Reasearch And Application Of Industrial Big Data Time Series Forecasting Methods

With the development of computer technology and the applications of advanced control system,modern industrial system equipment and structures become increasingly complex.Industrial processes have accumulated a large amount of historical data,which contain abundant information such as operation rules of the process,operator experience,product quality and problems in the process.How to obtain useful information from the data and provide the basis for the control and decision of production process when it is difficult to establish a precise mechanism model has become a problem that has been widely concerned by researchers.The time series prediction based on industrial big data which refers to the use of some key performance indicators in the process of massive data prediction to provide supports for industrial process monitoring,model identification,fault diagnosis and prediction,has important theoretical significance and application values.With the continuous improvement of the scale and refinement of the industrial processes,a large amount of data collected have complex characteristics,including incompleteness,nonlinearity,mutual correlation and multi-modal characteristics.These complex characteristics of industrial big data challenge existing data-driven prediction models.In order to solve these problems,this paper studies the prediction of industrial time series with the above complex characteristics under the framework of data driven.The research works of this thesis mainly includes:(1)A prediction model of neural network construction algorithm based on feature extraction after sequence decomposition is proposed for univariate time series forecasting in industrial big data.After the most relevant input is constructed by the feature extraction algorithm,a neural network model based on the construction algorithm is established.Because the change of network topology will cause the problem of network performance oscillation,the algorithm of neural network construction is further improved.The simplification of model input and the reduction of model structure complexity make it possible to build a concise model and improve network performance while ensuring the prediction accuracy.The experimental results show that the proposed model can effectively improve the prediction performance when considering the accuracy and efficiency of the prediction,and is superior to other comparison models.(2)A neural network model based on feature selection is proposed for multivariate time series prediction.In order to avoid directly estimating the probability density of mutual information and achieve high dimensional mutual information calculation,k-nearest neighbor mutual information is used for feature selection of input variables.Since the selection of k value has a great impact on the mutual information results,this paper proposes a method to determine the neighbor number k based on the permutation test and cross validation,so as to achieve the improvement of the k-neighbor mutual information method,so as to improve the performance of the mutual information used in the correlation analysis.Experimental results show that the neural network model based on the proposed feature selection method can capture useful information in multivariable data,thus providing effective input for the model and improving the performance of the model.(3)A multi-model time series prediction method based on Gaussian process regression is proposed based on multi-stage statistical modeling strategy to solve various modal problems in production process due to changes in production schemes or product types.In this method,different modes in the data are identified first,and Gaussian mixture model is used to classify the modes of the data,and the probability of sample belonging to a certain mode is determined.In the parameter learning process,the online expectation maximization algorithm is adopted to derive the learning process,and the cluster model parameters are updated online.Then the local GPR model is established.The selection of prediction model parameters has great influence on the accuracy of the model,the particle swarm optimization algorithm is introduced to replace the traditional stochastic gradient descent method to optimize the regression covariance matrix parameters in the Gaussian process.Finally,the results of local GPR models are combined with Bayesian method.Experimental results show that the model performance can be effectively improved by partitioning the data mode and optimizing the model parameters.In order to verify that the experiment is significantly better than the comparison model,the non-parametric hypothesis testing method is used to test the experimental results statistically.(4)Aiming at the problem of time series prediction with missing values in industrial,a modeling method based on multiple imputation method and GPR model is proposed.The parameters of the mixed Gaussian model are derived for the incomplete dataset under the expected maximum learning framework,and the data distribution of the dataset with different miss rates is described.The GPR model is established for each new data set.By using the combined model idea,the prediction results of multiple models are weighted average,so that the prediction results of each sub-model are reasonably integrated.In order to verify the effectiveness of the proposed method,the model is used to predict the wind power of a wind farm.Experimental results show that the proposed method can effectively resolve the uncertainty and variability caused by missing data in the application of processing missing data.Compared with other models,the proposed model can effectively deal with the problem of incomplete time series prediction.