Estimating streamflow of the Kızılırmak River, Turkey with single- and multi-station datasets using Random Forests.

Abstract

Predicting missing historical or forecasting streamflows for future periods is a challenging task. This paper presents open-source data-driven machine learning models for streamflow prediction. The Random Forests algorithm is employed and the results are compared with other machine learning algorithms. The developed models are applied to the Kızılırmak River, Turkey. First model is built with streamflow of a single station (SS), and the second model is built with streamflows of multiple stations (MS). The SS model uses input parameters derived from one streamflow station. The MS model uses streamflow observations of nearby stations. Both models are tested to estimate missing historical and predict future streamflows. Model prediction performances are measured by root mean squared error (RMSE), Nash-Sutcliffe efficiency (NSE), coefficient of determination (R2), and percent bias (PBIAS). The SS model has an RMSE of 8.54, NSE and R2 of 0.98, and PBIAS of 0.7% for the historical period. The MS model has an RMSE of 17.65, NSE of 0.91, R2 of 0.93, and PBIAS of -13.64% for the future period. The SS model is useful to estimate missing historical streamflows, while the MS model provides better predictions for future periods, with its ability to better catch flow trends.