Load Forecasting of Electric Vehicle Charging Stations: Attention Based Spatiotemporal Multi–Graph Convolutional Networks

Abstract

The charging load forecasting is of significant importance to the economic operation of charging stations and the stable operation of power systems. The charging stations couple power systems with transportation systems. Their charging loads are not only affected by the driver’s driving and charging behavior simultaneously, but also by the above two networks. There is much literature reporting load forecasting combined with historical charging loads in temporal dimension. However, the spatial data of charging stations in the neighboring areas is helpful for load forecasting, as they share common conditions, including traffic and weather factors. This paper proposes load forecasting of electric vehicle charging stations based on the spatiotemporal multi-graph convolutional networks (STMGCN). STMGCN contains three components: gated dilated causal convolution, spatiotemporal attention mechanism, and multi-graph convolutional layer. Firstly, the load model based on graph structure is established according to the historical load and geographical information of charging stations. Then, a load forecasting method of multiple charging stations is proposed, which can effectively share spatiotemporal relationships among stations. Finally, experiments on real-world dataset illustrate that STMGCN is capable of improving the accuracy of charging stations load forecasting compared with the baselines, which shows the effectiveness of the model.