10.3 Tropical Cyclone Storm Surge Prediction Using Artificial Neural Network

Wednesday, 15 January 2020: 2:00 PM
158 (Boston Convention and Exhibition Center)
Mahmoud Ayyad, Stevens Institute of Technology, Hoboken, NJ; and R. Marsooli and M. Hajj

Tropical cyclones, such as hurricanes and tropical storms, can cause significant storm surge with devastating impacts in terms of loss of life, property destruction, damage to infrastructure, water, energy and transportation systems, spread of contaminants and long-term economic disruption. Today’s computational tools and power have significantly improved the capability to predict, with some uncertainties, the path and characteristics of storm events, which mitigated their impacts and resulted in saving human lives and significant reductions in economic losses. Still, there remains many challenges including the need for reducing the uncertainties of the predictions, evaluating more scenarios for specific storms, accounting for increased human development in coastal areas, and assessing impact of global climate change and rise in sea level. Addressing these needs holistically would require more computational power and capabilities, especially that nearshore zones are dynamic in nature and the human impact over time adds to these variations. On the other hand, deep learning opens the door for enhanced prediction capabilities of storm surge levels that can meet some of these needs. Particularly, artificial neural network can reduce the computational time required to account for a broad range of storm scenarios including impact of climate change and rise in sea level.

Towards this objective, we implement neural network model to predict storm surge on shoreline by considering 36909 storm scenarios obtained from synthetic storm simulations. Eighty percent of these storms were used for training to determine the best set of weights to minimize the loss function. Ten percent of the data set were used to validate the neural network model to tune its hyperparameters through minimizing the mean square error and maximizing the correlation coefficient. The last ten percent were used to provide an unbiased evaluation of the final model. Predictions based on six, twelve and eighteen hours prior to the landfall were performed. The results show correlation coefficients of near 99% for all storm surge levels and about 98% for levels above 1 m between predicted values using neural networks and those obtained from numerical simulations. The difference is that the time required for the neural network predictions was of the order of few seconds in comparison to the numerical simulations, which required few hours; pointing to the advantage of combining artificial neural network with numerical simulations for predicting storm surge levels.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner