Predictions of flow and temperature fields in a T-junction based on dynamic mode decomposition and deep learning

Abstract

Accurate flow field prediction methods are needed for the analysis of complex flows in energy and power field. Flow field and temperature field prediction methods combining Dynamic Mode Decomposition (DMD) and deep learning are proposed. A Convolutional Long Short-Term Memory (ConvLSTM) neural network model is built by adjusting the network structure reasonably. The DMD method, the ConvLSTM method and the method combining DMD and ConvLSTM are compared by the flow field and temperature field prediction results in a T-junction, which is widely used in energy industry. The time series dataset of the velocity, pressure and temperature field of a wall jet in a T-junction are obtained through large eddy simulation (LES). The overall relative errors in the predictions of velocity, pressure and temperature fields remained about 4%, 60% and 0.13% for the DMD method, 3%, 10% and 0.08% for the ConvLSTM method, and 2%, 10% and 0.06% for the method combining DMD and ConvLSTM, respectively. The combining method is the most accurate and stable prediction method. Its information loss rates of the velocity, pressure and temperature fields are the smallest and 2.21%, 13.38% and 0.11%, respectively, and will not increase significantly with the increase of the prediction duration.