Skip to Content
SustainabilitySustainability
  • Article
  • Open Access

18 September 2020

38 Pages

Evaluating the Adaptability of Reinforcement Learning Based HVAC Control for Residential Houses

,
,
,
,
,
,
,
,
and
1
Computational Sciences and Engineering Division, Oak Ridge National Laboratory, Oak Ridge, TN 37831, USA
2
Energy and Transportation Science Division, Oak Ridge National Laboratory, Oak Ridge, TN 37831, USA
3
Computer Science and Mathematics Division, Oak Ridge National Laboratory, Oak Ridge, TN 37831, USA
4
Department of Electrical Engineering and Computer Science, University of Tennessee, Knoxville, TN 37996, USA

Abstract

Intelligent Heating, Ventilation, and Air Conditioning (HVAC) control using deep reinforcement learning (DRL) has recently gained a lot of attention due to its ability to optimally control the complex behavior of the HVAC system. However, more exploration is needed on understanding the adaptability challenges that the DRL agent could face during the deployment phase. Using online learning for such applications is not realistic due to the long learning period and likely poor comfort control during the learning process. Alternatively, DRL can be pre-trained using a building model prior to deployment. However, developing an accurate building model for every house and deploying a pre-trained DRL model for HVAC control would not be cost-effective. In this study, we focus on evaluating the ability of DRL-based HVAC control to provide cost savings when pre-trained on one building model and deployed on different house models with varying user comforts. We observed around 30% of cost reduction by pre-trained model over baseline when validated in a simulation environment and achieved up to 21% cost reduction when deployed in the real house. This finding provides experimental evidence that the pre-trained DRL has the potential to adapt to different house environments and comfort settings.

Article Metrics

Citations

Article Access Statistics

Multiple requests from the same IP address are counted as one view.