Application of Data Mining Tools for Long-Term Quantitative and Qualitative Prediction of Streamflow

Abstract

This paper evaluates the performances of two long-term prediction approaches for streamflow and riverine total dissolve solids (TDS) and compares their results with observed data and with short-term predicted values. The future values predicted by the first, long-term, prediction approach (Approach 1) depend on data corresponding to time steps prior to the prediction time step. The future values predicted by the second, long-term, prediction approach (Approach 2) depend on data comprised within the observational period. Each long-term prediction approach calculates streamflow and TDS over a 12-month period ranging from April through March (Scheme 1) and by agricultural water year (December through November, Scheme 2). Genetic programming (GP) is implemented for long-term prediction. Prediction is applied to the streamflow and TDS of the Karoon River in southwestern Iran. The long-term Approach 1 was found to be more accurate than the long-term Approach 2 judged by the values of several diagnostic statistics. The root mean square error (RMSE), correlation coefficient (R2), and Nash-Sutcliffe efficiency (E) statistics of long-term predictions of streamflow and TDS with Approach 1 are lower than those obtained with the long-term prediction Approach 2 for April–March and for the agricultural water-year predictions. It is concluded that prediction of the Karoon River’s streamflow and TDS is best accomplished using GP in combination with the long-term prediction Approach 1.