基于SCAD-ESN的时间序列预测模型

发布时间：2018-05-14 11:14

本文选题：混沌时间序列预测 + 回声状态网络　；参考：《工程科学与技术》2017年06期

【摘要】：回声状态网络(ESN)是一种重要的时间序列预测方法,但在训练数据存在噪声或野点情况下,ESN将会出现过拟合问题。针对该问题,提出基于平滑消边绝对偏离罚函数的回声状态网络(SCAD-ESN)模型。不同于在模型中加入岭回归、L1范数罚函数及小波降噪等常规方法,该模型利用SCAD罚函数对变量进行选择,将小变量置为零以满足变量稀疏性,将大变量直接置为常数,从而能够很好地解决ESN过拟合问题并满足近似无偏估计。对于SCAD罚函数的非凸函数优化问题,提出基于局部二次近似(LQA)的求解方法,将最小角回归(LQR)方法用于SCAD罚函数求解,避免了计算量巨大的问题。使用基于粒子群优化(PSO)的超参数选取方法快速确定平滑消边绝对偏离回声状态网络模型的超参数,克服利用经验选取超参数时存在的盲目性较大且难以确定整体最优的超参数问题。混沌系统数值仿真和网络流量仿真结果表明,相对于常规模型,该模型能有效地降低测试误差,从而克服过拟合问题。
[Abstract]:Echo state network (ESNN) is an important time series prediction method, but if there is noise or outliers in the training data, the ESNN will be overfitted. To solve this problem, a SCAD-ESN-based echo state network model based on smooth edge-elimination and absolute deviation penalty function is proposed. Different from the conventional methods such as ridge regression L1 norm penalty function and wavelet denoising, the model uses SCAD penalty function to select variables, sets small variables to zero to satisfy the sparsity of variables, and places large variables directly as constants. Thus, the problem of ESN overfitting can be solved well and the approximate unbiased estimation can be satisfied. For the non-convex function optimization problem of SCAD penalty function, a method based on local quadratic approximation (LQA) is proposed. The minimum angle regression method is used to solve the SCAD penalty function, which avoids the problem of huge computation. Based on particle swarm optimization (PSO), the super-parameter selection method is used to quickly determine the superparameters of the smooth edge-free absolute deviation from the echo state network model. In order to overcome the problem of blind and difficult to determine the global optimal hyperparameter when using experience to select superparameters. The simulation results of chaotic system and network traffic show that compared with the conventional model, the proposed model can effectively reduce the test error and overcome the problem of over-fitting.
【作者单位】：河南科技大学网络与通信技术研究所;河南科技大学网络信息中心;河南科技大学数学与统计学院;
【基金】：国家自然科学基金资助项目(11501067) 赛尔网络下一代互联网技术创新项目资助(NGII20150508)
【分类号】：TP18

【相似文献】