A transfer learning method for electric vehicles charging strategy based on deep reinforcement learning

Reinforcement learning (RL) is popularly used for the development of an orderly charging strategy for electric vehicles (EVs). However, a new environment (e.g., charging areas and times) will cause EV users' driving behaviors and electricity prices to change, which leads to the trained RL-based charging strategy is not suitable. Besides, developing a new RL-based charging strategy for the new environment will cost too much time and data samples. In this paper, a deep transfer reinforcement learning (DTRL)-based charging method for EVs is proposed to realize the transfer of trained RL-based charging strategy to the new environment. Firstly, we formulate the uncertainty problem of EV charging behaviors as a Markov Decision Process (MDP) with an unknown state transfer function. Furthermore, an RL-based charging strategy based on deep deterministic policy gradient (DDPG) is well-trained by using massive driving and environmental data samples. Finally, an EV charging method based on transfer learning (TL) and DDPG is proposed to perform the knowledge transfer on the trained RL-based charging strategy to the new environment. The proposed method is verified by numerous simulations. The results show that the proposed approach can reduce the outliers to meet the user charging demands and shorten the EV charging strategy development time in the new environment.

Automated summary

This paper proposes a deep transfer reinforcement learning (DTRL)-based charging method for electric vehicles (EVs) to transfer the trained RL-based charging strategy to a new environment. The uncertainty problem of EV charging behaviors is formulated as a Markov Decision Process (MDP) with an unknown state transfer function. An RL-based charging strategy using deep deterministic policy gradient (DDPG) is trained with extensive driving and environmental data samples. A charging method based on transfer learning (TL) and DDPG is proposed to transfer the trained RL-based charging strategy to the new environment. Simulations demonstrate that the proposed approach reduces outliers and shortens the development time of the EV charging strategy in the new environment.