基于18种核函数映射的孪生回归支持向量机月径流预测OA

The reasonable selection of kernel functions and kernel function parameters is of great significance for improving the performance of Twin Support Vector Regression(TWSVR).To improve the prediction accuracy of monthly runoff time series and compare and verify the effectiveness of TWSVR with different kernel function mappings,Wavelet Packet Transform(WPT),18 kernel functions(such as linear kernel function),Superb Fairy-wren Optimization Algorithm(SFOA),and TWSVR were used to propose the WPT-SFOA-TWSVR model with 18 kernel function mappings.Five common kernel function mappings of WPT-SFOA-SVR models were constructed for comparative analysis.A total of 23 models were validated through monthly runoff prediction examples at the Dishui,Nandong,Mengda,and Nankanghe hydrological stations in Yunnan Province.Firstly,WPT is used to decompose and process the monthly runoff time-series data of the instance,dividing it into a training set and a validation set.Then,SFOA is applied to optimize the TWSVR/SVR hyperparameters of different kernel function mappings.Finally,using the optimal hyperparameters,a WPT-SFOA-TWSVR/SVR model with different kernel function mappings was established to train,predict,and reconstruct each component of monthly runoff for the four instances.The results show that:① The WPT-SFOA-TWSVR model based on linear kernel function,Gaussian kernel function,polynomial kernel function,wavelet kernel function,Sigmoid kernel function,and neural kernel function mapping has the smallest prediction error and the best performance.The WPT-SFOA-TWSVR model based on ANOVA kernel function,Bessel kernel function,logarithmic kernel function,multiple quadratic kernel function,and power-law kernel function mapping follows closely.The WPT-SFOA-TWSVR model based on T-Student kernel function,Cauchy kernel function,and rational quadratic kernel function mapping has relatively larger prediction errors;The WPT-SFOA-TWSVR model based on Laplace kernel function,Fourier kernel function,chi square kernel function,and spherical kernel function mapping has the largest prediction error.② Under the same WPT decomposition and SFOA optimization conditions,the TWSVR model performs significantly better than SVR.③ Optimizing TWSVR hyperparameters using SFOA can significantly improve model performance and computational efficiency.④ The WPT-SFOA-TWSVR model with different kernel function mappings has good universality,providing reference and inspiration for the selection and optimization of TWSVR kernel functions.