Thermal simulation trained deep neural networks for fast and accurate prediction of thermal distribution and heat losses of building structures

Abstract

• Thermal simulation of building structures was trained by deep learning method. • Thermal behavior of building was predicted by deep neural network in a fast time. • Thermal distribution image was trained by convolutional encoding–decoding network. • Heat loss of building was trained by multi-layer perceptron. In this study, state-of-the art deep neural networks to train and predict the heat transfer in building structures were proposed. Today, many of studies analyze thermal energy performance of buildings by analytical or numerical methods. Although building energy performance can be predicted effectively by finite element method, it is still time-consuming to calculate and solve the heat transfer problem. Moreover, expert engineer is required and complicate process to set simulation model is essential. In this work, a novel deep-learning method, which was pre-trained by the thermal simulation data, was developed to predict the thermal behavior of building structures in a fast time without complicated process. Heat transfer simulations of the slab wall building structure depending on its thermal properties and geometries were carried out to get training datasets for deep learning. The database of thermal simulation results was used for deep learning training. The image of temperature and heat flow distribution was trained by convolutional encoding–decoding network and the value of total heat loss through building and thermal bridge coefficient was trained by multi-layer perceptron. After train completed, the thermal behavior could be predicted in a second by just feeding information such as blueprint image and thermal properties of constructions into deep-learning architecture. There was no need to set a new simulation model at each time which consumes time and effort for modeling, meshing and calculating. With the developed network, the prediction of thermal behavior with high accuracy was possible in a super-fast time.