文献类型：期刊文献

英文题名：Evaluating a coupled discrete wavelet transform and support vector regression for daily and monthly streamflow forecasting

作者：Liu, Zhiyong[1] Zhou, Ping[2] Chen, Gang[3] Guo, Ledong[2]

第一作者：Liu, Zhiyong

通信作者：Zhou, P[1]

机构：[1]Heidelberg Univ, Inst Geog, D-69120 Heidelberg, Germany;[2]Guangdong Acad Forestry, Res Stn Doneangyuan Forest Ecosyst, Guangzhou 510520, Guangdong, Peoples R China;[3]Chinese Acad Forestry, Inst Forest Resource Informat Tech, Being 100091, Peoples R China

年份：2014

卷号：519

期号：PD

起止页码：2822-2831

外文期刊名：JOURNAL OF HYDROLOGY

收录：;EI(收录号：20145100353269);Scopus(收录号：2-s2.0-84918773671);WOS:【SCI-EXPANDED(收录号:WOS:000347589600014)】；

基金：The study was financially supported by the State Forestry Administration Public Benefit Research Foundation of China (201204104, 201104005) and State Forestry Administration 948 Innovative Significant Project of China (2011-76). Special appreciation is given to the USGS for providing the streamflow data. We gratefully acknowledge the helpful comments of the editor and anonymous reviewers.

语种：英文

外文关键词：Wavelet analysis; Support vector regression; Streamflow forecasting; Model averaging; Indiana

摘要：This study investigated the performance and potential of a hybrid model that combined the discrete wavelet transform and support vector regression (the DWT-SVR model) for daily and monthly stream-flow forecasting. Three key factors of the wavelet decomposition phase (mother wavelet, decomposition level, and edge effect) were proposed to consider for improving the accuracy of the DWT-SVR model. The performance of DWT-SVR models with different combinations of these three factors was compared with the regular SVR model. The effectiveness of these models was evaluated using the root-mean-squared error (RMSE) and Nash-Sutcliffe model efficiency coefficient (NSE). Daily and monthly streamflow data observed at two stations in Indiana, United States, were used to test the forecasting skill of these models. The results demonstrated that the different hybrid models did not always outperform the SVR model for 1-day and 1-month lead time streamflow forecasting. This suggests that it is crucial to consider and compare the three key factors when using the DVVT-SVR model (or other machine learning methods coupled with the wavelet transform), rather than choosing them based on personal preferences. We then combined forecasts from multiple candidate DWT-SVR models using a model averaging technique based upon Akaike's information criterion (AIC). This ensemble prediction was superior to the single best DWT-SVR model and regular SVR model for both 1-day and 1-month ahead predictions. With respect to longer lead times (i.e., 2- and 3-day and 2-month), the ensemble predictions using the AIC averaging technique were consistently better than the best DWT-SVR model and SVR model. Therefore, integrating model averaging techniques with the hybrid DVVT-SVR model would be a promising approach for daily and monthly streamflow forecasting. Additionally, we strongly recommend considering these three key factors when using wavelet-based SVR models (or other wavelet-based forecasting models). (C) 2014 Elsevier B.V. All rights reserved.