New hurricane impact level ranking system using artificial neural networks, A

Abstract

Tropical cyclones are intense storm systems that form over warm water but have the potential to bring multiple related hazards ashore. While significant advancements have been made for forecasting of such extreme weather, the estimation for the resulting damage and impact to society is significantly complex and requires substantial improvements. This is primarily due to the intricate interaction of multiple variables contributing to the socio-economic damage on multiple scales. Subsequently, this makes communicating the risk, location vulnerability, and the resulting impact of such an event inherently difficult. To date, the Saffir-Simpson Scale, based off of wind speed, is the main ranking system used in the United States to describe an oncoming tropical cyclone event. There are models currently in use to predict loss by using more parameters than just wind speed. However, they are not actively used as a means to concisely categorize these events. This is likely due to the scrutiny the model would be placed under for possibly outputting an incorrect damage total. These models use parameters such as; wind speed, wind driven rain, and building stock to determine losses. The relationships between meteorological and locational parameters (population, infrastructure, and geography) are well recognized, which is why many models attempt to account for so many variables. With the help of machine learning, in the form of artificial neural networks, these intuitive connections could be recreated. Neural networks form patterns for nonlinear problems much as the human brain would, based off of historical data. By using 66 historical hurricane events, this research will attempt to establish these connections through machine learning. In order to link these variables to a concise output, the proposed Impact Level Ranking System will be introduced. This categorization system will use levels, or thresholds, of economic damage to group historical events in order to provide a comparative level for a new tropical cyclone event within the United States. Discussed herein, are the effects of multiple parameters contributing to the impact of hurricane events, the use and application of artificial neural networks, the development of six possible neural network models for hurricane impact prediction, the importance of each parameter to the neural network process, the determination of the type of neural network problem, and finally the proposed Impact Level Ranking System Model and its potential applications.