Predicting electric vehicle charging demand to support public charging infrastructure location planning

• We discuss the relevant design parameters of charging demand prediction models. • Predictive performance is mostly determined by the metric used to measure demand. • Predictions are compared with a dataset of geo-referenced citizen requests for CI. • Observed charging demand is an imperfect proxy for true charging demand. • EVCI roll-out models should also consider coverage, besides demand predictions. Roll-out plans for public Electric Vehicle Charging Infrastructure (EVCI) often rely on demand predictions to locate charging stations. However, we find that most existing prediction models are not well-designed to support a location decision, as they: (1) are constructed at spatially aggregated levels, (2) do not consider the longitudinal robustness of their predictions, and (3) are often based on observed charging demand without considering its limitations. In this study, we present a prediction model that is trained on real-world charging data from Brussels and discuss its relevant design parameters to support location planning. Our results demonstrate that even at very detailed spatial levels (e.g., building block level), prediction models possess significant predictive power. However, the predictive performance is largely determined by the metric that is used to measure demand. We compare the predictions with a unique dataset of georeferenced citizen’s requests for charging stations, which demonstrates the limitations of solely using observed charging data to predict charging demand. We advise future research and practitioners to also consider the spatial coverage of the EVCI network, besides demand predictions, when deciding on the locations of new charging stations.