多元混沌时间序列的因子回声状态网络预测模型

Factor Echo State Network for Multivariate Chaotic Time Series Prediction

针对采用回声状态网络预测多元混沌时间序列时存在的病态解问题,本文建立了因子回声状态网络模型,通过因子分析(Factor analysis,FA)方法提取高维储备池状态矩阵的公因子,去除冗余和噪声成分.利用降维后的因子变量与期望输出之间的线性回归关系,求解网络未知参数.基于Lorenz序列和大连月平均气温–降雨量的仿真实验验证了本文所提模型的有效性.

When an echo state network is used to predict mul-tivariate time series, there may exist ill-posed problem. This pa-per proposes a novel prediction model, named factor echo state network, to solve the problem. It uses a factor analysis （FA） al-gorithm to extract the common factors of the reservoir matrix, and to remove the redundancies and noises. Then the unknown output weights are calculated by linear regression of the output and common factors. The model is used to predict Lorenz series and monthly average temperature-rainfall time series in Dalian, and simulation results substantiate its effectiveness.