Fourier‐Enhanced Graph Spatiotemporal Learning for Multi‐Site Photovoltaic Power Forecasting

Abstract

This chapter tackles core challenges in multi-site photovoltaic power forecasting, where coarse-resolution numerical weather prediction (NWP) data and decoupled modeling of spatiotemporal dependencies limit performance. While leveraging spatial correlations across sites improves regional forecast reliability, existing methods are often hindered by systematic meteorological biases and architectural inefficiencies. To address this, we introduce a novel, two-stage integrated framework. The first stage focuses on input refinement. A dedicated correction module seamlessly fuses historical NWP forecasts with high-resolution local meteorological data. This process generates site-specific, bias-corrected weather variables, effectively mitigating spatial mismatches and systematic errors while preserving large-scale atmospheric trends and injecting fine-grained local context. The second stage enables unified spatiotemporal learning. We employ a graph neural network that operates in the Fourier spectral domain. This design allows the model to jointly capture both spatial and temporal dependencies across the entire network of photovoltaic (PV) sites in a cohesive manner. By learning in the frequency domain, it efficiently models complex, long-range interactions and global dependency patterns that are critical for accurate regional forecasting. The combined framework demonstrably enhances prediction accuracy, robustness against input noise, and scalability for large, geographically dispersed PV fleets. It provides a more reliable, consistent foundation for grid planning, supporting the secure integration of high-penetration renewable energy into modern power systems.