Abstract

Residential electrification is accelerating in cold-climate regions such as Quebec, where approximately 60% of space heating demand is supplied by electricity. As a result, winter morning and evening peak periods place increasing strain on the electrical grid as heating demand coincides with household appliance use and residential electric vehicle (EV) charging. This challenge is expected to intensify with increasing EV adoption. Addressing peak demand while preserving occupant comfort and mobility requires intelligent, scalable control strategies capable of managing heterogeneous residential loads.

The objective of this dissertation is to quantify and exploit the energy flexibility of residential buildings through automated control of distributed energy resources, with the aim of shifting electricity demand away from peak hours while accounting for varied occupant preferences and behaviors. The work investigates reinforcement learning (RL)-based control of residential energy storage systems (ESS), heat pump heating supply power, and bidirectional EV charging in residential neighborhoods. For distributed ESS, decentralized RL achieves a 45% reduction in electricity cost and a 51% reduction in peak-period imports. Decentralized RL delivers more than double the savings of a rule-based controller (20% cost reduction) and approaches the performance of a day-ahead MILP benchmark (66% cost reduction), demonstrating competitive performance despite relying only on observable states and near-term forecasts. The work then examines residential heating flexibility during winter demand response events by explicitly modeling occupant thermostat setpoint preferences and override behavior. The multi-agent RL controller reduces electricity consumption by 17% during demand response events while maintaining acceptable indoor air temperatures and limiting thermostat overrides. Finally, the dissertation assesses the flexibility potential of residential EVs under diverse commuting patterns using RL-based bidirectional charging control. Compared with business-as-usual charging, RL significantly reduces peak-period electricity use and costs while meeting state-of-charge requirements at vehicle departure.

Overall, this dissertation demonstrates that RL-based control can coordinate ESS, heat pumps, and EVs to provide meaningful grid services in residential neighborhoods. By integrating occupant behavior into the analysis, the work shows that peak load reductions and economic benefits can be achieved without compromising comfort or mobility, supporting intelligent residential control as a pathway to more reliable, electrified energy systems.