Fault diagnosis framework for photovoltaic system using fully connected neural network-based back surface thermal analysis

Photovoltaic (PV) systems play a vital role in advancing sustainable energy, but their performance and reliability are often compromised by undetected internal faults. This study presents a fault detection and classification framework that utilizes back-surface thermal data from the junction box along with voltage measurements to identify common faults such as short circuits, ground faults, and open connections. Unlike conventional surface thermography, this method addresses the limitations posed by varying environmental conditions, enabling consistent detection of hidden anomalies. The framework is implemented using a fully connected neural network enhanced with batch normalization and ReLU activation functions, ensuring efficient processing of complex thermal and electrical data. Experimental validation demonstrates an accuracy of 99.4% across key metrics, including precision, recall, and F1-score, even under conditions of environmental noise. The proposed approach not only ensures high reliability but is also scalable for real-time deployment in large PV installations. It reduces reliance on specialized diagnostic equipment and supports proactive fault identification.