Enhancing solar power forecasts through spatiotemporal learning and physics-based anomaly detection

Accurate very-short-term solar irradiance forecasting is crucial for enhancing grid stability, reducing operational costs, and facilitating the high-penetration integration of photovoltaic power. While deep learning models have surpassed traditional methods, spatially-oriented architectures often neglect temporal dependencies, and a comprehensive diagnosis of model failures is lacking. This study develops a high-resolution global horizontal irradiance (GHI) forecasting model based on the predictive recurrent neural network (PredRNN) to address these gaps. A rigorous comparison against a strong spatial-attention baseline, namely CBAM-UNet, elucidated the relative importance of spatiotemporal dynamics versus spatial refinement. Furthermore, a novel, physically-informed anomaly diagnosis framework was introduced to systematically localize and analyze forecast failures. Results demonstrated that PredRNN consistently outperforms CBAM-UNet, maintaining lower errors, higher structural similarity, and reduced systematic bias, particularly beyond 180-minute forecasts. The anomaly diagnosis revealed that forecast failures are not random but are spatiotemporally correlated with the lifecycle of deep convective clouds—a highly nonlinear process that pure data-driven models struggle to capture. This work confirmed the superiority of spatiotemporal modeling for GHI forecasting and establishes an advanced diagnostic paradigm, ultimately highlighting the inherent limitations of statistical learning and pointing towards hybrid physical-data-driven approaches as the path forward for securing renewable-energy-centric power systems.