Are hybrid models integrated with data preprocessing techniques suitable for monthly streamflow forecasting? Some experiment evidences

Abstract

Summary A number of hydrological studies have proven the superior prediction performance of hybrid models coupled with data preprocessing techniques. However, many studies first decompose the entire data series into components and later divide each component into calibration and validation datasets to establish models, which sends some amount of future information into the decomposition and reconstruction processes. As a consequence, the resulting components used to forecast the value of a particular moment are computed using information from future values, which are not available at that particular moment in a forecasting exercise. Since most papers don’t present their model framework in detail, it is difficult to identify whether they are performing a real forecast or not. Even though several other papers have explicitly stated which experiment they are performing, a comparison between results in the hindcast and forecast experiments is still missing. Therefore, it is necessary to investigate and compare the performance of these hybrid models in the two experiments in order to estimate whether they are suitable for real forecasting. With the combination of three preprocessing techniques, such as wavelet analysis (WA), empirical mode decomposition (EMD) and singular spectrum analysis (SSA), and two modeling methods (i.e. ANN model and ARMA model), six hybrid models are developed in this study, including WA-ANN, WA-ARMA, EMD-ANN, EMD-ARMA, SSA-ANN and SSA-ARMA. Preprocessing techniques are used to decompose the data series into sub-series, and then these sub-series are modeled using ANN and ARMA models. These models are examined in hindcasting and forecasting of the monthly streamflow of two sites in the Yangtze River of China. The results of this study indicate that the six hybrid models perform better in the hindcast experiment compared with the original ANN and ARMA models, while the hybrid models in the forecast experiment perform worse than the original models and the performances of WA-based and EMD-based models vary largely across different extension methods. It can be concluded that the hybrid models are not suitable for monthly streamflow forecasting in this study. New extension methods and modified preprocessing techniques can improve the prediction performance of these hybrid models in forecast experiments.