基于RF-LSTM模型的长江上游汛期日含沙量预测

Prediction of Daily Sediment Concentrations in the Upper Reaches of Yangtze River During Flood Season Based on RF-LSTM Model

含沙量预测对流域泥沙治理、水沙调控以及水质与水环境管理等具有重要意义。长江上游地区幅员广阔,支流众多,水沙来源复杂,对准确预测三峡入库含沙量过程构成了挑战。针对长江上游区间降雨和干支流来水来沙对寸滩站含沙量产生不同程度的影响,提出了一种基于随机森林(Random Forest,RF)与长短时记忆(Long Short Term Memory,LSTM)神经网络结合的日含沙量预测深度学习模型RF-LSTM。首先,该模型利用RF算法筛选出与寸滩站日含沙量相关性强的水沙因子,然后将这些因子作为LSTM神经网络的输入变量,进一步识别出优选水沙因子与寸滩含沙量之间的映射关系,最后以长江上游向家坝至寸滩区间为研究区域,应用该模型对不同预见期下的寸滩站汛期日含沙量进行了预测,结果表明:与LSTM模型相比,RF-LSTM模型能较好地考虑预测因子对含沙量影响的滞后效应,且有效捕获与寸滩站日含沙量相关性强的特征,四种预见期下其在预测精度和性能方面均有较好表现,其中无预见期和预见期1 d时两种模型预测精度均较高,验证期的纳什效率系数均大于0.82,无预见期下RF-LSTM模型的纳什效率系数可达到0.91,相应的均方根误差和平均绝对误差分别较LSTM模型降低了13%和8%,且两种预见期下RF-LSTM模型可以较为准确捕获沙峰及峰现时间;当预见期增加至2 d和3 d时两种模型精度均有明显下降,但RF-LSTM模型计算精度仍优于LSTM模型。研究结果可为长江上游含沙量预测提供参考。

The prediction of sediment concentrations is of significant importance for watershed sediment control,water-sediment regulation,as well as water quality and environmental management.The upper reaches of the Yangtze River,vast area in size and abundant in tributaries with complex water and sediment sources,pose a great challenge to accurately predict the process of suspended sediment concentration(SSC)entering the Three Gorges Reservoir.In this study,we propose a deep learning model named RF-LSTM,which combines Random Forest(RF)algorithm and Long Short-Term Memory(LSTM)neural network for daily SSC prediction at Cuntan station.This model addresses the varying impacts of rainfall in the upper reaches of the Yangtze River,as well as the inflow of water and sediment from its main stream and tributaries,on the daily SSC observed at the station.Firstly,the RF algorithm is employed to identify water and sediment factors that exhibit a strong correlation with the SSC at Cuntan.These factors are then utilized as input variables for the LSTM neural network to discern the mapping relationship between the optimized set of factors and the SSC at Cuntan.Finally,the model is applied in the region spanning from Xiangjiaba to Cuntan of the upper Yangtze River to predict the daily SSC during flood season at Cuntan station under different forecast periods.Results show that,compared to the LSTM model,the RF-LSTM model can better account for the lagged effects of the predictor factors on the SSC,and effectively capture the features that are strongly correlated with the SSC at Cuntan station.Under all the four different forecast horizons considered,the RF-LSTM exhibits superior performance in terms of both prediction accuracy and overall capability.Specially,when considering no-forecast and 1-d forecast horizons,both models exhibit high prediction accuracy,with Nash-Sutcliffe efficiency coefficients exceeding 0.82 during the validation period.Notably,the RF-LSTM model achieves a Nash-Sutcliffe efficiency coefficient of 0.91 in the no-forecast horizon,outperforming the LSTM model by reducing mean absolute errors and root mean square errors by 8%and 13%,respectively.Furthermore,under both of these forecast horizons,the RF-LSTM model more precisely captures SSC peaks and their occurrence timings.However,as the forecast horizon increases to 2 days and 3 days,the accuracy of both models decrease significantly.Nevertheless,the RF-LSTM model continues to outperform the LSTM model in terms of computational accuracy,demonstrating its robustness and reliability across different forecast horizons.These findings highlight the potential of the RF-LSTM model as a valuable tool for SSC prediction in the upper reaches of the Yangtze River,offering a valuable reference for future studies in this domain.