The continuous training of machine learning-based energy management strategy for plug-in hybrid electric vehicle, part I: electric driving mode

Thanks to the adaptability of machine learning-based strategy to uncertainties, reinforcement learning (RL) has been widely adopted to optimize the performance of energy management strategy (EMS) in real driving conditions. Although the RL-based EMS is superior to others in terms of adaptability, the performance is still limited by the data used in strategy training, which is based on past behaviors. To optimize the performance of the energy management strategy in unknown conditions with uncertainties, this study proposed a RL-based EMS that is capable of continuous learning new data in the training strategy, rather than studying the new data via retraining episode in traditional RL strategy. By inserting several new cycles into the existed data during the training period, the results demonstrate the continuous learning capability of proposed strategy. The method of optimal rule-based is used as a benchmark in this paper. Compared with the optimal rule-based method, multi-condition paralleling training only increases energy consumption by 0.234 kWh, which is only 3.24 % higher than the optimal rule-based method. Furthermore, the simulation results show that the proposed parallel training-based continuous learning EMS outperforms other learning-based EMS, i.e., sequential and optimal, by 0.155 kWh in CHTCB cycle; The proposed method effectively improves the adaptability of RL in unknown condition with uncertainties by continuous adopting new data in the training process.