Application-oriented mode decision for energy management of range-extended electric vehicle based on reinforcement learning.

• An application-oriented DQN-based EMS mode decision-making method is proposed. • High-dimensional mappings can be extracted from trained agents for controllers. • The practical feasibility of the proposed EMS is verified using EGS test bench. • The method can train the customized strategies and save the calibration work time. Energy management strategies play an important role in range-extended electric vehicles. An application-oriented energy management strategy based on deep Q-network is proposed, which takes the deviation of the SOC from the reference curve, vehicle speed, and required power of the vehicle as the state, and takes the mode decision as the action. The designed concise Q agent can be well-trained under standard driving cycles and real-world driving cycle. The high-dimensional mappings can be extracted from trained agents and are applicable to actual vehicle controllers. Compared with the rule-based and Q learning-based EMS, the proposed strategy has more mode selection judgment conditions and continuous state space, which can better select the timing of mode switching, thus effectively improving the engine operating efficiency and optimizing fuel economy. The experimental tests are conducted to verify the practical feasibility. [ABSTRACT FROM AUTHOR]