Energy management strategy design for pure electric buses based on Adaptive-Advanced Neuro-Evolution of Augmenting Topologies

Learning-based energy management strategies (EMSs) show significant advantages in reducing the energy consumption by adopting multiple proper motors with appropriate power allocating algorithms. However, the adaptability of training-based strategy to unknown environments still face challenge. In this paper, a dual-motor four-speed electrified bus powertrain is selected as the benchmark to investigate the proposed adaptability improvement method for learning-based EMS. To determine the appropriate driving mode and the torque allocation coefficient, a topological Neuro-Evolution of Augmenting Topologies (NEAT) is adopted in the EMS. the generalization of the learning-based real-time EMS is improved. The results show that the proposed strategy is capable of updating the evolved network in the real-time, and adjust the network structure and weights automatically during training. Furthermore, to tackle the challenges of population fitness stagnation, fitness concentration and genetic diversity depletion of NEAT in exploring hybrid action space (discrete operating modes and continuous torque distribution coefficients), a dynamic mutation and roulette wheel selection method-based Adaptive-Advanced Neuro-Evolution of Augmenting Topologies (AANEAT) is proposed in this study. The simulation results from both tests, i.e. CHTC_B scenarios and real driving cycles, demonstrate that AANEAT-based EMS consumes less energy and show similar behaviors to the global optimum EMS.