Statistical forecasting of tropical cyclones for Bangladesh

Abstract – Bangladesh experiences devastating tropical cyclones (TCs) in pre-monsoon (May-June) and post-monsoon (October-November) periods. The devastation caused by the TCs can be appreciated from the fact that the country has on record 49 percent of the world’s total fatalities from this natural calamity. Recent scientific and technological advances have made it possible for the people to be forewarned about an impending TC, but the lead time is only a few days. Such a short lead time, though undoubtedly is saving many lives, is inadequate in view of the country’s infrastructure and socio-economic considerations. Short-term forecasts have been made reliable through numerical modelling, satellite imagery analysis, and synoptic methods but these forecasts cannot be extended to longer lead times because TCs grow from a state of non-existence to full strength within a few days. In such circumstances as in many branches of science statistical methods remain as the only viable option for long-term forecasts. This paper develops the framework upon which a statistical formulation of the prediction problem can be based.In statistical modelling, it is necessary to identify the dependent and independent variables which can come from physical considerations. A literature survey is made to identify the hydro-meteorological variables which can be possible candidates for independent variables. The cyclones that hit Bangladesh are almost all formed in the Bay of Bengal in the main development region (MDR) of 850E-950E and 50N-150N. After a cyclone is formed, two aspects that are important in Bangladesh context are its intensity and the track. These are the dependent variables. The intensity is captured through the energy of the cyclone and the prediction of tracks is limited to the prediction of possible recurvature. Bangladesh coast would directly be in the path of a TC formed in the Bay of Bengal only if recurvature occurs to its track. The forecasting model is first formulated as a linear regression model gradually build by stepwise regression up to the number of independent variables equal to the significant number of principal components. The usefulness of the series of models in prediction is then tested by the techniques of artificial intelligence, specifically by artificial neural networks and support vector machines. In model testing, data are split into 70% for training and 30% for testing which is the standard industrial practice. Finally, the model variables which have most predictive capability in the context of Bay of Bengal cyclones are identified. All the statistical analyses are done using the software package WEKA which is an open source free software product.