Predicting power consumption of shared electric vehicles using regression and cluster analysis

Abstract

When planing a trip with an electric vehicle, one has to predict how much power will be consumed in advance in order not to risk running out of power during the trip. This problem is especially relevant for car-sharing services, as they want to schedule multiple trips in close succession. Current methods for estimating power consumption all rely on information regarding the exact route that will be followed, which in a car-sharing service is not always known beforehand. This thesis presents prediction models based on historical trip information capable of predicting power consumption for planned trips with electric vehicles without information about the specific route. The models are created by combining linear regression and unsupervised learning, in particular clustering with K-means. Kmeans is used to cluster users into groups based on previous driving habits. Separate regression models are then trained for each group in order to make more accurate predictions. We show that it is possible to predict power consumption using stand alone regression models and a few simple features such as booking duration and estimated distance. We also show that these predictions can be improved by the use of clustering. While these results are satisfactory on their own, we conclude that improvements can be made to the way distances are estimated in order to further improve predictions.