AI-DRIVEN THERMAL INTELLIGENCE: A DEEP LEARNING FRAMEWORK FOR REAL-TIME THERMAL MANAGEMENT IN HIGH-ENERGY DENSITY ELECTRIC VEHICLE BATTERY SYSTEMS

The rising global demand for sustainable energy has accelerated the adoption of thermal photovoltaic (PV) systems as a clean and renewable power source. These systems rely on continuous operation and optimal performance to ensure efficient energy generation. However, PV systems are susceptible to various faults, such as Maximum Power Point Tracking (MPPT) failures, Low Power Point Tracking (LPPT) issues, partial shading, and hardware degradation, which can significantly reduce their efficiency and lifespan. Therefore, accurate and timely fault detection and classification have become essential for enhancing system reliability and minimizing maintenance costs. Traditionally, fault detection in EV batteries has relied on manual inspections and threshold-based monitoring using Supervisory Control and Data Acquisition (SCADA) systems. These approaches are often limited by their inability to handle large volumes of high-frequency data and detect subtle or complex fault patterns. They are time-consuming, errorprone, and lack scalability, especially in largescale or distributed solar installations. The need for intelligent, automated, and scalable fault detection mechanisms has led to the exploration of machine learning (ML) techniques in this domain. This project proposes an AI-driven fault detection and classification system that leverages highfrequency operational data from EV batteries and machine learning algorithms to identify and categorize faults automatically. The system incorporates a complete pipeline that includes data preprocessing, class balancing using the SMOTE algorithm, feature reduction using Principal Component Analysis (PCA), and model training using Light Gradient Boosting Machine (LGBM), CatBoost Classifier, and a proposed Random Forest Classifier (RFC). A user-friendly GUI, built with Tkinter, allows seamless interaction with the system and facilitates real-time visualization of predictions and performance metrics. Extensive evaluation shows that the proposed RFC model outperforms other models with an accuracy of 99.82%, precision of 99.86%, recall of 99.86%, and F1-score of 99.86%, compared to the LGBM and CatBoost models, which achieved accuracies of 82.39% and 82.48% respectively. These results highlight the effectiveness of the proposed system in delivering a robust, efficient, and scalable solution for fault detection in thermal photovoltaic applications