Minute-Scale Solar Irradiance Nowcasting with Infrared Sky Imagery

Accurate one-minute-ahead solar irradiance forecasts are essential for real-time photovoltaic control and grid balancing. While persistence is strong on stable days, it tends to degrade during the most critical and highly variable periods. This work presents the Simple Irradiance-forecasting Model, which predicts Beam Horizontal Irradiance from sequences of long-wavelength infrared all-sky images and co-located meteorological measurements. The model combines a Convolutional Long Short-Term Memory encoder to capture cloud-motion dynamics with a ConvNeXt backbone, a modern CNN architecture, and a regression head conditioned on the solar elevation angle. To improve operational stability, the network predicts Beam Horizontal Irradiance, which is then converted into one-minute increments of Global Horizontal Irradiance through a physics-based postprocessing scheme that enforces admissible irradiance bounds. The proposed approach outperforms persistence on both a full dataset and a high-variability subset, achieving root mean square errors of 49.51 (12.15% forecast skill) and 69.71 (15.20% forecast skill), respectively. Evaluation on an additional dataset with lower Beam Horizontal Irradiance variability further demonstrates the robustness of the proposed pipeline, achieving performance comparable to persistence under conditions with limited scope for forecast improvement. Ablation results show that postprocessing and temporal encoding are critical, and case-study days confirm lower cumulative error under rapidly changing cloud conditions.