PREDICTION OF WIND PRESSURE TIME SERIES ON SQUARE CYLINDER UPON LONG SHORT-TERM MEMORY NEURAL NETWORK

Based on Long Short-Term Memory (LSTM) neural network, this paper predicts the wind pressure time series on unknown circumferential locations of the structure using measured data from only a few pressure taps. Using the wind pressure data measured from the wind tunnel test of uniform flow past a square cylinder at various incidences, the LSTM model has effectively predicted the time series of pressure on the cylinder. When choosing the correlation length of model data, the temporal correlation of wind pressure should be considered to achieve better prediction results. The multiple layers of neural network can improve the precision. More training taps produce better results, but a balance between precision and efficiency should be considered. The predicted results of mean and fluctuating pressure coefficients, as well as the pressure time series of typical taps, meet well with those of experiments. However, the prediction error near cylinder’s corners is relatively large, which might be related to the strong non-Gaussian characteristic of pressure.