Degradation-constrained multi-agent reinforcement learning with centralized training and decentralized execution for...

High renewable penetration and large-scale electric vehicle integration impose voltage instability, frequency deviation, and network congestion challenges in distribution systems while accelerating battery degradation. This study proposes a degradation constrained multi agent reinforcement learning framework based on centralized training with decentralized execution for coordinated vehicle to grid optimization. The method integrates DC power flow constraints, stochastic renewable uncertainty, and electrochemical battery aging dynamics within a unified control architecture to ensure network aware and lifecycle aware scheduling. The framework is evaluated on the IEEE 33 bus distribution network with electric vehicle penetration up to 50% and compared against two baselines: a rule based conventional V2G scheduler and a single agent deep reinforcement learning controller without coordinated network constraints or degradation penalization. Grid stability is quantified using a normalized composite index derived from frequency and voltage magnitude deviations. Across repeated simulation trials, the proposed approach improves the mean stability index by approximately 10% relative to the rule-based method and 7% relative to the single agent baseline. Renewable utilization increases by about 18% age points, peak load reduction reaches 40% under high penetration scenarios, and cumulative battery aging decreases by nearly 13% over a 24-hour horizon, demonstrating enhanced coordinated control performance within standardized simulation environments.