Comparative performance of support vector regressions for accurate streamflow predictions

Abstract

Obtaining accurate streamflow predictions can be challenging due to the inherent variabilities and complex nonlinear nature in streamflow generation processes. Support vector regression model is an effective forecasting tool to forecast streamflow as it is able to capture the nonlinearity in the data and attain the global optimum parameters in the forecasted model. However, the efficiency of SVR might be hindered by noise that typically exists in any hydrological time series data through random influences and inaccuracies in recording. Thus, this condition could compromise the quality of input data into SVR. In this study, we investigate the effectiveness of forecasting monthly streamflow data using different settings of SVR in two ways. First, we use different variations of wavelet denoising technique using different selections of wavelet decomposition levels and mother wavelets in order to preserve information and reduce distortion of the original time series. For this purpose, we measured the impact of six different wavelets on SVR namely Daubechies of type db3, db4, db5, db6 and db7 with two different levels of decomposition which are level 3 and level 4. There is more information that may contribute to better performance of the model when the decomposition level is increase. Then, the data are applied using radial basis function (RBF) by performing K-fold cross-validation to obtain the optimal parameter for kernel function in forecasting streamflow. We illustrate the methods using the monthly streamflow data observed at Segamat River in the state of Johor. The results demonstrated that SVR based wavelet denoising for 1-month lead time streamflow forecasting of type db5 with level 3 give better results using Gaussian (RBF) kernel function based on K-fold cross-validation compared to regular SVR. This implies that reduced variance in the denoising procedure and obtain optimal parameter in kernel function may improve forecasting accuracy.