论文部分内容阅读
为克服最小二乘支持向量机(LSSVM)依赖人为经验选择学习参数,以及标准遗传算法(GA)存在早熟收敛等不足,提出自适应随机遗传优化算法(ARGA),利用ARGA算法优化选择LSSVM惩罚因子C和核函数参数σ~2,构建ARGA-LSSVM年径流预测模型,并与自适应遗传算法(AGA)、标准遗传算法(GA)优化选择LSSVM学习参数的AGA-LSSVM、GA-LSSVM模型以及GA-BP、RBF模型进行对比分析。以云南省龙潭站年径流预测为例,利用龙潭站前34a和后20a资料分别对各模型进行训练和预测。结果表明,ARGALSSVM模型对实例后20a年径流预测的平均相对误差绝对值和最大相对误差绝对值分别为2.42%、5.85%,预测精度优于AGA-LSSVM、GA-LSSVM模型,大幅优于GA-BP和RBF模型。ARGA算法全局寻优能力强、收敛速度快,利用ARGA算法优化得到的LSSVM学习参数可有效提高LSSVM模型的预测精度和泛化能力。
In order to overcome the shortcomings of Least Squares Support Vector Machines (LSSVM), which depend on man-made experience to learn learning parameters and the premature convergence of standard genetic algorithm (GA), an adaptive random genetic algorithm (ARGA) is proposed to optimize the LSSVM penalty factor (AGA) and standard genetic algorithm (GA), the AGA-LSSVM and GA-LSSVM models of LSSVM learning parameters and GA -BP, RBF model for comparative analysis. Taking the annual runoff forecast of Longtan station in Yunnan Province as an example, the models were trained and predicted by the data of 34a and 20a after Longtan station. The results show that the average absolute relative error and the maximum relative error of ARGALSSVM model are 2.42% and 5.85%, respectively, and the prediction accuracy is better than that of AGA-LSSVM and GA-LSSVM models in the last 20 years. BP and RBF models. ARGA algorithm has strong global search ability and fast convergence speed. The LSSVM learning parameters optimized by ARGA algorithm can effectively improve the prediction accuracy and generalization ability of LSSVM model.