Cold-start Power Prediction of Regional Photovoltaic Clusters Integrating Spatiotemporal Graph Convolution and Transfer Learning

Aiming at the cold-start problem that newly-built photovoltaic (PV) stations lack historical data and can hardly achieve accurate power prediction, a regional PV cluster power prediction method combining spatiotemporal graph convolutional network and transfer learning is proposed. Taking PV stations in the region as nodes, a weighted adjacency matrix is built based on geographical distance and meteorological similarity to establish a spatiotemporal graph structure. A spatiotemporal graph convolutional network (ST-GCN) integrating spatial graph convolution, temporal convolution and attention mechanism is then constructed to extract the spatiotemporal features of meteorology-power coupling. A transfer strategy based on parameter sharing and maximum mean discrepancy (MMD) domain adaptation is further adopted to transfer knowledge from source-domain stations to newly-built stations. A case study on 16 PV stations in a provincial region shows that the root mean square error of the proposed method in cold-start scenarios is reduced by 52.5% and 39.8% compared with LSTM and GCN, respectively.