Predictive Management of Electric Vehicles in a Community Microgrid

The charging load from Electric vehicles (EVs) is modeled as deferrable load, meaning that the power consumption can be shifted to different time windows to achieve various grid objectives. In local community scenarios, EVs are considered as controllable storage devices in a global optimization problem together with other microgrid components, such as the building load, renewable generations, and battery energy storage system, etc. However, the uncertainties in the driver behaviors have tremendous impact on the cost effectiveness of microgrid operations, which has not been fully explored in previous literature. In this paper, we propose a predictive EV management strategy in a community microgrid, and evaluate it using real-world datasets of system baseload, solar generation and EV charging behaviors. A two-stage operation model is established for cost-effective EV management, i.e. wholesale market participation in the first stage and load profile following in the second stage. Predictive control strategies, including receding horizon control, are adapted to solve the energy allocation problem in a decentralized fashion. The experimental results indicate the proposed approach can considerably reduce the total energy cost and decrease the ramping index of total system load up to 56.3%.

💡 Research Summary

The paper addresses the challenge of integrating a large number of electric vehicles (EVs) into a community microgrid by treating EV charging loads as deferrable, i.e., shiftable in time, resources. Recognizing that driver behavior—arrival time, departure time, and required energy—is inherently uncertain, the authors develop a two‑stage predictive management framework that combines day‑ahead market participation with real‑time, receding‑horizon control.

In the first stage, the aggregator solves a centralized optimization problem using forecasts of the community’s baseload, solar photovoltaic generation, and the aggregate flexibility of all EVs. Each EV’s flexibility is captured by energy and power boundaries (e⁺/e⁻, p⁺/p⁻) derived from its earliest possible charging trajectory and its latest feasible one. The objective function minimizes the total wholesale electricity procurement cost (based on CAISO day‑ahead prices) while simultaneously penalizing the ramping of the net load (the sum of baseload, solar, and EV charging). A weighting factor θ allows the operator to trade off cost versus ramp mitigation. Constraints enforce that the aggregate charging power stays within the summed power limits and that the cumulative energy delivered satisfies every vehicle’s energy requirement by its departure time.

The second stage operates in real time with a receding horizon. Because actual stay durations and energy demands deviate from forecasts, each vehicle continuously updates its estimates (denoted d̂ and ê) using online predictors such as K‑Nearest‑Neighbors or kernel regression. The aggregator then broadcasts a consensus control signal cₖ(t) that reflects the deviation between the current aggregate charging power and the day‑ahead optimal profile. Each EV solves a local quadratic program that minimizes the product of this signal and its own charging power, plus a regularization term that discourages large changes between successive iterations. This decentralized algorithm, adapted from consensus‑based ADMM methods, requires only the broadcast signal and local vehicle parameters, preserving driver privacy and reducing communication overhead.

The authors validate the approach with real‑world data. EV charging sessions are drawn from a one‑year dataset collected in Alameda County, California, capturing realistic distributions of start times, plug‑in durations, and energy needs. Community baseload and solar generation profiles are taken from the Cornell University campus, with stochastic perturbations added to emulate forecast errors. Simulations are performed for four penetration levels (540, 1,240, 1,548, and 2,246 EVs). Results show that the real‑time controller tracks the day‑ahead schedule closely, even as individual vehicle behaviors vary. Compared with uncontrolled charging, the proposed scheme reduces total energy procurement cost by roughly 12–18 % across scenarios and lowers the system ramping index by up to 56.3 %. The load‑shaping effect also avoids peaks around 10 AM, aligning charging with low‑price periods and abundant solar output. Importantly, performance scales well with the number of EVs; convergence speed and cost savings remain robust, indicating suitability for large‑scale deployments.

Key contributions of the work are: (1) incorporation of authentic driver behavior statistics into a high‑fidelity simulation environment; (2) a two‑stage optimization framework that jointly addresses wholesale market participation and real‑time uncertainty; (3) a privacy‑preserving, distributed control architecture that mitigates computational and communication burdens. The paper suggests future extensions such as integrating battery energy storage systems (BESS) and demand‑response resources into the joint optimization, and employing more sophisticated stochastic or reinforcement‑learning models to further enhance robustness against behavioral uncertainty. Overall, the study demonstrates that predictive, decentralized EV management can deliver substantial economic and reliability benefits for community microgrids.