EECinsights - Probabilistic Occupancy Forecasting for Risk-aware Ventilation

Predicting future occupancy profiles for ventilation predictive control could be a promising approach to reducing the latency impacts between sensors and control systems. Forecasting occupancy in a probabilistic manner with uncertainty bound or confidence interval enables the decision-makers to make risk-informed building energy management and ventilation operations. We are developing Bayesian deep neural networks for probabilistic occupancy prediction and ventilation operations.

Our research addressed the following challenges:

How are the different uncertain sources taken into account in model development?

The complete architecture of the developed Bayesian deep neural networks framework for occupancy forecasting contains two layers: (i) an unsupervised learning layer, autoencoder, that captures the inherent pattern in the time series and is learned during pre-training, and (ii) a supervised prediction network that receives input both from the learned representative embeddings within the autoencoder framework as well as external features (e.g., calendar information) to predict the occupancy. Three types of prediction uncertainty are considered in this deep learning framework: model misspecification, epistemic uncertainty, and aleatoric uncertainty.

The proposed model can be suitable to use for the risk-aware decision-making process, as its predictive confidence interval covered almost all observed occupant values.

How are the predicted probabilistic occupancy profiles utilised for risk-aware ventilation decision making?

Based on the probabilistic occupancy profiles, we provide decision-making schemes for two operation modes, normal demand-controlled ventilation (DCV) mode and anti-infection mode under pandemic and non-pandemic periods, respectively. The normal DCV mode is mainly intended to save energy with the satisfaction of indoor air quality in response to changes in occupant number. The anti-infection mode, which is a special form of DCV, aims to reduce the risk of individual infection by introducing the outdoor air to dilute the indoor air contaminant based on occupancy density. Below two figures show the probability of ventilation inadequacy under normal DCV mode and individual infection risk for 1-hour exposure under anti-infection mode respectively. To provide indoor spaces with low infection probability and high energy efficiency, decision-makers can make their feasible decisions by setting proper risk thresholds.

More detailed information can be found in our recent publication in Building and Environment.