Improvement of Support Vector Machine and Random Forest Algorithm in Predicting Khorramabad River Flow Using Non-uniform De-Noising of data and Simplex Algorithm

In this study, in order to simulate the monthly flow of the Khorramabad River, the time series of this river was decomposed into three levels using the wavelet of Daubechies-3, during the period of 1955-2014. Based on this, it was found that there is a Non-uniform noise that includes two periods of time in this signal, with the October 2008 border which required that the signal become non-uniform de-noising. Subsequently, by using two models of support vector machine (ɛ and Nu) and random forest algorithm (RF), the main signal and non-uniform de-noising signal of the river flow were simulated separately. The results validation criteria of the model showed that, with the non-uniform de-noising of the river flow signal, the error of the models dropped ɛ from 5.7 to 3.1, Nu from 5.8 to 3.2 and RF from 5 to 2.9 m3/s. Also, the comparison and testing of the computational error ɛ: ɛD; Nu: NuD; RF: RFD were obtained by using the MGN test (-15, -15, -107.6), which indicates a significant improvement in the performance of the models used as a result of the non-uniform de-noising signal. In the following, using optimization simplex algorithm in the of three models ɛD, NuD and RFD, the mean flow of the river was very high in all three models.