Multi-scale fused Graph Convolutional Network for multi-site photovoltaic power forecasting

Multi-site photovoltaic power forecasting with refined spatiotemporal relationship mining has recently gained significant attention due to its potential to reduce modeling costs and improve accuracy. However, existing approaches often overlook the complex and varying spatiotemporal correlations across different time scales among multiple sites in real-world scenarios. To address this limitation, this study proposes a novel and effective model from a multi-scale perspective: the Multi-Scale Fused Graph Convolutional Neural Network (MSF-GCN). The MSF-GCN incorporates a Multi-Graph Convolution (MGCN) block that utilizes both predefined and adaptive learnable graphs to capture diverse spatial dependencies between photovoltaic sites based on data observed across different time scales. Additionally, a lightweight Decomposed-Bidirectional-Fusion (DBF) block is designed to extract inter- and intra-scale correlations. This block allows fine-grained information from low scales to enhance the extraction of microscopic features at higher scales, while coarse temporal variations from high scales provide lower ones with a macroscopic view of power generation patterns. Furthermore, the model employs multi-predictors with identical structures but unshared weights to leverage both distinct features and complementary forecasting capabilities from multi-scale data simultaneously. Experimental results on two open-access datasets demonstrate that the proposed MSF-GCN consistently outperforms existing methods in terms of accuracy while maintaining favorable run-time efficiency. In terms of prediction accuracy, our model outperforms the state-of-the-art spatiotemporal model by an average of 13.21% for MAE and 28.48% for MSE. The average increase in MAE of 55.86%, 47.49%, and 44.55% resulting from the ablation of the multi-scale, the MGCN, and the DBF in MSF-GCN, respectively, further justifies the effectiveness of the designed structures.