DEST-GNN: A double-explored spatio-temporal graph neural network for multi-site intra-hour PV power forecasting

Accurate forecasting of photovoltaic (PV) power is crucial for real-time grid balancing and storage system optimization. However, due to the intermittent and fluctuating nature of PV power generation , achieving accurate PV power forecasting remains a challenge. In this paper, we propose a novel approach for multi-site intra-hour PV power forecasting. Different from current work which predicts the power of each PV station independently, we predict the power of each PV station simultaneously by considering the inherent spatio-temporal correlation with other PV stations and develop a novel graph network named DEST-GNN. In DEST-GNN, an undirected graph is used to represent the dependence of these PV stations. Each PV station is represented by a node and the spatio-temporal correlation of any two PV stations is represented by an edge between them. To improve the accuracy of prediction, sparse spatio-temporal attention is adopted to filter out the weak associations of these PV stations. We then develop an adaptive graph convolution network (GCN) that adopts an adaptive adjacency matrix and a temporal convolution network to capture the hidden spatio-temporal dependency of these PV stations. Experimental studies using datasets from Alabama and California, provided by the National Renewable Energy Laboratory (NREL), demonstrate the effectiveness of DEST-GNN. For the Alabama dataset, DEST-GNN achieves a mean absolute error (MAE) of 0.49 over a 12-mon training scale. Furthermore, DEST-GNN attains an MAE of 0.42 on the California dataset, continuing to exhibit its strong forecasting capabilities. • A novel DEST-GNN that captures spatio-temporal correlations for multi-site PV power forecasting. • Leveraged sparse spatio-temporal attention and adaptive GCN to improve prediction accuracy. • Achieved performance with MAEs of 0.49 and 0.42 for Alabama and California datasets, respectively.