A Hybrid GCN-GRU Model with Attention Mechanism for Ultra-Short-Term Photovoltaic Power Forecasting

Accurate photovoltaic (PV) power forecasting is crucial for maintaining real-time grid stability. Single-site prediction methods fail to utilize spatial correlations effectively and are highly susceptible to outliers. This study proposes a hybrid model designed for multi-site PV power forecasting. The proposed model addresses the limitations of single-site methods. It also tackles the challenges of complex parameters in traditional models. Specifically, the model integrates Graph Convolutional Networks (GCN), Gated Recurrent Units (GRU), and an Attention Mechanism (AM). Initially, an undirected graph is constructed based on Pearson correlation, and subsequently, GCN is employed to capture spatial correlations among multiple PV power plants. GRU models complex temporal correlations, while AM dynamically highlights key features, achieving a comprehensive and unified spatiotemporal representation. The model was trained using a multi-site PV power generation dataset and validated under diverse weather conditions. The results demonstrate that the model achieves a coefficient of determination (R2) of 0.9855, a mean absolute error (MAE) of 2.7136, and a root mean square error (RMSE) of 3.7457, indicating both high accuracy and practical utility.