Fault Diagnosis for Grid-Connected Photovoltaic Systems Based on Digital Twins

In the fault diagnosis of grid-connected photovoltaic (PV) systems, practical applications face challenges such as difficulty in fault data collection, data imbalance, and poor diagnostic accuracy, which severely impact the reliability of PV system operation. To address these challenges, this paper proposes a fine-grained fault diagnosis framework for grid-connected PV systems based on digital twins (DT-FGFD). The framework comprises three main stages: First, digital twin models of the gridconnected PV system are constructed, expanding the fault database from both mechanistic and data-driven perspectives. Second, one-dimensional time-series data from the PV system is preprocessed and transformed into two-dimensional timefrequency scale maps using Continuous Wavelet Transform (CWT), facilitating adaptation to two-dimensional convolutional neural networks for more accurate fault diagnosis. Finally, a finegrained fault diagnosis method is introduced, which utilizes ResNet50 to extract time-frequency features and semantic features, merging these two feature sets into fine-grained characteristics for precise fault classification. The proposed method achieves a fault classification accuracy of 99.17%, demonstrating the advantages of fine-grained features in fault diagnosis.