Physics-Guided Residual Learning for Short-Term Solar Irradiance Forecasting

Abstract

Accurate short-term solar irradiance forecasting is essential for reliable operation of renewable-dominated energy systems, where scheduling, flexibility management, and grid operation depend on prediction stability. Solar irradiance prediction remains challenging because it is influenced by both deterministic solar radiation behavior and uncertain atmospheric variations. Existing physical and data-driven approaches often struggle to represent both regular irradiance patterns and rapid weather-related fluctuations simultaneously. This paper proposes a physics-guided residual learning framework that separates these components through a sequential forecasting strategy. A deep temporal learning model first captures regular radiative behavior using meteorological observations and solar geometry features, while a secondary learning stage models the remaining structured forecasting errors. The final prediction combines the initial forecast with the learned residual correction. The proposed method is evaluated on a long-term, high-latitude solar dataset with chronologically separated data. The results demonstrate improved forecasting reliability across seasonal and weather-dependent conditions, providing a practical approach for renewable energy operation and digital energy applications.