Energy Management Strategy Research Based On Reinforcement Learning And Driving Cycle Prediction

The hybrid vehicle is one of the important ways to solve the problem of energy shortage and environmental pollution.The energy management strategy,as one of the key technologies of hybrid vehicles,can reduce fuel consumption and improve fuel economy of vehicles,by optimizing the energy distribution between multiple power sources.Aiming at the series hybrid tracked vehicle,this paper develops a speedy Q-learning based real-time energy management strategy and driving condition prediction based energy management strategy for hybrid vehicles.The hardware-in-loop simulation and real vehicle test are conducted to validate the performance of proposed strategy.In this paper,the research contents mainly include:Aiming at the series hybrid tracked vehicle,the mathematical model and forward simulation model of driving system is established,which includes the engine-generator model,battery pack model,motor model and vehicle dynamic model.The simulation is conducted,as the vehicle speed gathered from real vehicle test is used as the target driving cycle of simulation.The accuracy of the vehicle driving system model is verified by comparison of simulation test and real vehicle test resultsThe online-update energy management strategy based on reinforcement learning is proposed.Based on the recursive algorithm,the online update of the transition probability matrix of power-request is realized.The Kullback-Leibler divergence rate is used as the evaluation index of the difference of diverse transition probability matrix to determine the update of control policy.A new variant of Q-learning,called speedy Q-learning,is adopted to improve the computing speed of control strategy.The simulations based on online-update algorithm and offline algorithm are carried out.The results show that online-update energy management strategy based on speedy Q-learning can reduce the fuel consumption to improve the fuel economy of vehicles.Based on the multi-scale driving cycle prediction method,the driving cycle prediction based control algorithm is proposed.In order to avoid the influence of prediction error on optimization performance,the reinforcement learning energy management strategy as real-time correction control is integrated.The prediction root-mean-square error and error standard deviation as input,the fuzzy control is used to adjust the proportion of final control between the predictive energy management and reinforcement learning control,to effectively improve vehicle fuel economy.Based on Moto Tron controller and RT-LAB real-time simulation system,the hardware-in-loop is built to validate the performance of adaptive energy management strategy proposed in Chapter 3.The vehicle control strategy is developed in Moto Tron controller.The vehicle model is built in RT-LAB.The CAN communication interface is applied to realize real-time communication.The results show that the real-time energy management strategy can effectively save fuel consumption and improve the fuel economy of vehicles.Based on the data gathered by real vehicle,the hardware-in-loop simulation is used to replace the real vehicle test to validate the optimization of the energy management strategy proposed.