Evaluation of random forest and hybrid time series models in bivariate simulation of electrical conductivity

The quality parameters of the river, including electrical conductivity, are highly dependent on changes in flow rate. Adding the flow rate parameter to the simulation of this parameter can increase the certainty of the simulation results. For this reason, in this study, random forest, CARMA and CARMA-GARCH models were used to model the electrical conductivity values in Gerdyaghoub, Kutar and Bitas stations in Mahabadchai basin, taking into account the flow rates. In this regard, the monthly values of electrical conductivity and flow discharge in the statistical period 1986-2018 were used. The results were analyzed using Nash-Sutcliffe statistics, root mean square error and violin plot. The results of evaluation the root mean square error and Nash-Sutcliffe statistics showed that the simulation results of CARMA-GARCH model compared to CARMA model in Bitas and Kuter stations as well as the training step in Gerdyaghoub station were improved. The results showed that the combination of nonlinear and linear models could improve the modeling error in three stations, Gerdyaghoub, Kotar and Bitas in the training step of 9.56, 9.70 and 21.68 percent. By examining the violin plots, the results showed acceptable accuracy and performance of CARMA and CARMA-GARCH models compared to the random forest model. In general, the results showed that time series models have higher accuracy in bivariate simulating of electrical conductivity values in the study area.