Logo image
Back
SOCDASH: a scalable occupant centric data predictive control framework for residential air source heat pump-based systems
Dissertation   Open access

SOCDASH: a scalable occupant centric data predictive control framework for residential air source heat pump-based systems

Richard Lamont Kimball III
Doctor of Philosophy (Ph.D.), Drexel University
Mar 2026
DOI:
https://doi.org/10.17918/00011314
pdf
Kimball_Richard_20266.56 MBDownloadView
PDF Open Access Open Access (License Unspecified)

Abstract

Class imbalance problem Data predictive control Multi-objective control Occupant centric control Computer Science
Many residential homes are transitioning to electric heat pumps for both heating and cooling as these systems produce no on-site carbon emissions. However, 40% of residential electricity is still generated from non-renewable sources that emit carbon. Thus, even homes that adopt heat pumps need to reduce electricity demand and operate more efficiently. Model predictive control (MPC) offers an effective strategy for optimal HVAC system operation to reduce electricity consumption especially when data-driven machine learning (ML) models are used, allowing for data-driven MPC or DPC. Yet, occupants' thermal behaviors could have significant impact on the effectiveness of MPC or DPC. Thus, the inclusion of occupant thermal comfort in a DPC strategy, or occupant centric control (OCC), can improve the performance of MPC/DPC, while improving the wellbeing of occupants. Nevertheless, occupant comfort data often suffers from class imbalance problem (CIP) which leads to poor performance for forecasting models trained upon it and there is a lack of OCC strategies for residential homes with air-source heat pump systems. This study develops SOCDASH, a scalable occupant centric DPC framework for detached, single family homes with air source heat pump-based HVAC systems. To support the development and evaluation of this framework, a virtual testbed is firstly created to simulate of a residential building with occupants exhibiting stochastic thermal comfort behaviors, such as thermostat setpoint adjustments. This testbed is used to generate training data for two core components in the SOCDASH framework: whole building electric energy forecasting models and occupant thermal comfort vote forecasting models. It also provides evaluation data for assessing the SOCDASH framework performance. A systematic comparison of synthetic data resampling methods found success in addressing CIP in occupant comfort data and resulted in an average 10% increase in occupant thermal comfort forecasting vote model performance trained on resampled vs imbalanced data. A multi-criteria fuzzy model structure score was developed to aid in selection of the best ML model structure for whole building electric energy forecasting models with inclusion of stochastic occupant behaviors. The SOCDASH framework is initially evaluated under Philadelphia, PA weather conditions for both heating and cooling seasons, using occupants with typical thermal preferences. To evaluate the scalability of the framework, two additional evaluation cases are developed: one case with a more extreme weather pattern, and the other case with skewed occupant thermal preferences. The developed SOCDASH framework on average reduced whole building electric energy consumption by 15% and occupant discomfort votes by 55%, when compared to a baseline control strategy.

Metrics

7 File views/ downloads
2 Record Views

Details

Logo image