An adaptive spatio-temporal graph recurrent network for short-term electric vehicle charging demand prediction

Predicting Electric vehicle (EV) charging demand can facilitate the efficient operation and management of the smart power grid and intelligent transportation systems. We propose an adaptive spatial-temporal graph recurrent network (ASTGRN) to predict the EV charging demand in short term at the charging station level. Specifically, we design an adaptive graph learning layer that learns the spatial correlations in a data-driven manner. Additionally, an embedding project layer is integrated to enhance the graph learning layer. Subsequently, a graph recurrent layer consisting graph convolutional kernel and gated recurrent unit is employed to extract spatial-temporal features from the observations. We evaluate the proposed ASTGRN model using a real-world EV GPS trajectory dataset containing charging information of over 76,000 EVs in Beijing. The experiment results suggest that ASTGRN achieves state-of-the-art performance compared to those advanced spatial-temporal prediction models (e.g., Temporal Graph Convolutional Network and GraphWave Net). The effectiveness of the proposed model in charging demand prediction indicates that the spatial correlation between different charging stations may not be related to geographical distance in the charging demand prediction task, and the use of prior knowledge of geographical location may undermine model performance.