改进粒子群算法优化下的Lasso-Lssvm预测模型

Lasso-Lssvm Prediction Model Optimized by Improved Particle Swarm Algorithm

文章首先针对最小二乘支持向量机(Lssvm)对高维输入特征较为敏感的问题,提出以Lasso算法筛选输入特征,建立Lasso-Lssvm预测模型,并通过粒子群算法对Lssvm中核函数的未知参数进行优化。然后,针对粒子群算法容易陷入局部最优和后期收敛较慢等问题,提出了改进的粒子群算法(IPSO):基于网格划分方法,完成粒子初始化;在设定粒子速度更新的惯性权重时,基于Sigmod函数提出种群对比自适应动态惯性系数;针对粒子所处位置的优劣,动态变化学习因子。最后,基于1985—2018年能源排放相关数据建立改进的粒子群算法优化下的Lasso-Lssvm模型。结果表明,Lasso方法可以有效解决Lssvm对高维输入特征敏感的问题,所提出的粒子群算法有更好的寻优能力和鲁棒性,改进粒子群算法优化下的Lasso-Lssvm模型拥有更好的拟合效果和预测精度,验证了该方法的适用性和优越性。

In order to solve the problem that Lssvm is sensitive to the high-dimensional input characteristics, this paper proposes the Lasso algorithm to screen the input characteristics, establishes the Lasso-Lssvm prediction model, and optimizes the unknown parameters of the kernel function in Lssvm by particle swarm algorithm. Then, the paper proposes an Improved Particle Swarm Optimization(IPSO) algorithm to solve the problem of particle swarm optimization(PSO) easily falling into local optimum and of slow convergence in late stage: based on the meshing method, particle initialization is completed;when the inertia weight of particle velocity update is set, the adaptive dynamic inertia coefficient of population comparison is proposed based on Sigmod function;the learning factor is changed dynamically according to the position of the particle. Finally, based on the data of energy emissions from 1985 to 2018, the paper establishes the Lasso-Lssvm model optimized by the IPSO algorithm. The results show that the Lasso method can effectively solve the problem that Lssvm is sensitive to high-dimensional input characteristics, that the proposed PSO algorithm has better optimization ability and robustness, and that the Lasso-Lssvm model optimized by the IPSO algorithm has better fitting effect and prediction accuracy, which reflects the applicability and superiority of this method.