Quantification of aleatoric and epistemic uncertainty in Data-Driven occupant behavior model for building performance simulation

Occupant behavior (OB) is a major source of uncertainty in building performance simulations, significantly influencing energy consumption and indoor environmental conditions. While machine learning-based OB models have shown strong predictive capabilities, their reliability is often undermined by data limitations, extrapolation errors, and a lack of explicit uncertainty quantification. This study introduces a Bayesian Deep Learning (BDL) framework to address these challenges by quantifying OB-related uncertainties. Using Monte Carlo (MC) dropout, the framework distinguishes between aleatoric uncertainty (inherent randomness) and epistemic uncertainty (knowledge limitations).

Experiments were conducted using data from six residential households in Seoul, South Korea, over a three-month summer period. This study examines how the training period affects uncertainty and evaluates its impact on energy predictions through a co-simulation with EnergyPlus and BCVTB. Results indicate that aleatoric uncertainty was the dominant factor during model validation, primarily due to sensor noise and unpredictable occupant behavior. However, epistemic uncertainty increased in the co-simulation, especially under extrapolated conditions, leading to greater variability in energy predictions. Extending the training period reduced epistemic uncertainty, lowering the coefficient of variation in energy predictions from 54.3% to 20.4%, but had no noticeable effect on aleatoric uncertainty, which substantiates the inherent unpredictability of OB.

These findings highlight the need to incorporate uncertainty metrics alongside accuracy metrics when evaluating data-driven OB models, as prediction confidence is crucial for simulation-based decision-making. Moreover, the observed uncertainty propagation in energy predictions underscores the advantages of probabilistic modeling over deterministic approaches. This study provides a systematic framework for integrating uncertainty analysis into data-driven OB modeling, offering insights into improving model robustness, generalizability, and practical applicability in building performance simulations.