<p>Accurate prediction of suspended sediment concentration in rivers is essential for water resources management, hydraulic structure design, and environmental protection. This study evaluates the performance of four deep learning-based modeling frameworks, including two standalone models (gated recurrent unit (GRU) and quasi-recurrent neural network (QRNN)), a sequential hybrid model (GRU-QRNN), and a hybrid model optimized with an improved atom search algorithm (IASO-GRU-QRNN), in daily sediment prediction at two stations with different hydrological conditions (Albuquerque at Rio Grande River and Omaha at Missouri River). Daily discharge and sediment concentration data over 26 years (1998–2024) were used, and six structured input scenarios were designed to examine the impact of historical memory and instantaneous data. The results showed that the performance of the models is strongly influenced by basin characteristics. At the Omaha station with the adjusted regime, the GRU model achieved the best performance (coefficient of determination 0.92, root mean square error 89 mg/l). In contrast, at the more complex Rio Grande station, the IASO-GRU-QRNN hybrid model with automatic hyperparameter optimization outperformed, achieving a coefficient of determination of 0.63 and a root mean square error of 780 mg/l, representing a 25% improvement in error over the baseline GRU model. Scenario analysis clearly revealed the determining role of the current discharge variable, with its inclusion resulting in an average 20% improvement in the coefficient of determination across all models. This study demonstrates that in more complex environments, intelligent hybrid approaches that combine advanced architectures and metaheuristic optimization can achieve higher reliability in sediment prediction.</p>