Designing a decomposition-based multi-phase pre-processing strategy coupled with EDBi-LSTM deep learning approach for sediment load forecasting

Forecasting accurately suspended sediment load (SSL) in the basin is one of the most critical issues for river engineering, environment, and water resources management which effectively reduces flood damages. In this study, a new multi-criteria hybrid expert system comprised of empirical wavelet decomposition (EWT) integrated with Encoder-Decoder Bidirectional long short-term memory (EDBi-LSTM), supported by five feature selection (FS) methods was developed for the first time to forecast daily SSL at two study sites (Bamini and Ashti) of Godavari river basin, India. The employed FS schemes are including Boruta-Random forest (BRF), simulated annealing (SA), Relief algorithm, Ridge regression (RR), and Mutual information (MI) where the BRF coupled with EWT and EDBi-LSTM (i.e., E_WT-E_DBi-LSTM-Boruta) is identified as the main forecasting paradigm. Here the original SSL signals in the monsoon season (2001–2015) as the only input information were considered to forecast SSL events at a daily time scale in both study zones. The SSL signals were decomposed using the EWT technique considering the significant antecedent time-lagged inputs based on partial auto-correlation function (PACF). In the next stage, five FS strategies were addressed to specify the significant sub-sequences to reduce computational cost and enhance forecasting accuracy. Besides, the extreme gradient boosting (XGB) approach was implemented to compare the potential of the hybrid EDBi-LSTM and standalone counterpart models for both study sites. According to several goodness-of-fit indices and validation tools, the outcomes at the Bamini and Ashti sites demonstrated that the E_WT-E_DBi-LSTM-Boruta as the main model, achieved the best accuracy, followed by E_WT-X_GB-Boruta, E_WT-E_DBi-LSTM-SA, and E_WT-X_GB-SA, respectively. Comparing all the hybrid models showed that the BRF, SA, and RR strategies performed better in integration with machine learning (ML) models.