Neural Networks for Real Time Catchment Flow Modeling and Prediction

Abstract

Accurate prediction of catchment flow has been recognized as an important measures for effective flood-risk management strategy. A neural network modeling approach was used to construct a real time catchment flow prediction model for a river basin. Two types of neural network architectures i.e. feed forward and recurrent neural networks, and three types of training algorithm i.e. Levenberg–Marquardt, Bayesian regularization, and Gradient descent with momentum and adaptive learning rate backpropagation algorithms were examined in this study. A total of six different neural network configurations were developed and examined in terms of optimum results for 1 to 5-h ahead prediction. The methods were used to predict flow in the Cilalawi River in Indonesia, and their performances were evaluated using various statistical indices. The modeling results indicate that reasonable prediction accuracy was achieved for most of models for 1-h ahead forecast with correlation >0.91. However, the model accuracy deteriorates as the lead-time increases. When compared, a 4-10-1 recurrent network and 4-4-1 feed forward network, both trained with the Levenberg–Marquardt algorithm has produced a better performances on indicators related to average goodness of prediction for the 1 to 5-h ahead river flow forecasts compared to other models. Feed forward network trained with gradient descent with momentum and adaptive learning rate backpropagation algorithm model appears to be the worst of the adaptive techniques investigated in terms of modeling performances. Thus, the results of the study suggest that recurrent and feed forward network trained with Levenberg–Marquardt are able to forecast the catchment flow up to 5 h in advance with reasonable prediction accuracy.