基于WPT-ITTA-RELM/ELM/LSSVM模型的日径流预测研究  

Research on Daily Runoff Prediction Based on WPT-ITTA-RELM/ELM/LSSVM Model

作　　者：董欣林 崔东文 DONG Xinlin;CUI Dongwen(Wenshan Prefecture Qingshui River Water Conservancy Hub Engineering Management Bureau,Yunnan Province,Wenshan 663000,China;Yunnan Provincial Wenshan Water Bureau,Wenshan 663000,China)

机构地区：[1]文山州清水河水利枢纽工程管理局,云南文山663000 [2]云南省文山州水务局,云南文山663000

出　　处：《三峡大学学报（自然科学版）》2024年第4期16-24,共9页Journal of China Three Gorges University：Natural Sciences

基　　金：国家高分辨率对地观测系统重大科技专项(89-Y50-G31-9001-22);云南省创新团队建设专项(2018HC024);国家澜湄合作基金项目(2018-1177-02)。

摘　　要：为提高日径流预测精度,验证改进足球战术算法(ITTA)寻优正则化极限学习机(RELM)、极限学习机(ELM)、最小二乘支持向量机(LSSVM)超参数对日径流预测精度的影响,提出小波包分解(WPT)-ITTA-RELM/ELM/LSSVM时间序列预测模型,并通过德厚大型水库入库日径流预测实例进行验证.首先,利用WPT分解处理日径流时序数据,以获得更具规律的子序列分量;其次,通过典型测试函数和RELM/ELM/LSSVM超参数寻优适应度函数对ITTA寻优能力进行检验,并与基本足球战术算法(TTA)、灰狼优化(GWO)算法、倭黑猩猩优化(BO)算法、黏菌算法(SMA)、鲸鱼优化算法(WOA)的优化结果作对比;最后,建立WPT-ITTA-RELM/ELM/LSSVM模型对实例日径流进行预测,并构建WPT-TTA/GWO/BO/SMA/WOA-RELM、WPT-TTA/GWO/BO/SMA/WOA-ELM、WPT-TTA/GWO/BO/SMA/WOA-LSSVM、WPT-RELM/ELM/LSSVM作对比分析模型.结果表明:对于高维和低维优化问题,ITTA寻优精度均优于TTA、GWO、BO、SMA、WOA,表明通过Levy飞行策略及平衡系数等的改进,可有效提高ITTA全局搜索性能和全局、局部平衡能力.WPT-ITTA-RELM、WPT-ITTA-ELM模型对实例日径流预测的平均绝对百分比误差(E_(MAP))分别为0.521%与0.604%,平均绝对误差(E MA)分别为0.024 m^(3)/s与0.025 m^(3)/s,纳什效率系数(E_(NS))均为0.9992,优于其他对比模型;其中WPT-ITTA-ELM模型运行时间较长,不利于大容量样本的预测研究.对于RELM/ELM超参数高维寻优,ITTA优化效果最好,SMA、TTA次之,GWO、BO、WOA优化效果较差;对于LSSVM超参数低维寻优,由于优化维度低、问题简单,ITTA等6种算法均具有较好的优化效果,但ITTA优化效果最好.To improve the accuracy of daily runoff prediction and verify the influence of hyperparameters,such as regularized extreme learning machine(RELM),extreme learning machine(ELM),and least squares support vector machine(LSSVM),optimized by the Improved Football Tactical Algorithm(ITTA)on the accuracy of daily runoff prediction,a wavelet packet decomposition(WPT)-ITTA-RELM/ELM/LSSVM time series prediction model is proposed and verified through an example of daily runoff prediction in Dehou Large Reservoir.Firstly,WPT decomposition was used to process daily runoff time series data to obtain more regular sub sequence components.Secondly,the optimization ability of ITTA was tested using typical test functions and RELM/ELM/LSSVM hyperparameter optimization fitness functions,and compared with the optimization results of basic football tactical algorithms(TTA),grey wolf optimization(GWO),bonobo optimization(BO),slime mold algorithm(SMA),and whale optimization algorithm(WOA).Finally,a WPT-ITTA-RELM/ELM/LSSVM model was established to predict the daily runoff of the instance,and comparative analysis models were constructed,i.e.,WPT-TTA/GWO/BO/SMA/WOA-RELM,WPT-TTA/GWO/BO/SMA/WOA-ELM,WPT-TTA/GWO/BO/SMA/WOA-LSSVM,and WPT-RELM/ELM/LSSVM.The results show that:for high-dimensional and low-dimensional optimization problems,the optimization accuracy of ITTA is better than TTA,GWO,BO,SMA,and WOA,indicating that the improvement of Levy flight strategy and balance coefficient can effectively improve the global search performance and global/local balance ability of ITTA.The E_(MAP) of the WPT-ITTA-RELM and WPT-ITTA-ELM models for predicting daily runoff of the instance are 0.521%and 0.604%,respectively.The E MA is 0.024 m^(3)/s and 0.025 m^(3)/s,respectively.The E_(NS) is 0.9992,which is better than other comparative models.The WPT-ITTA-ELM model has a long running time,which is not conducive to the prediction research of large sample sizes.For high-dimensional optimization of RELM/ELM hyperparameters,ITTA has the best optimization effec

关 键 词：日径流预测 极限学习机 最小二乘支持向量机 改进足球战术算法 小波包变换 超参数优化 

分 类 号：P338[天文地球—水文科学]