Streamflow forecasting by combining neural networks and fuzzy models using advanced methods of input variable selection

In this paper the performance of extreme learning machine (ELM) training method of radial basis function artificial neural network (RBF-ANN) is evaluated using monthly hydrological data from Ajichai Basin. ELM is a newly introduced fast method and here we show a novel application of this method in monthly streamflow forecasting. ELM may not work well for a large number of input variables. Therefore, an input selection is applied to overcome this problem. The Nash-Sutcliffe efficiency (NSE) of ANN trained by backpropagation (BP) and ELM algorithm using initial input selection was found to be 0.66 and 0.72, respectively, for the test period. However, when wavelet transform, and then genetic algorithm (GA)-based input selection are applied, the test NSE increase to 0.76 and 0.86, respectively, for ANN-BP and ANN-ELM. Similarly, using singular spectral analysis (SSA) instead, the coefficients are found to be 0.88 and 0.90, respectively, for the test period. These results show the importance of input selection and superiority of ELM and SSA over BP and wavelet transform. Finally, a proposed multistep method shows an outstanding NSE value of 0.97, which is near perfect and well above the performance of the previous methods.