Development of a DNN model using calibrated simulation for the optimal control of HVAC systems

Abstract

Acquiring actual data for developing data-based prediction models for the optimization of control setpoints (hereinafter referred to as optimal control (OC)) in heating, ventilation, and air conditioning (HVAC) systems is challenging, especially when changing control setpoints while considering all possible boundary conditions. Thus, considering the scalability of OC, it is a promising solution to generate sufficient learning data according to the control settings using a highly accurate simulation model, build a source model, and expand the learning model based on the actual data of the target system. This is the first study to improve the scalability of OC aiming to develop a deep neural network (DNN) model to predict the power consumption of a heat source system. First, this study will construct a calibrated heat source system simulation model and then use the generated learning data to train a DNN model. The actual operation conditions of the target building and heating system located in Vietnam were analyzed using operation data. A physics-based simulation model of the target heat source system was developed and calibrated using actual operation data to improve accuracy. By using the calibrated simulation, sufficient learning data was generated for all control settings allowing for the development of a learning model based on DNN that predicted the power consumption of the heat source system. In addition, a highly accurate prediction model through hyperparameter optimization was secured.