ESTIMATION OF EXTREME FLOWS FOR UNGAUGED CATCHMENTS

Abstract

The estimation of flood magnitude corresponding to a selected return period is one of the most common problems being faced by the hydrologists. The problem becomes more serious, when there are no flood records at the site of interest. In India, like most of the developing countries in the world, the majority of the basins are either sparsely gauged or not gauged at all. The gauged records are also of short lengths (generally 15 to 30 years). This necessitates the development of robust flood estimation models for such conditions. The research on flood frequency analysis has taken place with varying intensity over couple of decades. During seventies and eighties much effort was directed on developing efficient at-site flood frequency procedures. Many new distributions and estimation methods were developed and reported in literature. Research in nineties was mainly dominated by Lmoments. During nineties, number of technical papers based on application of artificial neural network (ANN) also appeared in hydrologic literature. These applications mainly included rainfall-runoff modeling, precipitation estimation, modeling of stream flows, ground water and water quality modeling, stage-discharge relationships, rainfall dis-aggregation, evapo-transpiration estimation, and classification of river basins. Most of these applications have been documented in the report of ASCE task committee on ANN and volume 5 (2), 5 (3) and 5 (4) of Journal of Hydrologic Engineering, ASCE (2000). However, in flood frequency analysis only limited efforts have been reported in literature. The present study tries to (i) explore further the application of artificial neural network in flood frequency analysis and (ii) address various issues related to regional homogeneity and flood quantile estimation. Annual flood data and physiographic ii characteristics of 116 sites of Zone 3 of Central India as well as synthetically generated flood like data have been used for performance evaluation of various methods. Finally, the flood frequency formulae for the use of field engineers have been recommended. For regional homogeneity, four clustering techniques namely Ward's method, Kmeans method, Fuzzy K-means and Kohonen's network have been used. The homogeneity of various clusters formed by these techniques has been evaluated on the basis of L-moments based homogeneity criteria. Various parametric and artificial neural network based techniques have been used for at-site flood quantile estimation. To compare the performance of these techniques, longterm synthetic data from 8 Wakeby parents having different parameters were generated. These data were divided in 200 samples of varied length ranging from 30 to 75. The flood quantiles of different return periods, ranging from 10 to 200 years were estimated from these samples using various methods. These flood quantiles given by various techniques were compared on the basis of various loss functions such as bias, variance, mean square error, underdesign loss and overdesign loss. To evaluate the performance of ANN in regional flood frequency analysis (RFFA), the analysis has been carried out for different clusters. For these clusters, relationships have been developed between (i) mean annual flood and physiographic and climatic characteristics ofcatchments (ii) QT IQ and Ti.e., ratio ofTyears return period flood (QT) and mean annual flood (Q)and return period (T). For the development of first relationship, multiple regression and ANN based approaches have been used. For the development of second relationship, methods based on standardized L-moments and ANN have been used. For developing the ANN models several trials were carried out in order to determine the best configuration of hidden layers and nodes in hidden layer. The study finally develops six in clusters of catchments and suggests a methodology for flood quantile estimation and allocation of gauged / ungauged catchment to one of the clusters. The major findings of the study are summarized below: 1. Pooling of catchments on the basis of catchment characteristics is more logical as compared to geographic location. On evaluating the performance of different clustering techniques, it is seen that almost 50 to 100 percent catchments are common in the clusters formed by different techniques. Out of four techniques studied, none of the techniques is perfect in ensuring the regional homogeneity of clusters. However, the regional homogeneity of different clusters can be achieved by heuristic rearrangement of catchments. The heuristic rearrangement gives additional advantage of more data length as some of the catchments may be overlapped in different clusters. The clusters so formed are recommended. 2. The artificial neural network (ANN) can be used as a satisfactory alternative tool for quantile estimation up to the return period equal to the length of data. The extrapolation beyond this period is not satisfactory. However, performance of ANN improves quite substantially if historical information is incorporated in the sample. The use of ANN for flood quantile estimation is recommended for at-site flood quantile estimation if historical information at the site of interest is available. 3. For different clusters, regional distributions have been established. The regional equations developed for these distributions are recommended for use by field engineers, for flood quantile estimation. However, if the facilities of iv detailed ANN analysis are available, ANN based methods should also be tried as alternate methods in regional flood frequency analysis for estimation of flood quantiles of small return periods.