Short-term photovoltaic power prediction based on KAN model

Photovoltaic (PV) power generation has developed rapidly as a promising clean energy application in recent years. Reliable PV prediction contributes to the stable operation of the power system and promotes the widespread application of PV power. In this paper, a hybrid model, which combines a new deep learning framework Kolmogorov-Arnold Network (KAN) with traditional models (such as long short-term memory(LSTM), gate recurrent unit (GRU), and bi-directional long short-term memory (BiLSTM)), is proposed. The hybrid model not only combines the advantages of LSTM, GUR, and BiLSTM in processing time-series data, but also incorporates the faster neural scaling features of KAN networks, which result in faster performance improvement along with increasing the model size. To verify the prediction performance of the hybrid model, three evaluation indices are applied: mean absolute error(MAE), root mean square error(RMSE), and mean absolute percentage error(MAPE). The case study demonstrates that compared to non-improved benchmark models including long short-term memory (LSTM), forecasting errors of the LSTM-KAN-V2 model decrease by 33.45% (MAE), 21.65% (MAPE), and 37.23% (RMSE), respectively.