Spatio-temporal Graph Neural Network with Fourier features for multi-site photovoltaic power forecasting

Multi-site photovoltaic (PV) power forecasting is crucial for modern power systems, enhancing grid stability and maximizing renewable energy utilization. However, accurate forecasting is challenging due to strong spatiotemporal dependencies, data heterogeneity, and the need for periodic pattern modeling. To address these challenges, we propose a novel forecasting framework integrating Graph Neural Networks (GNN) with Fourier-based feature extraction. First, we construct a hyper-variable graph to jointly model the spatiotemporal dependencies of multi-site PV generation and numerical weather prediction (NWP) data within a unified framework. This structure effectively captures intricate correlations across time and space. Second, we incorporate a Fourier transform module to project time-series data into the frequency domain, facilitating the extraction of key periodic features. A low-rank approximation is employed to compress representations, mitigate redundancy, and reduce computational complexity while preserving dominant periodic components and suppressing high-frequency noise. Extensive experiments on two real-world PV datasets demonstrate the effectiveness of our approach. The proposed model achieves 1-step RMSE values of 0.0267 and 0.0228, respectively, outperforming strong baselines such as LSTNet, Autoformer, and MTGNN by up to 35.3%. These results confirm the model’s superior forecasting accuracy, generalization ability, and scalability for practical deployment in modern power systems.