| With the vigorous development of hybrid vehicles at home and abroad,the hybrid systems of different hybrid configurations are also constantly evolving.The seriesparallel hybrid system based on multi-disc clutches is becoming more and more mature,there is an urgent need for a sufficiently real-time intelligent energy management strategy to optimize the distribution of fuel-electricity ratios at each moment in the driving conditions of hybrid vehicles,so that hybrid electric vehicles consume the least comprehensive energy and maximize the effective mileage of hybrid vehicles.The current mainstream deep reinforcement learning algorithms that can be used for energy management strategy optimization are facing the problems of low exploration efficiency and local optimal solutions,including SAC(Soft Actor-Critic)and TD3(Twin Delayed Deep Deterministic Policy Gradient),these two most cutting-edge Actor-Criticbased algorithms.In this paper,through the exploration of the energy management strategy of multi-disc clutch series-parallel hybrid electric vehicles,as well as the sampling method of the Actor-Critic algorithm,the problem of high estimated values,the design of the action space,and the improvement and optimization of action selection,a new method called S2A3C(Soft Twin Actor Triple Critic)algorithm improves the comprehensive fuel economy of the energy management strategy.The main work and contributions of this paper are:· Constitute the driving mode and engine speed of the hybrid vehicle into a hybrid action space,and introduce the Gumbel-Softmax reparameterization technique to deal with the mixed action output by the Actor network.Combining empirical rule-based and deep reinforcement learning-based optimization methods to implement energy management strategies for hybrid electric vehicles.· Introduce sampling methods Emphasizing Recent Experience without Forgetting the Past(ERE)and Prioritized Experience Replay(PER).Combine them with the SAC algorithm to form the SAC algorithm SAC-PER and SAC-ERE of the two sampling methods.· Proposed and implemented the S2A3C algorithm based on the SAC algorithm of Twin Actor and Triple Critic(Soft Twin Actor Triple Critic),and combined it with the sampling method of ERE to form a deep reinforcement learning algorithm named S2A3C-ERE.A hybrid system energy management strategy based on S2A3C-ERE,SAC-PER and SAC-ERE algorithms is implemented.In international conditions NEDC,CLTC-P and WLTC,compared with the SAC-PER algorithm,the comprehensive energy consumption of the S2A3C-ERE algorithm under the three working conditions was reduced by 6%,7.5% and 3.9% respectively,which verifies the propose the effectiveness of the algorithm. |
