Multi-Scale Gated Aggregation for solar irradiance forecasting: A general fusion method for intra-modal and cross-modal scenarios

Artificial intelligence-based solar irradiance forecasting is crucial for grid-integrated photovoltaic systems. While current methods range from textual time-series models to multimodal approaches combining text with sky images, they often grapple with high computational costs, limited feature extraction, and inefficient multimodal fusion. To address these issues, this paper proposes a novel Multi-Scale Gated Aggregation (MSGA) method—a unified, linear-complexity fusion mechanism applicable to both intra-modal (e.g., endogenous–exogenous variables) and cross-modal (e.g., text–image) scenarios. Instantiating this method, we present the Multi-Scale Gated Aggregation Transformer (MSGAformer) model, a Transformer-based time-series model centered on two dedicated components: the MSGA module for efficient dependency modeling and fusion, and the Dual-Gated Feature Synergy (DGFS) module for adaptive feature refinement. Experiments demonstrate that MSGAformer consistently outperforms state-of-the-art models, reducing root mean square error (RMSE) by 1.8%–3.8% across four ultra-short-term horizons (10–60 min). In multimodal settings, MSGA-based fusion also surpasses conventional dual cross-modal attention in accuracy and robustness. These results validate MSGA as an effective fusion paradigm, offering reliable guidance for future advancements in solar irradiance forecasting and broader multimodal time-series applications.