Accurate prediction of sediment concentration processes is crucial for effective reservoir flood regulation, sediment management, and ecological conservation in river basins with significant sediment loads, such as the Yellow River. The ability to predict sediment concentrations accurately is key to mitigating the negative impacts of sedimentation in reservoirs, optimizing flood control, and ensuring the safety of infrastructure and water quality management. In this study, we focus on the sediment concentration process at the Tongguan Hydrological Station, which is located in the middle and lower reaches of the Yellow River. The Yellow River, known for its high sediment load, plays a vital role in the sediment transport dynamics that influence the river"s water quality and sedimentation patterns. The Tongguan station is a critical monitoring point in the Yellow River basin because it marks the confluence of the Yellow River mainstream with major tributaries such as the Weihe and Beiluohe Rivers, and it is located just before sediment enters the Sanmenxia Reservoir, which significantly influences downstream sedimentation and flood management. This study proposes a Parallel Spatio-Temporal Attention Long Short-Term Memory (PSTA-LSTM) model designed for sediment concentration forecasting, specifically tailored to handle the dynamic and complex flow-sediment conditions of the Yellow River. The PSTA-LSTM model integrates a parallel spatio-temporal attention mechanism that allows it to jointly capture multiscale temporal dependencies and spatial correlations among different watershed sites, significantly improving its ability to model sediment transport processes. In addition, the model incorporates an adaptive segmented rectified linear unit (SReLU) in the hidden layers to enhance the model’s capacity to learn complex nonlinear features and better handle rapid fluctuations in sediment concentrations, particularly during peak sediment events. This adaptive function helps the model manage the large variability in sediment loads commonly observed in rivers like the Yellow River, which is subject to varying flow conditions, tributary contributions, and local erosion-deposition dynamics. Experiments were conducted using measured hydrological and sediment data from 2000 to 2024. The results show that compared to the traditional serial LSTM structure, introducing the parallel spatio-temporal attention mechanism reduces the overall Root Mean Square Error (RMSE) by approximately 25.6%, and improves Peak Sediment Prediction Accuracy (PRE) by about 12.7%. Incorporating the SReLU activation function significantly enhances peak prediction accuracy, with the Nash-Sutcliffe Efficiency (NSE) improving by over 9%, showing that SReLU can more effectively handle peak sediment values. The study focuses on the sediment concentration process at the Tongguan hydrological station, located in the confluence area of the middle Yellow River. Tongguan Station is situated at the junction of the Yellow River mainstream and major tributaries such as the Weihe and Beiluohe Rivers. It serves as a critical sediment control section before the Yellow River enters the Sanmenxia Reservoir, playing an important role in reservoir regulation and downstream sediment transport. The upstream water and sediment mainly originate from the Yellow River basin above the Longmen Station, as well as the Weihe River basin at Huaxian Station and the Beiluohe River basin at Zhuangtou Station. Among these, the mainstream floods are characterized by sharp rises and falls with high sediment concentration; the Weihe River floods have a longer duration and a more blunt peak, while the Beiluohe River floods have sharp, narrow peaks, high sediment concentration, and rapid sediment wave propagation. The confluence and superposition of floods from different sources at the Tongguan confluence area not only affect sediment transport intensity but also determine the temporal variation characteristics of the sediment concentration process at Tongguan Station. Experiments were conducted using long-term observed hydrological and sediment data from 2000 to 2024, with the samples classified into five flow-sediment regimes: high-flow/high-sediment, medium-flow/medium-sediment, low-flow/low-sediment, high-flow/low-sediment, and low-flow/high-sediment. The PSTA-LSTM model was trained and evaluated separately under each regime and further compared between flood and non-flood seasons to examine its adaptability to varying hydrological conditions. The results show that introducing the parallel spatio-temporal attention mechanism improves the model’s performance significantly, with the Root Mean Square Error (RMSE) decreasing by approximately 25.6% and Peak Sediment Prediction Accuracy (PRE) improving by about 12.7%. Additionally, incorporating the SReLU activation function led to an increase in Nash-Sutcliffe Efficiency (NSE) by 6-11% compared to the traditional ReLU and softplus activation functions, showing its enhanced capability to handle peak sediment values effectively. The results also demonstrate that regime-based training based on flow-sediment types improves the prediction accuracy, with RMSE decreasing by approximately 15.7%, and NSE reaching over 80%. The comparison between flood and non-flood seasons shows that the PSTA-LSTM model exhibits stronger responsiveness during the flood season, especially during periods with significant peak sediment concentrations and rapid short-term fluctuations. These results highlight the model"s ability to adapt to dynamic and complex sediment transport conditions and its potential for real-time sediment concentration forecasting in large river systems like the Yellow River.