Decentralized off-grid vehicle-to-building (V2B) operation: A reinforcement learning approach for optimal charge–discharge control of energy storage systems

Recent advances in vehicle-to-everything (V2X) technology position electric-vehicle (EV) batteries as mobile energy-storage systems (ESSs), promising a new paradigm for power supply flexibility. However, prevailing centralized V2X operation scenarios and dynamic compensation schemes offer limited economic incentives for EV owners, hindering large-scale adoption. This study proposes a decentralized V2X scenario that treats EV owners as independent economic agents. It introduces a predetermined discharge tariff set at 70% of the contemporaneous time-of-use (ToU) electricity price, thereby offering clear, ex-ante profit expectations for all participants. Participations in V2X transactions are modeled as a time-dependent Poisson process. Within the proposed framework, we formulate an optimal ESS scheduling problem that minimizes the sum of electricity costs and ESS degradation costs for a commercial building equipped with photovoltaic (PV) generation, stationary ESS, and vehicle-to-building (V2B) interfaces. To solve this problem, we develop a world models-based reinforcement learning framework that performs multi-horizon forecasting of PV output and building load, using these forecasts as states to learn an optimal control policy. Compared with benchmark strategies, the proposed approach achieves a total cost reduction of up to 9.65% and attains near-global performance—within 1.57% of the optimality gap from the ideal strategy derived from perfect foresight data.