Multi-scale Reconstruction Based Lstm Prediction Methods For Industrial Multivariate Time Series

With continuous developments of industrializations and increasing demands of product qualities,it is in desperate need for monitoring and prediction of key industrial process variables.Due to multi-level dynamics and structural complexities of industrial processes,traditional time series prediction methods reveal limitations to meet the practical need of processes and monitoring tasks.Alternatively,multivariate time series predictions take into account the correlations of multivariate time series,which demonstrates potential applications in industrial process time series predictions.However,in face of non-stationary,strong correlation and disturbance characteristics of industrial process multivariate time series,it is very important to investigate novel methods accordingly.Motivated by these observations,this thesis has achieved the following main research results.1.Aiming at the non-stationarity of industrial process time series data,this study proposes a multi-scale reconstruction method for time series through improving the complete ensemble empirical mode decomposition.Firstly,original time series are decomposed into multiple modal components with complete ensemble empirical mode decompositions.Subsequently,fuzzy entropy-GK clustering methods are employed to decompose,combine and reconstruct multiple modes to obtain three elements in terms of trend,period and randomness,reducing the complexity and enhancing the analyzability and predictability of original time series.2.To deal with strong correlations and disturbances of industrial process multivariate time series,a multivariate time series prediction model with integrated stacked denoising autoencoders is established,formulating the basis of traditional long-term and short-term memory neural networks.This model can also adaptively extract useful information of multiple input variables and remove redundant information,helping improve the prediction performance of multivariate time series.3.The proposed methods are applied to an industrial methanol rectification process,predicting the control performance index of key control loops,helping monitor operating situations of control loops more effectively.Satisfactory results are achieved,which verifiy the effectiveness of the thesis contribution.