Beyond correlation: A causality-driven model for indoor temperature control

Recent advancements in data availability and computational resources have driven the widespread adoption of machine learning (ML) in control-oriented building simulations. To model dynamic relationships in building systems, many ML models, including artificial neural networks (ANNs), rely on statistical associations between the input and output variables (i.e., correlations). However, these methods often fail to provide reliable predictions under unfamiliar conditions. This study highlights how such correlational approaches can misinterpret confounding factors and operational biases in real-world datasets, leading to the misrepresentation of the underlying mechanisms in building systems. To address this, this study proposes a novel causality-driven modeling framework for a 38.5 m 2 office space equipped with a heat pump (HP) cooling system. The model estimates the heterogeneous treatment effects (HTE) of the HP cooling set point temperature (SET) on the indoor air temperature (IAT) under varying control conditions as well as generates counterfactual IAT outcomes under different HP controls. To achieve this, outdoor weather, indoor environment , and HP operational data were collected over 28 days. Double machine learning (DML) was employed to mitigate biases within the dataset and isolate the causal relationships between the control and its outcome. In this framework, IAT was treated as the outcome variable, SET as the treatment variable, and outdoor air temperature , solar radiation, and prior thermal conditions (e.g., supply air temperature) were included as key confounders. The proposed DML model trained on causal relations outperformed an ANN model trained on correlations, both of which were developed using the same dataset. The DML model demonstrated a superior capability to capture causal mechanisms under varying control conditions, closely aligning with physical principles and realistic operational behaviors . Its validity was further supported by a statistically significant p -value (0.048) for HTE on IAT. In contrast, even though the ANN model achieved a low MSE (0.19 °C) in IAT prediction, it often failed to maintain physical consistency under unfamiliar control options. These findings highlight the importance of transitioning from correlation-driven to causality-driven approaches for developing reliable and scalable building control systems based on true causal mechanisms.