Variational spatiotemporal factorized graph neural networks for integrated electric vehicle usage prediction and pattern recognition with missing data on regional networks

Analyzing and modeling network-level electric vehicle (EV) usage data are pivotal for the continuous development of the EV industry. A unified prediction and pattern recognition framework contributes to a comprehensive view of spatiotemporal EV usage dynamics, informing both regular and real-time charging facility assignments. Therefore, a variational spatiotemporal factorized graph neural network (VSTF-GNN) framework is developed in this paper to integrate regional EV usage prediction and pattern recognition. It generalizes the prevailing predictors to explicitly predict the latent EV usage patterns. Specifically, graph-based and time-series models are applied to predict the spatial and temporal patterns, respectively, during the prediction horizons. The variational Bayesian method is introduced for spatial pattern learning to capture the uncertainty due to missing values. Future observations are derived based on the recognized patterns, which not only enhances the prediction interpretability, but also boosts model scalability to longer prediction horizons as the spatiotemporal correlations are decoupled. The framework was evaluated on a real-world freeway EV flow dataset and an urban EV charging dataset with multiple prediction horizons and missing data rates. Based on pattern-informed prediction, experiments reveal significant improvements in accuracy, which are positively correlated to the predetermined pattern numbers. The superiority in accuracy also increases with the missing rates, indicating the robustness of the patterns inferred from incomplete data. Extended prediction horizon tests further validate the framework's scalability and interpretability, stemming from the recognized patterns that capture both primary and secondary EV usage characteristics during a relatively long period.