Aleatoric and epistemic uncertainties in occupant behavior models for residential buildings: real-life cases

Even though occupant Behavior (OB) has been known as one of the significant sources of uncertainty in Building Performance Simulation (BPS) for the last decade, the reliable modeling of OB and its impact on building design and operations still demands deeper investigation. In general, the OB models inherently carry uncertainties due to occupant diversity, temporal and spatial variation, insufficient data of monitored OB responding to multi-modal (thermal, luminous, acoustic, IAQ) indoor environments, and many other unknown confounding factors.

In particular, data-driven OB models, one of the popular OB modeling approaches, utilize machine-learning algorithms owing to their high prediction accuracy. However, data-driven OB models heavily depend on training data and this data dependence limits the models’ generalizability and scalability. With this in mind, this study presents quantified two uncertainties in machine-learning-based OB models: aleatoric and epistemic.

For this purpose, the authors used measured data including indoor and outdoor temperatures and humidities, indoor illuminance, CO2 concentration, and windows opening out of four residential households at a sampling time of 1 min for 12 months (Aug 2001-Sep 2002). Subsequently, we developed five OB prediction models—presence state, window state, light switch, AC switch, and boiler switch. After applying Bayesian deep learning techniques to the aforementioned five OB models, it was found that the aleatoric uncertainty, caused by unpredictable occupant behavior, is not easily mitigated by supplementing the training data. In contrast, the epistemic uncertainty, generally stemming from data insufficiency, can be reduced by enhancing the dataset. However, it is surprising that the aleatoric uncertainty tends to have larger values compared to the epistemic uncertainty. This implies that when employing the data-driven OB models, it needs to investigate the aforementioned two uncertainties beyond the accuracy metrics (CVRMSE, MAPE, MAE, etc.). This will enhance the reliability and interpretability of the OB models.