Application of artificial neural networks using sequential prediction approach in indoor airflow prediction. (15th June 2023)

• Record Type: Journal Article
• Authors: Kim, MinHo; Park, Hyung-Jun
• Is Part Of: Journal of building engineering. Volume 69(2023)
• Publication Date: 2023-06-15
• Subjects: Artificial neural networks (ANN) -- Computational fluid dynamics (CFD) -- Indoor environment -- Displacement ventilation -- Natural convection -- Richardson number; Building -- Periodicals; 690.05
• DOI: 10.1016/j.jobe.2023.106319
• Languages: English
• ISSNs: 2352-7102

Abstract

Abstract: The present study proposes a new approach based on artificial neural network (ANN) for predicting thermodynamic parameters in an indoor environment. The proposed prediction approach consists of two independently trained ANN models, where some output variables from the previous model are used as inputs to the subsequent model. To obtain the target variables, which are pressure and temperature distribution, the velocity distribution is defined as the influencing variable that has a significant impact on the target variables. The predicted velocity distribution from the first model is employed as an additional input for the second model. The training, validation, and testing datasets are obtained from the results of computational fluid dynamics (CFD) analysis, and it contains a total of 100 case scenarios. To evaluate the accuracy of the prediction model, several performance indices are utilized, such as R 2 and root mean square, etc., which are based on mathematical statistics methods. The prediction results show that adding variables affecting the output value, such as the velocity, to the input node improves the performance of the model. That is, the proposed approach outperforms existing ANN models and provides a reasonable solution for indoor airflow prediction. Highlights: An artificial neural networks (ANN) model to predict thermodynamic parameters in an indoor environment is proposed. Two independent ANN models are sequentially connected to increase prediction accuracy. Datasets for training, validation, and testing are obtained from the results of CFD analysis. The proposed model shows improved results in predicting velocity, pressure, and temperature distribution. Performance is demonstrated through various indices based on mathematical statics methods.