Unified Fourier Graph-Based Spatiotemporal Learning and Corrected NWP for Multi-Site Ultra-Short Term Photovoltaic Power Forecasting

Accurate multi-site ultra-short term photovoltaic (PV) power forecasting is essential for grid stability and efficient energy management. Existing methods are limited by coarse-resolution weather data and insufficient modeling of spatiotemporal dependencies. We propose a novel framework that combines bias-corrected high-resolution weather data with an Adaptive Fourier Graph Neural Network (PV-AFGNN) to capture complex spatial and temporal patterns. A Gated recurrent unit (GRU)-based encoder-decoder module refines Numerical Weather Prediction forecasts using local meteorological data, generating site-specific inputs for PV-AFGNN, which operates in the Fourier domain to model correlations efficiently. By jointly optimizing weather data correction and spatiotemporal learning, the framework achieves high accuracy even under challenging forecasting scenarios. Experiments on real-world multi-site PV datasets show that our method consistently outperforms state-of-the-art benchmarks, offering a scalable and robust solution for reliable energy scheduling, and enhanced integration of PV power into modern electricity systems.