Learning-Assisted Predictive Frequency Stabilization Using Bidirectional Electric Vehicles

High renewable penetration reduces effective inertia and increases frequency variability in microgrids, thereby limiting the performance of purely reactive frequency regulation. This paper presents a two-timescale frequency-support strategy based on bidirectional electric vehicles. The main novelty lies in introducing a learning-assisted correction layer between forecast-based aggregate regulation and final EV-level dispatch. Rather than replacing the predictive controller with an end-to-end data-driven policy, this layer uses measured fleet-state information to correct the supervisory aggregate request online before a final feasibility-preserving dispatch stage converts it into executable vehicle-level commands under concurrent power, energy, plug-in, and departure constraints. A supervisory predictive layer determines the aggregate support action from forecasted photovoltaic and load disturbances, whereas a lower real-time dispatch layer redistributes that action across the available fleet. Feasibility is enforced through an explicit projection stage prior to actuation. The method is assessed in simulation using measured campus operating profiles of irradiance, temperature, demand, frequency, and electric-vehicle availability. Across four representative operating days, the proposed strategy reduced the mean cumulative frequency deviation by 30.3% relative to droop control and by 24.7% relative to predictive-only operation, while reducing the mean time outside the admissible frequency band by 22.2% and 20.0%, respectively. Zero post-projection constraint violations were observed in all evaluated cases. These gains were obtained at the expense of higher actuation usage, thereby making the regulation–usage trade-off explicit.