Coordinated Scheduling of Virtual Shared Energy Storage in Multi-station Electric Vehicle Charging Systems Based on Multi-agent Reinforcement Learning and Traffic-aware Forecasting

With the rapid proliferation of urban electric-vehicle charging stations, coordinated scheduling across multiple sites has become critical for improving distribution-grid operational efficiency. Addressing limitations of existing approaches in multi-station coordination, demand forecasting, and resource synergy, this paper proposes a coordination framework that integrates multi-agent deep reinforcement learning with traffic awareness. The framework introduces an integrated “multi-station—virtual shared storage—traffic awareness” architecture: without relying on cross-station DC interconnections, a public shadow-price/virtual shared-cost signal serves as a global cue to align the marginal decisions of local storage at each station, thereby achieving strategy-level coordination via information and economic signals. Methodologically, we adopt a centralized training, decentralized execution (CTDE) multi-agent reinforcement learning paradigm within a decentralized partially observable Markov decision process (Dec-POMDP) to learn inter-station cooperative patterns. For forecasting, an attention-based spatiotemporal graph convolutional network predicts traffic flows and outputs parameterized arrival rates (mean and uncertainty); combined with vehicle state-of-charge and charging-propensity models, these are transformed into charging demand, enabling proactive scheduling under uncertainty. Simulations driven by real traffic data in three-station scenario show improvements in operational metrics, validating the effectiveness and practicality of the proposed virtual-sharing and distributed-intelligence coordination approach.