Keywords: flood estimation, artificial neural networks, flood frequency, ungauged catchment, catchment attributes.

Flood risk assessment is required for the planning and design of water infrastructure. Design flood estimation can ideally be made by analysing recorded streamflow data. There are many rivers in Australia where recorded streamflow data is limited or completely unavailable (ungauged catchments). In the past, different regional flood frequency analysis (RFFA) methods have been proposed for Australia. Most of the traditional RFFA methods are based on linear models. Application of artificial neural networks (ANNs) in regional flood frequency analysis may be an alternate to traditional methods. The ANN do not impose a model structure on the data and can better deal with non-linearity of the input and output relationships. This paper focuses on the ANN-based RFFA methods in eastern Australia based on application of an artificial neural network (ANN). The study uses data from 452 gauging stations from eastern Australia.

In applying the ANN in RFFA, optimum regions need to be identified. Previous studies by Aziz et al. [1,2] examined formation of optimum regions based on the state and geographic boundaries in the application of ANN to Australia. The focus of this study is on the formation of regions in catchment characteristics data space. Different regions were formed on the basis of catchment attributes, e.g. catchment area, mean annual rainfall, main stream slope, mean annual evaporation and design rainfall intensity. To identify groups of catchments in catchment characteristics data space, two methods are adopted: cluster analysis and principal component analysis. Out of different cluster groupings, two best groupings (based on certain criterion) were selected for the ANN-based RFFA modeling. A split-sample validation technique is adopted and the relative accuracy of various ANN-based RFFA models is assessed using median relative error and median values of ratio between the predicted and observed flood quantiles.

It has been found that that K-means cluster analysis generates the best performing regions in the catchment characteristics data space. When compared with the geographic regions, some state-based groupings perform more poorly than the K-means cluster groupings. Overall, the best ANN-based RFFA model is achieved when all the data set of 452 catchments in eastern Australia are combined together, which gives a RFFA model with median relative error of 39 to 56%.

1

K. Aziz, A. Rahman, G. Fang, K. Haddad, S. Shrestha, "Design flood estimation for ungauged catchments: Application of Artificial Neural Networks for eastern Australia", World Environment and Water Resource Congress, ASCE, 16-20 May 2010, Providence, USA. doi:10.1061/41114(371)293

2

K. Aziz, A. Rahman, G. Fang, S. Shrestha, "Artificial Neural Networks Based Regional Flood Estimation Methods for Eastern Australia: Identification of Optimum Regions", 33rd Hydrology and Water Resources Symposium, 26 June-1 July 2011, Brisbane, Australia.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £70 +P&P)