Robust Classification of Electrical Faults in Photovoltaic Systems

Photovoltaic (PV) faults can arise from electrical, physical, or environmental issues, making accurate current-voltage analysis crucial for detecting electrical faults under varying conditions. A labeled dataset containing current, voltage, irradiance, temperature, and weather information was used to train machine learning (ML) models for fault detection. Distinguishing between normal and faulty currents under low irradiance was challenging. Principal component analysis (PCA) improved visualization, showing that 3 principal components capture nearly $99 \%$ of the data’s information. Feature mutual information highlighted current values in the PV array as the most critical features. While traditional ML models, such as support vector machines (SVM), achieved $99.6 \%$ accuracy, they struggled to fully differentiate between normal and line-to-line faults, potentially leading to power losses. A probabilistic neural network (PNN) provided utterly accurate predictions for electrical faults along with other classical ML techniques with nearly perfect prediction. The impact of dataset size reveals that 25% of the total data sample is sufficient to train a machine learning model for accurately predicting faults.