Application of rule based methods to predicting storm surge

Abstract

The accurate forecast of storm surge, the long wavelength sea level response to meteorological forcing, is imperative for flood warning purposes. There remain regions of the world where operational forecast systems have not been developed and in these locations it is worthwhile considering numerically simpler, data-driven techniques to provide operational services. In this paper, we investigate the applicability of a class of data driven methods referred to as rule based models to the problem of forecasting storm surge. The accuracy of the rule based model is found to be comparable to several alternative data-driven techniques, all of which result in marginally worse but acceptable forecasts compared with the UK's operational hydrodynamic forecast model, given the reduction in computational effort. Promisingly, the rule based model is considered to be skillful in forecasting total water levels above a given flood warning threshold, with a Brier Skill Score of 0.58 against a climatological forecast (the operational storm surge system has a Brier Skill Score of up to 0.75 for the same data set).The structure of the model can be interrogated as IF–THEN rules and we find that the model structure in this case is consistent with our understanding of the physical system. Furthermore, the rule based approach provides probabilistic forecasts of storm surge, which is much more informative to flood warning managers than alternative approaches.Therefore, the rule based model provides reasonably skillful forecasts in comparison with the operational forecast model, for a significant reduction in development and run time, and is therefore considered to be an appropriate data driven approach that could be employed to forecast storm surge in regions of the world where a fully fledged hydrodynamic forecast system does not exist, provided a good observational and meteorological forecast can be made.