Improving deep learning-based streamflow forecasting under trend varying conditions through evaluation of new wavelet preprocessing technique

Accurate machine learning streamflow prediction often requires coupling data-driven models with preprocessing techniques. This study aims to improve the performance of deep learning (DL) models, including long short-term memory, recurrent neural network (RNN), and gated recurrent unit (GRU) by incorporating maximal overlap discrete wavelet entropy transform (MODWET) techniques for streamflow forecasting. The merit of MODWET over maximal overlap discrete wavelet transform (MODWT) is that MODWET utilizes Entropy to determine the optimal decomposition level and suitable wavelet function, which was an unaddressed problem in wavelet-based decomposition models. Suitable decomposition level prevents providing unnecessary information or missing essential information. In this study we show that a unique decomposition level and wavelet filter is not suitable for any dataset. The research focuses on monthly streamflow data from three case studies in the CAMEL dataset in the USA. The accuracy of the models is evaluated using statistical measures such as Nash–Sutcliffe efficiency (NSE), root-mean-squared error, percent bias, and correlation coefficient (r). To determine the optimal model, a Taylor diagram is utilized. The results demonstrate the effectiveness of coupling MODWET with DL models in flood forecasting. Furthermore, genetic programming (GP) and partial correlation index (PCI) are employed for predictor selection. Hybrid models, namely MODWET-GP-GRU (NSE of 0.83), MODWET-GP-RNN (NSE of 0.95), and MODWET-PCI-GRU (NSE of 0.95), outperform simple DL models in terms of NSE and Taylor diagram evaluation. This study emphasizes the potential of hybrid models that combine DL algorithms with the recently proposed MODWET technique for streamflow prediction.