MS-Mixer: A multiscale hybrid model for solar irradiance forecasting

Solar energy, due to its clean and renewable characteristics, plays an increasingly important role in the global energy supply system. High-precision forecasting of solar irradiance can not only improve the reliability and efficiency of solar power systems but also provide important references for grid scheduling and energy management. However, existing deep learning methods struggle to effectively decouple macro-trends and micro-fluctuations across different time scales, and typically model trends and seasonal patterns independently, resulting in underutilization of key information. To address this issue, this paper proposes a novel multi-scale mixed model (MS-Mixer) for solar irradiance forecasting. The model first enhances inter-component interactions to strengthen seasonal-trend fusion. Then, it applies multi-scale frequency filtering to integrate features across different scales. Finally, it jointly models temporal and meteorological variables to capture complex temporal and cross-variable dependencies. Experimental results on three public datasets verify the superior performance of the proposed model. Compared with the second-best baseline model, iTransformer, MS-Mixer achieves average RMSE reductions of 2.54%, 10.81%, and 6.80% on the three datasets. The results show that MS-Mixer provides a novel and effective solution for high-accuracy solar irradiance orecasting.