Generative AI For Intra-Hour DNI Forecasts

Abstract

Sudden irradiance drops of cloud-induced ramp events are a challenge for solar energy systems. High-resolution local cloud observations from ground-based all-sky imagers (ASI) and irradiance measurements enable intra-hour forecasts to mitigate these fluctuations. While physical models excel at mesoscale predictions, they struggle with microscale dynamics driving local irradiance changes. Data-driven machine learning offers a solution, analysing sky image datasets without explicit physical modelling. However, many models optimized for root-mean-squared error (RMSE), produce overly smooth forecasts, missing ramp events in variable conditions. Recent advances in ASI-based forecasting use video prediction (VP) to model cloud dynamics in image space, aligning cloud patterns with irradiance for improved accuracy. Our approach enhances intra-hour solar forecasting with generative AI video prediction, addressing limitations in existing ASI-based models. A separate classifier model boosts ramp event detection.