Ventilation plays a noteworthy role in maintaining a healthy, comfortable and energy-efficient indoor environment and mitigating the risk of aerosol transmission and disease infection (e.g., SARS-COV-2). In most commercial and office buildings, demand-controlled ventilation (DCV) systems are widely utilized to conserve energy based on occupancy. However, as the presence of occupants is often inherently stochastic, accurate occupancy prediction is challenging. This study, therefore, proposes an autoencoder Bayesian Long Short-term Memory neural network (LSTM) model for probabilistic occupancy prediction, taking account of model misspecification, epistemic uncertainty, and aleatoric uncertainty. Performances of the proposed models are evaluated using real data in an educational building at the University of Cambridge, UK. The models trained on data of one open-plan space are used to predict occupant numbers for other spaces (with similar layout and function) in the same building. The probabilistic occupant profiles are then used for estimating optimal ventilation rates for two scenarios (i.e., normal DCV mode for energy conservation and anti-infection mode for virus transmission prevention). Results show that, during the test period, for the 1-h ahead prediction, the proposed model achieved better performance with up to 5.8% mean absolute percentage error reduction than the traditional LSTM model. More flexible alternatives for ventilation can be offered by the proposed risk-aware decision-making schemes serving different purposes under real operation. The findings from this study provide new occupancy forecasting solutions and explore the potential of probabilistic decision making for building ventilation optimization.