Performance Optimization of Machine-Learning Algorithms for Fault Detection and Diagnosis in PV Systems

Abstract

The early detection of faults in photovoltaic (PV) systems is crucial for ensuring efficiency, minimizing energy losses, and extending operational lifespan. This study evaluates and compares multiple machine-learning models for fault diagnosis in PV systems, analyzing their performance across different fault types and operational modes. A dataset comprising 2.2 million measurements from a laboratory-based PV model, covering seven fault categories—including inverter failures, partial shading, and sensor faults—is used for training and validation. Models are assessed under both Maximum Power Point Tracking (MPPT) and Limited Power Point Tracking (LPPT) conditions to determine their adaptability. The results indicate that the ensemble bagged tree classifier achieves the highest accuracy (92.2%) across all fault scenarios, while neural network-based models perform better under MPPT conditions. Additionally, the study highlights variations in model performance based on power mode, suggesting the potential for adaptive diagnostic approaches. The findings reinforce the feasibility of machine learning for predictive maintenance in PV systems, offering a cost-effective, sensor-free method for real-time fault detection. Future research should explore hybrid models that dynamically switch between classifiers based on system conditions, as well as validation using real-world PV installations.