Bayesian Reasoning for Tropical Cyclone Intensity Forecasting and Risk Analysis

Improved methods for tropical cyclone (TC) intensity forecasting and risk analysis are the end products of this project leading to the wider use in the mitigation of the devastating impacts of tropical cyclones. Singledisciplined approaches in TC intensity forecasting by meteorologists and in risk analysis by social scientists so far have not seen satisfactory results. This is because of the intrinsically high degree of complexity in both the modelling and the problem domain parts of the project. The current techniques of TC intensity forecasting are limited to linear regression methods, which may not provide the maximum accuracy for the available information. Bayesian inductive inference using the Minimum Message Length (MML) principle (Wallace Freeman 1987) is used to investigate the likelihood that various different models (linear and non-linear) can be applied to TC intensity data. The features of these models may well suggest new insights into the formation of TC, facilitating better explanation of the most probable underlying functions of TC intensity, with the aim of providing improved forecasting capacity. In the past, Rissanen’s Minimum Description Length (MDL) (Rissanen 1987), a related form of Bayesian inference technique, has been successfully applied to time-series analysis in a different problem domain. An MML regression algorithm has been implemented to build forecasting models using data for the Atlantic basin. The promising results have warranted further development and testing expansions to different TC basins with varying degrees of data quality. The next step will be to consider whether the forecasts can be improved by incorporating a wider range of meteorological data into the analysis. Tropical cyclone risk analysis encompasses variables ranging from the cyclone threat through to vulnerability in the form of social factors and building standards. Much of the interactions amongst these variables have either not been investigated or only been done in a highly subjective manner. A naive approach to setting up a Bayesian causal network for risk factors would be to try to capture all the interactions and consequences