HyMTSolar: A Hybrid Mamba-transformer Method for Enhancing Day-ahead Solar Irradiance Forecasting with Satellite Images

Accurate day-ahead solar irradiance forecasting is crucial for grid stability and photovoltaic optimization. This work introduces HyMTSolar (Hybrid mamba-transformer for Solar), a multi-modal framework that integrates ground-based meteorological time series with satellite imagery. By harmonizing the efficient sequence modeling of mamba with the global receptive field of transformer, the model captures complex temporal dependencies, thereby significantly enhancing the accuracy of solar irradiance forecasting. A physics-aware rotary position embedding scheme is introduced, which explicitly encodes geographic coordinates to capture cloud dynamics. Experiments on three Baseline Surface Radiation Network stations demonstrate that HyMTSolar outperforms competitive deep learning baselines. Specifically, it reduces mean absolute error by 5.4% to 25.6% compared to the strongest competitor and exhibits superior robustness under varying weather conditions.