Model predictive HVAC control with online occupancy model

• We develop a stochastic single-zone occupancy model using a Markov chain.
• Bayesian inference with automatic forgetting trains the Markov chain on-line.
• Occupancy predictions are applied to the MPC discomfort cost function.
• Simulation demonstrates energy savings and comfort performance between those of occupancy-triggered and scheduled controllers.

This paper presents an occupancy-predicting control algorithm for heating, ventilation, and air conditioning (HVAC) systems in buildings. It incorporates the building's thermal properties, local weather predictions, and a self-tuning stochastic occupancy model to reduce energy consumption while maintaining occupant comfort. Contrasting with existing approaches, the occupancy model requires no manual training and adapts to changes in occupancy patterns during operation. A prediction-weighted cost function provides conditioning of thermal zones before occupancy begins and reduces system output before occupancy ends. Simulation results with real-world occupancy data demonstrate the algorithm's effectiveness.