Comparison of different deep neural network architectures for isothermal indoor airflow prediction

Abstract

The rising awareness about energy conservation calls for more energy-efficient designs of heating, ventilation and air-conditioning (HVAC) systems and advanced control strategies. A rapid and accurate prediction method for indoor environment is thus of great importance. As one of the popular artificial intelligence models, a deep neural network (DNN) performs well in establishing a non-linear relationship between variables. This study aims to explore the feasibility of adopting DNN for predicting indoor airflow distribution. The detailed process of computational fluid dynamics (CFD) database establishment and construction of DNN are presented herein. To reveal the influence of DNN architecture on prediction performance, two DNNs, namely DNN A and DNN B, were constructed in consideration of different prediction strategies, and their performances on both training dataset and test dataset were compared. DNN A represents a DNN that outputs velocity values of the whole domain simultaneously, whereas DNN B outputs one velocity value of a target location for each run. Results indicate that the performance of DNN A and DNN B on the training dataset show slightly difference; however, DNN A performs much better than DNN B on test dataset. DNN A shows similar accuracies on predicting airflow distributions with different resolutions, whereas the prediction accuracy of DNN B deteriorated for airflow distributions with higher resolution. Both DNNs would output unreliable predictions for cases of which inputs are out of the range of those of training cases. The results confirm the possibility of rapid prediction of indoor airflow via a well-trained DNN which requires about 320 μs for each case, resulting in about 1.9 million times faster than CFD simulation in this study.