POD-LSTM model for predicting pressure time series on structures

Abstract

A novel method for predicting wind pressure time series on structures is proposed by combining long short-term memory network (LSTM) with proper orthogonal decomposition (POD). POD-LSTM predicts the pressure time series on any circumferential locations of a structure using data from limited pressure taps. The wind tunnel test data of pressure on a single square cylinder and the downstream cylinder of two tandem square cylinders is utilized to evaluate the performances of POD-LSTM, LSTM, and POD-BPNN (back-propagation neural network). Results of pressure time series, aerodynamic parameters, and pressure moments are presented. POD-LSTM takes advantage of LSTM in time series prediction and POD in extracting essential features, resulting in a better performance than POD-BPNN and LSTM. The similarity of pressure on nearby taps affects the accuracy of POD-LSTM. A larger error is observed at the rear corners of a single cylinder where intermittent flow reattachment occurs. The predicted results for mean and fluctuating pressure coefficients are satisfactory, but POD-LSTM underestimates the absolute value of skewness and kurtosis at structural surfaces where the non-Gaussian property of pressure is significant. Compared with the uniform flow condition, the prediction accuracy decreases when the structure is in the wake of upstream structures.