Research On Just-in-time Learning Modeling For Nonlinear Systems

Process industry system are characterized by nonlinear,inherent time-varying behavior and uncertainty,a single global modeling method is often unable to satisfy the needs of system modeling and optimization.Local modeling based on the thought of divide and conquer provides an effective strategy to solve this problem.This paper proposes a just-in-time learning modeling algorithm combining information entropy and time characteristics of sample.The specific work is summarized as follows:Similarity criterion determines the accuracy of just-in-time learning modeling,however,the common similarity criterion only considers the similarity between samples,and does not consider the correlation between input variables and output variables,thus affecting the prediction accuracy.Therefore,this paper proposes a just-in-time learning modeling algorithm for improving similarity metrics based on information entropy.It utilizes mutual information to evaluate the correlation between input variables and output variables,and constructs latent modeling spaces,then defines similarity indices in this potential space.A numerical simulation and the application of penicillin fermentation show that the accuracy of the method proposed in this paper is obviously improved compared with the traditional just-in-time learning method.In the process industry,the characteristics of process variables tend to change over time,i.e.,with obvious temporal characteristics.The traditional just-in-time learning method adopts the global search strategy and ignores the time characteristic of the data.Therefore,in this paper,the time information is incorporated into the similarity criterion,and a just-in-time learning modeling algorithm for time-varying system is proposed.Firstly,clustering algorithm is used to divide database samples into corresponding time periods,and the similarity criterion used by clustering is integrated into the temporal information,and then the local modeling samples are searched by using hierarchical search strategy.On the basis of this,the local model updating strategy is proposed,which corrects the output of the model with offset compensation algorithm,and reduces the on-line computation quantity of just-in-time learning.Finally,numerical simulation verifies the validity of the proposed algorithm.