A Study of Reinforcement Learning Based Predictive Control Framework for Optimal Bidirectional Electric Vehicle Charging in a Small Office Building
Year: 2026
Type: dissertation
Abstract: Vehicle-to-Building (V2B) operation enables electric vehicles (EVs) to provide flexible energy resources for building energy management, but bidirectional charging requires coordinated management of EV availability, building demand, photovoltaic (PV) generation, demand-response (DR) events, and electricity prices. This thesis proposes a reinforcement learning-based predictive control (RLPC) framework for bidirectional EV charging to ensure EV fulfillment, reduce energy costs, and mitigate peak grid demand. A modified Soft Actor-Critic (SAC) supervisory controller generates 24-hour charging and discharging power envelopes for charging-station groups using forecasts of building demand and PV generation, DR notifications, and pricing signals. These envelopes are converted into per-EV power setpoints by a rule-based regulatory controller that accounts for real-time EV availability, energy requirements, and battery limits. EVnet-Gym is developed for training and evaluation under stochastic EV behavior, building demand, PV generation, electricity prices, and battery constraints. Simulation results show that RLPC reduces accumulated energy cost by approximately 57% compared with as-fast-as-possible (AFAP) charging and 17% compared with limited as-fast-as-possible (LAFAP) charging, while maintaining EV battery fulfillment above 98% relative to user-specified departure targets. Based on daily peak comparisons, RLPC reduces all-day peak grid import by approximately 86% relative to AFAP. Relative to LAFAP, RLPC reduces morning on-DR peak demand by approximately 7%, but increases afternoon on-DR and off-DR peak demand by approximately 27% and 9%, respectively. This reflects a trade-off between energy-cost reduction through bidirectional V2B utilization and higher peak demand during afternoon on-DR and off-DR periods. Overall, RLPC reduces operating costs, supports peak-demand mitigation, and satisfies EV charging requirements.
Author: Phi, Gia-Khanh
Institution: Concordia University
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7221492104
