Cooperative multi-agent deep reinforcement learning-based eco-driving strategy for hybrid electric vehicles at multi-intersection scenarios

Eco-driving strategies, as a key component of intelligent transportation systems (ITS), offer an effective approach to enhancing vehicle performance across different driving scenarios through optimized driving behavior. This paper presents an eco-driving strategy framework for hybrid electric vehicles (HEVs) in signalized intersection scenarios, developed based on two modified multi-agent deep reinforcement learning (MADRL) algorithms: the multi-agent deep deterministic policy gradient (MADDPG) and the multi-agent twin delayed deep deterministic policy gradient (MATD3). The proposed framework incorporates two specialized agents with distinct tasks: the intersection speed planning agent dynamically computes optimal velocity trajectories within green light windows, while the energy management agent optimizes real-time power allocation between hybrid power sources. We formulate the eco-driving problem as a multi-objective optimization task that simultaneously addresses energy consumption, battery health, driving safety, and traffic efficiency. By embedding these objectives into the multidimensional reward structure of the agent training framework, the agents can select actions that effectively balance the competing objectives. The proposed strategies are trained and validated across three distinct signal phase configurations in multi-intersection scenarios. Experimental results demonstrate that both strategies exhibit good eco-driving performance under multiple complex traffic conditions, effectively balancing energy optimization and traffic efficiency. Furthermore, the MATD3-based eco-driving strategy outperforms the MADDPG-based approach in terms of convergence speed and training stability during the learning process. • A unified MADRL-based eco-driving framework is proposed for HEVs in Multi-intersection Scenarios. • A comprehensive comparative analysis is conducted between the MADRL-based eco-driving strategies and conventional benchmark methods. • Multi-objective rewards balance energy consumption, driving safety, traffic efficiency and battery health. • MADRL with cyclic learning rates and prioritized experience replay mechanism improves training stability and efficiency.