Spatio-temporal multi-head graph attention network for power forecasting of regional photovoltaic plants

Journal article, 2026

Improved prediction accuracy of regional photovoltaic (PV) generation significantly enhances multi-area coordination efficiency in modern power systems. This paper proposes a novel PV power forecasting model IIMGAT that integrates Improved Variational Mode Decomposition (IVMD), an improved Time-series Mixer architecture (Tsmixer), and a multi-head Graph Attention Network (GAT). First, the IVMD effectively extracts the key trend components from PV data as model inputs to mitigate interference from abrupt cloud cover changes. An innovative spatiotemporal dual-design framework is then employed to capture temporal patterns and extract spatial features. The proposed IIMGAT model achieves high precision forecasting by comprehensively capturing spatio-temporal correlations among regional PV power stations. The proposed model achieves R² values of 0.981 and 0.991 on the PVOD and DKASC datasets, respectively. The values of R² indicate the model’s high robustness.