Short horizon learning-based speed prediction for electric vehicles

Abstract

The automotive industry is moving more to the development of electric vehicles to meet environmental and emissions restrictions. As a result, much work is be- ing done to optimize the efficiency of these vehicles through the use of various control methods such as model predictive control. These efforts often rely on the knowledge of future vehicle speed, however, this information is difficult to predict beyond a trivially small horizon. This work proposes including route in- formation with onboard vehicle data to make longer speed predictions. This is done through the use of a new B-spline prediction concept in conjunction with a custom temporal-spatial neural network (TSNN) structure. The B-Spline pre- diction method is demonstrated first on a simple identification task and then the TSNN is trained on test vehicle data combined with route information from HERE maps. The TSNN was successfully shown to benefit from inclusion of the route information and outperform simple existing prediction methods.