Hourly stream flow prediction in tropical rivers by multi-layer perceptron network

Accurate stream flow (SF) prediction is considered among the basic requirements in dealing with several problems associated with the planning, designing and management of water resources and river systems. Multi-layer perceptron network (MLP), was employed to develop artificial intelligence-based models to predict the hourly SF in downstream areas from upstream water level and rainfall records in the Selangor River basin which is a paradigm of humid tropical rivers. Hourly SF, rainfall and water level records of a one-year period (2011) were applied to train and test the MLP-based models which comprise six different combinations of input variables. The developed models' performance was evaluated via the training, testing data set and overall data. The results of the performance evaluation criteria, i.e. correlation co-efficient (R) and mean absolute error (MAE) indicates that high prediction accuracy was attained. The best fit MLP model is M6-MLP with the highest R and lowest MAE. The R between the observed and predicted hourly SF by the M6-MLP model is 0.898 and 0.904, while the MAE is 10.922 and 10.83 for the training and testing data sets, respectively. The results demonstrate that the MLP technique is successfully applied with high accuracy for hourly SF prediction.