This research evaluated the application of artificial neural networks and hydrodynamic models for river flood modelling and prediction. The study has been completed in three main parts: First part focused on the application of artificial neural networks for flood prediction in ungauged catchments. Catchment descriptors were used as input data and the index flood was the output of the model. Different types and numbers of catchment descriptors (17 descriptors and more than 1000 catchments) were used to choose those that gave the best relationship with the hydrological behaviour and flood magnitude. ANN models with different architectures were developed and applied to training and validation sets of data to find the best type of ANN for this application. Selection of pooling groups of catchments either randomly or according to geographical proximity did not produce desirable results. Therefore hydrologically similar catchments were clustered using the FEH-Software before entering descriptors into the ANN model. This improved the accuracy of predicted floods. The second part of the research aimed to model river flow in a multi-gauging station catchment and provide real-time prediction of peak flow downstream. Three types of ANN (Multi-Layer Perception (MLP), Recurrent and Time Lag Recurrent) were adapted to evaluate the applicability of this technique. The study area covers the Upper Derwent River, a tributary of the River Trent in the UK. River flow was predicted at the subject site with lead times of 3, 6, 9 and 12 hours. Tests were completed using different lengths of input data to evaluate the effect of input data size in model outputs. The number of gauging sites to be used as data sources in the model as also evaluated. In the final part, the application of artificial neural networks (ANN) to optimise the results obtained from a hydrodynamic model of river flow as evaluated. The study area is Reynolds Creek Experimental Watershed in southwest Idaho, USA.