Abstract
Buildings contribute to 40% of global energy consumption, with 36% attributed to heating, ventilation, and air-conditioning (HVAC). Therefore, optimizing HVAC control in buildings is crucial in the transition to a more sustainable society. Model predictive control (MPC) and (deep) reinforcement learning (DRL) have been explored for optimal control strategies, producing promising results. However, their performance depends on the underlying simulation model's accuracy, which is why an accurate model throughout the building's lifecycle is important. Physics-based models introduce discrepancies due to necessary simplifications, called the sim-to-real-gap. Closing this gap requires expert knowledge to increase the model's complexity, which is often not feasible. Given the emergence of smart, sensor-equipped buildings, data-driven solutions are possible, enabling a hybrid model that exploits the advantages of data-driven and physics-based models. First, data-driven models, like for example deep neural networks (DNNs), are added to the physics-based model on the level of the components to close the sim-to-real-gap. Second, as components degrade, the sim-to-real-gap will grow again and is closed using the same approach. Third, the hybrid model facilitates automatic fault detection and diagnosis (AFDD) using results from the adjustment process. Finally, an assessment of energy loss due to component degradation guides cost-optimal maintenance strategies.
Researcher(s)
Research team(s)
Project type(s)