A data-driven model based on Fourier transform and support vector regression for monthly reservoir inflow forecasting

The recent trend for data-driven streamflow forecasting is to hybridize artificial intelligence with decomposition pre-processing. In this paper, a decomposition-based data-driven model called FT-SVR that exploits both Fourier transform (FT) and support vector regression (SVR) techniques is proposed for monthly reservoir inflow forecasting and the Three Gorges Dam (TGD) located on the Yangtze River in China is taken as the case for study. As the inflow time series contains oscillations of disparate scales, FT-SVR uses FT to appropriately decompose the series into multiple decomposed components, with each component comprising of neighboring frequencies and having a clear physical meaning. SVR is employed to develop an independent forecasting model for each decomposed component. The development of each SVR model involves data normalization, input selection based on autocorrelation function and partial autocorrelation function analysis, and parameter calibration by a metaheuristic. FT-SVR is compared with three other models which are the same with FT-SVR except that one uses ensemble empirical mode decomposition, one uses singular spectrum analysis for decomposition, and the other one performs no decomposition. Experimental results demonstrate that FT-SVR is able to give almost perfect monthly inflow forecasting for the TGD and significantly outperforms the three other models, in terms of evaluation criteria including root mean squared error, correlation coefficient, mean average percentage error, Nash-Sutcliffe efficiency coefficient, and relative error of maximum/minimum monthly inflow.