Enhanced energy management for fuel cell hybrid electric tracked vehicles: Integrating reinforcement learning with data augmentation and imitation learning

Data-driven techniques through deep reinforcement learning (DRL) are at the forefront of developing energy management strategies (EMSs) in electrified vehicles. The overall effectiveness of DRL-based EMSs hinges on both the quality and quantity of training data and the efficiency of the training process. In response to these, this paper develops an enhanced DRL-based EMS for a fuel cell hybrid electric tracked vehicle that integrates data augmentation and imitation learning. The soft actor-critic algorithm is first employed to develop an EMS focused on fuel cell longevity and hydrogen conservation. Subsequently, the data augmentation algorithm, the time-series generative adversarial network, is applied to generate velocity segments, producing diverse driving profiles that form a stochastic training environment. After that, pre-training through imitation learning is utilized to mimic globally optimal trajectories and initialize the policy network. Optimal samples from dynamic programming are then incorporated into the experience replay buffer to accelerate convergence. Experimental simulations show a 69.77 % increase in convergence speed, a 30.30 % improvement in learning ability, and a 9.43 % enhancement in overall energy management performance. The processor-in-the-loop test further confirms the proposed EMS's potential for real-time applications, contributing to energy savings for electrified tracked vehicles under practical scenarios.