Abstract:

Aiming at the problems of insufficient prediction accuracy and poor seasonal adaptability of the existing methods in the temperature prediction of photovoltaic modules, this paper proposes a hybrid prediction method that integrates a transient thermal model with a TRANSFORMER. This method, by integrating physical mechanisms with data-driven modeling, significantly improves the performance while enhancing the seasonal adaptability and robustness. Initially, a transient integrated heat transfer model for the PV module-ventilation channel-roof system is established based on the energy balance equation, and the transient thermal process between the PV module and its surrounding environment is solved. Subsequently, the calculated backplane temperature and ventilation channel temperature are used as feature input for the TRANSFORMER, which is jointly trained with meteorological data to correct the systematic errors of the transient model and enhance the nonlinear time-series modeling capability. The prediction performance of empirical models, theoretical models, neural network models, and the hybrid model is reviewed. The results indicate that the proposed hybrid model reduces the root mean square error (RMSE) and mean absolute error (MAE) in photovoltaic module temperature prediction by 25.81–66.24% and 10.53–65.63%, respectively, compared to the neural network model without physical feature inputs. This method can provide more reliable support for the precise prediction of photovoltaic power and the optimization and adjustment of module operation strategies.

Key words: photovoltaic module temperature, hybrid model, unsteady state model, feature fusion, short-term forecast