Research On Multivariable Time Series Model Identification Method

The theory of multivariable model identification has been the basis and key in the field of time series prediction and industrial control.Therefore,the research on multivariable model identification method was carried out in this paper has very important practical significance.The main contents of this paper are as follows:Firstly,the process of multivariate model identification is analyzed.In terms of data characteristics,the stability,nonlinearity,chaos and correlation of time series are tested and analyzed.In terms of model identification,the linear regression model,the nonlinear time series model and the MIMO linear system are mainly analyzed from the aspects of applicability,In the aspect of parameter identification,the least squares and its evolutionary algorithms are generally identified from the aspects of model accuracy,convergence property,storage capacity and calculation time.In order to solve the problem of multi-step prediction for high dimensional chaotic complex systems,we proposed a multivariate local multi-step prediction model based on cluster analysis of adjacent phase points to identify the nearest neighbor point in local area and improve prediction efficiency.Finally,the simulation results of Lorenz chaotic data showed that this model has good prediction performance.For the identification of multivariable linear state space model,the state space sub-model is mainly identified and analyzed.When the state is known,Multivariate least squares(MRLS)can be used to identify the model parameters.When the state is not known,a sub-model identification algorithm based on multivariate least squares and Kalman filter state estimation is proposed(S-KF-MRLS).The simulation results show that the S-KF-MRLS algorithm can get higher recognition efficiency than the KF-MRLS algorithm and greatly reduce the computational complexity.For the identification of multivariable nonlinear state-space model,the recurrent neural network model under nonlinear state-space is mainly identified and analyzed.Aiming at the problem of inexact solution,a nonlinear neural network model based on the differential technique is proposed.The Taylor series expansion and automatic differential(AD)techniques are used to solve ordinary differential equations and dynamic sensitivity of the nonlinear model.Experimental results show that compared with the traditional linear state space model,the proposed model has higher predictive control precision and superior control performance.