A systematic synthesis of sky image enhancement techniques for ground-based solar irradiance forecasting

The development of the renewable energy sector, especially solar energy, is essential to mitigate the rapidly progressing climate change. However, the inherent uncertainty related to the stability of solar energy production capabilities prevents large-scale integration of solar-based sources into power grids. Therefore, this work addresses this practical challenge by presenting a systematic, large-scale synthesis and validation of a comprehensive neural network-based image enhancement pipeline for ground-based solar forecasting. This consolidated pipeline includes a Sun mask, optical flow, a cloud channel, and a novel irradiance value channel for early-stage spatio-temporal fusion. As a second key methodological novelty, we introduce the mean adaptive BerHu penalty as a loss function, which is shown to be uniquely suited for robustly handling the high dynamic range of solar irradiance values. The proposed comprehensive pipeline, when evaluated on a spatio-temporal model, outperforms existing methods from the literature. This strong performance is not architecture-specific, as demonstrated by our cross-architecture study, in which the pipeline’s portability achieves an average forecast skill of 19.35% among evaluated models for the 15-minute forecasting horizon. This is due to physically referenced, representation-level enhancements that act as an architecture-agnostic driver of forecasting skill going beyond incremental, model-specific tuning. Its practical applicability is further confirmed by its effectiveness across multiple intra-hour time horizons, with significant skill improvements for both ultra-short 5-minute (16.22%) and longer 30-minute (22.52%) periods. Finally, its practical deployability was confirmed through extensive validation on resource-constrained edge devices. The source code used in this study is publicly available to promote research reproducibility.