Performance evaluation of ANNs and an M5 model tree in Sattarkhan Reservoir inflow prediction

Abstract

In this research, done from 2002 to 2012 at the Sattarkhan Reservoir in Iran, different combinations of values of a time series related to precipitation, evaporation, and discharge at upstream stations were used as the inputs and the next-day discharge was used as the output, in order to evaluate an artificial neural network (ANN), Support Vector Regression (SVR), Wavelet Neural Networks (WANN) and M5 model tree. Twelve scenarios consisting of input variables with different combinations of flow and climatological data by different lag times were created to study the role played by input variables in the accuracy of the model. Results of the various scenarios showed that when the inputs were consisted of the amount of temperature, precipitation and previous discharges (scenario 2), the next-day discharge could be estimated with higher accuracy by the WANN model with a RMSE value of 0.31 m3/s in comparison with the SVR model with RMSE value of 0.873 m3/s, M5 model tree with RMSE value of 0.880 m3/s and ANN model with RMSE value of 0.896 m3/s. Conclusively, the obtained results revealed that the wavelet transformation creates significant effect for increasing the prediction accuracies of the ANN model.