Dynamic and personalized battery-swapping recommendations for electric micro-mobility vehicles: Leveraging deep reinforcement learning

Battery-swapping service for electric micro-mobility vehicles (EMVs) provides users with a convenient way to replenish energy to extend travel range. However, massive integration of battery-swapping EMVs potentially causes disorderly swaps and negatively impacts user experience and operational efficiency. Considering shared battery-swapping stations (BSSs) for both ordinary and delivery EMVs, this study proposes a real-time BSS recommendation system based on deep reinforcement learning to enhance the quality of battery-swapping service. A reservation-based recommendation mode is designed to ensure a first-reserved, first-served swapping service, for which the long short-term memory (LSTM) and dynamic graph attention networks (DGAT) modules are developed to capture the dynamic correlation between EMVs and BSSs for predicting and representing upcoming swapping demands to avoid unintended competition for swapping resources. This recommendation system incorporates a utility function for swapping choices into the traditional system optimization-based reward function to ensure optimal system performance while enabling personalized recommendations. An improved Rainbow algorithm is suggested to enhance the performance of the battery-swapping recommendation strategy by integrating techniques such as prioritized experience replay and double Q-learning. Extensive numerical experiments demonstrate that the proposed LDRFusion algorithm outperforms several baseline methods.