基于PSO的最小二乘支持向量机稀疏化算法

Optimal sparseness approach for least square support vector machine based on PSO

针对最小二乘支持向量机(LSSVM)失去稀疏特性及经典迭代剪切稀疏化算法容易陷入性能指标函数局部收敛的问题,提出一种基于粒子群优化(PSO)的LSSVM稀疏化算法.将LSSVM稀疏化过程描述为一个最优化问题,以校验样本和预测输出之间的均方根误差RMSE为优化目标,以模型训练样本剪切率ε(%)为优化变量.并针对此非线性优化问题提出基于PSO的求解方法.以大型电厂飞灰含碳量LSSVM模型为例,对此算法进行了实例研究.结果表明,该方法能有效解决经典算法的局部收敛问题获得最优剪切率,具有更好的预测和泛化能力.

In term of the lack of sparseness of the least square support vector machine （LSSVM）, some classic pruning methods are successively proposed. However, the local convergence characteristic of the performance index may hinder the classical methods to get an optimal pruning rate of the training data set. This paper proposes a particle swarm optimization （PSO） based optimal sparseness approach for LSSVM models. The new approach firstly formulates the sparseness of LSSVM as a general optimization problem, where root-mean-square error （RMSE） between predicted values and real values is taken as an objective function for minimization; and then the pruning rate is employed as optimization variable, PSO is further proposed to solve this nonlinear optimization problem. A LSSVM model of carbon content in fly ash is taken as a case study. The operation data of a large-scale coal-fired power plant is collected for comparative investigation. The results show that the newly proposed approach has the ability to conquer the local conver- gence problem; and consequently an optimal pruning rate of the training data set is obtained. Compared with the classic methods, the new approach is better both in prediction performance and generalization performance.