Prediction of River Discharge by Using Gaussian Basis Function

For design of water resources engineering related project such as hydraulic structures
like dam, barrage and weirs river discharge data is vital. However, prediction of river
discharge is complicated by variations in geometry and boundary roughness. The
conventional method of estimation of river discharge tends to be inaccurate because
river discharge is nonlinear but the method is linear. Therefore, an alternative method
to overcome problem to predict river discharge is required. Soft computing technique
such as artificial neural network (ANN) was able to predict nonlinear parameter such
as river discharge. In this study, prediction of river discharge in Pari River is
predicted using soft computing technique, specifically gaussian basis function. Water
level raw data from year 2011 to 2012 is used as input. The data divided into two
section, training dataset and testing dataset. From 314 data, 200 are allocated as
training data and the remaining 100 are used as testing data. After that, the data will
be run by using Matlab software. Three input variables used in this study were
current water level, 1-antecendent water level, and 2-antecendent water level. 19
numbers of hidden neurons with spread value of 0.69106 was the best choice which
creates the best result for model architecture after numbers of trial. The output
variable was river discharge. Performance evaluation measures such as root mean
square error, mean absolute error, correlation of efficiency (CE) and coefficient of
determination (R2) was used to indicate the overall performance of the selected
network. R2 for training dataset was 0.983 which showed predicted discharge is
highly correlated with observed discharge value. However, testing stage performance
is decline from training stage as R2 obtained was 0.775 consequently presence of
outliers have affect scattering of whole data of testing and resulted in less accuracy
as the R2 obtained much lower compared to training dataset. This happened because
less number of input loaded into testing than training. RMSE and MSE recorded for
training much lower than testing indicated that the better the performance of the
model since the error is lesser. The comparison of with other types of neural network
showed that Gaussian basis function is recommended to be used for river discharge
prediction in Pari river.