Flood Forecasting for Klang River at Kuala Lumpur using Artificial Neural Networks

Abstract

This study evaluates the use of Multi-Layer Perceptron (MLP) neural network models to forecast water levels of a gauging station located at the Kuala Lumpur city centre in Malaysia using records of upstream multiple stations. Cross correlation analysis of water level data was performed to determine the input vectors which include the current and antecedent water level data of the upstream stations to ensure that of the data available, the most influential values corresponding to different lags are selected for analysis. Twelve well recorded storm events were used to train, test and validate the MLP models. The best performance based on MSE, MAE and R² was achieved with a model of 15 input vectors of upstream current and antecedent water levels, 7 hidden nodes and an output vector for the station at Kuala Lumpur centre. The R² values for training, testing and validation datasets are 0.81,0.85 and 0.85 respectively.