Data-driven Multiscale Modelling With Application To Fluid Catalytic Cracking Process

With rapid development of smart manufacturing,which has been bringing massive process data to factories,it gradually becomes a prevailing trend to model chemical process via data-driven method,which could tackle problems caused by nonlinear and multiscale issues during modeling period.Based on the multiscale structure decomposed by HHT method,one fluid catalytic cracking unit(FCCU)was investigated in present work in order to developing a suite of scientific analyzing method and modeling tools for timeseries model,feature selection model and multi-input single-output model,which could be used in terms of site monitoring,key variables screening and yield predicting respectively.Fast and accurate prediction of timeseries is required to enable the monitoring of key sites.To illustrate,the prediction of timeseries was carried out on two timeseries,the pressure and temperature at the top of disengager of the FCCU.The results give the information about that there was the coexistence of mesoscales and microscales,chaos and stability,together with the driving and coupling relation between scales.The inferences made by predecessors about the multi-scale nature in FCCU could be verified from an industrial data mining point of view.Comparing the single-scale model and multi-scale model,based on LSTM model which could representing the time memory effect,the prediction errors of pressure and temperature series reduced 29.8% and 32.8% severally with nearly no extra time expenditure,by which I mean that only by recognizing the multiscale structure of process,the accurate and high efficiency data-driven model could be developed.A Filter-Wrapper mixed feature selection model was developed for the sake of finding the key features for target variables rapidly.Important monitoring variables showing the operating status,such as pressure and temperature at the top of disengager,and indicators for the productivity,such as gasoline yield,were choosen as target variables.By recognizing the higher relevance between candidate features and target variables,irrelevant and weak relevant features,accounting for circa 90%~98%,were removed.This simplifying procedure led to an explicit and clear understanding with FCC process,which would not only reduce the complexity of modeling,but also dig out the latent related nature in a practical process by the knowledge discovery form the data itself.As for gasoline yield prediction,predictive machine learning model was developed based on multiscale decomposition method and feature selection method.There was a comparison between three types of input variables,containing original scale input,selected original scale input and multiscale input,through linear model as well as nonlinear model.By that indicated,the performance under nonlinear model was all higher than the linear counterpart,as well as the best prediction results showed over multiscale model,with 15% estimating errors was decreased.Conversely,the selected original scale model was the worst one in this case,unveiling the fact that the reason for improved performance was due to multiscale decomposition rather than feature selection,which emphasized that present model has the ability to capture the multiscale nature in FCC process.