Interpretable ultra-short-term photovoltaic power forecasting with multi-scale temporal modeling and variable-wise attention

Accurate photovoltaic (PV) power forecasting is essential for grid operation but remains difficult due to nonlinear multi-scale dynamics and seasonal distribution shifts. This work presents MKAN-iTransformer, a cascaded framework that integrates two existing components—the Multi-Scale Kolmogorov–Arnold Network (MKAN) for scale-aware temporal representation learning and iTransformer for variable-wise attention and inter-variable dependency modeling—under a 15-minute single-step setting. Experiments on a real-world 30 MW PV plant dataset from the Chinese State Grid Renewable Energy Generation Forecasting Competition use chronological splits within each season. MKAN-iTransformer achieves the best overall performance in spring, autumn, and winter. In spring, it reaches MSE=2.892, RMSE=1.701, MAE=0.864, and $${R^{2}}=0.947$$, improving over LSTM by 23.5%/12.5%/20.5% (MSE/RMSE/MAE). In autumn, it attains MSE=2.884, RMSE=1.698, MAE=0.774, and $${R^{2}}=0.962$$, reducing errors vs. iTransformer by 16.5%/8.7%/12.4%. In winter, it achieves MSE=1.721, RMSE=1.312, MAE=0.443, and $${R^{2}}=0.969$$, yielding 81.6%/57.1%/71.9% error reductions vs. Transformer. Ablation further confirms the complementarity between MKAN and iTransformer and shows that direct KAN integration can be unstable under winter shifts (KAN-iTransformer: MSE=7.082, $${R^{2}}=0.872$$).