Machine learning diagnosis (ANN vs KNN) of short-circuit, open-circuit, partial shading and line-to-ground faults in PV modules: Case study of Bejaia, Algeria

To ensure optimal energy production and operational safety in solar installations, the implementation of dependable fault identification mechanisms is critical, especially for smaller arrays vulnerable to local weather fluctuations. This research details a diagnostic approach using Artificial Neural Networks (ANN) for a 2×2 series-parallel PV configuration, simulated within the MATLAB/Simulink environment (Simscape Electrical) and driven by ERA5 meteorological records for Bejaia, Algeria. Four electrical fault types were considered: short-circuit, open-circuit, partial shading, and line-to-ground. Time-series simulation outputs were resampled to a one-minute grid and processed to produce a balanced dataset of 2,128 labeled samples. The primary classifier is a feedforward artificial neural network with one hidden layer of 10 neurons; a K-nearest neighbors (KNN) model (baseline, default parameters) was used for comparison. The ANN achieved an overall accuracy of =99.6% (independent test =99.7%) and a fault detection rate of =99.2%, with faults detected within the sampling interval and stable classification during dynamic two-day fault scenarios. Compared to KNN, the ANN exhibited fewer boundary misclassifications and no false positives in the reported experiments. Limitations include reliance on simulation driven by reanalysis data and lack of compound-fault cases; future work will validate the approach on measured field data and investigate richer feature sets and deployment on embedded hardware.