Causal modelling of occupant window behaviour for indoor temperature prediction

Data-driven models have been used to capture human–building interactions from measured data. However, correlation alone doesn’t always explain why occupants act or how their actions alter indoor thermal dynamics. This study develops a causality-based model that represents window operation as a behaviour driven by indoor conditions and as an intervention modifying outdoor–indoor built environment. Using data from four residential households in Seoul, South Korea, a structural causal model (SCM) incorporating logistic regression for window-opening behaviour was compared with a correlation-based model. Although both models showed comparable goodness-of-fit, only the SCM separated thermal continuity from behaviour-dependent exchange and enabled counterfactual evaluation of window states. Household-specific window-opening thresholds ranged from 21.60 to 28.02 °C, and a counterfactual case indicated a 0.154 °C change under alternative behaviour. The SCM further distinguished epistemic uncertainty from behaviour-driven aleatoric uncertainty, which was greatest when window-opening probability approached 0.5, where occupant behaviour was most uncertain.