This paper presents a simple yet highly accurate method for classification of three PV faults—Partial Shading (PS), Line-to-Line (LL), and Open-Circuit (OC). Each of these faults requires different maintenance strategies. While PS faults typically self-resolve without human actions, LL and OC faults can cause permanent damage if not addressed promptly. The proposed method begins by generating residual voltage and current signals through subtraction of the expected output calculated from a reference module under normal operating conditions. This removes the effect of natural irradiation variations. After that, an adaptive Discrete Wavelet Transform (DWT) (db4, level 5) eliminates measurement noise using an energy-based threshold. This threshold is calculated from D1 coefficients, preserving genuine fault transients regardless of noise level. From the denoised residuals, 36 features (12 time-domain and 24 timefrequency) are extracted, and the five most discriminative features are selected using one-way ANOVA (p < 0.001), which are fed to a machine learning method, a decision tree classifier. The resulting model uses only four decision nodes and achieves 100 % accuracy on training scenarios and 98.44 % accuracy on unseen test data. This white-box machine learning has low computational burden and full interpretability.

Keywords