Comparative studies of predictive models for unsteady flow fields based on deep learning and proper orthogonal decomposition

Abstract

With the application of data-driven technologies in fluid mechanics, models based on Proper Orthogonal Decomposition (POD) and Deep Learning (DL) for predicting flow fields have gradually been used in the rapid analysis of flow fields. However, differences among predictive models have not been widely concerned. In this work, based on the dataset of the unsteady flow around a cylinder, POD-LSTM (Proper Orthogonal Decomposition - Long short-term memory) and CAE-LSTM (Convolutional Autoencoder - Long short-term memory) were designed to compare their prediction performances, which could predict flow fields of the following eight time steps from that of eight time steps before. In a prediction period, the predictive accuracy of models was compared to analyze sources of error. In the long-term, it is noticed that the predictive accuracy of POD-LSTM decreased significantly with the increase of time steps, while CAE-LSTM maintained good robustness until the last time step. In addition, the phenomena of random predictions are presented in high-order temporal coefficients, which led to POD-LSTM's worse performance. Due to the capability of fusing spatiotemporal features, CAE-LSTM shows better accuracy and robustness in the long-term prediction period.