Performance estimator of photovoltaic modules by integrating deep learning network with physical model

Knowing whether the real-time performance of photovoltaic (PV) modules during operation is consistent with the expected conditions is crucial for managers to develop operation and maintenance strategies. Therefore, in this study, a PV module performance estimator by integrating deep learning network with physical model was proposed. Firstly, the original I-V curves were regionally divided and extracted according to the two-dimensional space composed of solar irradiance and module temperature. The training database was constructed by the enhanced Lévy flight bat algorithm, and two data cleaning methods were applied to eliminate anomalous data. Then, the constructed CNN-LSTM network was connected in series with the physical model of PV modules to form a performance estimator. This combines PV field knowledge with deep learning network to enable the system to have physically data-effective, generalizable, and interpretable features. Meanwhile, multi-output learning and Bayesian optimization theory were applied in optimization and hyperparameter tuning of CNN-LSTM network, respectively. Finally, for the PV modules in the NREL dataset, the physical model parameter estimation capability, performance parameter estimation capability, and fault diagnosis capability of the system were simulated and validated. The performance estimator can be used for low-cost performance monitoring and fault diagnosis of PV power plants.