Optimal operation of natural ventilation using a combined method of machine learning and physical simulation

Natural ventilation is an effective means of saving energy in office buildings. However, to maximize the effectiveness of natural ventilation, advanced operational techniques are required. Natural ventilation window should be opened as much as possible without compromising comfort, and the operating hours of HVAC system must be minimized as much as possible. Therefore, it is necessary to make sequential decisions on whether or not to open natural ventilation window based on indoor and outdoor environment. In recent years, machine learning methods have been proposed to optimize operations based on operational data. However, many issues have been pointed out, such as low accuracy in the stage immediately after construction is completed, when there is little operational data. In this study, we propose a method that combines classical physical simulation and machine learning to complement the weakness of optimal control using machine learning. Specifically, we attempted to make decisions on natural ventilation control using a new composite model that combines both a neural network model as a machine learning method and a thermal network model as a physical simulation method to predict the indoor environment after natural ventilation. In a case study for a virtual office building model in Tokyo, the proposed model showed higher prediction accuracy than the neural network model alone during the period shortly after construction was completed, when operational data was scarce, and was able to reduce RMSE (Root Mean Squared Error) by about 97% in predicting room temperature three hours after the start of natural ventilation. This study suggests that a newly proposed model for predicting the indoor environment can be used to achieve more accurate natural ventilation control.