CH-24-C104 - Bayesian Causal Inference for Occupant-Centric Building System Operation

The decarbonization of the building sector is important to achieve a carbon-neutral society since buildings account for 30 % of the total greenhouse gas emissions. Since it is widely acknowledged that the behavior and decision-making of building occupants significantly affect the level of carbon emissions from buildings, occupant-centric building operational strategies have been proposed to minimize energy consumption but satisfy their needs. Recent studies have introduced two issues in previous solutions with data-driven occupant modeling: (i) the importance of ensuring causal relationships between independent and dependent variables for building system operations and (ii) the vulnerability in prediction over unseen and heterogeneous data. To improve building energy solutions’ reliability considering the issues, implying causal structures between variables in occupant models is important for occupant-centric building operation. Nevertheless, identifying causal relationships between occupants’ perception and behavior and other factors is challenging in buildings because of the difficulty in disentangling the direct causal impact of one factor from other influencing factors. There have been two major reasons that make the quantification of the causal impacts challenging: difficulties in (i) conducting fully controlled experiments with real occupants and (ii) considering potential causal relationships with conventional statistical analysis over observational data. To deal with these challenges, modern data-driven causal inference methods have been proposed as promising tools to identify causal relationships and estimate causal effects from observed data. Implying causal inference methods in occupant modeling can improve the robustness of further occupant-centric building operations. In this regard, this study aims to (i) introduce a causal inference method with a Bayesian perspective to identify a causal structure from a synthetic dataset, (ii) develop a causal model based on the inferred causal structure, and (iii) evaluate the causal model’s performance compared with the performance of the association-based (non-causal) model. In the model evaluation, we considered two critical aspects: the estimation of true causal effects whereby independent variables impact a dependent variable and the assessment of prediction robustness to dataset shift. While the causal model estimated the true causal effects properly, the non-causal model fails to capture the causal impacts from the independent to dependent variables. Within the in-distribution dataset, both causal and non-causal models demonstrated comparable prediction performance. However, when evaluated on the Out-of-distribution dataset, the causal model showed an RMSE of 0.16 while the non-causal model indicated an RMSE of 2.86. The results show that the prediction model based on correct causal structure can outperform the non-causal model in two realms, (i) correct estimation of causal effects and (i) more robust prediction over dataset shift. Therefore, the causal models should be used for occupant-centric building operations such as optimal controllers or occupant-centric thermostat interfaces to ensure the causal effect of adjusting independent variables on dependent variables such as occupant comfort and behavior.