A BiLSTM-based digital twin model for photovoltaic strings under current mismatch condition

Abstract

The reliability of photovoltaic (PV) systems is increasingly challenged by string-level faults affecting both performance and safety. To address this issue, this study proposes a four-layer digital twin (DT) framework for intelligent monitoring and fault diagnosis of PV strings under mismatch conditions. In the virtual layer, the Sandia Array Performance Model and the Perez model are employed to estimate module temperature and plane-of-array irradiance, which are then input into a bidirectional long short-term memory (BiLSTM) network for current prediction. To enhance adaptability, a solar-elevation-based Current Mismatch Ratio (CMR) is introduced as an auxiliary correction factor, enabling dynamic modeling of mismatch behavior. The CMR-assisted BiLSTM achieves a root mean square error (RMSE) of 0.4306 and a coefficient of determination ($R^2$) of 0.9594, demonstrating high predictive accuracy. In the decision layer, a sliding-window mechanism combined with a support vector machine classifier distinguishes bypass diode short-circuit faults from mismatch phenomena using statistical features of $R^2$ and RMSE. Validation based on operational data from actual PV power plants shows that the proposed DT-based approach achieves an accuracy of 96.76%, precision of 93.39%, recall of 97.96%, and an F1-score of 95.63%, outperforming traditional reference string–based methods by 1.22%, 3.12%, and 1.59% in accuracy, precision, and F1-score, respectively. These results confirm that the proposed DT framework provides real-time fault diagnosis and predictive maintenance, significantly improving the operational reliability of PV systems under dynamic environmental conditions.