Abstract Reliable fault detection in photovoltaic systems is critical for maximizing energy yield, ensuring operational safety, and enabling scalable deployment of large photovoltaic assets. This paper presents a hybrid deep learning framework for string-level photovoltaic fault detection and classification under real operating conditions. A comprehensive dataset was collected from a six-string photovoltaic test facility at the Technical University of Denmark, comprising 24-module strings monitored over extended periods (≥3 months per condition) under six operating states: normal operation, partial shading, bypass diode failure, open circuit, short circuit, and degradation. Five synchronized electrical and environmental features (direct-current voltage, direct-current current, direct-current power, irradiance, and temperature) were used as model inputs. The proposed architecture combines convolutional neural network layers for structured feature extraction with gated recurrent units to capture temporal photovoltaic fault dynamics. When evaluated on a strictly unseen test dataset, the model achieved an overall accuracy of 94%, with precision, recall, and F1-score of 0.93, 0.92, and 0.92, respectively. Perfect classification (F1 = 1.0) was obtained for open- and short-circuit faults, while challenging, highly correlated conditions such as shading, bypass diode failure, and degradation were detected with F1-scores of 0.90, 0.87, and 0.78, respectively.
Dhimish et al. (Sat,) studied this question.