Physics informed neural network modelling for storm surge forecasting — A case study in the Bohai Sea, China

Abstract

Storm surges have a great impact on ocean engineering and design complex physical changes. Numerical simulation methods are often used for prediction, but they face problems such as long calculation time. Machine learning avoids these, but it also faces some problems, such as delays in predicting results, short prediction durations, and large data demands. Therefore, we built a PINN model to integrate storm surge physics with neural networks to reduce the need for data and improve the accuracy of storm surge forecasting. Using ADCIRC as a smaller dataset, the cold wave storm surge in Bohai Bay during 2018–2022 was simulated. In the storm surge process prediction experiment, the overall error of PINN is small, RMSE is 0.163. In a 48-h prediction experiments, RMSE of PINN's result is 0.241, which is more accurate than DNN. It is revealed that PINN has a strong physical mechanism learning ability. PINN can predict the storm surge of strong cold wave more accurately, the calculation speed is nearly one thousand times faster than ADCIRC, and it has broad application prospect in disaster prevention and reduction.