基于改进优化算法的WELM月径流预测模型研究  

Study on weighted extreme learning machine model with improved optimal algorithms for monthly runoff prediction

作　　者：王应武 华春莉 茶建帮 WANG Yingwu;HUA Chunli;CHA Jianbang(Yunnan Institute of Water&Hydropower Engineering Investigation,Design and Research,Kunming 650021,China;Water Bureau of Linxiang District,Lincang City,Yunnan Province,Lincang 677000,China)

机构地区：[1]云南省水利水电勘测设计研究院,云南昆明650021 [2]云南省临沧市临翔区水务局,云南临沧677000

出　　处：《人民长江》2025年第2期82-90,共9页Yangtze River

基　　金：国家自然科学基金项目(41702278);中国地质调查局地质调查项目(DD20221758,DD20190326);国家重点研发计划项目(2019YFC0507500)。

摘　　要：针对在月径流预测中将传统数据分解技术直接应用于整个时间序列时,在模型训练过程中会提前使用“未来信息”从而导致预测结果“不可信”的问题,提出两种不引入“未来信息”的小波包变换(WPT)-改进蝴蝶优化算法(IBOA)/改进海马优化(ISHO)算法-加权极限学习机(WELM)月径流时间序列预测模型。首先,将月径流时间序列划分为训练集和预测集,利用WPT分别对训练集和预测集进行分解处理,避免在训练过程中提前使用“预测集信息”;其次,通过6个典型函数验证IBOA/ISHO的寻优能力,利用IBOA/ISHO优化WELM输入层权值和隐含层偏差(简称“超参数”),建立WPT-IBOA/ISHO-WELM模型对各分解分量进行预测和重构;同时构建基于整个时间序列分解的WPT-IBOA/ISHO-WELM(全)模型,与其他4种优化算法和未经分解、未经优化的IBOA/ISHO-WELM、WPT-WELM模型作对比分析;最后,通过云南省李仙江流域把边、景东水文站月径流时间序列预测实例对各模型进行检验。结果表明:①WPT-IBOA-WELM、WPT-ISHO-WELM模型对把边、景东站月径流预测的平均绝对百分比误差(MAPE)为1.649%~1.897%,预测精度优于其他对比模型,具有更好的预测效果。②WPT-IBOA-WELM、WPT-ISHO-WELM模型的预测精度基本不受“未来信息”的影响,能客观真实反映出月径流预测效果,具有较好的实用意义。③IBOA/ISHO仿真精度和WELM超参数优化效果均优于其他优化算法,表明通过logistic映射等策略可以显著提升IBOA/ISHO优化性能。The traditional data decomposition techniques are directly applied to the entire time series in monthly runoff prediction,in which"future information"is used in advance during model training,this would lead unreliable results.Aiming at this problem,two types of monthly runoff time series prediction models without"future information"are proposed,it is wavelet packet transform(WPT)-Improved Butterfly Optimization Algorithm(IBOA),and Improve Sea-horse Optimization Algorithm(ISHO)-Weighted Extreme Learning Machine(WELM).Firstly,we divide the monthly runoff time series into a training set and a prediction set,and use WPT to decompose the training set and prediction set separately,avoiding the use of"prediction set information"in advance during the training process.Secondly,we validate the optimization capability of IBOA/ISHO through six typical functions;optimize the input layer weights and hidden layer biases(or"hyper-parameters"for short)of WELM using IBOA/ISHO,establish WPT-IBOA/ISHO-WELM model to predict and reconstruct various decomposition components;construct a WPT-IBOA/ISHO-WELM(full)model based on the decomposition of the entire time series,and compare with four other optimization algorithms,the undifferentiated and unoptimized IBOA/ISHO-WELM and WPT-WELM models.Finally,the monthly runoff time series prediction examples of the Babian and Jingdong hydrological stations in the Lixianjiang River Basin of Yunnan Province were used to test each model.The results show that:①The average absolute percentage error(MAPE)of the WPT-IBOA-WELM and WPT-ISHO-WELM models for predicting monthly runoff at the Babian and Jingdong stations range from 1.649%to 1.897%,with better prediction accuracy than other comparative models and better prediction performance.②The prediction accuracy of WPT-IBOA-WELM and WPT-ISHO-WELM models are basically not affected by"future information",and can objectively and truly reflect the monthly runoff prediction effect,which has good practical significance.③The simulation accuracy of IBOA/ISHO

关 键 词：月径流预测 小波包变换 改进蝴蝶优化算法 改进海马优化算法 加权极限学习机 超参数优化 把边水文站 景东水文站 李仙江流域 

分 类 号：P338[天文地球—水文科学]