Stochastic event set generation for tropical cyclone using machine‐learning approach guided by environmental data

Abstract

AbstractTropical cyclone stochastic eventset is a critical component in any cyclone risk assessment model. In this study, a novel methodology is proposed for the development of basin wide tropical cyclone stochastic eventset. The proposed methodology utilizes the reanalysis data to represent the environmental conditions of cyclones, which in turn governs the cyclone behaviour and variability in space and time. The basic assumption in the proposed methodology is that, given the high correlation between cyclone behaviour and environmental conditions, the cyclone may respond similarly if the environmental conditions are identical. Thus, it is possible to model the response of a cyclone based on identifying a group of events with similar state in history. The state of cyclone at any given instance is defined using cyclone track characteristics and environmental parameters from ocean and atmospheric reanalysis data. The methodology presented herein addresses some key open questions in the stochastic eventset development about capturing the seasonal and spatial variability of cyclone, and modelling of events in historically data sparse region. The k nearest neighbour machine learning algorithm was used to train the track movement, intensification, and inland decay model. The proposed methodology is illustrated through a case study consisting of the development of stochastic event set for South West Pacific basin and compared with historical observations. The methodology introduced in this article represents a step toward a rational approach to capture the seasonality of cyclone; and effective use of environmental information together with machine learning to generate realistic stochastic events.