Flood forecasting by committees of models : case of Dapoling-Wangjiaba catchment in Huai river basin

The Huai River is very vulnerable to floods causing serious damages, so it is very importantfor authorities and public to be informed of floods in advance. Hydrological models are widely used in simulating river discharge for flood forecasting. Such a model may be accurate on average, but there is big chance that it will miss the extreme values, such as peaks, which is very essential in flood forecasting. Compared with single model, a multi-model(committees of models) can produce more accurate predictions that are better than a single model. The objective of this thesis is to explore the use of such models on several case studies.There are a number of traditional methods to combine local models. In this thesis, fuzzy committee method, Bayesian Model Averaging method, and a neural network method are researched. They are applied on six catchments and the performance of each method is evaluated and compared. The results show that all of the three mult-model combination methods can improve the model performance (however to different degrees) compared with single hydrological model. Additonally, the results from the six catchments also indicate that the fuzzy committee method is better than Bayesian Model Averaging method, and neural network method also performs better than Bayesian Model Averaging method.The multi-models respectively combined with the three methods which are fuzzy committee method, BMA method and NN method are applied in the Huai River to simulate the discharge at Wangjiaba Hydrological Station. The rainfall-runoff model used is HBV model based on the precipitation of the upper catchment which is Dapoling-Wangjiaba catchment. Generally speaking, the discharges simulated from the three models are quite close to the measured ones. All of them can forecasts the correct day of peak discharge in most ccases, and the forecasted values of peak discharge are similar to the measured one. Among the combination methods, the neural network-rainfall runoff model has the best performance, the specilised model combined by fuzzy committee method comes second, and the specilised modelscombined by BMA method is the least one