Development of an efficient input selection method for NN based streamflow model

In this paper, using a neural network-based streamflow simulation model (NNSSM), we simulate the runoff of the Ajichai River. The selection of suitable inputs is an essential step toward developing NNSSM. For this aim, we investigate a novel application of the Genetic Classification Algorithm (GCA) as an input variable selection (IVS) method in comparison with the Self-Organizing Map (SOM) and Binary Fully Informed Particle Swarm Optimization (BFIPSO). In another innovative application, we establish Social Choice (SC) for the final ranking of selected data using SOM. Next, the model was improved by adding seasonality indexes. The results indicate the superiority of GCA. The average (maximum) Nash-Sutcliffe index for GCA was found to be 0.63(0.84), while it was 0.55(0.71) and 0.58(0.77) for SOM-SC and BFIPSO, respectively. Moreover, GCA took less than 30 min for each run, while for SOM-SC and BFIPSO at least 3 and 48 h were needed under the same circumstances.