Hidden neuron variation in multi-layer perceptron for flood water level prediction at Kusial station

Abstract

In Malaysia, east coast of peninsular is experiencing the rainy season between mid - October until March every year. Heavy seasonal rains cause the Kelantan River to overflow and flood the surroundings area. In this paper, the application of feed forward multi-layer perceptron (FFMLP) in neural networks for flood water level prediction is presented. The method focused on the neuron variation in hidden layer. By using measured data of three stations; Tualang, Kuala Krai and Kusial, FFMLP neural networks was developed. The inputs are water level river at three stations and output is water level river at Kusial station. The numbers of neuron in hidden layer were varied from one to ten and Levenberg Marquadt algorithm is used to train the network. The performance of network was evaluated using Mean Square Error (MSE). It is shown that three neurons in hidden layer afforded the lowest MSE, 0.043. The Regression, R for training network is closed to 1 (0.991), supports that the model is acceptable and able in predicting water level at Kusial station.