Predicting Key Parameters Of Dynamic Processes Based On The Spatio-temporal Bidirectional De-redundancy

A large amount of operation data in the industry process contains rich and useful information reflecting the process characteristics.It is an extremely effective way to predict key parameters based on these data.The long production process and large reaction equipment,making the data be distributed in the different space,and there is the information redundancy among the data at different spatial points.At the same time,due to the fluctuation of production conditions,the time series of data should be used to describe the dynamic characteristics,and there is the redundancy among the data at different time points.Therefore,retaining the spatiotemporal two-dimensional structure of input data,this paper proposes to study the bidirectional de-redundancy based on the two-dimensional principal component analysis to realize the dynamic feature extraction,and applies it to the prediction of key process parameters.The main contents and innovations are as follows:(1)Aiming at the problem that input variables from different time and space points have the different importance on the key parameter,a weighted two-dimensional principal component analysis is proposed.In the dynamic feature extraction,it preserves the two-dimensional structure of input data,and analyzes the correlation between the key parameter and the input variables to obtain the weights,which can be used to measure the importance;then uses the weighted two-dimensional input to solve the projection matrix to achieve the dynamic feature extraction,increasing the correlation between the dynamic features and the key parameter,and improving the prediction accuracy.(2)Aiming at the multi-step prediction of the key parameter in the future,a two-dimensional probabilistic principal component analysis is proposed to extract the dynamic feature with better characterization ability to the future data.In this method,each training sample includes not only the input time series composed of the current data and the past data,but also the future time series composed of the future data;then,the probability relationship among the input time series,the dynamic features,and the future time series is established.Therefore,the dynamic features extracted from the input time series through the model have better representation ability for the future data,improving the accuracy of multi-step prediction of key parameter.(3)Finally,aiming at the problem of the spatiotemporal bidirectional redundancy in a two-dimensional input,the idea of the dynamic feature extraction by bidirectional de-redundancy is proposed.In this idea,the methods proposed above are used to calculate the projection matrix for the feature extraction from the time and the space dimension,then the input is projected in the two directions to get the features with bidirectional redundancy removed.The proposed idea is applied to the debutanizer process and the leaching process of zinc smelting,respectively.Simulation results show that the proposed method greatly reduces the dimension of dynamic features and improves the prediction accuracy of the key parameters.