Two classes of methods are widely used to describe natural phenomena, knowledge driven and data driven. Knowledge driven methods are developed and applied based on our understanding of the physical world, commonly expressed in physical laws or approximations to these laws. Data driven methods extract patterns from the observed data without direct reference to the underlying physics. Both approaches are widely applied in time series analyses. The prediction of major flood events like those experienced recently in China (1996,1998), Bangladesh (1998) and Poland and the Czech Republic (1997) presents an extremely challenging prediction problem. The accompanying tragic loss of life only serves to highlight the need to develop and improve flood forecasting techniques. It must be recognised that no prediction model can perform flawlessly, yet the consequences of poor predictions may be catastrophic. Differences between the modelled and the real worlds will always exist. In order to obtain accurate predictions of the future behaviour of modelled systems, a so-called data assimilation technique may be employed so as to minimise the errors between the observed and the simulated data. This study explores the application of different data driven approaches to the general problem of flood forecasting. The purpose of the study is to examine and evaluate the performance of these different approaches to specific flood forecasting problems. The study focuses firstly on the treatment of phase errors. An evaluation of the different methods is carried out for the Bird Creek catchment that has been used in a previous WMO intercomparison study. Several sources of phase error can be identified and the ability of several data assimilation methods to treat significant phase errors is examined. These methods are: 1) MIKE l l FF updating procedure, which works as an advanced error prediction model that includes the identification and correction of phase errors; 2) Ensemble Kalman Filter, which updates the model by representing the statistical properties of the state vector with an ensemble of domain state vectors. The vectors are propagated according to the dynamic system subjected to internal models, whereby the resulting ensemble provides estimates of the forecast state vector and the associated uncertainty; 3) Artificial Neural Networks, in particular Time Lag Recurrent Networks (TLRN), have been used here because past behaviour is accommodated inherently; 4) Chaotic theory, which has been developed for dealing with the unpredictability of dynamic systems using an embedding theorem. The identification and correction of significant phase errors has proven to be an extremely difficult task. Data assimilation using ANN proved the most effective, while several limitations were found for the other methods that were investigated. A second goal of this study was to examine the extrapolation properties of several of the data driven methods. There is always a risk of extreme flood events larger than those previously recorded. The extreme flooding that occurred on the Yangtze River during 1998 has been used to examine the extrapolation properties of the data driven techniques. An important conclusion of this study is that forecast accuracy and extrapolation behaviour is improved when data driven methods are combined with knowledge. driven methods via data assimilation.
Back to the list of MSc abstracts