SHORT-TERM PREDICTION OF ELECTRIC VEHICLE CHARGING STATION AVAILABILITY USING CASCADED MACHINE LEARNING MODELS

Driving long distances with battery electric vehicles is becoming possible thanks to increasing battery capacities and a growing network of fast-charging stations. During peak usage hours, multiple users may require recharging, thereby exceeding available charge points resulting in a queue. An algorithm is necessary to predict when a charging station is likely to be occupied and how long the waiting times at such a station would be to avoid such waiting times. This paper presents a methodology to cascade two machine-learning models to create such an algorithm. The first of these submodels predicts the likelihood that a current occupant is still at the charge point for any time in the future. It is implemented by training an ensemble learner with past charge events, thus learning station-specific and general usage characteristics. The second submodel predicts the probability of new visitors coming to the station and the occupation probability. Both achieve high accuracies in their respective domains. By mathematically combining both models, it is possible to construct an overarching model able to predict future charging station occupation likelihood based on the current occupation level of the station.