SolarFM: Spatio-temporal solar irradiance forecasting based on time series foundation model

Accurate solar irradiance prediction plays a pivotal role in optimizing photovoltaic system operations and enhancing grid integration efficiency. Existing deep learning methods use diverse encoder architectures for spatiotemporal irradiance forecasting. However, they often fail to effectively decouple cross-channel temporal dependencies, limiting their capacity to capture complex irradiance variation patterns. To overcome this challenge, SolarFM, is introduced as an innovative spatiotemporal forecasting framework inspired by time series foundation models. The proposed method begins with zero-shot forecasting using localized meteorological sequences. It capitalizes on foundation models’ generalization ability, gained through pretraining on vast temporal data, to correlate irradiance patterns with weather variables. A dual-branch spatiotemporal encoder is subsequently designed to synergistically process historical observations and forecasted sequences, where a tensor factorization network extracts high-dimensional temporal representations from future sequence projections. The framework ultimately integrates multimodal features through adaptive fusion to generate synchronized multi-region irradiance predictions. Comprehensive evaluations on public datasets demonstrate that SolarFM achieves 2.6%–15.1% accuracy improvements in single-step and multi-step forecasting tasks compared to baseline methods. Furthermore, systematic experiments reveal remarkable robustness under diverse meteorological conditions and superior performance consistency in high-irradiance scenarios, establishing its potential as a paradigm-shifting solution for solar energy forecasting systems.