Advanced control algorithms promise to reduce residential heating and cooling energy consumption, but practical deployment in real homes remains rare. Researchers conducted a month-long field trial comparing two approaches—Model Predictive Control (MPC) and Reinforcement Learning (RL)—operating an air-to-air heat pump in an occupied house in a cold climate.
Both algorithms dynamically adjusted the thermostat setpoint based on indoor temperature and real-time power consumption. Relative to constant-setpoint operation, MPC delivered 18.1% energy savings while RL achieved 20.9% savings, both figures weather-normalized and statistically significant. These results validate that data-driven control can meaningfully reduce residential heating demand.
However, the algorithms diverged in occupant experience. MPC maintained thermal comfort throughout the trial. RL, by contrast, kept the house cooler than preferred during its initial learning phase, generating three discomfort complaints. This reflects a fundamental RL challenge: the algorithm must explore the state-action space while guaranteeing acceptable conditions—a tension difficult to resolve without pre-training or safety constraints.
On engineering effort, RL required approximately one-third less work to deploy than MPC, primarily because RL avoids the need for detailed system modeling. Both methods consumed similar amounts of operational data, suggesting data availability is not a limiting factor.
The study reveals important trade-offs for grid operators and building managers. RL's lower deployment cost could accelerate adoption of demand-response-capable HVAC across millions of homes. Yet the comfort issues encountered underscore the need for safer initialization strategies and better alignment between controller assumptions and actual building dynamics.
These field results bridge a critical gap: laboratory comparisons alone cannot reveal how algorithms perform when weather varies, buildings have unknown thermal characteristics, and occupants report satisfaction in real time. The findings suggest that hybrid approaches—combining RL's ease of deployment with MPC's comfort guarantees—may offer the path forward.
