Photovoltaic power forecasting algorithm based on LSTM-KAN hybrid architecture

Accurate photovoltaic (PV) power generation forecasting plays a pivotal role in enabling optimal grid dispatch operations by providing essential technical references for power system scheduling. To address the challenges of complex meteorological patterns and sudden environmental variations, a photovoltaic power forecasting model named the LSTM-KAN (Long Short-Term Memory-Kolmogorov-Arnold Network) Hybrid Architecture is proposed, with an attention mechanism introduced to enhance model robustness. To investigate the correlation between Numerical Weather Prediction (NWP) features and power output, a multi-dimensional feature selection mechanism was integrated into the data preprocessing phase. To validate the superior performance of the proposed model, photovoltaic operational data from the Alice Springs region in Australia was employed for experimental verification, with comparative analysis conducted against conventional LSTM models in terms of both prediction error and operational stability. Experimental results demonstrated that the LSTM-KAN model achieves 12.04% and 13.26% reductions in MAE and RMSE respectively compared to the baseline LSTM model when evaluated on the Dual dataset.