On Optima Selection Strategies of Adversarial Swarm Learning for Energy Management of PHEVs

Developing robust optimization and learning methods is necessary for intelligent vehicles since an increasing number of critical control functions will be handled by artificial intelligence. This paper proposes an adversary swarm learning (ASL) system and an optima selection strategy for robust energy management of plug-in hybrid electric vehicles (PHEVs). The proposed ASL system comprises an attacking swarm and a defending swarm, which compete against each other iteratively to derive the most robust equivalent consumption minimization strategy (ECMS) for PHEV energy management. During the attacking rounds, the ECMS settings are fixed by the defender. Meanwhile, the attacker generates worst-case driving conditions by training a model in order to Maximize the equivalent energy consumption. During the defending rounds, the ECMS settings are optimized by the defender based on the driving scenarios generated by the attacker. The settings of robust ECMS are derived by introducing the concepts of “attacking rate” and “defending rate”, based on the ASL system’s minimum attack rate (MAR) and maximum defence rate (MDR). The proposed methods demonstrate the superiority over traditional ECMS, which was optimized across four standard driving cycles as a baseline, on both software-in-the-loop and hardware-in-the-loop platforms. The experimental result shows that the robust ECMS derived through the ASL system outperforms the baseline across all tested driving scenarios. Significantly, the ASL system employing the MAR strategy proves more effective. Its average cost-reducing rate is 10.58% higher than the baseline.