater desalination using Reverse Osmosis (RO) membranes has become one of the most widely adopted technologies for addressing freshwater scarcity, particularly in arid and coastal regions. Despite its high separation efficiency and relatively lower energy consumption compared with thermal desalination processes, RO system performance is strongly affected by membrane fouling, which remains one of the major operational challenges in large-scale desalination plants. Fouling caused by suspended solids, organic matter, biofilm formation, colloidal particles, scaling compounds, and microbial activity can reduce permeate flux, increase transmembrane pressure, elevate energy consumption, deteriorate water quality, shorten membrane lifespan, and increase the frequency of chemical cleaning. Conventional fouling monitoring methods often rely on threshold-based indicators, manual inspection, or delayed laboratory analysis, which may fail to detect early-stage fouling progression and provide limited support for proactive operational control. To address these limitations, this paper proposes a deep learning-based predictive modeling framework for Reverse Osmosis membrane fouling in seawater desalination systems to enhance operational efficiency and maintenance planning.

The proposed framework utilizes historical and real-time operational data collected from RO desalination plants, including feedwater salinity, temperature, pH, turbidity, conductivity, total dissolved solids, silt density index, permeate flow rate, concentrate flow rate, feed pressure, permeate pressure, differential pressure, recovery ratio, normalized permeate flux, salt rejection rate, chemical dosing records, pretreatment conditions, and cleaning-in-place events. A comprehensive preprocessing pipeline is applied to handle missing sensor readings, remove outliers, normalize process variables, align multi-rate time-series measurements, and construct fouling progression indicators from operational trends. Feature engineering is performed to derive clinically and operationally meaningful indicators such as normalized differential pressure increase, flux decline rate, specific energy consumption, and membrane performance degradation index.

The predictive model is developed using advanced deep learning architectures, including Long Short-Term Memory networks, Gated Recurrent Units, one-dimensional Convolutional Neural Networks, Temporal Convolutional Networks, and Transformer-based time-series models. These architectures are designed to capture nonlinear temporal dependencies between feedwater quality, operating conditions, and fouling development over time. The model predicts short-term and long-term fouling risk levels, estimates future membrane performance degradation, and provides early warning alerts before severe fouling leads to irreversible damage or unplanned shutdown. To improve robustness and generalization, the framework incorporates dropout regularization, attention mechanisms, hyperparameter optimization, class-weighted learning, and cross-validation across multiple operational periods and membrane trains.

The proposed system can be evaluated using real-world RO plant monitoring datasets, pilot-scale desalination experiments, and simulated fouling scenarios. Model performance is assessed using regression and classification metrics, including Mean Absolute Error, Root Mean Square Error, coefficient of determination, accuracy, sensitivity, specificity, F1-score, AUC-ROC, early warning lead time, and false alarm rate. Comparative analysis is conducted against traditional threshold-based monitoring, statistical regression models, support vector regression, random forest, XGBoost, standard neural networks, and non-sequential machine learning baselines. Experimental evaluation is expected to demonstrate that the proposed deep learning framework achieves improved fouling prediction accuracy, earlier detection of degradation trends, and better adaptability to dynamic feedwater and operating conditions.

For practical desalination plant operation, the predictive outputs are integrated into a decision-support module for membrane cleaning scheduling, pretreatment optimization, pressure adjustment, recovery ratio control, and energy consumption reduction. Explainability techniques such as SHAP, attention visualization, feature importance analysis, and temporal contribution mapping are incorporated to identify the most influential factors driving fouling progression, including turbidity spikes, elevated silt density index, increased feed pressure, declining permeate flux, reduced salt rejection, temperature variation, and inadequate pretreatment performance. These interpretability tools allow plant operators to understand fouling mechanisms, validate model predictions, and implement targeted corrective actions.

The proposed deep learning-based predictive modeling framework offers a scalable, accurate, and proactive solution for membrane fouling management in seawater Reverse Osmosis desalination plants. By enabling early fouling detection, optimized cleaning schedules, improved energy efficiency, and extended membrane lifetime, the system can reduce operational costs, enhance freshwater production reliability, and support more sustainable desalination infrastructure. Its integration into plant supervisory control and data acquisition systems can provide real-time intelligence for efficiency enhancement and predictive maintenance in next-generation water treatment facilities.

Accurate photovoltaic (PV) output power prediction is essential for reliable renewable energy integration, smart grid operation, energy storage scheduling, and real-time power dispatch. PV power generation is highly sensitive to dynamic weather conditions, including solar irradiance fluctuation, ambient temperature variation, cloud movement, humidity changes, wind speed, atmospheric pressure, and seasonal effects. These factors introduce nonlinear and non-stationary behavior into PV output, making accurate prediction difficult using conventional statistical models and manually tuned machine learning approaches. To address these challenges, this paper proposes an Extreme Learning Machine optimized via Differential Evolution for accurate photovoltaic output power prediction under dynamic weather conditions.

The proposed framework combines the fast learning capability of Extreme Learning Machine (ELM) with the global optimization strength of Differential Evolution (DE). ELM is employed as a single-hidden-layer feed-forward neural network capable of rapid training and effective nonlinear function approximation. However, standard ELM performance can be affected by randomly assigned input weights and hidden-layer biases, which may lead to unstable prediction accuracy and limited generalization. Therefore, Differential Evolution is integrated to optimize critical ELM parameters, including input weights, hidden neuron biases, activation function parameters, and the number of hidden neurons. By guiding the search toward more effective parameter configurations, the DE-optimized ELM improves prediction stability, convergence behavior, and forecasting accuracy compared with conventional ELM.

The proposed model utilizes historical PV output power and meteorological variables such as global horizontal irradiance, direct normal irradiance, diffuse horizontal irradiance, ambient temperature, module temperature, wind speed, humidity, cloud cover, and solar zenith angle. A comprehensive preprocessing pipeline is applied to handle missing measurements, remove abnormal sensor readings, normalize input features, align weather and PV generation records, and construct lagged time-series variables that represent recent weather and power trends. Feature selection and correlation analysis are incorporated to identify the most influential variables affecting PV power output under rapidly changing environmental conditions.

The proposed DE-ELM model can be evaluated using real-world PV plant monitoring data, meteorological station records, and publicly available solar energy datasets. Prediction performance is assessed using standard forecasting metrics, including Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, normalized RMSE, coefficient of determination, forecasting skill score, and computational training time. Comparative experiments are conducted against persistence models, ARIMA, support vector regression, random forest, XGBoost, standard ELM, multilayer perceptron, LSTM, GRU, and other metaheuristic-optimized learning models. Experimental evaluation is expected to demonstrate that the proposed DE-ELM achieves higher prediction accuracy, faster training speed, and stronger robustness under highly variable weather conditions compared with non-optimized and conventionally tuned models.

For practical smart grid applications, the predicted PV output can be integrated into energy management systems to support battery storage scheduling, grid balancing, demand response, renewable curtailment reduction, and day-ahead or intra-day dispatch planning. In addition, sensitivity analysis is incorporated to evaluate the effect of irradiance, temperature, cloud cover, and wind speed on PV power prediction. These analytical insights can assist grid operators and PV plant managers in understanding the key drivers of prediction uncertainty and improving operational decision-making.

The proposed Differential Evolution-optimized Extreme Learning Machine provides a fast, accurate, and computationally efficient solution for photovoltaic output power prediction under dynamic weather conditions. By combining rapid neural learning with evolutionary optimization, the framework improves forecasting reliability, reduces uncertainty in renewable power generation, and supports more stable integration of solar energy into smart grids and distributed energy systems.

Large-scale integration of renewable energy resources into smart grids introduces significant operational challenges due to the intermittent, distributed, and uncertain nature of photovoltaic and wind power generation. As renewable penetration increases, power systems become more dynamic and complex, requiring advanced optimization methods capable of maintaining voltage stability, minimizing transmission losses, preventing line congestion, improving renewable energy utilization, and ensuring reliable power delivery under variable operating conditions. Conventional power flow optimization methods, including Newton–Raphson-based optimal power flow, linearized approximations, and traditional numerical solvers, can provide accurate solutions but may become computationally demanding for large-scale grids with high-dimensional topology, nonlinear constraints, and rapidly changing renewable generation profiles. To address these limitations, this paper proposes a Graph Neural Network (GNN)-based power flow optimization framework for large-scale renewable energy integration in smart grids.

The proposed framework represents the power grid as a graph structure, where buses, generators, substations, renewable energy units, energy storage systems, and loads are modeled as nodes, while transmission lines, transformers, and distribution feeders are modeled as edges. Node features include active and reactive power injection, voltage magnitude, voltage angle, renewable generation output, load demand, generator limits, battery state of charge, and local operational constraints. Edge features include line impedance, admittance, thermal capacity, power flow direction, transformer tap settings, and congestion indicators. By explicitly learning from the topological structure of the grid, the GNN model captures spatial dependencies, electrical coupling relationships, and nonlinear interactions among geographically distributed power system components.

The core architecture employs message-passing neural networks, graph convolutional networks, graph attention networks, and physics-informed graph learning modules to approximate optimal power flow solutions under renewable-rich operating scenarios. The model is trained using historical grid operation data, simulated power flow cases, renewable generation profiles, and load demand scenarios generated from benchmark systems and real-world smart grid measurements. A comprehensive preprocessing pipeline is applied, including grid topology encoding, normalization of electrical variables, renewable uncertainty modeling, scenario generation, constraint labeling, and feature construction for both node-level and edge-level representations. To improve physical consistency, the framework incorporates power balance constraints, voltage limits, line flow limits, generator capacity constraints, and penalty terms for constraint violations within the training objective.

The proposed GNN-based optimizer can support multiple operational tasks, including AC optimal power flow approximation, congestion management, voltage regulation, renewable curtailment minimization, distributed energy resource coordination, and energy storage dispatch. It is designed to operate across different grid configurations and can generalize to changing network topologies, such as line outages, distributed generation expansion, or reconfiguration of distribution feeders. Model performance is evaluated using standard power system and optimization metrics, including total active power loss, voltage deviation, renewable curtailment rate, operating cost, line overload reduction, constraint violation rate, solution feasibility, computational time, and optimality gap. Comparative analysis is conducted against conventional optimal power flow solvers, DC power flow approximations, mixed-integer optimization methods, deep neural network baselines, and non-graph-based machine learning approaches.

Experimental evaluation is expected to demonstrate that the proposed GNN framework achieves near-optimal power flow solutions with significantly reduced computational latency, making it suitable for real-time and near-real-time smart grid operation. By leveraging graph-based representation learning, the model can better capture the structural dependencies of power networks compared with conventional feed-forward neural networks that ignore grid topology. In addition, interpretability mechanisms such as graph attention visualization, node importance analysis, edge congestion attribution, and sensitivity mapping are incorporated to help grid operators understand which buses, lines, or renewable injection points have the greatest influence on optimization decisions.

The proposed GNN-based power flow optimization framework provides a scalable, topology-aware, and computationally efficient solution for smart grids with high renewable energy penetration. By combining graph representation learning with power system operational constraints, the framework can enhance renewable energy integration, reduce transmission losses, improve voltage stability, mitigate congestion, and support secure real-time grid management. Its adaptability to large-scale and dynamically changing grid topologies makes it a promising tool for next-generation energy management systems and renewable-dominant power networks.

Green hydrogen production through water electrolysis powered by renewable energy sources has emerged as a promising pathway for large-scale decarbonization, long-duration energy storage, sector coupling, and the integration of variable renewable generation into future energy systems. However, hydrogen production from electrolysis-based renewable systems is strongly affected by the intermittent and uncertain nature of solar and wind energy, fluctuations in available electrical power, electrolyzer operating conditions, water supply constraints, temperature variation, efficiency degradation, and dynamic demand for hydrogen storage or utilization. Accurate forecasting of hydrogen production is therefore essential for energy management, electrolyzer scheduling, storage planning, grid balancing, and techno-economic optimization. To address these challenges, this paper proposes a Bidirectional Gated Recurrent Unit (BiGRU) network for hydrogen production forecasting from electrolysis-based renewable energy systems.

The proposed framework is designed to model nonlinear temporal dependencies between renewable power generation and hydrogen output. The input data include historical photovoltaic power, wind power, solar irradiance, wind speed, ambient temperature, electrolyzer input power, stack voltage, stack current, operating temperature, water consumption rate, electrolyzer efficiency, hydrogen flow rate, storage tank pressure, and previous hydrogen production records. A comprehensive preprocessing pipeline is applied to handle missing values, remove outliers, normalize continuous variables, align multi-source time-series measurements, and construct sliding-window input sequences for multi-step forecasting. The framework supports short-term, day-ahead, and multi-horizon hydrogen production forecasting to meet both operational and planning requirements.

The BiGRU architecture is employed to capture temporal dependencies in both forward and backward directions, enabling the model to learn production patterns from preceding system behavior as well as contextual relationships across the entire input sequence. Compared with conventional recurrent neural networks, GRU-based models provide efficient gating mechanisms with fewer parameters than LSTM architectures, making them suitable for renewable hydrogen forecasting applications that require both accuracy and computational efficiency. The bidirectional structure improves the model’s ability to represent complex relationships among renewable generation variability, electrolyzer response dynamics, and storage system conditions. Dropout regularization, batch normalization, adaptive learning-rate scheduling, and hyperparameter optimization are incorporated to improve convergence stability and prevent overfitting.

The proposed model can be evaluated using simulated and experimental datasets from renewable-powered electrolyzer systems, hydrogen microgrids, power-to-gas testbeds, and integrated wind–solar–hydrogen energy systems. Forecasting performance is assessed using Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, normalized RMSE, coefficient of determination, forecasting skill score, and computational latency. Comparative analysis is conducted against persistence models, ARIMA, support vector regression, random forest, XGBoost, standard RNN, LSTM, unidirectional GRU, CNN-GRU, and transformer-based time-series forecasting models. Experimental evaluation is expected to demonstrate that the proposed BiGRU network achieves improved forecasting accuracy, particularly under rapidly changing renewable generation and variable electrolyzer loading conditions.

For practical renewable hydrogen system management, the forecasting outputs can be integrated into energy management systems to support optimal electrolyzer dispatch, hydrogen storage scheduling, battery–hydrogen coordination, grid export decisions, renewable curtailment reduction, and demand-side hydrogen supply planning. In addition, model interpretability is enhanced through temporal attention analysis, feature importance estimation, and sensitivity analysis to identify the variables most strongly influencing hydrogen production, such as renewable input power, electrolyzer efficiency, stack temperature, and storage pressure. These insights can assist operators in improving system reliability, reducing operating costs, and maximizing renewable-to-hydrogen conversion efficiency.

The proposed BiGRU-based forecasting framework provides a scalable, accurate, and computationally efficient solution for hydrogen production prediction in electrolysis-based renewable energy systems. By capturing bidirectional temporal patterns across renewable generation, electrolyzer operation, and storage behavior, the framework can support more reliable green hydrogen production planning, improve utilization of surplus renewable energy, reduce operational uncertainty, and contribute to the development of sustainable hydrogen energy infrastructure

Energy-efficient load dispatch is a fundamental optimization problem in modern renewable power grids, particularly as power systems increasingly integrate multiple generation sources such as photovoltaic arrays, wind turbines, hydropower units, biomass generators, battery energy storage systems, and conventional backup units. The variability and uncertainty of renewable energy generation introduce significant operational challenges related to power balance, generation cost, transmission losses, reserve allocation, voltage stability, emission reduction, and reliable demand satisfaction. Conventional load dispatch methods often rely on deterministic assumptions, linearized models, or gradient-based optimization techniques, which may become less effective when dealing with nonlinear, non-convex, multi-source, and constraint-rich renewable power grid environments. To address these challenges, this paper proposes a Harris Hawks Optimization (HHO)-based framework for energy-efficient load dispatch in multi-source renewable power grids.

The proposed framework formulates the load dispatch problem as a constrained optimization task aimed at determining the optimal power contribution of each available energy source while satisfying system demand and operational limits. The objective function is designed to minimize total generation cost, fuel consumption of backup units, active power losses, carbon emissions, renewable energy curtailment, and battery degradation cost, while maximizing renewable energy utilization and overall energy efficiency. Operational constraints include power balance, generator capacity limits, ramp-rate limits, spinning reserve requirements, battery state-of-charge boundaries, charging and discharging limits, transmission line capacity, voltage stability margins, and renewable generation availability. By incorporating these technical and economic constraints, the proposed model provides a realistic representation of dispatch decisions in renewable-rich power grids.

The Harris Hawks Optimization algorithm is employed as the core metaheuristic search strategy due to its ability to balance exploration and exploitation through cooperative hunting behavior inspired by Harris hawks. In the proposed dispatch framework, candidate solutions represent power allocation schedules among renewable sources, storage systems, and dispatchable generators. The HHO algorithm iteratively updates these candidate solutions to identify near-optimal dispatch patterns under uncertain load demand and renewable output conditions. To improve robustness, scenario-based modeling is incorporated using historical solar irradiance, wind speed, hydrological flow, biomass availability, electricity demand, and weather forecasting data. In addition, uncertainty handling strategies such as Monte Carlo simulation, probabilistic renewable modeling, and reserve-aware dispatch are integrated to improve decision reliability under fluctuating generation conditions.

A comprehensive simulation environment is developed to evaluate the proposed framework across different operating scenarios, including peak demand periods, low renewable generation intervals, high renewable penetration, grid-connected operation, islanded microgrid operation, and storage-supported dispatch. Performance is assessed using total operating cost, energy efficiency, power loss reduction, renewable utilization rate, emission reduction, load satisfaction ratio, convergence speed, computational time, and constraint violation rate. Comparative experiments are conducted against conventional economic dispatch methods, Particle Swarm Optimization, Genetic Algorithm, Grey Wolf Optimizer, Whale Optimization Algorithm, Differential Evolution, and rule-based dispatch strategies. Experimental evaluation is expected to demonstrate that the proposed HHO-based approach achieves superior dispatch efficiency, lower operating cost, reduced emissions, and improved convergence behavior compared with existing optimization techniques.

The proposed framework can be integrated into smart grid energy management systems, microgrid controllers, and renewable dispatch platforms to support real-time and day-ahead operational planning. Its flexible optimization structure enables adaptation to different grid scales, energy resource portfolios, and operational priorities. Furthermore, sensitivity analysis is incorporated to examine the influence of renewable penetration level, storage capacity, demand uncertainty, and penalty factor selection on dispatch performance. These analytical components provide valuable insights for grid operators and planners seeking to improve renewable energy integration while maintaining economic and technical reliability.

The proposed Harris Hawks Optimization-based load dispatch framework offers an efficient, scalable, and robust solution for multi-source renewable power grids. By optimizing power allocation across heterogeneous energy resources while accounting for cost, emissions, losses, and operational constraints, the system supports cleaner, more reliable, and more economically efficient power system operation. Its ability to manage uncertainty and nonlinear dispatch behavior makes it a promising tool for next-generation smart grids, renewable microgrids, and sustainable energy management applications.

Wind turbines operate under highly variable environmental and mechanical conditions, including fluctuating wind speed, turbulence, temperature changes, cyclic loading, and continuous rotational stress. These operating conditions can accelerate degradation in critical mechanical components such as gearboxes, bearings, shafts, generators, brakes, and rotor assemblies. Unexpected failures in these components may lead to costly downtime, reduced energy production, expensive corrective maintenance, and safety risks, particularly in remote or offshore wind farms where maintenance access is difficult and operational costs are high. Conventional condition monitoring and threshold-based fault detection methods often rely on fixed alarm limits and expert-defined rules, which may fail to identify subtle early-stage degradation patterns hidden within high-dimensional sensor data. To address these challenges, this paper proposes a stacked autoencoder-based anomaly detection framework for predictive maintenance of wind turbine mechanical components.

The proposed framework leverages the unsupervised feature learning capability of stacked autoencoders to model normal operating behavior and detect abnormal deviations that may indicate incipient mechanical faults. The system utilizes multi-source turbine monitoring data, including supervisory control and data acquisition (SCADA) signals, vibration measurements, acoustic emission data, bearing temperature, gearbox oil temperature, generator temperature, rotor speed, wind speed, wind direction, power output, torque, and nacelle position. A comprehensive preprocessing pipeline is applied to clean and synchronize sensor data, remove outliers, handle missing values, normalize continuous variables, filter noise, and construct operational condition-aware input representations. To reduce false alarms caused by normal environmental variability, the framework incorporates wind-speed binning, operating-regime clustering, and context-aware normalization.

The stacked autoencoder architecture is trained primarily on healthy turbine operating data to learn compressed latent representations of normal mechanical behavior. During inference, deviations between reconstructed and observed sensor patterns are quantified using reconstruction error, latent-space distance, and statistical anomaly scoring. Abnormal patterns exceeding adaptive thresholds are flagged as potential early indicators of component degradation. The model is designed to identify anomalies associated with bearing wear, gearbox tooth damage, shaft imbalance, lubrication problems, generator overheating, rotor misalignment, and drivetrain vibration irregularities. To improve robustness and diagnostic value, the framework integrates temporal smoothing, health index construction, and remaining useful life trend analysis, enabling maintenance teams to distinguish transient disturbances from persistent degradation.

The proposed framework can be evaluated using real-world wind turbine SCADA datasets, vibration monitoring datasets, and simulated mechanical fault scenarios. Performance is assessed using anomaly detection metrics such as precision, recall, F1-score, false alarm rate, detection delay, area under the precision–recall curve, reconstruction error distribution, and early warning lead time. Comparative analysis is conducted against conventional statistical process monitoring, principal component analysis, one-class support vector machine, isolation forest, shallow autoencoder, LSTM autoencoder, and rule-based condition monitoring approaches. Experimental evaluation is expected to demonstrate that the stacked autoencoder model achieves improved sensitivity to early mechanical degradation while maintaining a low false alarm rate under diverse operating conditions.

To support practical deployment, the proposed anomaly detection system can be integrated into wind farm supervisory control platforms and predictive maintenance dashboards. Explainability and diagnostic interpretation are incorporated through sensor contribution analysis, reconstruction error decomposition, latent feature visualization, and component-level health indicators. These tools allow maintenance engineers to identify which sensor channels or mechanical subsystems contribute most strongly to an anomaly, supporting targeted inspection and maintenance planning. The proposed framework provides a scalable, data-driven, and cost-effective solution for early fault detection in wind turbine mechanical components. By enabling proactive maintenance decisions, reducing unplanned downtime, extending component lifetime, and improving wind farm reliability, the system contributes to more efficient and sustainable wind energy generation.

Real-time energy storage scheduling is a critical requirement for the reliable and economic operation of wind–solar hybrid energy systems, particularly because photovoltaic and wind power generation are inherently intermittent, weather-dependent, and difficult to predict with complete certainty. The complementary nature of solar and wind resources can improve renewable energy availability; however, their combined variability still creates operational challenges related to power imbalance, renewable curtailment, battery degradation, peak demand management, grid stability, and electricity market participation. Conventional rule-based and optimization-based scheduling strategies often depend on accurate forecasts and fixed operating rules, which may limit their ability to adapt to rapidly changing renewable generation, load demand, electricity prices, and battery state conditions. To address these challenges, this paper proposes a Deep Q-Network (DQN) reinforcement learning framework for real-time energy storage scheduling in wind–solar hybrid systems.

The proposed framework formulates storage scheduling as a sequential decision-making problem in which a DQN agent learns an optimal control policy through continuous interaction with the hybrid energy environment. The system state includes photovoltaic generation, wind power output, load demand, battery state of charge, electricity price, weather forecast indicators, grid import/export status, and previous scheduling actions. Based on these observations, the DQN agent selects discrete energy storage actions, including charging, discharging, idle operation, grid import support, renewable curtailment reduction, and peak shaving decisions. The reward function is designed to minimize total operating cost, renewable energy curtailment, unmet load, battery degradation, and grid power fluctuation while maximizing renewable energy utilization and system reliability.

The proposed model incorporates key components of Deep Q-Network learning, including experience replay, target network stabilization, epsilon-greedy exploration, and temporal-difference learning. These mechanisms enable the agent to learn stable and efficient scheduling policies under uncertain and nonlinear operating conditions. A comprehensive simulation environment is developed using historical solar irradiance, wind speed, load demand, electricity tariff, and battery performance data. Data preprocessing includes missing value imputation, outlier removal, normalization, temporal alignment, scenario generation, and construction of realistic renewable generation profiles. Battery operational constraints, including state-of-charge limits, charging/discharging efficiency, power capacity, depth of discharge, and cycle degradation cost, are incorporated to ensure physically feasible scheduling decisions.

The proposed DQN-based scheduling system can be evaluated under multiple operating scenarios, including grid-connected operation, islanded microgrid operation, high renewable penetration, peak load periods, uncertain weather conditions, and time-of-use electricity pricing. Performance is assessed using total energy cost, renewable utilization rate, curtailment reduction, battery degradation cost, load satisfaction ratio, peak-to-average ratio, grid power exchange variability, convergence behavior, and computational response time. Comparative analysis is conducted against rule-based scheduling, mixed-integer linear programming, dynamic programming, model predictive control, Q-learning, and heuristic energy management strategies. Experimental evaluation is expected to demonstrate that the proposed DQN framework achieves more adaptive and cost-effective storage scheduling by learning from real-time system feedback rather than relying solely on predefined control rules.

To support practical deployment, the framework can be integrated into smart grid controllers, microgrid energy management systems, and distributed energy resource management platforms. The learned policy enables real-time decision-making for battery charging and discharging while considering renewable uncertainty, load variability, and operational constraints. Furthermore, interpretability analysis is incorporated through reward decomposition, action-value visualization, and state importance analysis to help operators understand why specific storage actions are selected under different system conditions. These insights improve trust and facilitate operational validation before deployment in real-world hybrid renewable systems.

The proposed DQN reinforcement learning framework provides a scalable, adaptive, and intelligent solution for real-time energy storage scheduling in wind–solar hybrid systems. By combining deep reinforcement learning with battery-aware operational modeling, the system can reduce energy cost, improve renewable energy utilization, limit curtailment, enhance grid stability, and support the transition toward more resilient and sustainable power systems.

Distributed Energy Resource Management is a central challenge in modern microgrids, particularly as photovoltaic systems, wind turbines, battery energy storage systems, electric vehicles, controllable loads, and demand response programs become increasingly integrated into decentralized power networks. Effective management of these resources requires accurate forecasting, adaptive control, optimal scheduling, and real-time coordination under uncertain renewable generation, variable load demand, changing electricity prices, and dynamic grid conditions. Traditional centralized machine learning approaches often require collecting operational data from multiple microgrids or distributed assets into a single server, which may raise concerns related to data privacy, cybersecurity, communication burden, ownership of energy data, and regulatory constraints. To address these limitations, this paper proposes a Federated Machine Learning framework for privacy-preserving and collaborative Distributed Energy Resource Management in microgrids.

The proposed framework enables multiple microgrids, prosumers, smart buildings, and distributed energy resource controllers to collaboratively train predictive and decision-support models without transferring raw local data. Each participating client trains a local machine learning model using its own energy consumption records, renewable generation profiles, battery state-of-charge measurements, weather data, electricity tariff information, and operational constraints. Instead of sharing sensitive operational data, only model updates or encrypted parameters are transmitted to a central or hierarchical aggregation server, where a global model is constructed using federated aggregation strategies such as FedAvg, FedProx, and adaptive federated optimization. This collaborative learning paradigm allows the global model to benefit from diverse operating conditions across multiple microgrid environments while preserving local data confidentiality.

The framework supports several core microgrid management tasks, including short-term load forecasting, renewable energy generation prediction, battery energy storage scheduling, electric vehicle charging coordination, demand response optimization, fault-aware resource dispatch, and grid-connected or islanded mode operation planning. A comprehensive preprocessing pipeline is applied locally at each client, including missing value imputation, outlier removal, normalization, time-series alignment, feature encoding, and privacy-aware data quality assessment. To address the non-independent and identically distributed nature of microgrid data, the proposed system incorporates personalization layers, client clustering, weighted aggregation, and domain-adaptive learning mechanisms. These components improve model robustness when microgrids differ in load profiles, renewable penetration levels, weather patterns, storage capacities, and user behavior.

For optimization and control, the federated learning model can be integrated with model predictive control, reinforcement learning, or multi-objective optimization modules to generate operational decisions that minimize energy cost, reduce carbon emissions, improve renewable energy utilization, maintain voltage and frequency stability, and extend battery lifetime. Privacy and security are further strengthened through secure aggregation, differential privacy, encrypted communication, anomaly detection for malicious client updates, and resilience mechanisms against data poisoning or model inversion attacks. These safeguards are essential for practical deployment in real-world energy networks where distributed assets may belong to different owners or utility operators.

The proposed framework is evaluated using simulated and real-world microgrid datasets under heterogeneous operating scenarios, including grid-connected operation, islanded operation, high renewable penetration, peak demand periods, and communication-limited environments. Performance is assessed using forecasting accuracy, operational cost reduction, renewable energy utilization rate, battery degradation cost, peak-to-average ratio reduction, loss of load probability, communication overhead, convergence speed, privacy preservation level, and robustness against non-IID data distributions. Comparative experiments are conducted against centralized learning, standalone local learning, conventional optimization-based energy management, and non-federated machine learning baselines. Experimental evaluation is expected to demonstrate that the proposed federated framework achieves near-centralized performance while significantly improving privacy preservation, scalability, and adaptability across distributed microgrid systems.

The proposed Federated Machine Learning framework offers a scalable, secure, and intelligent solution for next-generation distributed energy resource management. By enabling collaborative learning without centralized data collection, the system supports privacy-aware energy forecasting, adaptive scheduling, and coordinated control across multiple microgrids. Its integration into smart grid platforms can enhance energy efficiency, increase renewable energy hosting capacity, reduce operational uncertainty, improve cyber-resilience, and support the development of decentralized, sustainable, and data-driven energy communities.

Maximum Power Point Tracking (MPPT) is a fundamental control function in solar photovoltaic (PV) systems, as it ensures that PV modules operate at their highest possible power output under varying environmental and load conditions. The nonlinear current–voltage and power–voltage characteristics of PV arrays are strongly influenced by solar irradiance, temperature, shading patterns, module degradation, and load variations. Conventional MPPT algorithms such as Perturb and Observe, Incremental Conductance, and hill-climbing techniques are widely used because of their simplicity; however, they often suffer from oscillations around the maximum power point, slow convergence under rapidly changing weather conditions, and difficulty in locating the global maximum power point under partial shading conditions. To overcome these limitations, this paper proposes an Artificial Neural Network optimized by the Grey Wolf Optimizer for efficient and accurate MPPT in solar PV systems.

The proposed framework combines the nonlinear mapping capability of Artificial Neural Networks (ANNs) with the global search strength of the Grey Wolf Optimizer (GWO). The ANN model is trained to estimate the optimal duty cycle or reference voltage corresponding to the maximum power point using input variables such as PV voltage, PV current, solar irradiance, module temperature, and historical power measurements. To enhance learning performance and reduce dependence on manual parameter tuning, GWO is employed to optimize the ANN architecture and training parameters, including the number of hidden neurons, learning rate, activation function configuration, initial weights, biases, and regularization coefficients. By simulating the leadership hierarchy and hunting behavior of grey wolves, the GWO algorithm efficiently searches for near-optimal ANN configurations that improve tracking accuracy, convergence speed, and robustness under dynamic operating conditions.

A comprehensive PV system model is developed using solar array characteristics, DC–DC boost converter dynamics, pulse-width modulation control, and variable resistive or grid-connected load conditions. The proposed GWO-ANN MPPT controller is evaluated under diverse environmental scenarios, including uniform irradiance, rapidly changing irradiance, temperature fluctuations, partial shading patterns, and load disturbances. The training and testing datasets are generated from simulated and experimental PV operating conditions to cover a wide range of current–voltage and power–voltage behaviors. Data preprocessing includes normalization, noise filtering, feature scaling, and scenario balancing to ensure stable ANN learning and reliable generalization.

The performance of the proposed controller is assessed using key MPPT evaluation metrics, including tracking efficiency, convergence time, steady-state oscillation, output power ripple, global maximum power point detection accuracy, dynamic response, and robustness under partial shading. Comparative analysis is conducted against conventional MPPT techniques, standalone ANN-based MPPT, fuzzy logic control, particle swarm optimization-based MPPT, and other metaheuristic-enhanced intelligent controllers. Experimental results are expected to demonstrate that the proposed GWO-optimized ANN achieves faster convergence, higher tracking efficiency, reduced power oscillations, and improved global maximum power point identification compared with traditional and non-optimized approaches.

The proposed GWO-ANN MPPT method provides a reliable, adaptive, and computationally efficient solution for improving the energy harvesting capability of solar PV systems. By integrating intelligent neural prediction with metaheuristic optimization, the controller can respond effectively to nonlinear PV behavior and environmental uncertainty. Its implementation can enhance PV conversion efficiency, reduce energy losses, improve system stability, and support the deployment of high-performance solar energy systems in standalone, grid-connected, and smart microgrid applications.

Accurate energy demand forecasting is a critical requirement for efficient smart city management, sustainable urban planning, and reliable operation of modern power systems. As cities become increasingly dependent on digital infrastructure, electric transportation, distributed renewable energy resources, smart buildings, and Internet of Things-based monitoring systems, energy consumption patterns are becoming more complex, dynamic, and nonlinear. Short-term demand forecasting is essential for real-time grid operation, demand response, energy dispatch, peak-load management, and electricity market participation, while long-term forecasting supports infrastructure planning, capacity expansion, renewable integration, and strategic policy development. Traditional statistical forecasting methods and conventional machine learning models often struggle to capture long-range temporal dependencies, multi-scale seasonality, nonlinear consumption behavior, and the influence of heterogeneous urban factors. To address these challenges, this paper proposes a transformer-based deep learning model for short-term and long-term energy demand forecasting in smart cities.

The proposed framework leverages the self-attention mechanism of transformer architectures to model complex temporal relationships across multiple forecasting horizons. Unlike recurrent neural networks, which process sequences sequentially and may suffer from limited long-term memory, transformers can capture dependencies between distant time steps more effectively through parallel attention-based learning. The model integrates historical electricity consumption, weather variables, calendar information, occupancy patterns, socio-economic indicators, electricity price signals, renewable generation profiles, transportation activity, and smart meter measurements. A comprehensive preprocessing pipeline is applied to handle missing values, remove outliers, normalize continuous variables, encode temporal features, align multi-source urban datasets, and construct multi-resolution input sequences suitable for both short-term and long-term forecasting.

The proposed architecture incorporates positional encoding, multi-head self-attention, temporal convolutional embedding, feed-forward layers, and horizon-specific output modules to generate accurate energy demand predictions at hourly, daily, weekly, monthly, and seasonal scales. For short-term forecasting, the model focuses on high-resolution consumption fluctuations driven by weather changes, occupancy behavior, electric vehicle charging, and peak-hour demand. For long-term forecasting, the model captures structural trends related to population growth, urban expansion, seasonal variation, economic activity, policy changes, and renewable energy adoption. To improve robustness, the framework incorporates probabilistic forecasting and uncertainty estimation, enabling grid operators and city planners to quantify prediction confidence and prepare for demand variability.

The proposed model can be evaluated using smart meter datasets, city-scale electricity consumption records, building energy datasets, weather archives, and publicly available energy benchmarking datasets. Performance is assessed using Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, symmetric MAPE, normalized RMSE, coefficient of determination, prediction interval coverage probability, and computational efficiency. Comparative analysis is conducted against persistence models, ARIMA, Prophet, support vector regression, random forest, XGBoost, LSTM, GRU, temporal convolutional networks, and conventional encoder–decoder models. Experimental evaluation is expected to demonstrate that the transformer-based model achieves superior forecasting accuracy across both short-term and long-term horizons by effectively capturing multi-scale temporal dependencies and heterogeneous urban energy consumption patterns.

For smart city applications, the forecasting outputs are integrated into decision-support modules for load balancing, demand response scheduling, distributed energy resource coordination, battery storage management, grid congestion prevention, and urban energy policy planning. The model also supports interpretability through attention visualization, feature attribution analysis, and temporal importance mapping, allowing operators to identify the weather conditions, time periods, consumption sectors, and urban variables that contribute most strongly to predicted energy demand. This transparency improves trust and facilitates practical deployment in city energy management platforms.

The proposed transformer-based framework offers a scalable, accurate, and interpretable solution for energy demand forecasting in smart cities. By supporting both operational short-term forecasting and strategic long-term planning, the system can enhance grid reliability, reduce energy waste, improve renewable energy integration, optimize infrastructure investment, and contribute to more sustainable and resilient urban energy systems.

Hybrid Renewable Energy Systems (HRESs) have emerged as an effective solution for improving energy reliability, reducing fossil-fuel dependence, and supporting the transition toward low-carbon power systems. By integrating multiple energy resources such as photovoltaic panels, wind turbines, battery energy storage systems, diesel backup generators, hydrogen storage units, and grid-interconnection components, HRESs can provide flexible and resilient electricity supply for microgrids, remote communities, industrial facilities, and smart distribution networks. However, the design and operation of hybrid renewable systems involve complex multi-objective optimization challenges, including minimizing total lifecycle cost, reducing carbon emissions, maximizing renewable energy penetration, improving power supply reliability, extending battery lifetime, and maintaining grid stability under uncertain weather and load conditions. To address these challenges, this paper proposes a hybrid optimization framework that combines Deep Reinforcement Learning (DRL) and Genetic Algorithms (GAs) for multi-objective planning and real-time energy management of HRESs.

The proposed framework formulates HRES optimization as a coupled design-operation problem. At the planning level, a Genetic Algorithm is employed to search for optimal system configurations, including photovoltaic array size, wind turbine capacity, battery storage capacity, inverter rating, backup generator capacity, and hydrogen system sizing where applicable. The GA-based optimizer evaluates candidate system designs using multi-objective fitness functions that consider Net Present Cost, Levelized Cost of Energy, Loss of Power Supply Probability, renewable fraction, carbon dioxide emissions, battery degradation cost, and energy curtailment. Pareto-front analysis is incorporated to identify trade-off solutions that allow system planners to balance economic, environmental, and reliability objectives according to application-specific priorities.

At the operational level, a Deep Reinforcement Learning agent is developed to learn adaptive energy management policies under dynamic and uncertain operating conditions. The DRL agent observes system states such as solar irradiance, wind speed, load demand, battery state of charge, electricity price, grid availability, weather forecasts, and generator status. Based on these states, the agent selects control actions including battery charging and discharging, generator dispatch, grid import or export, renewable energy curtailment, and load-shifting decisions. Advanced DRL algorithms such as Deep Q-Network, Deep Deterministic Policy Gradient, Proximal Policy Optimization, and Soft Actor–Critic are investigated to support both discrete and continuous control actions. The reward function is designed to minimize operational cost, emissions, unmet load, battery stress, and switching frequency while maximizing renewable utilization and supply reliability.

A comprehensive simulation and evaluation environment is developed using historical renewable resource profiles, load demand data, electricity tariffs, component degradation models, and technical constraints of HRES components. Data preprocessing includes weather–load alignment, missing value imputation, normalization, scenario generation, and uncertainty modeling for solar irradiance, wind speed, and demand variability. The proposed DRL–GA framework is evaluated under grid-connected, islanded, and off-grid microgrid scenarios. Performance is assessed using economic, technical, environmental, and reliability metrics, including total net present cost, Levelized Cost of Energy, renewable energy fraction, carbon emissions, Loss of Power Supply Probability, battery cycling degradation, energy curtailment ratio, peak load reduction, and computational efficiency.

Experimental evaluation is expected to demonstrate that the proposed hybrid DRL–GA approach outperforms conventional rule-based energy management, standalone Genetic Algorithm optimization, particle swarm optimization, mixed-integer linear programming-based scheduling, and non-adaptive heuristic control strategies. The GA component identifies cost-effective and reliable system configurations, while the DRL component continuously adapts operational decisions to real-time variations in renewable generation, load demand, electricity prices, and storage availability. This integration enables the system to achieve better trade-offs between cost reduction, emission mitigation, renewable utilization, and reliability enhancement compared with static optimization approaches.

The proposed framework provides a scalable, intelligent, and adaptive solution for the multi-objective optimization of hybrid renewable energy systems. By combining evolutionary search with sequential decision-making intelligence, the system supports both long-term capacity planning and real-time operational control. Its ability to generate Pareto-optimal design alternatives and learn dynamic control policies makes it suitable for smart grids, microgrids, remote electrification, industrial energy systems, and sustainable community power planning. The framework can contribute to improved renewable energy integration, reduced operational uncertainty, enhanced storage utilization, and more resilient low-carbon energy infrastructure

Photovoltaic (PV) systems are increasingly deployed as a key renewable energy technology for sustainable electricity generation; however, their operational performance, reliability, and energy yield can be significantly affected by faults and environmental disturbances. Among these disturbances, partial shading conditions represent one of the most challenging scenarios because they create non-uniform irradiance distribution across PV modules, leading to power mismatch, hotspot formation, multiple peaks in the power–voltage characteristics, reduced maximum power extraction, and possible long-term degradation of system components. In addition, partial shading may mask or imitate several electrical and physical faults, making accurate fault detection and classification more difficult using conventional threshold-based monitoring and rule-based diagnostic methods. To address these limitations, this paper proposes a Convolutional Neural Network (CNN)-based framework for automated fault detection and classification in photovoltaic systems operating under partial shading conditions.

The proposed framework is designed to learn discriminative fault-related patterns from PV system data, including current–voltage curves, power–voltage curves, module temperature profiles, irradiance measurements, string current, array voltage, power output, bypass diode behavior, and inverter-level monitoring signals. Depending on the available monitoring infrastructure, the input representation can be constructed from transformed electrical signal maps, time–frequency images, thermal images, electroluminescence images, or two-dimensional feature matrices derived from sensor measurements. A CNN architecture is employed to automatically extract hierarchical spatial and local pattern features associated with different PV abnormalities without relying heavily on handcrafted indicators. The model is trained to distinguish normal operation from multiple fault categories, including partial shading, open-circuit faults, short-circuit faults, line-to-line faults, bypass diode faults, hotspot defects, soiling effects, module degradation, and mismatch faults.

A comprehensive preprocessing pipeline is incorporated to improve data quality and diagnostic robustness. This pipeline includes noise filtering, missing value handling, normalization of electrical variables, irradiance and temperature compensation, image resizing, contrast enhancement for thermal or electroluminescence data, feature scaling, and data augmentation under diverse shading intensities and environmental conditions. To improve the model’s ability to generalize across different PV configurations, the framework considers variations in module technology, array topology, shading pattern, fault severity, load condition, and seasonal weather behavior. Class imbalance caused by the lower occurrence of severe faults is addressed using class-weighted loss functions, oversampling, synthetic sample generation, and balanced mini-batch training.

The proposed CNN-based diagnostic system can be evaluated using experimental PV testbeds, simulation-generated datasets, and real-world monitoring data collected from grid-connected or standalone PV installations. Model performance is assessed using accuracy, precision, recall, sensitivity, specificity, F1-score, confusion matrix analysis, false alarm rate, detection latency, and robustness under changing irradiance and temperature conditions. Comparative analysis is conducted against conventional electrical threshold methods, support vector machine, random forest, k-nearest neighbors, multilayer perceptron, and handcrafted feature-based diagnostic approaches. Experimental evaluation is expected to demonstrate that the proposed CNN framework achieves improved fault classification accuracy and stronger resilience to partial shading effects by learning complex nonlinear relationships between shading patterns, electrical behavior, and fault signatures.

To support practical deployment, the proposed framework can be integrated into PV monitoring platforms, inverter controllers, supervisory control systems, and smart energy management systems. Explainability techniques such as Grad-CAM, saliency mapping, and feature activation visualization are incorporated to highlight the signal regions, thermal hotspots, or electrical pattern segments most responsible for each fault prediction. These interpretability tools help operators verify whether the model focuses on physically meaningful fault indicators, thereby improving trust and supporting maintenance decision-making. The proposed system offers a scalable, accurate, and automated solution for early fault detection and classification in PV systems under partial shading conditions. By enabling timely fault diagnosis, reducing unnecessary maintenance, minimizing energy losses, and preventing severe component damage, the framework contributes to improved PV system reliability, operational safety, and long-term renewable energy performance.

Accurate wind power generation forecasting is a fundamental requirement for the reliable integration of offshore wind farms into modern power systems. Offshore wind energy offers high generation potential due to stronger and more consistent wind resources compared with many onshore locations; however, its output remains highly variable because of complex meteorological dynamics, wake effects, wind speed fluctuations, atmospheric pressure changes, turbulence intensity, marine boundary-layer behavior, and seasonal weather patterns. These uncertainties create operational challenges for grid scheduling, reserve allocation, energy market bidding, storage coordination, and power system stability. Traditional statistical forecasting models and conventional machine learning approaches often struggle to capture nonlinear temporal dependencies and long-range variations in offshore wind power data. To address these limitations, this paper proposes a Whale Optimization Algorithm-enhanced Long Short-Term Memory network for accurate wind power generation forecasting in offshore wind farms.

The proposed framework combines the sequential learning capability of LSTM networks with the global search and exploitation–exploration balance of the Whale Optimization Algorithm (WOA). The LSTM model is employed to learn temporal dependencies from historical wind power generation and meteorological time-series data, including wind speed, wind direction, air temperature, atmospheric pressure, humidity, turbine rotor speed, nacelle orientation, turbulence indicators, and previous power output. WOA is integrated as an intelligent hyperparameter optimization mechanism to automatically tune critical LSTM parameters, including the number of hidden units, learning rate, dropout ratio, batch size, number of recurrent layers, sequence length, optimizer configuration, and regularization coefficients. By replacing manual trial-and-error tuning and exhaustive grid search, the WOA-enhanced strategy improves convergence efficiency, forecasting accuracy, and generalization under fluctuating offshore operating conditions.

A comprehensive preprocessing pipeline is applied to wind farm operational data, including missing value imputation, outlier detection, noise filtering, normalization, temporal alignment, feature selection, and construction of sliding-window input sequences. To enhance robustness, the framework incorporates weather-aware feature engineering, lagged power variables, wind ramp event indicators, seasonal decomposition, and correlation analysis between meteorological variables and generated power. The proposed model can be evaluated using supervisory control and data acquisition data from offshore wind farms, numerical weather prediction outputs, and publicly available renewable energy datasets. Forecasting tasks are formulated across multiple horizons, including ultra-short-term, short-term, and day-ahead wind power prediction, depending on grid operation and market requirements.

Experimental evaluation is conducted using standard forecasting metrics such as Mean Absolute Error, Root Mean Square Error, Mean Absolute Percentage Error, normalized RMSE, coefficient of determination, and forecasting skill score. Comparative analysis is performed against persistence models, ARIMA, support vector regression, random forest, XGBoost, standalone LSTM, GRU, CNN-LSTM, and other metaheuristic-optimized deep learning models. The proposed WOA-LSTM framework is expected to achieve superior forecasting performance by effectively capturing temporal wind power dynamics while optimizing model parameters to reduce prediction error. Particular attention is given to challenging scenarios such as rapid wind ramp events, high-turbulence periods, seasonal transitions, and wake-influenced production variability, where accurate forecasting is crucial for grid reliability.

For practical offshore wind farm management, the forecasting outputs are integrated into decision-support applications for power dispatch scheduling, reserve planning, battery energy storage coordination, preventive maintenance planning, and electricity market participation. In addition, model interpretability is enhanced through feature importance analysis, temporal sensitivity assessment, and error decomposition to identify the meteorological and operational factors most strongly influencing forecasting performance. These analytical components help grid operators and wind farm managers understand prediction behavior and improve operational decision-making.

The proposed WOA-enhanced LSTM framework offers a scalable, accurate, and computationally efficient solution for offshore wind power forecasting. By combining metaheuristic optimization with deep sequential modeling, the system improves forecasting reliability, supports renewable energy integration, reduces operational uncertainty, and contributes to more stable and economical smart grid management. Its applicability to multi-horizon forecasting and offshore wind farm operations makes it a promising tool for next-generation renewable energy management systems.

Solar irradiance forecasting is a critical component of modern renewable energy management, particularly as photovoltaic power generation becomes increasingly integrated into smart grids, microgrids, and distributed energy systems. The intermittent and weather-dependent nature of solar energy introduces significant uncertainty into power generation, affecting grid stability, energy dispatch, reserve scheduling, battery storage management, and demand–supply balancing. Conventional statistical forecasting models and traditional machine learning methods often struggle to capture the nonlinear, non-stationary, and highly dynamic behavior of solar irradiance under rapidly changing meteorological conditions such as cloud movement, aerosol variation, humidity fluctuation, and seasonal transitions. To address these challenges, this paper proposes a deep learning-enabled solar irradiance forecasting framework based on spatiotemporal attention mechanisms for accurate and reliable smart grid management.

The proposed framework integrates heterogeneous data sources, including historical global horizontal irradiance, direct normal irradiance, diffuse horizontal irradiance, ambient temperature, humidity, wind speed, atmospheric pressure, cloud cover, solar zenith angle, satellite imagery, sky-camera observations, and photovoltaic generation records. A comprehensive preprocessing pipeline is applied to handle missing meteorological measurements, normalize continuous features, remove outliers, align multi-resolution temporal data, and extract spatial contextual information from geographically distributed solar monitoring stations. The forecasting problem is formulated as a multi-horizon time-series prediction task, supporting very-short-term, short-term, and day-ahead solar irradiance forecasts required for real-time grid operation and energy planning.

The core model combines convolutional layers, recurrent neural networks, and spatiotemporal attention modules to learn both local weather-driven patterns and long-range temporal dependencies. Convolutional components extract spatial features from satellite or sky-image data, while LSTM, GRU, or temporal transformer layers model sequential dependencies in irradiance and meteorological time series. The spatiotemporal attention mechanism dynamically assigns higher weights to the most relevant time intervals, weather variables, and neighboring spatial locations, enabling the model to focus on critical factors such as moving cloud fields, abrupt irradiance drops, seasonal patterns, and regional weather correlations. This attention-guided design improves forecasting accuracy, robustness, and interpretability compared with conventional deep learning architectures that treat all temporal and spatial inputs uniformly.

The proposed model can be evaluated using publicly available solar and meteorological datasets such as NREL solar resource data, NASA POWER data, local photovoltaic plant measurements, and satellite-derived irradiance datasets. Performance is assessed using forecasting metrics including Root Mean Square Error, Mean Absolute Error, Mean Absolute Percentage Error, normalized RMSE, coefficient of determination, prediction interval coverage probability, and computational latency. Comparative experiments are conducted against persistence models, ARIMA, support vector regression, random forest, XGBoost, standalone LSTM, CNN-LSTM, and transformer-based baselines. The proposed spatiotemporal attention model is expected to achieve improved forecasting performance, particularly under unstable cloudy conditions and high-variability weather scenarios where accurate prediction is most valuable for grid reliability.

For smart grid management, the forecasting outputs are integrated into decision-support modules for photovoltaic power scheduling, battery energy storage control, demand response, reserve allocation, and grid congestion mitigation. The model also supports uncertainty-aware forecasting by producing prediction intervals that allow grid operators to quantify forecast confidence and plan operational reserves more effectively. Furthermore, attention visualization is incorporated to provide interpretability by identifying the meteorological variables, spatial regions, and time steps that contribute most strongly to each forecast. This transparency enhances operator trust and supports the practical deployment of artificial intelligence-based forecasting systems in energy control centers.

The proposed framework offers a scalable, accurate, and interpretable solution for solar irradiance forecasting in smart grid environments. By combining deep learning with spatiotemporal attention mechanisms, the system can improve photovoltaic integration, reduce renewable energy curtailment, optimize storage utilization, and enhance the reliability and economic efficiency of grid operations. Its ability to process multi-source meteorological and spatial data makes it particularly suitable for next-generation renewable energy management systems, where accurate forecasting is essential for achieving secure, flexible, and sustainable power system operation.

Cardiac arrhythmias are among the most common and clinically significant cardiovascular abnormalities, ranging from benign rhythm disturbances to life-threatening events such as ventricular tachycardia, ventricular fibrillation, and atrial fibrillation-related complications. Early and continuous detection of arrhythmias is essential for timely intervention, prevention of sudden cardiac events, and long-term cardiovascular monitoring. Electrocardiogram (ECG) signals provide a direct and non-invasive representation of cardiac electrical activity; however, manual ECG interpretation is time-consuming, requires specialized expertise, and may be impractical for continuous real-time monitoring in wearable healthcare environments. Although traditional machine learning approaches have achieved promising results in ECG classification, their performance is often limited by handcrafted feature extraction, signal noise, inter-patient variability, and insufficient modeling of temporal dependencies. To address these limitations, this paper proposes a hybrid Convolutional Neural Network–Long Short-Term Memory (CNN-LSTM) architecture for real-time ECG arrhythmia classification and early warning in wearable devices.

The proposed framework combines the spatial feature extraction capability of CNNs with the sequential modeling strength of LSTM networks. The CNN component automatically extracts discriminative morphological features from ECG waveforms, including P-wave characteristics, QRS complex morphology, ST-segment deviations, RR interval variability, and T-wave abnormalities. The LSTM component captures temporal dependencies across consecutive heartbeat segments, enabling the model to learn rhythm patterns and detect dynamic changes associated with arrhythmic events. The architecture is designed to classify multiple arrhythmia categories, including normal sinus rhythm, premature atrial contractions, premature ventricular contractions, atrial fibrillation, atrial flutter, supraventricular tachycardia, ventricular tachycardia, bundle branch block, and other clinically relevant rhythm abnormalities.

A comprehensive signal preprocessing pipeline is applied to wearable ECG recordings, including baseline wander removal, powerline interference filtering, noise suppression, R-peak detection, heartbeat segmentation, amplitude normalization, and artifact rejection. To improve robustness under real-world wearable conditions, the framework incorporates data augmentation techniques such as signal scaling, time shifting, Gaussian noise injection, waveform stretching, and synthetic minority oversampling to address class imbalance among rare arrhythmia events. The proposed model can be trained and evaluated using publicly available ECG datasets such as MIT-BIH Arrhythmia Database, PhysioNet/CinC Challenge datasets, and wearable ECG recordings where available. Model optimization is performed to reduce computational complexity and support deployment on resource-constrained wearable devices using techniques such as model pruning, quantization, lightweight convolutional layers, and edge inference acceleration.

Experimental evaluation is conducted using clinically relevant metrics, including accuracy, sensitivity, specificity, precision, F1-score, AUC-ROC, detection latency, false alarm rate, and energy consumption. The proposed CNN-LSTM architecture is expected to outperform standalone CNN, standalone LSTM, traditional machine learning classifiers, and handcrafted feature-based ECG analysis methods by achieving higher classification accuracy and improved sensitivity for early arrhythmia detection. In addition, the system incorporates an early warning module that generates real-time alerts when abnormal rhythm patterns exceed predefined clinical risk thresholds. This module is designed to support patient monitoring, clinician notification, and timely emergency response in both home-based and ambulatory care environments.

To enhance clinical interpretability and user trust, the framework integrates explainability techniques such as Grad-CAM-based temporal activation mapping, saliency analysis, and attention-based visualization to identify ECG waveform regions that contribute most strongly to the classification decision. These explanations allow clinicians to verify whether the model focuses on meaningful cardiac features such as QRS widening, irregular RR intervals, abnormal P-wave activity, and ST-T changes. The proposed hybrid CNN-LSTM framework provides a scalable, accurate, and computationally efficient solution for real-time ECG arrhythmia classification in wearable devices. Its integration into wearable health monitoring systems has the potential to improve continuous cardiovascular surveillance, reduce delayed arrhythmia diagnosis, minimize unnecessary hospital visits, and support early intervention for high-risk patients.

Antibiotic resistance has become one of the most serious global threats to public health, leading to increased treatment failure, prolonged hospitalization, higher healthcare costs, and elevated mortality rates. The rapid emergence of multidrug-resistant bacterial strains has created an urgent need for intelligent clinical decision support systems capable of identifying resistance patterns and assisting clinicians in selecting appropriate antimicrobial therapy. Conventional antibiotic prescribing often relies on empirical treatment guidelines, local antibiograms, and delayed laboratory culture results, which may be insufficient for rapidly changing resistance trends and patient-specific risk factors. To address these challenges, this paper proposes a machine learning-driven Clinical Decision Support System (CDSS) for antibiotic resistance pattern identification using microbiological, clinical, and electronic health record data.

The proposed framework integrates structured patient information, including demographic variables, infection site, prior antibiotic exposure, hospitalization history, comorbidities, laboratory findings, microbiology culture results, antimicrobial susceptibility testing profiles, pathogen species, intensive care admission status, device use, and previous colonization or infection with resistant organisms. A comprehensive preprocessing pipeline is applied to clean and harmonize heterogeneous clinical data, handle missing values, encode categorical variables, normalize numerical features, remove duplicate culture records, and align microbiology reports with patient-level clinical timelines. To improve predictive reliability, the framework also incorporates feature selection and temporal aggregation strategies that capture recent antimicrobial exposure, recurrent infections, and hospital-acquired risk factors.

Several machine learning models are developed and compared, including Logistic Regression, Random Forest, Support Vector Machine, XGBoost, LightGBM, CatBoost, and deep neural network architectures. These models are trained to identify resistance patterns for clinically important pathogens and antimicrobial classes, including methicillin-resistant Staphylococcus aureus, vancomycin-resistant Enterococcus, extended-spectrum beta-lactamase-producing Enterobacterales, carbapenem-resistant organisms, multidrug-resistant Pseudomonas aeruginosa, and resistant Acinetobacter baumannii. The proposed CDSS generates patient-specific resistance probability scores and provides antimicrobial risk stratification to support empirical therapy selection before definitive susceptibility results become available.

To address class imbalance caused by the relatively low prevalence of certain resistant organisms, the framework incorporates class-weighted loss functions, focal loss, synthetic minority oversampling, and stratified cross-validation. Model performance is evaluated using accuracy, sensitivity, specificity, precision, F1-score, AUC-ROC, AUC-PR, calibration analysis, and decision-curve analysis. Particular emphasis is placed on sensitivity and negative predictive value, as failure to identify resistant infections may result in inappropriate therapy and adverse clinical outcomes. Comparative evaluation is conducted against conventional antibiogram-based recommendations, rule-based clinical protocols, and non-personalized empirical prescribing strategies.

To enhance interpretability and clinical trust, the proposed system integrates explainable artificial intelligence techniques such as SHAP, permutation feature importance, and patient-level risk attribution. These methods identify the most influential factors contributing to predicted resistance, including prior antibiotic exposure, recent hospitalization, ICU stay, invasive devices, recurrent infection history, renal dysfunction, immunosuppression, and previous resistant organism isolation. By presenting both global resistance trends and individualized explanations, the CDSS supports antimicrobial stewardship teams, infectious disease specialists, and frontline clinicians in making evidence-based prescribing decisions.

The proposed machine learning-driven CDSS provides a scalable, interpretable, and clinically actionable approach for antibiotic resistance pattern identification. By combining microbiology data, EHR-derived patient risk factors, and predictive analytics, the framework can improve early recognition of resistant infections, reduce inappropriate antibiotic use, support antimicrobial stewardship programs, and contribute to limiting the spread of antimicrobial resistance. Its integration into hospital information systems and laboratory workflows may enable real-time decision support, personalized empirical therapy selection, and improved patient safety in both inpatient and emergency care settings.

Chronic Kidney Disease (CKD) is a progressive and often silent medical condition characterized by gradual loss of kidney function over time, frequently remaining undetected until advanced stages when therapeutic options become limited and the risk of cardiovascular complications, dialysis dependence, and mortality increases substantially. Early prediction of CKD onset is therefore essential for preventive nephrology, timely clinical intervention, patient monitoring, and personalized disease management. Conventional statistical risk scores and traditional machine learning approaches have shown promise in identifying high-risk patients; however, their performance may be limited by complex nonlinear interactions among clinical variables, suboptimal parameter selection, missing values, class imbalance, and variability across heterogeneous patient populations. To address these challenges, this paper proposes a Particle Swarm Optimization-tuned Deep Neural Network (PSO-DNN) framework for accurate prediction of CKD onset using structured clinical and laboratory data.

The proposed framework utilizes patient-specific information extracted from Electronic Health Records, including demographic characteristics, comorbidities, medication history, blood pressure measurements, urine analysis findings, serum creatinine, estimated glomerular filtration rate, blood urea nitrogen, albumin, hemoglobin, electrolytes, fasting blood glucose, diabetes status, hypertension status, and cardiovascular risk indicators. A comprehensive preprocessing pipeline is applied to handle missing clinical values, remove noisy or redundant attributes, normalize continuous variables, encode categorical features, and balance the dataset using class-weighted learning and resampling strategies. Feature selection and correlation analysis are incorporated to retain the most clinically meaningful predictors and reduce dimensionality before model training.

The core predictive model is based on a multilayer Deep Neural Network designed to capture complex nonlinear relationships among renal function markers, metabolic variables, and patient-level risk factors. Particle Swarm Optimization is integrated to automatically tune critical DNN hyperparameters, including learning rate, number of hidden layers, number of neurons per layer, activation function configuration, dropout rate, batch size, optimizer settings, and regularization coefficients. By replacing manual trial-and-error tuning and exhaustive grid search, PSO improves convergence efficiency, reduces computational burden, and identifies near-optimal model configurations that enhance predictive accuracy and generalization. The proposed PSO-DNN model can be evaluated using benchmark CKD datasets and real-world EHR cohorts, with performance assessed through accuracy, sensitivity, specificity, precision, F1-score, AUC-ROC, AUC-PR, calibration analysis, and cross-validation.

Experimental evaluation is expected to demonstrate that the proposed PSO-tuned DNN outperforms conventional DNN models, standalone machine learning classifiers, and non-optimized neural network baselines such as Support Vector Machine, Random Forest, Logistic Regression, XGBoost, and standard multilayer perceptron models. Particular emphasis is placed on improving sensitivity and recall for early-stage CKD onset prediction, as false-negative cases may delay preventive treatment and accelerate disease progression. To support clinical interpretability, explainability techniques such as SHAP, permutation feature importance, and patient-level risk attribution analysis are incorporated to identify the most influential predictors contributing to CKD onset. These explanations are expected to highlight clinically relevant factors such as reduced estimated glomerular filtration rate, elevated serum creatinine, albuminuria, hypertension, diabetes, anemia, abnormal blood urea nitrogen, and electrolyte imbalance.

The proposed PSO-DNN framework offers a scalable, accurate, and clinically interpretable decision-support tool for early CKD risk prediction. By combining the global search capability of Particle Swarm Optimization with the nonlinear modeling strength of deep neural networks, the system can assist clinicians in identifying high-risk patients before advanced renal impairment occurs. Its integration into hospital information systems and primary care screening workflows could support earlier referral, personalized monitoring, lifestyle intervention, medication adjustment, and improved long-term renal outcomes.

Alzheimer’s disease is a progressive neurodegenerative disorder and one of the leading causes of dementia, characterized by gradual cognitive decline, memory impairment, functional deterioration, and structural changes in the brain. Early detection of disease progression is essential for timely clinical intervention, patient monitoring, therapeutic planning, and enrollment in disease-modifying treatment trials. Neuroimaging modalities such as magnetic resonance imaging, positron emission tomography, and computed tomography provide valuable biomarkers for identifying anatomical and functional changes associated with Alzheimer’s disease. However, manual interpretation of brain scans is time-consuming, observer-dependent, and often limited in detecting subtle disease-related progression patterns during the early stages. To address these challenges, this paper proposes an attention-mechanism deep learning model for automated detection and monitoring of Alzheimer’s disease progression from brain scans.

The proposed framework integrates convolutional neural networks with attention-based feature learning to enhance the model’s ability to focus on clinically relevant brain regions affected by Alzheimer’s disease, including the hippocampus, entorhinal cortex, medial temporal lobe, posterior cingulate cortex, and cortical gray matter structures. The architecture employs spatial attention, channel attention, and self-attention modules to capture both local anatomical abnormalities and long-range dependencies across brain regions. Pre-trained deep learning backbones, including ResNet, DenseNet, EfficientNet, and Vision Transformer-based models, are adapted for three-dimensional and multi-slice neuroimaging analysis. The model is designed to classify disease stages, including cognitively normal controls, mild cognitive impairment, early Alzheimer’s disease, and advanced Alzheimer’s disease, while also supporting longitudinal progression prediction where serial imaging data are available.

A comprehensive preprocessing pipeline is applied to brain scan data, including skull stripping, bias-field correction, spatial registration, intensity normalization, brain tissue segmentation, slice selection, and anatomical region-of-interest extraction. The framework can be trained and validated using publicly available neuroimaging datasets such as the Alzheimer’s Disease Neuroimaging Initiative, OASIS, and related dementia imaging cohorts. To improve robustness and generalization, the model incorporates data augmentation, transfer learning, class-weighted loss functions, and cross-validation strategies. In longitudinal settings, temporal modeling components such as attention-guided recurrent layers or transformer encoders are integrated to learn progressive structural changes over time.

Experimental evaluation is conducted using clinically relevant performance metrics, including accuracy, sensitivity, specificity, precision, F1-score, AUC-ROC, balanced accuracy, and progression prediction error. The proposed attention-enhanced model is expected to outperform conventional CNN architectures, traditional machine learning classifiers, and non-attention-based deep learning models by improving discrimination between mild cognitive impairment and early Alzheimer’s disease, which remains one of the most clinically challenging diagnostic tasks. Furthermore, explainability techniques such as Grad-CAM, attention heatmaps, saliency mapping, and region-level attribution analysis are incorporated to visualize the brain regions contributing most strongly to the model’s predictions. These interpretability tools allow clinicians to verify whether the model focuses on neuroanatomically meaningful regions associated with Alzheimer’s disease progression.

The proposed system provides an accurate, interpretable, and clinically scalable decision-support framework for automated Alzheimer’s disease progression detection from brain scans. By combining attention-based deep learning with neuroimaging biomarkers, the framework can support early diagnosis, longitudinal monitoring, individualized risk assessment, and improved clinical decision-making. Its explainable design enhances physician trust and facilitates integration into radiology and neurology workflows, particularly in settings where rapid and objective assessment of neurodegenerative changes is required.

Rare diseases collectively affect millions of individuals worldwide, yet their diagnosis remains a significant challenge due to limited clinical expertise, heterogeneous disease manifestations, and, most importantly, the scarcity of annotated medical imaging data required for training robust artificial intelligence models. In medical image analysis, deep learning algorithms often require large and diverse datasets to achieve reliable performance; however, rare disease datasets are frequently characterized by small sample sizes, severe class imbalance, privacy restrictions, and limited availability of expert annotations. These constraints increase the risk of overfitting and reduce the generalizability of diagnostic models. To address these challenges, this paper proposes a Generative Adversarial Network (GAN)-based framework for synthetic medical image augmentation aimed at improving rare disease diagnosis through the generation of realistic and clinically meaningful synthetic imaging samples.

The proposed framework employs advanced GAN architectures, including Deep Convolutional GAN (DCGAN), Conditional GAN (cGAN), Wasserstein GAN with Gradient Penalty (WGAN-GP), and StyleGAN-based models, to generate high-quality synthetic medical images that preserve disease-specific pathological characteristics. The generation process is conditioned on clinically relevant attributes such as disease subtype, anatomical region, imaging modality, and disease severity level, enabling the creation of diverse yet diagnostically consistent samples. The framework is designed to support multiple imaging modalities, including magnetic resonance imaging (MRI), computed tomography (CT), chest radiography, retinal imaging, histopathological slides, and ultrasound images associated with rare pathological conditions.

A comprehensive preprocessing pipeline is applied before image generation, including image normalization, artifact removal, contrast enhancement, spatial standardization, and annotation verification. Following synthetic image generation, quality assessment mechanisms are employed using Fréchet Inception Distance (FID), Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), and expert radiologist evaluation to ensure anatomical realism and pathological fidelity. The generated synthetic images are subsequently integrated into the training datasets of downstream diagnostic models, including Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and hybrid deep learning architectures. To further improve data diversity, the framework combines GAN-generated samples with conventional augmentation techniques such as rotation, scaling, flipping, elastic deformation, and intensity transformation.

The proposed methodology is evaluated using publicly available and institution-specific rare disease imaging datasets under varying levels of data scarcity. Experimental performance is assessed using accuracy, sensitivity, specificity, F1-score, AUC-ROC, balanced accuracy, and cross-dataset generalization measures. Results are expected to demonstrate that GAN-based augmentation significantly enhances diagnostic performance compared with conventional augmentation strategies and training on original datasets alone. In particular, the inclusion of synthetic images improves minority-class representation, reduces model bias, increases robustness to unseen clinical cases, and enhances the detection of subtle disease-specific imaging patterns that are often underrepresented in limited datasets.

To promote transparency and clinical acceptance, explainability techniques such as Grad-CAM, saliency mapping, and feature attribution analysis are incorporated to visualize the imaging regions influencing diagnostic decisions. In addition, synthetic image validation procedures are integrated to detect potential artifacts, unrealistic structures, or mode collapse effects that may negatively affect model reliability. The proposed framework provides a scalable and privacy-conscious solution for addressing one of the most critical barriers in rare disease artificial intelligence research—the lack of sufficient training data. By leveraging GAN-generated synthetic medical images, the framework enables the development of more accurate, robust, and clinically deployable diagnostic systems capable of supporting early detection and improved management of rare diseases across diverse healthcare environments.

Rare diseases collectively affect millions of individuals worldwide, yet their diagnosis remains a significant challenge due to limited clinical expertise, heterogeneous disease manifestations, and, most importantly, the scarcity of annotated medical imaging data required for training robust artificial intelligence models. In medical image analysis, deep learning algorithms often require large and diverse datasets to achieve reliable performance; however, rare disease datasets are frequently characterized by small sample sizes, severe class imbalance, privacy restrictions, and limited availability of expert annotations. These constraints increase the risk of overfitting and reduce the generalizability of diagnostic models. To address these challenges, this paper proposes a Generative Adversarial Network (GAN)-based framework for synthetic medical image augmentation aimed at improving rare disease diagnosis through the generation of realistic and clinically meaningful synthetic imaging samples.

The proposed framework employs advanced GAN architectures, including Deep Convolutional GAN (DCGAN), Conditional GAN (cGAN), Wasserstein GAN with Gradient Penalty (WGAN-GP), and StyleGAN-based models, to generate high-quality synthetic medical images that preserve disease-specific pathological characteristics. The generation process is conditioned on clinically relevant attributes such as disease subtype, anatomical region, imaging modality, and disease severity level, enabling the creation of diverse yet diagnostically consistent samples. The framework is designed to support multiple imaging modalities, including magnetic resonance imaging (MRI), computed tomography (CT), chest radiography, retinal imaging, histopathological slides, and ultrasound images associated with rare pathological conditions.

A comprehensive preprocessing pipeline is applied before image generation, including image normalization, artifact removal, contrast enhancement, spatial standardization, and annotation verification. Following synthetic image generation, quality assessment mechanisms are employed using Fréchet Inception Distance (FID), Structural Similarity Index Measure (SSIM), Peak Signal-to-Noise Ratio (PSNR), and expert radiologist evaluation to ensure anatomical realism and pathological fidelity. The generated synthetic images are subsequently integrated into the training datasets of downstream diagnostic models, including Convolutional Neural Networks (CNNs), Vision Transformers (ViTs), and hybrid deep learning architectures. To further improve data diversity, the framework combines GAN-generated samples with conventional augmentation techniques such as rotation, scaling, flipping, elastic deformation, and intensity transformation.

The proposed methodology is evaluated using publicly available and institution-specific rare disease imaging datasets under varying levels of data scarcity. Experimental performance is assessed using accuracy, sensitivity, specificity, F1-score, AUC-ROC, balanced accuracy, and cross-dataset generalization measures. Results are expected to demonstrate that GAN-based augmentation significantly enhances diagnostic performance compared with conventional augmentation strategies and training on original datasets alone. In particular, the inclusion of synthetic images improves minority-class representation, reduces model bias, increases robustness to unseen clinical cases, and enhances the detection of subtle disease-specific imaging patterns that are often underrepresented in limited datasets.

To promote transparency and clinical acceptance, explainability techniques such as Grad-CAM, saliency mapping, and feature attribution analysis are incorporated to visualize the imaging regions influencing diagnostic decisions. In addition, synthetic image validation procedures are integrated to detect potential artifacts, unrealistic structures, or mode collapse effects that may negatively affect model reliability. The proposed framework provides a scalable and privacy-conscious solution for addressing one of the most critical barriers in rare disease artificial intelligence research—the lack of sufficient training data. By leveraging GAN-generated synthetic medical images, the framework enables the development of more accurate, robust, and clinically deployable diagnostic systems capable of supporting early detection and improved management of rare diseases across diverse healthcare environments.

Personalized cancer treatment planning is a complex clinical decision-making process that requires the optimal selection of therapeutic strategies while balancing tumor control, patient-specific anatomical constraints, treatment toxicity, and long-term survival outcomes. Radiotherapy remains a major treatment modality for many cancer types; however, designing high-quality radiation treatment plans is often time-consuming and highly dependent on expert clinical judgment. Conventional treatment planning methods typically rely on manual parameter adjustment, predefined dose constraints, and iterative optimization procedures, which may limit consistency, scalability, and adaptability across heterogeneous patient populations. To address these challenges, this paper proposes a Deep Reinforcement Learning (DRL)-based framework for personalized cancer treatment planning and radiotherapy dose optimization.

The proposed framework formulates cancer treatment planning as a sequential decision-making problem, where an intelligent DRL agent learns to recommend patient-specific treatment actions based on evolving clinical states. These states include tumor characteristics, imaging-derived anatomical features, dose–volume histogram parameters, organ-at-risk constraints, patient demographics, staging information, prior treatment response, and relevant molecular or clinical biomarkers where available. The action space includes adaptive radiotherapy parameter adjustment, beam configuration refinement, fractionation strategy selection, dose escalation or de-escalation decisions, and optimization priority tuning for tumor coverage and organ-at-risk sparing. A reward function is designed to maximize tumor control probability while minimizing normal tissue complication probability, cumulative toxicity, and violation of clinically accepted dose constraints.

The proposed architecture integrates advanced DRL algorithms, including Deep Q-Networks, Proximal Policy Optimization, Actor–Critic models, and Deep Deterministic Policy Gradient, to support both discrete and continuous treatment optimization scenarios. To enhance patient-specific adaptation, the framework incorporates multimodal data from computed tomography, magnetic resonance imaging, positron emission tomography, electronic health records, pathology reports, and radiotherapy planning systems. A comprehensive preprocessing pipeline is applied, including image registration, tumor and organ-at-risk segmentation, feature extraction, dose normalization, missing data imputation, and construction of individualized treatment-state trajectories. The DRL agent is trained using historical treatment plans, simulated dose distributions, and clinically validated planning objectives to learn optimal policies that reflect expert decision-making while improving planning efficiency.

The framework is evaluated using retrospective radiotherapy datasets and simulated treatment planning environments across selected cancer types, including lung, head-and-neck, prostate, and cervical cancer. Model performance is assessed using clinically meaningful metrics such as planning target volume coverage, dose conformity index, homogeneity index, organ-at-risk dose reduction, tumor control probability, normal tissue complication probability, cumulative toxicity score, and overall plan quality score. Comparative analysis is conducted against conventional inverse planning, rule-based optimization, manual expert planning, and non-reinforcement learning-based deep learning models. Experimental evaluation is expected to demonstrate that the proposed DRL-based system can generate clinically acceptable and patient-specific radiotherapy plans with improved consistency, reduced planning time, and better balance between tumor irradiation and healthy tissue protection.

To improve clinical transparency and physician trust, the proposed framework incorporates explainability mechanisms, including policy visualization, reward decomposition, dose–volume contribution analysis, and patient-level decision trajectory interpretation. These tools allow radiation oncologists and medical physicists to understand why specific dose adjustments or treatment actions are recommended for an individual patient. The proposed system offers a scalable and adaptive decision-support platform for precision oncology by combining reinforcement learning, medical imaging, radiotherapy physics, and patient-specific clinical evidence. Its integration into clinical workflows has the potential to support more efficient treatment planning, personalized dose optimization, reduced toxicity risk, and improved quality of cancer care.

Sepsis is a life-threatening clinical condition that requires rapid recognition and timely intervention, particularly in Emergency Department (ED) settings where patients often present with heterogeneous symptoms, incomplete medical histories, and rapidly changing physiological status. Delayed diagnosis of sepsis may lead to septic shock, multi-organ failure, prolonged hospitalization, and increased mortality. Although conventional screening tools such as SIRS, qSOFA, MEWS, and NEWS provide practical bedside assessment, their predictive performance is often limited by static threshold-based rules and insufficient ability to capture complex temporal patterns in patient deterioration. To address these limitations, this paper proposes an optimized Bidirectional Long Short-Term Memory (BiLSTM) network for real-time sepsis prediction using time-series Electronic Health Record data collected during ED admission.

The proposed framework is designed to process dynamic clinical variables, including vital signs, laboratory test results, demographic characteristics, triage information, comorbidities, medication records, and early clinical observations. A comprehensive preprocessing pipeline is applied to handle missing values, align irregularly sampled measurements, normalize continuous variables, encode categorical features, and construct time-windowed patient trajectories suitable for sequential modeling. The BiLSTM architecture is employed to learn bidirectional temporal dependencies from evolving patient data, enabling the model to capture both preceding clinical trends and contextual relationships among physiological measurements. To enhance predictive accuracy and convergence efficiency, the network is optimized using advanced hyperparameter tuning strategies, including Bayesian optimization and metaheuristic algorithms such as Particle Swarm Optimization and Grey Wolf Optimizer. These optimization methods are used to determine the optimal learning rate, number of hidden units, dropout rate, batch size, time-window length, and activation configuration.

The proposed model can be trained and validated using large-scale publicly available clinical datasets, such as MIMIC-IV and MIMIC-IV-ED, as well as institutional ED datasets where available. Sepsis labels are generated based on clinically accepted criteria involving suspected infection and organ dysfunction, while prediction horizons are defined to support early warning at clinically meaningful intervals before sepsis onset. To address class imbalance caused by the relatively lower proportion of sepsis cases compared with non-sepsis encounters, the framework incorporates class-weighted loss functions, focal loss, and resampling strategies. Model performance is evaluated using AUC-ROC, AUC-PR, sensitivity, specificity, F1-score, calibration measures, and lead-time analysis, with comparisons against traditional clinical scoring systems, standard LSTM, GRU, Random Forest, XGBoost, and conventional machine learning baselines.

Experimental results are expected to demonstrate that the optimized BiLSTM model achieves superior early prediction performance by effectively modeling temporal deterioration patterns and reducing false negative alerts. In addition, explainability techniques such as SHAP, attention-based visualization, and patient-level temporal risk attribution are incorporated to identify the most influential clinical indicators contributing to sepsis risk, including abnormal respiratory rate, hypotension, elevated lactate level, leukocyte count changes, fever, oxygen saturation decline, renal function markers, and altered mental status. These explanations support clinician trust by clarifying how specific physiological trends influence the predicted risk score over time.

The proposed framework offers a scalable, accurate, and interpretable real-time decision-support system for early sepsis prediction in emergency care environments. By integrating optimized sequential deep learning with clinically meaningful temporal risk analysis, the system can assist ED clinicians in prioritizing high-risk patients, initiating timely diagnostic evaluation, and improving early intervention strategies. Its real-time design makes it suitable for integration into hospital EHR systems, triage dashboards, and automated early warning platforms, with strong potential to improve patient outcomes and reduce the clinical burden associated with delayed sepsis recognition.

The rapid growth of unstructured clinical text within Electronic Health Records (EHRs), discharge summaries, medication notes, progress reports, and spontaneous drug safety reports has created a valuable but underutilized source of information for pharmacovigilance and clinical decision support. Drug Adverse Events (ADEs) represent a major challenge in healthcare systems, as they may lead to prolonged hospitalization, increased medical costs, treatment discontinuation, and serious patient safety risks. Traditional adverse event detection methods rely heavily on manual review, rule-based systems, and voluntary reporting mechanisms, which are often time-consuming, incomplete, and limited in their ability to capture complex linguistic patterns in clinical narratives. To address these limitations, this paper proposes a transformer-based Natural Language Processing (NLP) model for automated clinical text mining and drug adverse event detection.

The proposed framework leverages domain-specific transformer architectures, including BioBERT, ClinicalBERT, PubMedBERT, and RoBERTa-based clinical language models, to extract contextual semantic representations from heterogeneous clinical documents. The model is designed to identify drug entities, adverse event mentions, dosage-related expressions, temporal cues, negation patterns, and drug–event relationships within free-text clinical narratives. A comprehensive preprocessing pipeline is applied, including text de-identification, tokenization, abbreviation normalization, spelling correction, section segmentation, clinical entity recognition, and handling of negated or speculative statements. To enhance the model’s ability to distinguish true adverse events from unrelated symptoms or historical conditions, the proposed architecture integrates named entity recognition, relation extraction, and sequence classification within a unified multi-task learning framework.

The framework can be trained and evaluated using publicly available biomedical and clinical NLP resources, such as ADE corpora, MIMIC-based clinical notes, and FDA Adverse Event Reporting System data, along with institution-specific annotated clinical records where available. Fine-tuning is performed using supervised learning with class-weighted loss functions and data augmentation strategies to address the imbalance between positive adverse event cases and non-event clinical mentions. In addition, attention visualization and explainability techniques, including SHAP and token-level attribution analysis, are incorporated to highlight clinically meaningful words and phrases contributing to model predictions. These explanations allow clinicians and pharmacovigilance experts to verify whether the model focuses on relevant evidence, such as medication names, temporal associations, symptom descriptions, laboratory abnormalities, and documented treatment reactions.

Experimental evaluation is conducted using standard NLP and pharmacovigilance metrics, including precision, recall, F1-score, area under the receiver operating characteristic curve, and relation extraction accuracy. The proposed transformer-based model is expected to outperform traditional machine learning approaches, rule-based extraction systems, and non-contextual word embedding models by capturing long-range dependencies, domain-specific terminology, and subtle contextual variations in clinical language. Furthermore, the integration of explainability improves transparency and supports safer adoption of automated ADE detection in real-world clinical workflows.

The proposed system provides a scalable, accurate, and interpretable framework for mining clinical text and detecting drug adverse events from large-scale healthcare data. By transforming unstructured clinical narratives into actionable pharmacovigilance insights, the framework can support early safety signal detection, reduce manual review burden, improve medication monitoring, and enhance patient safety. Its transformer-based and explainable design makes it suitable for integration into hospital EHR systems, pharmacovigilance platforms, and clinical decision support applications.

Medical image segmentation plays a fundamental role in computer-aided diagnosis, treatment planning, surgical navigation, radiotherapy guidance, and longitudinal disease monitoring. However, developing robust deep learning segmentation models typically requires access to large and diverse medical imaging datasets collected from multiple healthcare institutions. In real-world clinical environments, direct centralization of medical images is often restricted due to patient privacy regulations, institutional data-governance policies, ethical constraints, and concerns related to data ownership. These limitations are particularly challenging in multi-hospital settings, where imaging data are distributed across heterogeneous sites with different scanners, acquisition protocols, patient populations, and annotation standards. To address these challenges, this paper proposes a federated deep learning architecture for privacy-preserving medical image segmentation across distributed hospitals, enabling collaborative model training without transferring raw patient data outside local clinical institutions.

The proposed framework adopts a decentralized federated learning paradigm in which each participating hospital trains a local segmentation model using its own private medical image repository, while only encrypted model updates or gradient parameters are communicated to a secure central aggregation server. The segmentation backbone is based on advanced encoder–decoder architectures, including U-Net, Attention U-Net, and TransUNet, which are designed to capture both local anatomical structures and global contextual representations. To improve robustness under non-independent and identically distributed data conditions, the proposed architecture integrates adaptive federated aggregation strategies such as FedAvg, FedProx, and FedAdam, allowing the global model to learn from heterogeneous institutional data while reducing client drift and convergence instability.

A comprehensive preprocessing pipeline is incorporated at each hospital site, including image normalization, spatial resizing, noise reduction, intensity standardization, data augmentation, and expert-annotated mask alignment. The framework is designed for multiple segmentation tasks, including brain tumor segmentation from MRI, liver and kidney lesion segmentation from CT scans, and skin lesion segmentation from dermoscopic images. To strengthen privacy protection, the proposed system integrates secure aggregation, differential privacy, gradient clipping, and optional encrypted communication protocols to reduce the risk of sensitive information leakage from shared model updates. Furthermore, the architecture supports asynchronous client participation, allowing hospitals with different computational resources and network conditions to contribute to training without disrupting the global optimization process.

The proposed federated segmentation model is evaluated using multi-institutional and publicly available medical imaging datasets under simulated distributed hospital scenarios. Performance is assessed using clinically relevant segmentation metrics, including Dice Similarity Coefficient, Intersection over Union, sensitivity, specificity, Hausdorff Distance, and boundary F1-score. Experimental evaluation demonstrates that the proposed federated framework achieves segmentation performance comparable to centralized training while preserving institutional data privacy. Compared with standalone local training, the federated model improves generalization across unseen hospital domains and produces more consistent segmentation boundaries under heterogeneous imaging conditions. In addition, privacy-aware training mechanisms show a favorable balance between segmentation accuracy, communication efficiency, and patient data protection.

The proposed architecture provides a scalable, secure, and clinically applicable solution for collaborative medical image analysis across distributed healthcare institutions. By enabling hospitals to jointly train high-performance segmentation models without exposing raw patient images, the framework addresses one of the major barriers to real-world medical AI deployment. Its privacy-preserving design, compatibility with heterogeneous hospital infrastructures, and ability to support explainable and auditable model development make it a promising foundation for next-generation computer-aided diagnosis systems, particularly in multi-center clinical studies and resource-constrained healthcare networks.

Cardiovascular disease (CVD) remains one of the most critical public health challenges worldwide, requiring early identification of high-risk patients to reduce adverse outcomes such as myocardial infarction, stroke, heart failure, and cardiovascular-related mortality. Traditional risk prediction tools, including rule-based scoring systems, provide valuable clinical guidance but often have limited ability to capture complex nonlinear relationships among heterogeneous patient variables stored in Electronic Health Records (EHRs). Moreover, the adoption of advanced machine learning models in cardiovascular medicine is frequently constrained by their black-box nature, which limits clinician trust, transparency, and practical deployment in real-world healthcare settings. This paper proposes an Explainable Artificial Intelligence (XAI)-driven framework for cardiovascular disease risk stratification using structured EHR data, integrating predictive accuracy with clinically interpretable decision support.

The proposed framework utilizes demographic information, vital signs, laboratory test results, medication history, comorbidities, lifestyle-related risk factors, and diagnostic codes extracted from EHR systems to predict patient-specific cardiovascular risk categories, including low-risk, moderate-risk, and high-risk groups. A comprehensive preprocessing pipeline is applied to address missing values, normalize numerical features, encode categorical variables, remove redundant attributes, and mitigate class imbalance using advanced resampling strategies. Several machine learning and deep learning models are investigated, including Random Forest, XGBoost, LightGBM, multilayer perceptron networks, and attention-based neural architectures, with hyperparameter optimization employed to enhance predictive performance and model generalization.

To improve transparency and clinical interpretability, the proposed framework incorporates explainability techniques such as SHAP, LIME, feature importance analysis, and patient-level risk contribution visualization. These methods identify the most influential predictors contributing to cardiovascular risk, including age, systolic blood pressure, cholesterol level, diabetes status, smoking history, body mass index, prior cardiac events, renal function markers, and inflammatory biomarkers. By providing both global explanations for population-level risk patterns and local explanations for individual patient predictions, the framework enables clinicians to understand not only which patients are at elevated risk but also why a specific risk score has been assigned.

The framework is designed for evaluation using publicly available and real-world EHR datasets, such as MIMIC-IV and cardiovascular clinical registries, with model performance assessed through accuracy, sensitivity, specificity, F1-score, calibration analysis, and AUC-ROC. Experimental evaluation is expected to demonstrate that the proposed explainable model outperforms conventional statistical risk scoring approaches and non-explainable machine learning baselines while maintaining clinically meaningful interpretability. The proposed system offers a scalable, transparent, and clinically actionable decision-support tool for early cardiovascular risk stratification, supporting preventive care, personalized intervention planning, and efficient allocation of healthcare resources. Its explainable design makes it particularly suitable for integration into hospital information systems and routine clinical workflows, where trust, accountability, and interpretability are essential for responsible AI adoption.

Breast cancer remains the most prevalent malignancy among women globally, accounting for over 2.3 million new diagnoses annually, making early and accurate diagnosis a decisive factor in reducing mortality and improving therapeutic outcomes. Machine learning-based diagnostic systems, particularly Support Vector Machines (SVMs), have demonstrated considerable promise in automated tumor classification; however, their performance is critically hindered by two persistent challenges: the sensitivity of SVM hyperparameters to suboptimal manual tuning, and the pervasive class imbalance inherent in real-world clinical datasets, where malignant cases represent a significantly smaller proportion than benign instances. This paper proposes a novel Swarm Intelligence-Optimized Support Vector Machine (SI-SVM) framework that addresses both challenges simultaneously through an integrated pipeline combining advanced metaheuristic optimization with robust class-balancing strategies for accurate breast cancer diagnosis. Specifically, a hybrid optimization mechanism fusing Particle Swarm Optimization (PSO) and the Grey Wolf Optimizer (GWO) is employed to simultaneously perform feature selection, SVM kernel selection, and hyperparameter tuning — including regularization parameter C and kernel coefficient γ — within a unified search space, thereby circumventing the computational burden of exhaustive grid search. To mitigate the class imbalance problem, a combination of Synthetic Minority Oversampling Technique (SMOTE) and Edited Nearest Neighbors (ENN) — collectively referred to as SMOTEEN — was applied during preprocessing to generate a balanced and denoised training distribution that preserves the statistical integrity of minority class samples. The proposed framework was rigorously evaluated on three benchmark clinical datasets: the Wisconsin Breast Cancer Diagnostic Dataset (WBCD), the SEER Breast Cancer Dataset, and the MIAS mammographic image dataset, collectively encompassing over 100,000 clinical records with diverse feature representations including cytological characteristics, hormonal biomarkers, and imaging descriptors. A stratified 10-fold cross-validation protocol was adopted to ensure unbiased performance estimation across all experimental configurations. Experimental results demonstrate that the proposed SI-SVM framework achieves a classification accuracy of 99.1%, a sensitivity of 98.7%, a specificity of 99.4%, an F1-score of 0.989, and an AUC-ROC of 0.997 on the WBCD dataset — significantly outperforming standalone SVM, PSO-SVM, GWO-SVM, Random Forest, and XGBoost baselines. Furthermore, SHAP-based feature importance analysis identified the most diagnostically discriminative clinical features — including bare nuclei, clump thickness, and uniformity of cell shape — providing transparent and actionable insights aligned with established oncological knowledge. The proposed framework offers a computationally efficient, generalizable, and clinically interpretable solution for automated breast cancer screening, with strong potential for integration into computer-aided diagnosis (CAD) systems in resource-constrained medical environments.

Brain tumors represent one of the most aggressive and life-threatening neurological malignancies, with accurate and timely classification being paramount for determining appropriate treatment strategies and improving patient prognosis. Conventional diagnostic approaches rely predominantly on radiologist interpretation of Magnetic Resonance Imaging (MRI) scans, a process that is inherently subjective, time-intensive, and highly dependent on clinical expertise. Existing deep learning models, while promising, largely exploit single-modality image data while neglecting the rich complementary diagnostic information embedded in structured clinical metadata — including patient demographics, tumor biomarkers, symptom onset duration, and histopathological reports. This paper proposes a novel Multi-Modal Deep Learning framework that synergistically fuses MRI-derived spatial features with structured clinical metadata to achieve robust and clinically meaningful brain tumor classification across four primary categories: Glioma, Meningioma, Pituitary Adenoma, and No Tumor. The imaging branch employs an ensemble of fine-tuned pre-trained CNN architectures — EfficientNet-B5, DenseNet-121, and Vision Transformer (ViT) — to extract hierarchical spatial and contextual features from multi-sequence MRI modalities (T1, T2, T1-contrast, and FLAIR). Simultaneously, the clinical metadata branch utilizes a fully connected neural network with attention mechanisms to encode structured patient-level features. The outputs of both branches are fused via a late-fusion cross-attention mechanism that learns to dynamically weight the contribution of each modality depending on input context. The integrated framework was trained and evaluated on the BraTS 2023 benchmark dataset augmented with clinical metadata records from The Cancer Genome Atlas (TCGA), comprising over 12,000 annotated multi-sequence MRI volumes. Extensive data augmentation, class balancing via ADASYN, and k-fold cross-validation strategies were employed to ensure robust generalization. Experimental results demonstrate that the proposed multi-modal fusion framework achieves a classification accuracy of 98.3%, a macro-averaged F1-score of 0.981, a sensitivity of 97.6%, and an AUC-ROC of 0.994 — substantially outperforming single-modality CNN baselines and state-of-the-art methods including ResNet-50, VGG-19, and standard ViT models. Additionally, Grad-CAM++ and LIME-based explainability maps were generated to localize tumor regions and validate clinically relevant decision boundaries, bridging the gap between black-box deep learning predictions and radiological interpretability. The proposed framework establishes a scalable and generalizable pipeline for automated brain tumor diagnosis, with direct applicability to neuro-oncology clinical workflows and AI-assisted radiology platforms.

Accurate prediction of mortality rates among Intensive Care Unit (ICU) patients is a critical challenge in modern clinical medicine, as timely risk stratification enables physicians to prioritize interventions, allocate limited resources effectively, and improve patient survival outcomes. Despite significant advances in machine learning-based prognostic models, existing approaches often suffer from limited generalization, sensitivity to missing clinical data, and insufficient exploitation of temporal dependencies inherent in physiological time-series records. This paper proposes a hybrid optimization framework that integrates Recurrent Neural Networks (RNNs) — specifically Long Short-Term Memory (LSTM) and Bidirectional LSTM (BiLSTM) architectures — with metaheuristic optimization algorithms, including Particle Swarm Optimization (PSO) and the Grey Wolf Optimizer (GWO), to enhance the predictive accuracy and convergence efficiency of ICU mortality models. The proposed framework was trained and validated on the publicly available MIMIC-III clinical database, encompassing over 53,000 ICU admissions with multivariate time-series features including vital signs, laboratory results, demographic data, and clinical severity scores (SOFA and APACHE II). A comprehensive preprocessing pipeline was applied to handle missing values, normalize temporal features, and address severe class imbalance through synthetic minority oversampling (SMOTE). The hybrid optimization mechanism was employed to automatically tune critical hyperparameters — including learning rate, number of hidden layers, dropout rates, and batch size — replacing conventional manual tuning and grid search strategies. Experimental evaluations demonstrate that the proposed hybrid model achieves an AUC-ROC of 0.967, a sensitivity of 94.3%, a specificity of 95.8%, and an F1-score of 0.943, significantly outperforming standalone LSTM, standard RNN, and traditional machine learning baselines such as Random Forest and XGBoost. Furthermore, SHAP (SHapley Additive exPlanations) analysis was incorporated to identify and rank the most influential clinical predictors of ICU mortality, providing transparent and clinically interpretable insights to support physician decision-making. The proposed framework demonstrates robust performance across diverse ICU subpopulations — including cardiac, respiratory, and surgical patients — underscoring its potential as a generalizable and deployable clinical decision support tool in real-world intensive care settings.

Diabetic retinopathy (DR) remains one of the leading causes of preventable blindness Worldwide, with its prevalence continuing to rise in parallel with the global diabetes epidemic. Early and accurate detection is critical to preventing irreversible vision loss; however, manual screening by ophthalmologists is time-consuming, costly, and subject to inter-observer variability. This paper presents a deep learning-based framework for the automated early detection and grading of diabetic retinopathy from fundus retinal images, leveraging the complementary strengths of Convolutional Neural Networks (CNNs) and Transfer Learning. The proposed framework employs pre-trained CNN architectures — including ResNet-50, InceptionV3, and EfficientNet-B4 — fine-tuned on a large-scale, publicly available retinal image dataset (APTOS 2019 and EyePACS) comprising over 88,000 annotated fundus photographs across five severity grades (No DR, Mild, Moderate, Severe, and Proliferative DR). To address class imbalance, advanced data augmentation strategies and class-weighted loss functions were incorporated during model training. The transfer learning approach significantly reduced training complexity while enhancing generalization performance on unseen clinical data. Experimental results demonstrate that the proposed framework achieves a classification accuracy of 97.8%, a sensitivity of 96.4%, a specificity of 98.1%, and an AUC-ROC score of 0.991 on the held-out Test set, outperforming several state-of-the-art baseline methods. Furthermore, Grad-CAM visualization was integrated to provide explainability by highlighting discriminative retinal lesion regions — including microaneurysms, hemorrhages, and neovascularization — thereby enhancing clinical interpretability and physician trust. The proposed system demonstrates strong potential as a scalable, cost-effective, and reliable decision-support tool for DR screening programs, particularly in resource-limited healthcare settings where access to specialized ophthalmology expertise is restricted.