Author(s): Zhengbang Zhou; Saiyu Yuan; Hongwu Tang

Linked Author(s):

Keywords: Deep learning; Physics-informed neural network; ConvLSTM; Shallow water equations

Abstract: Physics-Informed Neural Network (PINN) have emerged as a promising approach in hydrodynamic simulation recently. However, traditional PINN models based on Deep Neural Network (DNN) exhibit limitations in handling high-dimensional spatiotemporal nonlinear data and predictions. This study integrates the PINN strategy with Convolutional Long Short-Term Memory networks (ConvLSTM), addressing the challenges posed by the scarcity of flow velocity data and limited observational points in lake environments. This framework facilitates both the solution of forward problems and the resolution of flow velocity fields driven by water level fields, even the predictions for future time steps. By leveraging convolutional layers for the efficient extraction of complex spatial features from two-dimensional lake data and LSTM layers for capturing temporal continuity, this approach incorporates Partial Differential Equation (PDE) constraints to enhance the accuracy, physical interpretability, and extrapolation capabilities of predictions for both flow velocity and water level fields. This study demonstrates improved performance in current flow field solutions and future step predictions.

DOI: https://doi.org/10.64697/978-90-835589-7-4_41WC-P1688-cd

Year: 2025