Multi-Agent Deep Reinforcement Learning for Joint Scheduling of EV Charging, Solar PV, and BESS under Real-Time Tariff Feedback

The rapid growth of behind-the-meter renewable energy resources and electric vehicle (EV) adoption has increased uncertainty in scheduling and operational planning in modern smart grids. Under conditions of stochastic user behavior, time-varying tariffs, and operational constraints on distributed energy resources (DERs), conventional centralized optimization techniques become difficult to apply. This study presents a Multi-Agent Deep Reinforcement Learning (MADRL) framework for coordinating electric vehicle (EV) charging, photovoltaic (PV) generation, and battery energy storage systems (BESS) under time-varying tariff signals. A federated actor-critic approach with shared global incentive feedback is employed, through which each agent independently learns control actions, including charge/discharge levels, flexible load shifting, and allocation of surplus renewable energy. The learning process accounts for real-time price signals, transformer loading limits, and end-user comfort constraints. The methodology is implemented in the MATLAB Reinforcement Learning Toolbox using the IEEE-33 bus distribution test system and stochastic demand and solar generation datasets. The results show a reduction in peak-hour demand, increased utilization of renewable energy, and lower overall energy costs when compared with rule-based and singleagent control strategies. The proposed architecture demonstrates scalability and coordinated tariff-responsive operation, which are required for practical transactive energy systems.