Research On Energy Management Strategy Of Power-split Hybrid Electric Vehicle Based On Reinforcement Learning

Hybrid electric vehicles(HEV)can effectively alleviate the energy crisis and reduce air pollution,and gradually become a global hot spot.Energy management,as a key technology for hybrid electric vehicles,improves the fuel efficiency of vehicles by controlling the reasonable work area of various components,which is of great significance for energy conservation and emission reduction.This article has established an energy management strategy based on reinforcement learning by taking advantage of the good self-learning optimization ability of reinforcement learning.The main research contents are as follows :(1)Analyze the working mode of power split hybrid electric vehicles,and a rule-based logic threshold energy management strategy and an energy management strategy based on the dynamic programming are formulated.The threshold values of the rule-based logic threshold energy management strategy are optimized by analyzing the working mode of the vehicle when using the energy management strategy based on the dynamic programming,.The optimized rule-based strategy and dynamic programming strategy will be used as the comparison basis for the subsequent development of reinforcement learning strategies.(2)Analyze the Markov decision process in HEV energy management,an energy management strategy based on Q-learning reinforcement learning algorithm is established.The working process of the algorithm is optimized by means of initial value optimization,boundary constraint optimization,and value model structure optimization.(3)Analyze the adaptability and continuity of reinforcement learning energy management strategy.The JS divergence values between the probability transition matrices of different driving cycle are used to judge the change of driving cycle.Through targeted reinforcement learning for typical driving cycle,an energy management strategy for reinforcement learning based on adaptive driving cycle was formulated.Aiming at the discrete problem of the state and control variables in the Q-learning algorithm,a neural network is used to learn the Q-learning reinforcement learning strategy,and the control strategy is made continuous by the predictive ability of the neural network.(4)Combined with the V-mode development process of vehicle controller,model-in-the-loop and software-in-the-loop are used to further verify the reinforcement learning control strategy in practical applications.Established the vehicle and vehicle control software model,and conducted model-in-the-loop testing.Automatic code generation technology is used to transform the model into C code,and the C code software model is tested in software-in-the-loop.The research has shown that the energy management strategy based on reinforcement learning has better fuel economy and has certain practical application value.