Eco-driving-based multi-agent reinforcement learning cooperative control strategy for PHEV platoon

To address the poor adaptability of PHEV platoon control strategies under complex driving conditions and the insufficient consideration of the coupling between longitudinal control and energy management in hierarchical control strategies, eco-driving-based multi-agent reinforcement learning cooperative control strategy for PHEV platoon is proposed. Firstly, a sample database is established by selecting 22 standard driving cycles. A two-stage clustering method combining PCA and SOM-K-means is adopted to construct three types of typical driving cycles, namely urban, suburban and highway cycles, on the basis of which a BP neural network driving cycle recognizer is trained. Secondly, the states of platoon vehicles are obtained based on V2V communication. Each following vehicle is regarded as an agent, and the MADDPG policy network is trained separately for different typical driving cycles to realize the collaborative optimization of platoon longitudinal control and energy management. Finally, in real-time online implementation, the corresponding policy network is called to execute control based on the online driving cycle identification results. The results demonstrate that the proposed strategy further improves fuel economy and emission performance while ensuring platoon following performance, safety and comfort. Compared with the strategy without driving cycle recognition, the average fuel consumption and average emissions of the platoon are reduced by 3.73% and 7.84%, respectively. In contrast to the hierarchical control strategy, they are decreased by 2.53% and 2.18%, respectively.