基于BP和LSSVM的径流分频预测模型构建  

Construction of runoff frequency division prediction model based on BP and LSSVM

作　　者：张炳林 李俊[2] 宋松柏[1] ZHANG Binglin;LI Jun;SONG Songbai(College of Water Resources and Architectural Engineering,Northwest A&F University,Yangling,Shaanxi 712100,China;School of Ecology and Environment,Hainan University,Haikou,Hainan 570100,China)

机构地区：[1]西北农林科技大学水利与建筑工程学院,陕西杨凌712100 [2]海南大学生态与环境学院,海南海口570100

出　　处：《干旱地区农业研究》2024年第3期254-263,275,共11页Agricultural Research in the Arid Areas

基　　金：国家自然科学基金项目(52369002);海南省自然科学基金项目(423QN211)。

摘　　要：针对径流序列具有较强的随机性和波动性特征,提出一种短期月径流预测混合模型CEEMDAN-VMD-(BP,LSSVM)-LSSVM。首先利用自适应白噪声完整集成经验模态分解(CEEMDAN,complete ensemble empirical mode decomposition with adaptive noise)将径流序列分解为高频、中频和低频分量,再利用变分模态分解(VMD,variational mode decomposition)方法进一步分解高频分量,并根据样本熵对两次分解得到的子序列进行整合,采用麻雀搜索算法优化的反向传播神经网络(BP,back-propagation neural network)和最小二乘支持向量机(LSSVM,least square support vector machine)分别预测高频分量和中低频分量,最后将不同频率分量训练期的拟合值作为LSSVM的输入,进行二次预测得到最终的径流预测结果。将提出的模型应用于黑河流域莺落峡站和祁连站的月径流预测,验证期相关系数和纳什效率系数均达到0.99以上,对比其他8组对照模型,该模型具有更高的预测精度,可以应用于实际的短期月径流预测。This paper proposes a short-term monthly runoff prediction hybrid model,CEEMDAN-VMD-(BP,LSSVM)-LSSVM,to address the strong randomness and volatility characteristics of runoff sequences.Firstly,a complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN)was used to decompose the runoff sequence into high-frequency,mid-frequency,and low-frequency components.Then,the variational mode decomposition(VMD)method was further applied to decompose the high-frequency component,and the obtained sub-sequences from the two decompositions were integrated based on sample entropy.The back-propagation neural network(BP)optimized by the sparrow search algorithm and the least square support vector machine(LSSVM)were employed to predict the high-frequency and mid-low-frequency components,respectively.Finally,the fitting values of different frequency components during the training period were used as inputs for LSSVM to obtain the final runoff prediction.The proposed model was applied to the monthly runoff prediction at Yingluoxia and Qilian stations in the Heihe River Basin.The correlation coefficients and Nash efficiency coefficients during the verification period are both above 0.99.Compared with other eight models,this model demonstrates better prediction accuracy and can be applied to practical short-term monthly runoff prediction.

关 键 词：径流预测 经验模态分解 变分模态分解 样本熵 神经网络 支持向量机 

分 类 号：S273.29[农业科学—农业水土工程] P333.2[农业科学—农业工程]