Application of artificial neural networks and genetic algorithm for regional flood estimation in Eastern Australia

In Australia, design flood estimation in small ungauged catchments is often carried out using the Probabilistic Rational Method (PRM). However, in regional flood frequency analysis (RFFA), the application of regression-based techniques is currently being investigated as a part of the ongoing revision of Australian Rainfall and Runoff (ARR). It is likely that the PRM will be replaced by the regression based methods in the new ARR. In contrast to these traditional linear-model based RFFA methods, Artificial Neural Networks (ANN) and Genetic Algorithms (GA) can also be applied. These techniques do not impose a fixed model structure on the data and can better deal with non-linearity of the input and output relationship. These non-linear techniques have been applied successfully in many hydrological problems; however, there have been only limited applications of these in RFFA, particularly in Australia. This paper focuses on the development and testing of the ANN and GA based RFFA models for eastern parts of Australia. A data set comprising of 452 stations from New South Wales, Victoria, Queensland and Tasmania are used to develop the new RFFA models. Two different types of ANN paradigms are tested and compared. One using the traditional feed forward back propagation technique called BPANN, and the other using the Genetic Algorithm for ANN optimisation and is called GAANN. An independent testing is undertaken to assess the performance of both RFFA techniques. The results show that the ANN based RFFA outperforms the GAANN and QRT. The ANN can be used as an alternative method to check the validity of the ARR soon-to-be-recommended regression-based methods