A physics-informed spatiotemporal graph neural network for photovoltaic power forecasting and micro-anomaly localization
2949-821X
Abstract
Accurate power forecasting and precise fault diagnosis are critical for the reliable operation of large-scale photovoltaic systems. However, existing data-driven models often fail to capture complex spatiotemporal dependencies and lack physical interpretability, limiting their ability to detect subtle performance deviations. To address this, this study introduces a novel physics-informed spatiotemporal graph neural network framework that synergistically integrates physical energy conversion constraints with deep graph learning. Unlike conventional approaches, this unified architecture simultaneously optimizes forecasting accuracy and enables fine-grained micro-anomaly localization by leveraging physics-based residuals. Comprehensive evaluations demonstrate that the proposed framework significantly outperforms existing state-of-the-art models, reducing the root mean square error by 60.3% compared to traditional temporal baselines and achieving a spatial localization accuracy of 94.5%. Furthermore, the physics-informed constraints ensure robust predictive performance even under severe data scarcity. These findings confirm that embedding physical laws into spatiotemporal graph learning not only enhances forecasting precision but also provides a highly reliable, interpretable mechanism for early fault detection and predictive maintenance in modern renewable energy infrastructure.