Evaluation of Graph Convolutions for Spatio-Temporal Predictions of EV-Charge Availability

Abstract

When driving an electric vehicle (EV) it is necessary to plan recharging ahead, as the infrastructure for charging is still scarce. High quality predictions of charging availability can help avoid unnecessary searches for an available charger and provide better distribution of cars onto locations. This article analyzes EV-charge availability prediction performance for different types of information used as input for different neural network based prediction models. In the end, we present a sequence-to-sequence model for integration of such different types of information and evaluate the performance over different time horizons. In particular, we demonstrate that, on the one hand, temporal information on the history of charging interaction is important, and, on the other hand, integrating in addition spatial information from other stations increases prediction accuracy and can be leveraged into better long term predictions using graph convolutions.