Flood forecasting using artificial neural networks; Proefschrift Vrije Universiteit Amsterdam en het UNESCO-IHE Institute for Water Education, Delft.

Flood disasters continue to occur In many countries around the world and cause tremendous casualtles and properties damage. To mitigate the effects of floods, both structural and non-structural measures can be employed, such as dykes, channelisatlon, flood proofing of properties, land-use regulation and flood warning schemes. Sufficie͏̈nt advance warning time may save lives and property by allowing time to effect various structural and other adjustment. Earlier advanced warning can be achieved through advances in mathematical modelling. There are many rainfall-runoff models being developed and employed for flood forecasting which lead to the issue of flood warnings. Among these models, the Artificial Neural Network (ANN) has been found suitable for modelling the rainfall-runoff process in a wide variety of catchments under specific circumstances. An ANN is a quick and flexible approach which gives very promising results. Unfortunately, the inability to predict beyond the limits of the training range (extrapolation problem) was found to be a serious limitation of this approach. It is the solution to this problem that forms the one of the principal objectives of the present work. A method of creating room for an ANN to extrapolate has been proposed, which focuses on the setting of the range of the activation functions applied to each node in the network. This approach has been successfully applied to several case studies, i.e. The Silk Stream and Rivers Thrushel and Brue. Different scale factors are also investigated and the results are summarised. Recurrent Neural Networks, of both the Elman and the Jordan forms, are studied and compared to the basic Feed Forward Network. The results reveal that the Jordan recurrent network offers a possibility to recur the errors back to the context unit, which helps to improve the results. It is these recurred errors which leads to the further investigation of the forward shift operator. In an altemative approach, the First Derivative Transfer Function (FDTF) was employed to derive the forward shift operator that is input to the network to produce the forecast results. The results reveal that the forward shift operator provides useful information which inhibit the phase errors as forecasting proceeds in time