Suspended sediment concentration prediction using a hybrid VMD-LSTM-SA-ensemble model

Accurate prediction of suspended sediment concentration (SSC) is critically important for water quality assessments in marine and terrestrial aquatic environments. However, there is uncertainty associated with traditional predictive algorithms due to the non-stationarity of the SSC data; hence, studies to improve predictions are required. A Variational Mode Decomposition-Long Short-Term Memory-Self Attention-ensemble (VMD-LSTM-SA-ensemble) model is proposed for SSC prediction in this paper. The SSC data from the Rio Grande River, U.S., and the Belgian coastal zone were used to compare the performance of several models. It is found that the accuracy of the LSTM model is higher than the Convolutional Neural Network (CNN) and Gated Recurrent Unit (GRU) models. The integration of the LSTM model with a SA mechanism enhances prediction accuracy. Additionally, by applying VMD, the non-stationarity of the SSC data is effectively reduced. Thus, both the VMD-LSTM-SA-add and VMD-LSTM-SA-ensemble models perform better than the LSTM-SA model. The VMD-LSTM-SA-ensemble model effectively mitigates the accumulation of errors common in other models. Consequently, the VMD-LSTM-SA-ensemble model exhibits better performance in both single-step and multi-step predictions. These findings demonstrate the VMD-LSTM-SA-ensemble model exhibits superiority in predicting the SSC.