For ungauged catchments where there is no flow data it is always more difficult to find a method to obtain acceptable results of future flow prediction. In recent years the technique of Artificial Neural Networks (ANN) has been found as an efficient tool to solve variety of problems concern with water resources, hydrology and hydraulic. In this research, the applicability of the Artificial Neural Network (ANN) technique has been investigated for predicting the index flood for several ungauged catchments across the UK. Catchment descriptors have been used as input data and the index flood of the catchments as output. Different types and number of catchment descriptors were used as inputs to choose those that show the best relationship with the catchment hydrological behaviour and flood magnitude. A network with 7 inputs represented the best result although a network with only three inputs showed accuracy slightly less than the network with 7 inputs. ANN models with different architectures have been used with training and validation sets of data to find the best ANN for this application. It has been found that the Multi-Layer Perceptron network with three layers, and Tanh function for hidden layer and Sigmoid function for output layer is the most accurate network for this purpose. The best result has been obtained when the ANN was trained for a group of catchments, which are hydrologically similar. The classification of the catchments according to their similarity in hydrological behaviour was carried out using WINFAP-FEH software (developed by the UK Institute of Hydrology). It identifies the catchments, which are homogenous to the subject site using some parameters including drainage area, annual rainfall and base flow index. The outputs of the model where close enough to the measured values both in training and testing phases. Evaluation of pooling-group heterogeneity factor (H2) and closeness of predicted peak flow to measured values (R2) for several different pooling groups as well as plotting these two factors against each other shows that where heterogeneity factor raises accuracy of the results of ANN based predicted flows are decreased.

Keywords