摘要
【目的】解决传统河流径流量及含沙量预测方法精度不高的问题。【方法】以陕西泾河流域为例,以泾河景村水文站1981—2010年实测逐日径流量、含沙量系列数据为样本,提出了一种基于长短时记忆神经网络(LSTM)的时间序列预测方法。首先对基础数据进行分析及处理,消除测量误差及数据缺失对预测性能的影响,挖掘30a内逐日的径流量、含沙量之间的关系以及在年尺度中的变化规律。其次构建LSTM深度学习模型,通过不断训练及验证,揭示径流量和含沙量数据信息的时间序列相关性。【结果】利用LSTM模型预测的来水量及含沙量均方根误差(RMSE)、平均绝对误差(MAE)均远小于自回归滑动平均(ARMA)、支持向量回归(SVR)和线性回归(LR)等传统模型预测结果。【结论】该预测方法可以充分利用水文信息的时序性,提高预测精度。
【Objective】We aimed to solve the problem of low accuracy of traditional river runoff and sediment concentration prediction methods.【Method】Taking Jinghe River Basin in Shaanxi Province as an example,a time series prediction method based on long-term and short-term memory neural network(LSTM)was proposed in this paper,taking the series data of daily runoff and sediment concentration measured at Jinghe village hydrological station from 1981 to 2010 as samples.Firstly,the basic data were analyzed and processed to eliminate the influence of measurement error and data missing on the experimental performance,and the relationship between runoff and sediment concentration of each day in 30 years and the change rule in the annual scale were explored.Secondly,the LSTM deep learning model was constructed to reveal the time series correlation of runoff and sediment data through continuous training and verification.【Result】The results showed that the root mean square error(RMSE)and mean absolute error(MAE)of inflow and sediment concentration predicted by LSTM model were far less than those predicted by traditional models such as autoregressive moving average(ARMA),support vector regression(SVR)and linear regression(LR).【Conclusion】This method can make full use of the time sequence of hydrological information and improve the prediction accuracy.
作者
同套文
TONG Taowen(Shaanxi Jiaokou Pumping Irrigation Administration Bureau,Weinan 714000,China)
出处
《灌溉排水学报》
CSCD
北大核心
2021年第S01期1-4,共4页
Journal of Irrigation and Drainage
关键词
LSTM深度学习
相关性分析
预测
验证
LSTM deep learning
correlation analysis
prediction
verification
