基于循环神经网络和卡尔曼滤波器的多变量混沌时间序列预测

MULTIVARIATE CHAOTIC TIME SERIES PREDICTION BASED ON RECURRENT NEURAL NETWORK AND KALMAN FILTER

混沌时间序列的鞍点远多于极值点,且容易出现多重共线性问题。对此,提出基于循环深度神经网络和卡尔曼滤波器的时间序列预测算法。利用循环深度神经网络预测高维时间序列,使用实时递归学习算法搜索最小化预测误差的最优网络参数,采用Levenberg-Marquardt算法对神经网络进行迭代训练。在线更新循环神经网络的过程中,利用卡尔曼滤波器对循环神经网络的输出权重进行调节,解决多重共线性的问题。实验结果表明,该算法实现了较低的预测误差,并且缓解了高维时间序列的多重共线性问题。

There are much more saddle points than extreme value points in the chaotic time series,and it is easily to meet the multicollinearity problem.In view of this,we propose a time series prediction algorithm based on recurrent deep neural network and Kalman filter.It used recurrent deep neural network to predict the high dimensional time series,used real time recursion learning algorithm to search the optimal network parameters corresponding to the minimized prediction errors,and adopted Levenberg-Marquardt algorithm to train the neural network iteratively.In the process of online updating recurrent neural network,we took advantage of Kalman filter to adjust the output weights of recurrent neural network to handle the multicollinearity problem.The experimental results show that our algorithm achieves a low prediction error,and reduces the multicollinearity problem of high dimensional time series.