Battery swapping stations (BSS) are a crucial infrastructure for flexibility within the new power system. Their charging and discharging scheduling

however

is significantly impacted by fluctuating user demands and system complexity

making it challenging to simultaneously ensure both accuracy and economic efficiency. To address this

this paper proposes a grouping-based scheduling model leveraging the multi-agent twin delayed deep deterministic policy gradient algorithm to achieve the optimal scheduling of the BSS charging and discharging process. First

we establish a BSS scheduling mathematical model and a Markov decision process (MDP) model to describe the charging and swapping behaviors. Building upon this foundation

we tackle the curse of dimensionality inherent in the MDP model’s high- dimensional state and action spaces by proposing a centralized system decomposition method based on device grouping and inter-group collaborative decision-making. To address the collaborative scheduling challenge arising from information silos between decision-making subsystems

we design a bucket-based state encoding method and an unbalanced attention mechanism to enhance the BSS subsystems' ability to grasp global information

thereby facilitating state sharing and joint scheduling between multiple agents. Finally

we conduct simulations to validate the proposed MATD3 model and algorithm. The results demonstrate that

in large-scale BSS

the proposed model exhibits both scheduling economy and robustness. It achieves operational profits at least 2.37% higher than benchmark deep reinforcement learning algorithms and 7.95% higher than model-based algorithms

while also exhibiting higher solution efficiency.