Application of ANN in regional flood estimation : a case study for New South Wales, Australia

Flood estimation in ungauged catchments is often needed in hydrology. Regional flood frequency estimation (RFFE) methods can be used for this purpose. The RFFE models in Australia are mainly based on linear models, such as Index Flood Method, Quantile Regression Technique, Parameter Regression Technique and Probabilistic Rational Method. The application of non-linear RFFE techniques such as Artificial Neural Network (ANN) is quite limited in Australia. In this paper, an ANN based RFFE model is presented for New South Wales (NSW) State in Australia. It uses data from 88 gauged catchments in NSW. A total of eight predictor variables are considered and five different model forms are tested. It has been found that when all the eight predictor variables are used, the ANN based RFFE model performs the best generally; however, the gain in model accuracy from using only three predictor variables is marginal. Furthermore, the performances of ANN based RFFE models vary across the six AEPs, and there no model that is the best with respect to all the evaluation statistics adopted here. The result shows that increasing the number of predictor variables does not necessarily enhance the performance of the ANN based RFFE models. The results demonstrate the potentials of ANN based RFFE models; however, further testing is needed using a larger data set before ANN based RFFE model can be recommended for practice in NSW.