Research on Intelligent Fault Diagnosis and Monitoring of Photovoltaic Arrays Based on Data-Driven and Probabilistic Neural Networks

Aiming at the problems of limited diagnostic accuracy and insufficient result credibility caused by environmental noise and complex working conditions in photovoltaic array fault diagnosis, this study proposes an intelligent diagnosis method integrating data-driven signal processing and improved probabilistic neural networks. Signal enhancement is performed through variational mode decomposition and adaptive filtering, and a high-dimensional robust feature set is constructed by combining multi-domain feature fusion; a probabilistic neural network integrating an attention mechanism and a local density-aware adaptive smoothing factor is designed, and the evidence theory loss function is used to optimize its probability calibration. On the test set containing real power station data, the average diagnostic accuracy of this method reaches 98.7%, which is 2.1 percentage points higher than the optimal baseline model; the accuracy remains 91.2% in a 0dB strong noise environment; the expected calibration error of its probability output is 0.018, which is 65% lower than that of the standard probabilistic neural network. The results show that the proposed method can effectively improve diagnostic accuracy and noise robustness, and provide a highly reliable probabilistic fault decision basis.