Prioritized Expert-Based Multi-Agent Deep Reinforcement Learning for Eco-Driving of Plug-In Hybrid Electric Vehicles: A Model Predictive Control Framework

Integrating adaptive cruise control with energy management strategy represents a promising approach to improving the energy efficiency of plug-in hybrid electric vehicle within the domain of eco-driving. This article proposes a novel model predictive control (MPC) eco-driving framework based on multi-agent deep deterministic policy gradient (MADDPG). First, a MADDPG model is developed for car-following scenario to jointly optimize vehicle speed and power distribution. An expert-guided priority experience replay mechanism is incorporated to enhance agent training efficiency and overall performance. Subsequently, a leading-vehicle speed predictor is constructed using a Transformer bi-directional long short-term memory-based network. Second, the trained agents are integrated into the MPC framework to enable multi-step predictive optimization of the control sequence within the time horizon. Finally, comparative experiments under real-world driving cycles demonstrate that the proposed strategy outperforms state-of-the-art multi-agent deep reinforcement learning methods in terms of efficiency, optimization capability, and adaptability. Its practical feasibility is further validated through hardware-in-the-loop testing.