From reactive to proactive: A physically interpretable multi-site framework for very short-term solar forecasting

With the accelerating global deployment of solar photovoltaic (PV) systems, accurate very short-term solar irradiance forecasting has become critical for grid operation and real-time dispatch. However, conventional single-site models are inherently reactive and show pronounced prediction delays, while existing multi-site studies are often constrained by coarse spatiotemporal resolutions (e.g., 15-minute intervals over regional scales), which fail to capture micro-scale cloud-driven irradiance transients and lead to spatiotemporal mismatch. To address these limitations, this study proposes and evaluates a micro-scale multi-site layout (1–2 km spacing). By integrating high-resolution (5-sec) measurements of Global Horizontal Irradiance (GHI) and meteorological variables from three designed sites in Nottingham, UK, with both machine learning and deep learning models evaluated under single-site and multi-site configurations. Results demonstrate that wind-driven irradiance transients are captured, bridging the gap between statistical modeling and physically interpretable forecasting. The spatiotemporal information provided by the proposed multi-site layout is a more significant driver of forecasting accuracy than algorithmic complexity. Multi-site models consistently outperform the single-site LSTM baseline across all forecasting horizons, achieving a maximum forecast skill improvement of 6.97%. Typical day dynamics analysis further confirms that the CNN-LSTM and Dynamic GCN-LSTM models effectively reduce the response delay during periods of highly variable irradiance driven by cloud transients by leveraging spatiotemporal information. Furthermore, the framework exhibits strong operational feasibility, with all models achieving inference times below 2.0 ms. Overall, this work provides a scalable and physically interpretable framework for improving very short-term forecasting and enhancing power system resilience to rapid solar variability.