Prediction of suspended sediment concentration in fluvial flows using novel hybrid deep learning model

Accurately predicting suspended sediment concentration (SSC) in fluvial systems is essential for environmental monitoring, flood management, and riverine engineering applications. This study introduces a novel hybrid approach for forecasting SSC by leveraging advanced deep learning algorithms. Daily datasets from the U.S. Geological Survey, including discharge (Q) and SSC measurements, were analyzed from 2007 to 2017 at two key locations on the Mississippi River: Chester (CH) and Thebes (TH). The proposed framework integrates feedforward neural networks (FFNN), long short-term memory (LSTM) networks, stochastic gradient descent (SGD), and radial basis function (RBF) models, augmented with a first-order differencing technique. Additionally, hybrid models, including Supervised FFNN-LSTM and Supervised FFNN-SGD, were developed to enhance predictive performance. The dataset was partitioned into training (70%, 2,747 d) and testing (30%, 1,178 d) subsets, with daily temporal resolution. Six input scenarios incorporating lagged parameters were evaluated using performance metrics, including the correlation coefficient (CC), Nash–Sutcliffe efficiency (NSE), scatter index (SI), and Willmott’s index (WI). Sensitivity analysis identified SSCt-1 (i.e., one day before) as the most influential predictor for short-term forecasting. Among the models, the SFFNN-LSTM-6 achieved the highest performance, with CC values of 0.976 for CH and 0.960 for TH, demonstrating the ability to predict SSC effectively even in the absence of current-day discharge data. The proposed hybrid models exhibited exceptional robustness across diverse flow regimes, including extreme environmental conditions, establishing a reliable tool for SSC forecasting in complex fluvial systems.