Influence of data preprocessing on neural network performance for reproducing CFD simulations of non-isothermal indoor airflow distribution

• Published in: Energy and buildings Vol. 230; p. 110525
• Main Authors: Zhou, Qi, Ooka, Ryozo
• Format: Journal Article
• Language: English
• Published: Lausanne Elsevier B.V 01.01.2021; Elsevier BV
• Subjects: Air flow; Algorithms; Boundary conditions; Computer simulation; Data preprocessing; Indoor airflow prediction; Indoor environments; Inverse design; Learning algorithms; Machine learning; neural network (NN); Neural networks; Non-isothermal; Predictions; Preprocessing; Submergence; Surface temperature; Temperature distribution; Training; Two dimensional models; Velocity; neural network (NN); Non-isothermal; Data preprocessing; Indoor airflow prediction
• DOI: 10.1016/j.enbuild.2020.110525
• ISSN: 0378-7788

•Fast and accurate prediction of non-isothermal airflow distribution are achieved.•Suitable data preprocessing methods are proposed.•Data preprocessing and training dataset have impact on neural network generality.•The error submergence occurs when neural network output has no preprocessing.•Separate prediction of multiple variables without data preprocessing is possible. The indoor environment is important to the daily lives of humans. Fast and accurate prediction of indoor environments is desirable with regard to practical applications, such as coupled simulation, inverse design, and system control. Neural network (NN) is a popular machine learning model used to build mappings between target variables with nonlinear relations. To confirm the feasibility of an NN for fast and accurate prediction of indoor environments (including both velocity and temperature distributions), two-dimensional non-isothermal cases are set and an NN model is constructed in this study, where the inputs are boundary conditions (i.e. inlet velocity, temperature and window surface temperature) and outputs are velocity and temperature distributions. Various data preprocessing methods are utilized, and their results are compared to reveal the impact of data preprocessing on NN performance. The results show that, for most cases, different preprocessing methods can lead to similar NN performances with a prediction time of approximately 350 μs for each case and a prediction error of less than 10% for the maximum value and 5% for the mean value. Without data preprocessing, error submergence is likely to occur, and the gradient descent algorithm may fail to reduce errors of variables with smaller orders of magnitude during the training process. Separate prediction of multiple variables without data preprocessing can achieve accurate predictions as simultaneous prediction with data preprocessing; however, the computation cost for training multiple NNs for separate predictions should be considered.