基于STL-CEEMDAN-LSTM模型的月径流量预测  

Monthly Runoff Prediction Based on STL-CEEMDAN-LSTM Model

作　　者：汪海 沈延青 祁善胜 潘红忠[1,3] 霍建贞 王战策 WANG Hai;SHEN Yanqing;QI Shansheng;PAN Hongzhong;HUO Jianzhen;WANG Zhance(Key Laboratory of Oil and Gas Geochemistry and Environment,Yangtze University,Wuhan 430000,China;Qinghai Yellow River Upstream Hydropower Development Co.,Ltd.,Xi'ning 810000,China;China Yangtze Power Company Limited,Yichang 443000,China)

机构地区：[1]长江大学油气地球化学与环境湖北省重点实验室,湖北武汉430000 [2]青海黄河上游水电开发有限责任公司,青海西宁810000 [3]中国长江电力股份有限公司,湖北宜昌443000

出　　处：《人民珠江》2025年第4期39-46,共8页Pearl River

基　　金：智慧长江与水电科学湖北省重点实验室开放研究基金项目(242202000923)。

摘　　要：针对月径流序列非线性和非平稳性的特点,尝试将二次分解方法和机器学习相组合构建模型对月径流进行预测。该模型采用周期趋势分解(Seasonal and Trend Decomposition Vsing Loess,STL)将实测月径流序列分解为频率不同的趋势项、季节项和残差项,并再利用自适应噪声的完整集合经验模态分解算法(Complete Ensemble Empirical Mode Decomposition with Adaptive Noise,CEEMDAN)分解残差项得到不同频率成分的模态分量(Intrinsic Mode Function,IMF),最后将趋势项、季节项和各模态分量IMF作为长短时记忆神经网络模型(Long Short-Term Memory Network,LSTM)的输入进行训练、预测。以黄河上游唐乃亥水文站实测月径流数据进行模型验证,并与其他模型进行对比分析。结果表明:STL-CEEMDAN-LSTM预测模型模拟效果较好,模型预测期的NSE(Nash Sutcliffe Efficiency)、RMSE(Root Mean Square Error)、R2分别为0.813、239.02、0.810,其预测精度均优于单一模型和一次分解模型,STL-CEEMDAN二次分解可以有效提高模型预测精度。According to the nonlinear and non-stationary characteristics of monthly runoff sequences,the quadratic decomposition method was combined with machine learning to construct a model for predicting monthly runoff.This model uses a seasonal trend decomposition procedure based on loess(STL)to decompose the measured monthly runoff sequence into trend terms,seasonal terms,and residual terms with different frequencies.The complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)algorithm was then applied to decompose the residual terms to obtain intrinsic mode functions(IMFs)of different frequency components.Finally,the trend term,seasonal term,and each modal component IMF were used as inputs for the long short term memory network(LSTM)for training and prediction.The model was validated with measured monthly runoff data from Tangnaihai hydrological station in the upper reaches of the Yellow River and was compared and analyzed with other models.The results show that the STL-CEEMDAN-LSTM prediction model has a good simulation effect.The Nash Sutcliffe efficiency(NSE),root mean square error(RMSE),and R^(2) in the model prediction period are 0.813,239.02,and 0.810,respectively,with the prediction accuracy better than the single model and the primary decomposition model.The secondary decomposition of STL-CEEMDAN can effectively improve the prediction accuracy of the model.

关 键 词：径流预测 模态分解 长短时记忆神经网络 黄河上游 

分 类 号：TV21[水利工程—水文学及水资源]