Adaptive Energy Management Strategy Based on Deep Reinforcement Learning for Extended-range Electric Vehicles

Abstract

Energy management strategies play a crucial role in the economy and emission performance of hybrid electric vehicles. This paper proposes an adaptive energy management strategy based on pre-trained deep reinforcement learning (DRL) agents for extended-range electric vehicles. In this method, the experience learned by the DRL agents is integrated into the online controller. First, four deep deterministic policy gradient based agents are respectively trained under four standard driving cycles for optimal power split control. Then, in order to integrate the experience of different agents and avoid the over fitting problem of DRL, a power adaptive coefficient is introduced. It is calculated based on the experience of four pre-trained agents. Finally, a weighted evaluate strategy of power adaptive coefficient based on K-nearest neighbors algorithm is proposed to achieve optimal power split control which can determine the power adaptive coefficient in real time. The simulation results show that the proposed method has good adaptability and stability. Moreover, equivalent consumption of the proposed method under the test driving cycles is reduced compared with other online energy management strategies, which indicates the effect of proposed method.