House Energy Management System, for balancing Electricity Costs and Residential Comfort, based on Deep Reinforcement Learning

Abstract

Smart homes are becoming increasingly popular for their potential to reduce electricity costs, through device optimization. Balancing residential comfort with electricity cost reduction presents significant challenges. To tackle this problem, we developed a House Energy Management System (HEMS) using Deep Reinforcement Learning (DRL) to reduce electricity costs, by orchestrating device usage, without compromising residential comfort. The HEMS was trained on a smart home simulation powered by the Typhoon HIL application and supplemented with real-world data, from the Mainflux IoT Platform. The simulation included HEMS-controllable and uncontrollable devices, a solar panel and the electricity grid. We modelled a reward function that balances electricity cost with the residents' comfort and used it to train two DRL models: Double Deep Q Network (DDQN) and Proximal Policy Optimization (PPO). Our findings show that PPO maintains thermal comfort and reduces electricity costs more effectively than does DDQN, particularly in the colder season. As the PPO models’ behavior is season-dependent, it can reduce residential effort by automatically adjusting device schedules in response to changing weather conditions.