High Accuracy Identification Methods For Industrial MPC Systems

Dynamic system modeling plays the central role in industrial model predictive control systems.How to reduce modeling cost and/or increase modeling accuracy is the key issue that determines the control performance.System identification is an efficient way to obtain dynamic system models.This work takes the industrial model predictive control systems as background,and studies several problems that concern the identification accuracy in industrial systems.The main contributions are summarized as follows.1.The asymptotic theory of system identification under the over-sampling scheme is deduced,the mechanism of how over-sampling scheme achieves closed-loop identification without external excitation is found.The over-sampling scheme is an identification procedure that can identify systems in closed-loop tests without external excitation,which means it can reduce the cost of identification to zero.However,previous studies have not found the physical mechanism of the over-sampling scheme,nor can it explain why sometimes the accuracy of closed-loop identification without external excitation is not ideal.In this work,the over-sampling scheme is analyzed using the asymptotic theory of system identification,and the asymptotic theory under the over-sampling scheme is deduced.The asymptotic the-ory gives the asymptotic variance expression of the over-sampling scheme in frequency do-main.Based on the variance expression,this work finds the mechanism of the over-sampling scheme to achieve closed-loop identification without external excitation,that is,turns the high-frequency term in the output noise into identification excitation.2.High accuracy system identification method based on over-sampling scheme is proposed.The new method can improve the identification accuracy from two aspects:a)achieve anti-aliasing on the output noise,which means avoid the influence of high-frequency noise on the identification accuracy.This effect exists both in open-loop identification and closed-loop identification;b)in closed-loop identification,the high-frequency output noise can be turned into identification excitation to improve the identification accuracy.This work illustrates the performance of the new method by numerical simulations.3.A comprehensive theoretical analysis of anti-aliasing filtering scheme in system identifica-tion is conducted.The anti-aliasing filtering scheme is a method to avoid the effect of spectral aliasing on the output noise when sampling,but the correctness of this scheme is lack of theo-retical analysis,and has not been verified in applications.This work analyzes the anti-aliasing filtering scheme from two perspectives:time domain and frequency domain,and proves that it is impossible to obtain a consistent estimation of the real system using this scheme.The influence of using this scheme is also obtained:1.Increase the order of the system model;2.When the information of the real system is negligible on high frequency band,the frequency response of the identified model 'is an approximate estimate of the frequency response of the real system.These results are illustrated by numerical simulations.4.A globally convergent recursive identification method for parametric models is proposed.The new method can ensure the identification efficiency,i.e.minimum variance property,and global convergence at the same time,which makes it the most accurate recursive iden-tification method among all unbiased methods.This is achieved by recursive least-squares identification of high-order nonparametric model and weighted least-squares model reduc-tion.For the online selection of non-parametric model orders in the new method,this work also proposes a non-parametric model order adaptive adjustment scheme.Then the consis-tency and asymptotic efficiency of the new recursive method are theoretically proven.Fi-nally,the simulation results verify the performance of the new method and the order adaptive adjustment scheme.