Task-aware Latent-Space Coupling for Generative Solar Forecasting

Kurzfassung

Accurate intra-hour solar irradiance forecasting remains challenging due to rapid fluctuations caused by cloud dynamics, commonly referred to as ramp events. Ground-based all-sky imagers (ASI) provide high-resolution observations of local cloud fields and have enabled significant advances in data-driven forecasting approaches. However, conventional machine learning models optimized for mean squared error (MSE) tend to produce overly smooth predictions, limiting their ability to capture short-term variability. Recent generative video prediction models address this limitation by explicitly learning spatiotemporal cloud dynamics, thereby improving ramp event detection [1]. These approaches are inherently probabilistic and enable the generation of multiple future scenarios. Nevertheless, existing pipelines typically decouple video generation from irradiance estimation, relying on downstream models applied to synthetically generated images. This separation introduces domain shift, systematic biases, and error accumulation. This work proposes FarSky, a novel generative forecasting framework that integrates irradiance information directly into the representation learning of sky imagery, a concept referred to as task-aware latent-space coupling. By embedding irradiance data within the latent space, it is consistently preserved and propagated throughout the generative process, thereby overcoming key limitations of previously decoupled approaches.