Comparison of deep reinforcement learning-based energy management strategies for fuel cell vehicles considering economics, durability and adaptability

Abstract

The energy management strategy (EMS) is the top priority to ensure the safe and efficient operation of fuel cell hybrid vehicles. Nowadays, EMSs based on deep reinforcement learning (DRL) have become a research hotspot. However, most DRL-based EMSs have not discussed the impact of algorithm hyperparameters, and have not provided a comprehensive evaluation of indicators including fuel cost, aging, and efficiency. There is a lack of a unified performance metrics for different DRL algorithms. To solve this, a comparative study of EMSs based on five DRL methods is conducted in this paper, and a multi-objective reward function that integrates hydrogen consumption, fuel cell degradation, and battery state-of-charge fluctuation is designed. First, the hyperparameters and weight coefficients of the reward function are determined based on the algorithm convergence performance in the training process and average hydrogen consumption, respectively. Then the comprehensive performance of the above-mentioned DRL-based EMSs are compared horizontally. Finally, six driving conditions are used as test sets to explore the adaptability. The results show that the TD3-based EMS has the smallest equivalent hydrogen consumption and degradation per 100 km, which are 1165 g and 0.0651% respectively. This work can provide valid guidance for researchers to use DRL in EMS.