Application of Machine Learning to Predict the Timing and Duration of Flood Caused by Tropical Cyclones to Inform External Hazard Probabilistic Risk Assessment

Najarkolaie, Kaveh Faraji; Bensi, Michelle

Permalink

https://hdl.handle.net/2142/121848 Copy

Description

Title

Application of Machine Learning to Predict the Timing and Duration of Flood Caused by Tropical Cyclones to Inform External Hazard Probabilistic Risk Assessment

Author(s)

• Najarkolaie, Kaveh Faraji
• Bensi, Michelle

Issue Date

2023

Keyword(s)

• Machine learning
• Probabilistic risk assessment
• Flood duration
• Storm surge

Abstract

This study aims to predict the timing of storm surge-induced floods reaching a nuclear power plant (NPP) and the duration that the NPP will be under a certain level of flooding. Studies have shown that more intense hurricane events will occur due to climate change and higher ocean surface temperatures in the future. More frequent severe hurricanes lead to a higher risk of flooding due to storm surges in coastal areas, increasing the importance of assessing the risk of flooding due to the storm surge for NPPs. The extent of damage due to tropical cyclone (TC)-generated flooding on NPPs will depend on multiple factors such as TC features, NPP vulnerabilities and exposure to flood hazards, and protective actions that are taken to lessen the impact of flooding. Successful implementation of these protective actions partially depends on the time the NPP has before the TC impact and the adverse environmental effect that TC will cause. Here we are specifically interested in predicting the time that TC-generated flooding will start to affect the power plant, its variation by time, and the duration that the NPP will be flooded. For this purpose, we gathered data regarding the time series features and their modeled storm surge due to synthetic storms from the "Coastal Hazards System" database. The synthetic TC data includes information about TC spatiotemporal features like TC center location, TC heading direction, central pressure deficit, the radius of maximum wind, and forward speed. Furthermore, we collected time series storm surge elevation generated by the simulated TC. We use multiple machine learning methods that are suitable for time-series forecasts, including long short-term memory networks (LSTM), gated recurrent units (GRU), one-dimensional convolutional neural networks (1-D CNN), and transformer models. We train the models using temporal TC data to predict the storm surge timing reaching a certain elevation and the duration that storm surge height remains above that elevation for save points (SPs) along the U.S. coastal areas. We evaluate and compare the results of different models based on various metrics (e.g., RMSE). Additionally, we investigate the influence of different TC features, such as distance between TC and SPs, TC intensity, and TC forward speed, on the timing and duration of flooding. We provide the results of this study in the form of charts that can be used as input for external hazard probabilistic risk assessment.

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/121848

Owning Collections