The Capacity of the Hybridizing Wavelet Transformation Approach With Data-Driven Models for Modeling Monthly-Scale Streamflow

Hybrid models that combine wavelet transformation (WT) as a pre-processing tool with data-driven models (DDMs) as modeling approaches have been widely investigated for forecasting streamflow. The WT approach has been applied to original time series for decomposing processes prior to the application of DDM modeling. This procedure has been applied to eliminate redundant patterns or information that lead to a dramatic increase in the model performance. In this study, three experiments were implemented, including stand-alone data-driven modeling, hind cast decomposing using WT divided and entered into the extreme learning machine (ELM), and the extreme gradient boosting (XGB) model to forecast streamflow data. The WT method was applied in two forms: discrete and continuous (DWT and CWT). In this paper, a new hybrid model is proposed based on an integrative prediction model where XGB is used as an input selection tool for the importance attributes of the prediction matrix that are then supplied to the ELM model as a predictive model. The monthly streamflow, upstream flow, rainfall, temperature, and potential evapotranspiration of a basin named in 1805 and located in the south east of Turkey, are used for development of the model. The modeling results show that applying the WT method improved the performance in the hindcast experiment based on the CWT form with minimum root mean square error (RMSE = 4.910 m
³
/s). On the contrary, WT deteriorated the performance of the forecasting and the stand-alone models exhibited a better performance. WT increased the performance of the hindcast experiment due to the inclusion of future information caused by convolution of the time series. However, the forecast experiment experienced deterioration due to the border effect at the end of the time series. Hence, WT was found not to be a useful pre-processing technique in forecasting the streamflow.