Data-driven control of split heat-pump system based on thermal preference prediction

In this work, an experimental investigation was conducted in an office building involving a total of 16 participants. Subjective thermal sensation and thermal preference were assessed through periodic surveys. Several machine learning classification models were developed and evaluated, including Random Forest, K-Nearest Neighbors, Support Vector Machines, Logistic Regression, and Decision Trees. The Random Forest algorithm achieved the highest predictive performance, reaching an accuracy of 96.36% under random data splitting and 65.6% when chronological splitting was applied. The developed model was further evaluated under real operating conditions using a developed dynamic control framework for a split heat-pump system. The results demonstrate the feasibility of integrating thermal preference prediction into the control of split heat-pump systems under real office operating conditions. However, the reduced performance observed under chronological and user independent validation indicates that model performance may vary over time and across different users, highlighting the challenges of deploying thermal preference prediction models in real-world applications. Furthermore, differences between the heating and cooling operating periods suggest that seasonal operating conditions influence model performance. Overall, the findings support the potential of machine learning-based thermal preference prediction for occupant-centric heat-pump control while emphasizing that further research involving larger datasets, long-term field validation, quantitative energy performance assessment, and advanced or hybrid control strategies is required before practical large-scale deployment.

Energy and Buildings, 370, 118039