摘要
基于序列分解—参数优化—分项预测—结果叠加思想,构建奇异谱分析(SSA)-梯度优化(GBO)算法与相关向量机(RVM)、支持向量机(SVM)集成的中长期月径流预测模型。首先采用SSA方法对实例月径流数据进行处理,提取多个独立的子序列;其次介绍GBO算法原理,基于6个典型函数对GBO算法进行仿真测试。利用GBO算法优化RVM核宽度因子和超参数、SVM惩罚因子和核函数参数,分别建立SSA-GBO-RVM、SSA-GBO-SVM模型对各子序列进行预测,叠加后作为最终月径流预测结果;最后以云南省龙潭站65年共780个月月径流预测为例,选取实例前53年作为训练样本,后10年共120个月作预测样本对SSA-GBO-RVM、SSA-GBO-SVM模型进行检验。结果表明:GBO算法在不同维度条件下寻优效果优于MPA、PSO算法,具有较好的寻优精度和全局搜索能力。SSA-GBO-RVM、SSA-GBO-SVM模型对实例120个月月径流预测的平均绝对百分比误差分别为6.20%、7.82%,平均绝对误差分别为0.88、1.00 m^(3)/s,纳什系数分别为0.9926、0.9913,均具有较好的预测精度和较高的可信度。相对而言,SSA-GBO-RVM模型优于SSA-GBO-SVM。
According to the idea of sequence decomposition-parameter optimization-subitem prediction-result superposition,we construct a medium and long-term monthly runoff prediction model integrating singular spectrum analysis(SSA)-gradient-based optimization(GBO)algorithm with correlation vector machine(RVM)and support vector machine(SVM).To start with,SSA is conducted to process the monthly runoff data of the example and thereby extract multiple independent subsequences.Then,the principle of the GBO algorithm is expounded,and the GBO algorithm is simulated and tested with 6 typical functions.The GBO algorithm is applied to optimize the RVM kernel width factor and hyperparameters as well as the SVM penalty factor and kernel function parameters.SSA-GBO-RVM and SSA-GBO-SVM models are built to predict each subsequence,which is subsequently superimposed to serve as the final monthly runoff prediction result.Finally,the monthly runoff forecast for 65 years(780 months in total)at Longtan Station in Yunnan Province is discussed as an example.The first 53 years are selected as the training samples,and the next 10 years(120 months in total)are taken as the forecast samples to verify the SSA-GBO-RVM and SSA-GBO-SVM models.The results show that the GBO algorithm,with high optimization accuracy and great global search ability,is better than the marine predators algorithm(MPA)and the particle swarm optimization(PSO)algorithm in the optimization effect under different dimensional conditions.The SSA-GBO-RVM and SSA-GBO-SVM models have an average absolute percentage error of 6.20%and 7.82%,respectively,in the 120-month monthly runoff prediction for the example,respectively.The average absolute errors of the two models are 0.88 m^(3)/s and 1.00 m^(3)/s respectively,and the Nash coefficients are 0.9926 and 0.9913 respectively.This means the two models both have high prediction accuracy and reliability.Comparatively speaking,the SSA-GBO-RVM model is better than the SSA-GBO-SVM model.
作者
梁晓鑫
崔东文
LIANG Xiaoxin;CUI Dongwen(Wenshan Branch of Yunnan Hydrology and Water Resources Bureau,Wenshan 661100,China;Yunnan Province Wenshan Water Bureau,Wenshan 663000,China)
出处
《人民珠江》
2022年第1期111-118,共8页
Pearl River
关键词
月径流预测
奇异谱分析
梯度优化算法
相关向量机
支持向量机
仿真测试
monthly runoff forecast
singular spectrum analysis
gradient-based optimization algorithm
relevance vector machine
support vector machine
simulation test
