Cutting-Edge Generative AI For Intra-Hour Solar Forecasting

Abstract

Cloud passings pose significant challenges for photovoltaic (PV) power plants due to sudden drops in solar irradiance often referred to as ramp events. Anticipating these fluctuations requires local observations of clouds with high temporal resolution, which can be achieved using ground-based all-sky imagers (ASI) and irradiance measurements. With this information, models can be implemented to create intra-hour irradiance forecasts to respond to these fluctuations and mitigate power intermittencies. While physical cloud models excel at mesoscale predictions, such as those commonly used in numerical weather predictions for thunderstorms, they are limited in modeling microscale dynamics which determine local irradiance fluctuations. Data-driven approaches have emerged as a promising alternative. In this case, machine learning is used to analyze large datasets of sky images, bypassing the need to explicitly model physical processes. However, many of these models are optimized for the RMSE of the target variable, such as Global Horizontal Irradiance (GHI). As a result, these models often produce overly smooth forecasts, failing to capture ramp events during highly intermittent conditions. To address this, recent advances in ASI-based solar forecasting incorporate video prediction (VP) to model cloud dynamics in image space, forcing models to align cloud patterns with irradiance. Due to the additional optimization on the evolution of cloud patterns, the shortcut of predicting an expected average can be prevented. However, current VP models face challenges particularly regarding video quality. A major drawback is the image blurriness for increasing lead times, but also high model complexity which leads to time-consuming and costly model trainings and low image/time resolutions. Finally, these models often suffer from poor probabilistic forecasts and generalization, since they were trained on data from a single camera at one site. To address these limitations, we present a next-generation generative AI model for intra-hour solar forecasting, capable of forecasting realistic cloud scenes up to 30 minutes ahead, while enhancing spatial detail from higher image resolutions. Using generative diffusion processes, our VP model predicts more or less diverse future cloud patterns, depending on current conditions, thus enabling probabilistic forecasts to estimate uncertainty. Extensive validation on diverse datasets from multiple cameras demonstrates robust generalization and easy adaptability to new locations.