The potential of integrated hybrid data processing techniques for successive-station streamflow prediction

Streamflow is one of the most important issues in river engineering due to its impact on planning and operation of the water resources system. In this study, the capability of newly integrated hybrid prediction models based on artificial intelligence and data processing methods was assessed for monthly river streamflow modeling. In this regard, three successive hydrometric stations of Housatonic River were selected and based on the previous time steps of streamflow values during the period of 1941-2018 several models were developed. During the modeling process, two states based on stations own data (state 1) and upstream stations data (state 2) were considered. For data preprocessing, first temporal features of the streamflow series were decomposed via wavelet transform (WT). Then, the obtained subseries were further broken down into intrinsic mode functions using ensemble empirical mode decomposition (EEMD) to obtain features with higher stationary properties. Finally, the most efficient subseries were selected and used for artificial intelligence approaches [i.e., feed forward neural network (FFNN), kernel extreme learning machine (KELM), and support vector machine (SVM)] as inputs. Also, data post-proceeding was done using simple linear averaging (SLAM) and nonlinear neural ensemble (NNEM) methods. Based on the results, the pre- and post- processing data were more accurate as compared to single artificial intelligence method. The integrated pre-post-processing models improved the models efficiency by approximately 45%. It was observed that via the integrated approaches, the upstream stations data could be applied successfully for streamflow modeling when the stations own data were not available.

Automated summary

This study evaluated the capability of newly integrated hybrid prediction models based on artificial intelligence and data processing methods for monthly river streamflow modeling. The integrated pre-post-processing models improved the models efficiency by approximately 45% and allowed the successful application of upstream stations data for streamflow modeling.