Storm surge induced by tropical cyclones has consistently been the most devastating among all marine hazards in casualties and economic repercussions. Its accurate forecasting is of great significance for coastal economic and social development. In this study, a physics-informed neural network (PINN) which introduces the two-dimensional storm surge governing equation into a deep neural network (DNN), is proposed to forecast the storm surge events in the coastal areas of Zhejiang and Fujian. The storm surge levels from an advanced circulation (ADCIRC) model are used for training and evaluating the network. A storm surge event during Typhoon ‘Utor’ in 2001 is selected as the forecast case. Results indicate that the PINN-based model is able to accurately forecast storm surge levels with root mean square error (RMSE) below 0.2 m. The RMSE is smaller than that estimated by DNN, with the maximum reduction of 20%, indicating that the physics constraints can effectively improve the model performance in forecasting the storm surge.

storm surge, physics-constrained, deep learning, coastal areas of China