A Learning System for Identification and Ranking of Severe Storms

Abstract

Design of offshore structures is more and more often based upon design data computed by numerical hindcast models.

An important step in any hindcast study is the selection of storms to be hindcast. The subsequent statistical interpretation relies heavily on the correct identification of the most severe storms within the period covered by the study.

The present paper describes a program which is able to identify severe storms and to provide good first estimates of maximum sea-states during each storm. The program is based upon pattern recognition techniques. In the learning mode, the program is given the values of selected meteorological parameters and simultaneous measured wave heights. The meteorological parameters are read from a data base containing key parameters describing the air pressure patterns at three different length scales during a total of 17 years. The program is then able to formulate and test hypotheses regarding the relations between these meteorological patterns and the simultaneous sea-state at the location of the measured wave heights.

In the prediction mode, the acquired knowledge is used for predicting the sea-states during the complete 17 years period. These predictions form the basis for the storm selection for the subsequent hindcast study.

The program has been calibrated and validated using data from a number of locations in the North Sea. It has so far proven itself a useful tool in two major hindcast studies covering areas in the Norwegian, Danish, and UK North Sea sectors.