Correlation vs. causal graph-based analysis of heat pump control

Data-driven models are widely used to capture HVAC dynamics but often yield physically inconsistent and non-generalizable predictions, especially under unseen conditions, as they rely on statistical correlations rather than causal mechanisms. To address this limitation, this study proposes a shift from correlation-based to causality-driven reasoning for comprehending HVAC thermal dynamics. Using 14 days of data from an indoor zone with a heat pump (HP), a domain-informed causal graph was constructed to represent how HP controls shape the measured indoor thermal dynamics. Average Treatment Effect (ATE) along the target path was estimated using two approaches, followed by linear regression: a causality-driven approach, controlling for adjustment variables to block backdoor paths; a correlation-based approach, without adjustment. Results show that causality-driven approach yields more physically meaningful relationships. While correlation suggests a 1°C increase in HP return air temperature (RAT) is associated with a 2.1°C decrease in supply air temperature (SAT), the causality-driven approach reveals a 3.6°C decrease after blocking the backdoor path. Furthermore, the causal graph decomposes the total effect of setpoint temperature on indoor air temperature (0.513°C) into contributions like thermal inertia (0.511°C) and supply air convection (0.294°C). This study demonstrates that causality-driven analysis enhances the physical consistency and interpretability of data-driven HVAC analysis.