Complex Industrial Data-Driven Process Variables Prediction

Accurate prediction of production process variables can provide valuable guidance for the scheduling of industrial systems,and it is an important component in predictive control technology.At present,data-based prediction methods are widely employed in process variables forecasting.However,due to the high noise and the missing points in industrial data,it is difficult to build the complex nonlinear relationship among process variables accurately,making a great chellange in ensuring the accuracy of modeling.Therefore,this paper studies complex industrial data-driven process variables prediction methods.The specific contents are as follows.An embedded feature selection model based on variational inference with local linearization is proposed for industrial process variable prediction.In the relevance vector machines framework,as for the inference of the variational posterior distribution over the kernel parameters,a local linearization method is designed to approximate this posterior distribution as a Gaussian.Meanwhile,the posterior distributions of all other parameters in the model is derived,rather than their point estimate.Considering that the input samples are usually corrupted by high level noise,a relevant vector machines model with input noise is proposed.Due to the arrival of the input noise in a relevant vector machine,the exact solution for the marginal likelihood function is intractable.Thus,this study derives a Gaussian approximation for the marginal likelihood function based on global expectation and global variance criterion.The Markov chain Monte Carlo method is used to approximate the posterior distribution over the model weights.Considering that there are often missing points in the original time series,a relevant vector machine model is proposed for incomplete training dataset.The proposed model directly uses the data with missing points for model training without data imputation.To estimate the missing outputs,two approaches are employed,including the expectation maximization based algorithm and the marginal likelihood maximization method.The former can give a posterior distribution over the missing output points,while the latter one can only provide its point estimate.Aiming at constructing prediction intervals with incomplete testing sample in real time,a higher-order dynamic Bayesian network model based on relevance vector machines is proposed.The sparse Bayesian learning is employed to learn the parameters of continuous nodes.In the reasoning stage,for the missing points in the input samples,the likelihood weighting stochastic approximation algorithm is used to estimate the probability distribution of future nodes,based on which the prediction interval is constructed.In addition,considering that it is too time-consuming to perform the sampling steps in the likelihood weighting algorithm,a local linearization based variational inference algorithm is designed for dynamic Bayesian network reasoning.To verify the effectiveness of the proposed methods,the data of the by-product gas system of a steel enterprise in China and the data of a grinding plant in China are employed,as well as some artificial data sets and benchmark datasets.The experimental results show that better prediction results are obtained by the proposed embedded feature selection model in the industrial application.The proposed methods are also good at dealing with the situation of missing points in the data and the noise in the input samples,and can achieve better prediction results than the comparative methods.