Physics-informed neural network for chiller plant optimal control with structure-type and trend-type prior knowledge

Abstract

The development of advanced controller for heating, ventilation, and air conditioning (HVAC) system contributes significantly to building energy conservation. While the success of these optimal control technologies is highly relied on the accuracy of energy models. Existing energy models are mostly based on data-driven models, and their extrapolation/generalization ability is the major barrier for their real-world application. To solve this problem, this paper proposes a general framework of physics-informed neural network (PINN) to improve the extrapolation performance of energy models. The prior physics knowledge is divided into structure-type knowledge and trend-type knowledge, and they are embedded into neural network, forming the structure-type physics-informed neural network (S-PINN) and trend-type physics-informed neural network (T-PINN). The S-PINN aims at using known physics equation to guide the design of network architecture, while the T-PINN is to transform known trend relationship as physics loss function to ensure network output is consistent with physical trend. The overall idea of PINN is applied for the optimal control task of chiller plant in a real commercial building. The energy models of chilled water pump, cooling water pump, cooling tower and chiller are developed using both history data and physics knowledge. Comprehensive experiments are conducted to compare the extrapolation performance of gray-box model, pure data-driven model, and proposed PINN. The results demonstrate that both the structure-type knowledge and trend-type knowledge can significantly improve the model extrapolation performance. And the field experiments showed that the developed PINNs achieved 23.2 % improvement of energy efficiency by resetting system control setpoint.