Support vector machines for determination of an operational strategy for hybrid electric vehicles

Abstract

In this paper a data driven operational strategy for a hybrid electric vehicle (HEV) is developed. There are two big benefits of the proposed approach: The possibility of real-time implementation within embedded control units and the high potential for automated calibration. Starting point is a user defined set of fuel-optimized driving cycles for a hybrid vehicle, which is generated applying e.g. state of the art dynamic programming techniques. From this data the introduced methodology extracts a control strategy that determines the torque-split factor for a given driving situation. The approach is based on a combination of optimization and classification, as well as regression strategies. The data created by a dynamic programming algorithm (DP) is used to train support vector machines (SVMs) in order to get rid of the necessity of a-priori knowledge of the whole driving cycle. From the resulting functions a control law is derived that is able to identify a suitable torque-split factor, independent of the further driving course. Since reducing the information input into the control law will per definition reduce performance, validation of the methodology is based on comparison with optimized driving cycles generated by dynamic programming that use the whole driving cycle information.