Regional Flood Frequency Analysis for East Malaysian Catchments Using Artificial Neural Networks
Aunthors: Kee An Hong¹٭, Jer Lang Hong²
1Hong & Associates, Kuala Lumpur, Malaysia
2SNS Network, Kuala Lumpur, Malaysia
Corresponding author: Kee An Hong
Design flood peak estimation is needed at ungaged sites where there are no hydrological records available, especially for the design of hydraulic structures. In the absence of peak discharge data, flood estimation at ungaged sites is normally carried out using the method of regional flood frequency analysis (RFFA). The RFFA models in Malaysia are mainly based on linear models, for example, the index flood method, quantile regression method, parameter regression technique, and probabilistic rational method. The application of non-linear RFFA methods such as artificial neural networks (ANN) in this country is rare. In this study, artificial neural network models were used to derive the flood quantiles with return periods of 2, 20,50, and 100 years for East Malaysian (Sabah and Sarawak) catchments. Data from 40 gaged catchments from Sabah and Sarawak were used to derive the relationships between flood quantiles and catchment characteristics and climatic parameters. These flood statistics established can then be transformed to estimate the design peaks for ungaged catchments in the same hydrological region. A total of six predictor variables are used and 20 different ANN models are tested. The at-site flood quantiles, used as target outputs for ANN models were estimated using L-moments. The performances of the ANN RFFA models vary across the four design return periods and there is no model that is the best for all the quantiles considered based on the evaluation statistics. From the results obtained, it is found that increasing the number of predictor variables does not necessarily enhance the performance of the ANN model for certain quantiles. However, the ANN-based RFFA models have the potentials for regional flood estimation, and essentially the use of the method in practice still needs further testing for larger data set obtained.
Keywords: AN, RME, R, L-moment, RFFA, homogeneity, discordancy