ABSTRACT A schematic illustration of the river flow time-series prediction process, beginning with wavelet transformation and followed by artificial neural network (ANN) modeling. The workflow includes feature selection, optimization, and iterative processing steps. Time series of natural phenomena contain considerable information about the underlying mechanisms, and appropriate forecasting of such time series can be helpful for the management of related disasters. One of the most powerful and widely used approaches in this field is the hybridization of the discrete wavelet transform (DWT) with artificial intelligence (AI) models, including artificial neural networks (ANNs). In this study, group-discrete wavelet transforms (GDWTs) with parallel computation were employed to explore the unique features of wavelets and to extract distinct and hidden aspects of the time series. The NSGA-III algorithm was applied using a bi-objective optimization approach to select input variables and optimize model performance. Daily streamflow data from the Thames and Lee Rivers' catchments, near London, England, covering a 10-year period (2010–2020), were used to train and test the models. For example, the RMSE (m3/s) values in the verification stage for the Thames River catchment area were 14.29, 11.51, 9.20, and 4.15 for the ANN, DWT-ANN, GDWT-ANN, and GDWT-NSGA-III-ANN models, respectively. In conclusion, the proposed GDWT-NSGA-III-ANN model demonstrates reliable performance than the classic ANN, DWT-ANN, and GDWT-ANN models in river flow forecasting task.
Momeneh et al. (Fri,) studied this question.