Development of a Data-Driven Predictive Control Based on a Novel Physics-Informed Neural Network

Buildings account for a large portion of the total energy usage. HVAC systems take up the biggest part of energy usage in the building sector. Research has indicated that 15% of energy was wasted due to building control. Therefore, it is inevitable to adopt an advanced control strategy for HVAC systems to mitigate energy waste in the building sector. Among many control strategies, model predictive control (MPC) provides promising potential due to its ability to consider different objectives. Many research studies have demonstrated that MPC can significantly reduce HVAC systems’ energy usage with the guarantee of thermal comfort and indoor air quality. However, its broad implementation was impeded by the expertise required for the development of such strategies. According to a recent study, as one major effort required for MPC, model development and integration take 79 person-days. Therefore, a scalable, control-oriented, and physically reasonable model would greatly facilitate the implementation of predictive strategies. The machine learning approach is suitable for developing such a model, but traditional neural networks are not designed for control purposes. In this research, we proposed a novel physics-informed neural network with a special structure to facilitate optimal control of HVAC systems. The model was validated to capture building dynamics more accurately than the grey-box model while making physical sense. Coupling the model into a predictive control strategy, the simulation indicated more than 50% energy saving at the maximum in the cooling season.