Online Fault Diagnosis Method for Photovoltaic Arrays Based on Temporal Feature Learning

Abstract

To address the susceptibility of fault features to environmental fluctuations, insufficient use of temporal information in continuous operation, and class imbalance in online photovoltaic (PV) array fault diagnosis, a CNN-GRU-Attention-based method is proposed. First, relative-ratio features, an irradiance-normalized DC power feature, and derived meteorological features are constructed from the original DC-side electrical and meteorological variables and combined with the original measurements to form time-window samples. A class-sensitive sliding-window strategy adopts state-specific strides to increase the sampling density of fault segments while reducing redundant normal samples. Second, a CNN extracts local temporal variations, a bidirectional GRU models dynamic dependencies within each window, and temporal attention emphasizes critical time points; class-weighted cross-entropy further improves the learning of fault classes with fewer samples. Finally, validation using actual continuous operational data from a PV power plant yields 97.99 percent accuracy, 99.08 percent macro-averaged recall, and 98.61 percent macro-averaged F1-score, outperforming all compared models and demonstrating effective identification of PV array operating states under complex operating conditions.