Statistical Modeling for Prediction of Vehicle Usage

Abstract

Battery electric vehicles (BEVs) are increasingly being adopted in all parts of the world. However, challenges such as maximizing the battery lifespan and efficiency, as well as the need for smart energy management systems remain. Newer generation BEVs have greater computational capabilities, enabling them to host smart functions that can address these challenges in new ways. A crucial component is understanding vehicle usage patterns, i.e., when the vehicle will be driven or charged. For example, accurate predictions of departure times allows to optimally perform battery thermal preconditioning, and predictions of arrival times can enable more intelligent decisions around household energy usage with new vehicle-to-grid capabilities. If a vehicle is expected to arrive during a period of high electricity prices, it may be advantageous to charge it using energy stored in a home battery. Knowing driving times also facilitates the scheduling of major software updates and the delivery of relevant traffic information. In this thesis, statistical models for predicting both single and recurrent driving events are developed using time-to-event methods. The models are based on the flexible Royston–Parmar framework, which allows us to obtain a parametric version of the famous Cox model and the Prentice-Williams- Peterson model that is used to model recurrent events. The models are trained and evaluated on historical driving data collected from a fleet of BEVs during the last three years, with predictive performance assessed by typical metrics such as the root mean squared error. The best-performing models achieve a mean absolute error of 60 minutes for departure time prediction and 100 minutes when predicting all usage events during the day. Furthermore, their performance is evaluated against that of traditional machine learning models.