Day-ahead solar irradiance forecasting is critical for grid stability and renewable energy integration, yet existing methods often struggle to generalize across diverse climate regions without site-specific retraining. This study introduces the Similarity-based Forecasting Model (SIM), a day-ahead solar irradiance forecasting framework with early-day observational adjustments, built around a novel multi-dimensional Spatio-Temporal Index (STI) that characterizes cloud transition patterns, velocity fields, and trajectory dynamics to identify historically analogous atmospheric states from decade-scale geostationary satellite archives at 0. 5–2 km resolution. The STI captures dynamically evolving cloud motion and provides a geographically universal heuristic capable of global deployment without regional tuning. The model is validated across globally distributed BSRN ground stations during 2020, where SIM achieves an RMSE of 54. 6 W / m 2 and MAE of 31. 4 W / m 2, corresponding to 15%–30% lower errors than numerical weather prediction, time-series, cloud-motion-vector and persistence baseline approaches. The framework demonstrates consistent performance across cloud regimes, seasons, and climate zones (from polar to tropical environments). These results align with the broader trajectory of data-driven atmospheric modeling: recent advances such as Google DeepMind’s GraphCast corroborate the efficacy of similarity-based approaches at large scales, while simultaneously highlighting the need for domain-specific methods capable of resolving the fine spatial and temporal variability required for solar energy prediction. To quantify practical significance, order-of-magnitude estimates were made using a scenario-based analysis with global photovoltaic capacity of ≈ 1600 GW, low solar penetration (5%–15%), and a conservative 20% reduction in forecast error from SIM: indicating yields up to ≈ 1. 5 billion and ≈ 24 Mt CO 2 in annual cost and emissions savings. Together, these results establish SIM as a physically grounded, purpose-built framework for solar energy forecasting that achieves competitive accuracy across diverse conditions while remaining fully transferable across geographies without retraining. SIM thereby is a scalable, operationally deployable foundation well-suited for integration into next-generation global solar forecasting infrastructure alongside existing methods. • A multi-dimensional Spatio-Temporal Index (STI) encodes cloud transition dynamics and motion trajectories into a physically interpretable similarity metric, enabling atmospheric analogue retrieval from decade-scale satellite archives. • The Similarity-based Forecasting (SIM) framework achieves geographically universal day-ahead solar forecasting without site-specific tuning, offering a scalable and operationally deployable alternative to regionally constrained models. • Validated across 26 globally distributed BSRN stations spanning polar to tropical climate zones, SIM demonstrates competitive performance independent of regional atmospheric regime. • SIM reduces day-ahead RMSE by 15%–30% relative to NWP, cloud-motion-vector and time-series forecasting baselines.
Kosmopoulos et al. (Wed,) studied this question.