Real-time flood forecasting model intercomparison and parameter updating rain gauge and weather radar data

This thesis describes the development of real-time flood forecasting models at selected catchments in the three countries, using rain gauge and radar derived rainfall estimates and time-series analysis. An extended inter-comparison of real-time flood forecasting models has been carried out and an attempt has been made to rank the flood forecasting models. It was found that an increase in model complexity does not necessarily lead to an increase in forecast accuracy. An extensive analysis of group calibrated transfer function (TF) models on the basis of antecedent conditions of the catchment and storm characteristics has revealed that the use of group model resulted in a significant improvement in the quality of the forecast. A simple model to calculate the average pulse response has also been developed. The development of a hybrid genetic algorithm (HGA), applied to a physically realisable transfer function model is described. The techniques of interview selection and fitness scaling as well as random bit mutation and multiple crossover have been included, and both binary and real number encoding technique have been assessed. The HGA has been successfully applied for the identification and simulation of the dynamic TF model. Four software packages have been developed and extensive development and testing has proved the viability of the approach. Extensive research has been conducted to find the most important adjustment factor of the dynamic TF model. The impact of volume, shape and time adjustment factors on forecast quality has been evaluated. It has been concluded that the volume adjustment factor is the most important factor of the three. Furthermore, several attempts have been made to relate the adjustment factors to different elements. The interaction of adjustment factors has also been investigated. An autoregressive model has been used to develop a new updating technique for the dynamic TF model by the updating of the B parameters through the prediction of future volume adjustment factors over the forecast lead-time. An autoregressive error prediction model has also been combined with a static TF model. Testing has shown that the performance of both new TF models is superior to conventional procedures.