Complex Petrochemical Processes Index Data Analysis And Prediction Based On Time-series Data

Complex chemical manufacturing process has the characteristics of multi noise and complex relationship between variables.The traditional prediction method based on chemical reaction mechanism is often difficult to obtain the expected results that meet the requirements.The data-driven model collects the timing data of production devices in the actual production process and uses the method of deep learning to establish the prediction model.Compared with the mechanism model,the data-driven model has better prediction ability and flexibility.In this paper,polypropylene production equipment is used as the data source,and the feature learning ability of artificial neural network is used to mine the data change trend and other information from the time series data,so as to accurately predict the key index data in the chemical production process,and realize the research and application of the analysis and prediction method of time series data in the chemical process.The main research contents of this paper include:(1)In the process of chemical production,noise and other factors will lead to the abnormality and fluctuation of time series data,which cannot be directly applied to data modeling.To solve this problem,an anomaly detection and processing algorithm based on unsupervised learning is proposed.The improved tree model algorithm is used to find the abnormal data,which is processed by the automatic encoder for subsequent association analysis,modeling and prediction.(2)For the time series data generated in actual production has the characteristics of large quantity and fluctuation,a time series correlation analysis algorithm is proposed to calculate the fluctuation correlation of time series by the entropy characteristics of time series and segmentation characteristics,which is applied to the subsequent modeling and prediction work.Experimentally,it is proved that on polypropylene data set compared with the traditional algorithm this algorithm F1 index scores higher accuracy.(3)To address the problems of poor flexibility and low accuracy of traditional mechanistic models,based on the timing data after anomaly detection and its correlation analysis results,this paper proposes a data-driven timing data prediction model(MAtt-GRU)based on multi-headed attention mechanism and gated recurrent unit(GRU).Through validation experiments with actual production data sets of polypropylenes,the model can predict the trends and values of the timing data under single-step and multi-step prediction scenarios accurately.