Search

A hybrid electric vehicle (HEV) that uses multiple planetary gear units with clutches as transmission system is advanced for the powertrain performance, because the operation of the clutches can lead to distinct operating modes, and the induced possible operating modes provide additional freedom to deal with the energy optimal control problem. Under each operating mode, the powertrain mechanical system has specific dynamical behavior. In order to develop model-based optimization schemes that can tackle the transient operations of the vehicle, exact dynamical modeling is investigated focusing on a hybrid powertrain system that uses a two-planetary-gear transmission box with two clutches. It shows that according to the states of the two clutches, the powertrain system has the power-split mode, parallel mode and the electric vehicle (EV) mode. Finally, an analysis for the calculation of the desired driving torque and its application to the dynamic programming (DP)-based energy management indicate the significance of the developed exact dynamical models.

This paper presents a new architecture of real-time HEV’s energy management problem under a V2V and V2I environment using policy-based deep reinforcement learning. The ideal energy management controller that minimizes HEV energy costs needs to run engines most efficiently in the whole running considering battery SoC. The controller needs to predict the future vehicle speed and plan the power distribution to achieve it because the thermal efficiency of engines is more efficient when its rotational speed is higher. The future vehicle speed has relationship with connectivity information such as the behavior of the car in front, the traffic light signals, crowd of cars, and so on. This paper assumes the connectivity environment in the future and applies proximal policy optimization (PPO) [5] that is known as policy-based deep reinforcement learning algorithm to achieve the optimal power distribution predicting the future behavior by using connectivity information. In addition, this paper shows that locating the local controller in the reinforcement learning loop enables the AI controller to learn robustly. The local controller corrects against an exploration that is obviously not optimal or doesn’t satisfy the constraints.

We solve a optimization problem for HEV energy management. In this study, HEV system has three driving modes(EV, parallel, series parallel) with two motors, one engine and two clutches. We change this problem into a combination optimization problem by discretizing each torque as a variable. It is difficult to solve this problem because the combinations of variable are so many. To this hard problem, we solved it efficiently by using dynamic programming and applying steady-state approximation. It does not require much computation time, and a global approximate optimal solution can be obtained.