Value-Function Learning-based Solutions to Optimal Energy Management Problem of HEVs

Abstract

This paper presents two learning-based approaches to solve the optimal energy management problem for hybrid electric vehicles. It will be shown that by applying a learning algorithm to the interpolation of value-function, which is an optimal approximate value-function in continuous state space, the discretization error can be rejected when performing dynamic programming. Extreme Learning Machine and Gaussian Process Regression are exploited as learning tools. Finally, numerical simulation results with a parallel HEV will be demonstrated to show the effort of value-function learning.