Performance Evaluation of Machine Learning Models for Fault Diagnosis in PV Systems

This study proposes a methodology for performance evaluation of fault diagnosis in photovoltaic (PV) systems using machine learning techniques, including Random Forest, k-Nearest Neighbors (kNN), and Naive Bayes. Actual data were acquired from a 3kW PV testbed and categorized into eight classes representing normal and significant fault conditions. For each class, a confusion matrix and corresponding relevant metrics are utilized to conduct class-specific performance assessments. The results indicate that the Random Forest model outperforms other models in terms of accuracy, precision, recall, and F1 score. It demonstrates outstanding performance for normal conditions (Class 0 and Class 7) and maintains stable performance for major fault types (Class 1 and Class 6). While the kNN model delivers acceptable performance for Class 0, it shows limitations for certain fault types. The Naive Bayes model exhibits the lowest performance and faces significant challenges in accurately handling most fault types.