On the sediment concentration prediction at the Tongguan Section of the Yellow River based on parallel spatiotemporal attention LSTM
摘要
含沙量变化过程的准确预测对于水库防汛调度及泥沙管理具有关键作用.以黄河中下游潼关站为研究对象,构建了融合并行时空注意力机制的长短时记忆网络模型(PSTA-LSTM).模型在结构上集成并行时空注意力模块,同时捕捉含沙量时间序列中的长期时间依赖性和不同流域站点间的空间关联性;在隐层激活阶段引入改进的自适应分段线性函数(SReLU),提升模型对含沙量快速变化及峰值特征的表征能力.基于2000-2024年实测水沙数据进行实验验证,结果显示,相较于串行时空注意力LSTM,引入并行时空注意力模块使模型在测试期的均方根误差(RMSE)下降约8.5%,沙峰预测误差(PRE)降低约8.6%;引入SReLU激活函数显著提升了沙峰的预测精度,相较于传统的ReLU和softplus,SRe-LU能够更有效地处理沙峰值,纳什效率系数(NSE)提升超10%.将样本根据水沙类型划分为丰水丰沙、中水中沙、枯水枯沙、丰水枯沙和枯水丰沙5类,并在不同水沙条件下以及汛期与非汛期进行含沙量预测,验证了按水沙类型分类训练可进一步提高预测精度,RMSE和PRE均下降约15.7%,NSE达到80%以上;汛期与非汛期对比显示,PSTA-LSTM在含沙量峰值明显且短时变化幅度较大的阶段具有更强的响应能力.
Accurate prediction of sediment concentration dynamics is essential for reservoir flood control operation and sediment management.This study developed a parallel spatiotemporal attention long short-term memory(PSTA-LSTM)model for sediment concentration prediction at Tongguan Hydrological Station in the middle and lower reaches of the Yellow River.The proposed model integrates a parallel spatiotemporal attention mechanism to simultaneously capture long-term temporal dependencies in sediment concentration sequences and spatial correlations among upstream monitoring stations.In addition,an improved adaptive segmented rectified linear unit(SReLU)activation function was introduced into the hidden layers to enhance the representation of rapid sedi-ment concentration fluctuations and peak characteristics.The model was validated using observed hydrological and sediment data collected from 2000 to 2024.Compared with the sequential spatiotemporal attention LSTM(SSTA-LSTM),the proposed parallel attention mechanism reduced the root mean square error(RMSE)by approximately 8.5%and decreased the peak relative error(PRE)by approximately 8.6%on the test dataset.Furthermore,compared with the conventional ReLU and softplus activation functions,SReLU improved the Nash-Sutcliffe efficiency(NSE)by more than 10%,demonstrating superior capability in capturing sediment concentration peaks.The datasets were further classified into five hydrological-sediment regimes,namely high-flow/high-sediment,medium-flow/medium-sediment,low-flow/low-sediment,high-flow/low-sediment,and low-flow/high-sediment.Re-gime-specific training further improved the prediction accuracy,reducing both RMSE and PRE by approximately 15.7%,while a-chieving an NSE exceeding 0.80.Comparisons between flood and non-flood seasons further demonstrated that the proposed PSTA-LSTM model exhibited stronger responsiveness during periods characterized by pronounced sediment peaks and rapid short-term fluctuations.