Does causality matter? How incorporating cause-and-effect relationships benefits occupant-centric building control

Occupant-centric building control solutions have advanced alongside the development of data-driven modeling and control methodologies. Meanwhile, a growing body of literature has drawn renewed attention to the importance of considering causality when developing occupant-centric solutions, highlighting its potential to improve model robustness, which may in turn improve solution performance. Given the sparse empirical evidence, however, it remains unclear whether occupant-centric solutions based on causal relationships would provide practical benefits over conventional data-driven solutions primarily relying on associations in data. Through a simulation study, this paper demonstrates a tangible improvement in control performance when the cause-and-effect relationships behind occupant thermostat behavior are known and explicitly incorporated. To this end, we created five simulation environments, each consisting of a residential building and a synthetic occupant who stochastically overrides the thermostat. Using three months of data collected from each environment, we developed two data-driven occupant models (causal and non-causal) that predict thermostat override behavior. From each occupant model, we then trained corresponding optimal thermostat controllers via reinforcement learning with different reward functions and deployed them in the ground truth environments. The causal controllers achieved 0.06–81.9% higher mean electricity cost savings than the non-causal controllers. Moreover, the causal controllers resulted in 49.4–86.4% fewer mean thermostat overrides by the occupant, except for one case in which the occupant preferred cooler indoor conditions. In conclusion, this study establishes two findings. First, incorporating cause-and-effect relationships behind occupant behavior into occupant-centric control solutions can tangibly enhance the solutions' performance. Second, this improvement arises because causal occupant models provide more robust predictions under distribution shift, which may be unavoidable during solution deployment. This study therefore implies that discovering new cause-and-effect relationships behind occupant behavior is an important research task; data-driven causal learning approaches may help infer causal knowledge from observational settings.