Developing long short-term memory combined with numerical first order differential optimization and clockwork recurrent neural network to predict suspended sediment load

The analysis of aquatic ecosystems heavily relies on the parameter of suspended sediment load (SSL) in rivers, which affects the chemical and biological compositions of water. This is crucial for decision-making in water resource management and environmental conservation. This study aims to predict SSL in the Missouri and Mississippi rivers using daily SSL and discharge (Q) data from the Nebraska City (NC) and Below Grafton (BG) stations (1999–2017). We introduce an innovative strategy leveraging artificial intelligence techniques like Long Short-Term Memory (LSTM), Feedforward Neural Networks (FFNN), Random Forest (RF), and Clockwork Recurrent Neural Network (CWRNN) to enhance SSL forecasting. The integration of the first-order difference (DIFF) refines the input data, heightening sensitivity to temporal sediment variations. Models’ precision was assessed using correlation coefficient (CC), root mean square error (RMSE), relative absolute error (RAE), Bias Factor (BF), and Kling-Gupta Efficiency (KGE). Sensitivity analysis with Shapley Additive Explanations (SHAP) indicated SSLt-1 as the key parameter. The hybrid models, notably DIFF-CWRNN-LSTM, DIFF-FFNN-LSTM, and DIFF-RF-LSTM, displayed superior performance, with the DIFF-CWRNN-LSTM model exhibiting the highest accuracy: CC of 0.922 and 0.968, KGE of 0.980 and 0.995 for NC and BG stations, respectively. This novel approach consistently predicts SSL under varying conditions, marking it as a reliable and recommended model. The enhanced model effectively predicts SSL peak values, addressing a historical research gap, thus improving its applicability in environmental and hydrological impact assessments.