Solar Irradiation Nowcasting with Flow-Guided Cloud Dynamics Prediction by SimVP-Flow

Accurate prediction of solar irradiation is important for renewable energy integration, agriculture, and environmental studies. However, solar power output is highly intermittent, with rapid fluctuations driven by cloud advection, formation, and dissipation. This intermittency increases operational uncertainty for grid operators and can raise reserve requirements. We present a deep learning framework for ultra-short-term forecasting of cloud evolution and solar irradiation up to 7.5 hours ahead using a 5-hour morning history ending at 09:00 (30-minute sampling). The model is trained with GK-2A geostationary satellite observations and auxiliary meteorological information. Conventional video prediction models often under-represent early-stage advection signals and tend to produce overly smooth forecasts, which limits their utility for irradiation prediction. To address these issues, we propose SimVP-Flow (Simple Video Prediction) with three components. First, we use the GK-2A water vapour (WV) infrared channel, infrared window channel and solar zenith angle (SZA) as inputs to provide both mid-to-upper-tropospheric flow cues and physically consistent diurnal geometry during pre-dawn and post-sunrise periods. Second, we incorporate optical-flow-derived motion fields as an explicit constraint to encourage sharper and more advective-consistent forecasts. Third, the decoder is modified with hybrid skip connections and PixelShuffle-based upsampling to better retain high-frequency cloud boundaries and reduce blurring artifacts in long-horizon predictions. We evaluate the proposed method on GK-2A case studies and compare it against single-channel baselines and the original SimVP. Performance is assessed using image-based metrics for cloud fields (e.g., MSE and SSIM) and error statistics for irradiation. This work aims to improve physically consistent short-horizon solar forecasting in data-sparse regions using satellite imagery and lightweight auxiliary variables.

This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT)(RS-2025-00515357).