Prediction Method for Flow over Submarine Based on Multi-scale Deep Neural Network

In order to accurately predict the dramatic changes of pressure around submarine fore region, a method of prediction on continuous time series variables based a multi-scale network is proposed. The multi-scale network is constituted by long short-term memory network (LSTM). The pressure around submarine fore region under different time period are obtained through numerical simulation to establish the datasets as training samples and testing samples. Firstly, the dataset is decomposed into high frequency and low frequency by lowpass filter to train deep neural networks with two different scales respectively. Finally, A large-scale and a small-scale network are trained separately to achieve the response and capture of different process, the trained networks were then tested by predicting the flow fields in future time steps. This work has analyzed the influence of multi-scale neural network on the prediction accuracy. Meanwhile we verified the reliability of the network in transition zone. The results show that predicted flow fields using the multi-scale deep neural network are in good agreement with those calculated directly a computational fluid dynamic solver.