River flow simulation based on empirical mode function signals and random forest algorithm

To investigate the effect of time-series decomposition on the error rate and accuracy of simulations, two random forest, and hybrid CEEMD-RF models were used in simulating river discharge in the Mahabadchai, Nazlochai, Siminehrud, and Zolachai sub-basins over the Lake Urmia catchment from 1971 to 2019. The results showed that the simulated values by the random forest model are within the 95% confidence intervals, and the model's efficiency is deemed acceptable according to the Nash–Sutcliffe criterion. Besides, the average values of simulated flow are slightly higher than observed discharges. To simulate the river flow values using the CEEMD-RF model, the river flow data was first decomposed into nine empirical mode function values and a residual series. By choosing the best lag for each of the nine values, these values were simulated using the random forest algorithm in the training and testing phases, and the cumulative values were compared with the observed discharges. The research findings showed that the model's efficiency in the decomposed state had increased significantly. Based on the RMSE criterion, the simulation results of the river flow values on a monthly scale showed that the decomposition of the river flow values reduced the error rate in Mahabadchai, Nazlochai, Siminehrud, and Zolachai sub-basins in the training phase to about 36, 94, 29 and 20 percent, respectively, and in the testing phase, improve about 54, 86, 40, and 36 percent. The research results showed that by decomposing the observational values and choosing the best lag, it is possible to cover the range of observational data changes and simulate the minimum and maximum points of the data well.