Data-driven short-term photovoltaic power forecasting under extreme weather conditions using GRU-KAN model

The increasing frequency of extreme weather events can significantly reduce PV generation, threatening the security of the power system. To address this issue, this study proposes a data-driven short-term PV forecasting framework based on GRU-KAN model to enhance forecasting accuracy. First, historical PV data are preprocessed and key meteorological factors are extracted via Grey Relational Analysis. Second, normal weather conditions are clustered into sunny, cloudy, and overcast categories using K-Means++. Third, a data-driven method is used to identify extreme weather conditions characterized by limited data, and a similar-day selection method is employed to construct high-quality training sets. Fourth, a GRU-KAN model which integrates temporal modeling capability of GRU with the nonlinear approximation power of KAN is proposed for photovoltaic forecasting under both normal and extreme conditions. Finally, the study simplifies the neuron structure of the feature extraction layer to enhance forecasting performance when data are limited. Experimental results demonstrate that the proposed model achieves the lowest average RMSE of 3.86 and the highest average R2 of 96.90 % under normal conditions. In addition, the model achieves the lowest RMSEs of 0.900, 0.950, 1.157, and 1.438 across four representative extreme days.