CN-122020049-A - Characteristic decoupling self-coding polypropylene production prediction model based on residual error network

CN122020049ACN 122020049 ACN122020049 ACN 122020049ACN-122020049-A

Abstract

The application relates to a polypropylene production prediction model based on characteristic decoupling self-coding of a residual network, which comprises the steps of obtaining original data of polypropylene production, extracting trend characteristics and period characteristics of the original data through a trend period long-short-term memory neural network, extracting spatial characteristics of the original data through a dynamic self-attention convolution neural network, splicing the trend characteristics, the period characteristics and the spatial characteristics by utilizing characteristic decoupling self-coding to obtain decoupled multi-source characteristics, establishing a relation between the decoupled multi-source characteristics and output by using the residual network, and constructing the polypropylene production model according to the relation between the decoupled multi-source characteristics and the output. Therefore, the problems that in the related technology, the data-driven modeling method assumes that the production process is under a stable working condition, and the polypropylene production process is a non-static and nonlinear complex process, so that the model is difficult to adapt to complex dynamic changes, the generalization performance of the model is reduced and the like are solved.

Inventors

HAN YONGMING
LIU MIN
GENG ZHIQIANG
WU HAO
WANG MENGZHI
HU XUAN

Assignees

北京化工大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (10)

1. A residual network-based feature decoupling self-coding polypropylene production prediction model, comprising the steps of: Obtaining original data produced by polypropylene, extracting trend characteristics and period characteristics of the original data through a trend period long-term and short-term memory neural network, and extracting spatial characteristics of the original data through a dynamic self-attention convolutional neural network; Splicing the trend feature, the periodic feature and the space feature by utilizing feature decoupling self-coding to obtain a decoupled multi-source feature; And establishing a relation between the decoupled multi-source characteristic and the output by using a residual network, and constructing a polypropylene production model according to the relation between the decoupled multi-source characteristic and the output.
2. The residual network-based feature decoupling self-encoding polypropylene production prediction model of claim 1, wherein the extracting trend features and cycle features of the raw data by the trend cycle long-short term memory neural network comprises: decomposing the original data into a trend signal and a periodic signal by utilizing a variation modal decomposition strategy; and extracting features of the trend signal and the periodic signal based on the trend period long-term and short-term memory neural network so as to obtain the trend features and the periodic features.
3. The residual network-based feature decoupling self-encoding polypropylene production prediction model of claim 1, wherein the expression of the spatial features of the raw data is: , , , , Wherein, the In order to align the paths of the optical device, For an optimal calculation of the distance, Is vector quantity Sum vector The degree of similarity between the two, Is the attention weight.
4. The residual network-based feature decoupling self-encoding polypropylene production prediction model of claim 1, wherein the stitching the trend feature, the periodic feature, and the spatial feature with feature decoupling self-encoding to obtain a decoupled multi-source feature comprises: Learning the decoupled multi-source feature representation using a self-encoding training framework; Taking the trend periodic long-term memory neural network and the dynamic self-attention convolutional neural network as self-coding encoders, and taking a multi-layer perceptron as a self-coding decoder; And simultaneously training the encoder and the decoder, and splicing the trend characteristic, the periodic characteristic and the spatial characteristic to obtain a decoupled multi-source characteristic.
5. The residual network-based feature decoupling self-encoding polypropylene production prediction model of claim 1, wherein the stitching the trend feature, the periodic feature, and the spatial feature comprises: learning the trend features in a time domain through a long-short-term memory network, and learning the periodic features in a frequency domain through a learnable fourier layer; Constructing a frequency domain gate on the basis of the long-short-period memory network, and filtering the periodic characteristics through the frequency domain gate to obtain final periodic characteristics; and splicing the trend characteristic and the final period characteristic to obtain a decoupled trend characteristic and a decoupled period characteristic.
6. A framework for constructing a prediction model of polypropylene production based on characteristic decoupling self-coding of a residual network, comprising: The extraction module is used for obtaining the original data of polypropylene production, extracting trend characteristics and period characteristics of the original data through a trend period long-short-term memory neural network, and extracting spatial characteristics of the original data through a dynamic self-attention convolution neural network; The splicing module is used for splicing the trend characteristic, the periodic characteristic and the space characteristic by utilizing characteristic decoupling self-coding so as to obtain a decoupled multi-source characteristic; And the construction module is used for establishing the relation between the decoupled multi-source characteristics and the output by using a residual error network and constructing a polypropylene production model according to the relation between the decoupled multi-source characteristics and the output.
7. The frame of claim 6, wherein the extraction module comprises: The decomposition unit is used for decomposing the original data into a trend signal and a periodic signal by utilizing a variation modal decomposition strategy; And the acquisition unit is used for extracting the characteristics of the trend signal and the periodic signal based on the trend period long-short-term memory neural network so as to acquire the trend characteristics and the periodic characteristics.
8. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor executing the program to implement the residual network-based feature decoupling self-encoding polypropylene production prediction model of any one of claims 1-5.
9. A non-transitory computer readable storage medium having stored thereon a computer program, wherein the program is executed by a processor for implementing a residual network based feature decoupling self-encoding polypropylene production prediction model according to any one of claims 1-5.
10. A computer program product comprising a computer program, characterized in that the computer program is executed for implementing a residual network based feature decoupling self-encoding polypropylene production prediction model according to any one of claims 1-5.

Description

Characteristic decoupling self-coding polypropylene production prediction model based on residual error network Technical Field The application relates to the technical field of polypropylene production, in particular to a characteristic decoupling self-coding polypropylene production prediction model based on a residual error network. Background At present, the propylene polymerization generation process is used as one of the taps in the petrochemical industry, and the refined high-quality production is the key for guaranteeing national life. The grades of industrial polypropylene products are divided by melt index values, and the quick and accurate prediction of the melt index in the propylene polymerization process plays an important role in product quality control, grade conversion, new product development, transition material reduction and the like of polypropylene production devices. At present, the common melt index detection means is manual sampling and laboratory off-line analysis, so that the requirement of industrial production on real-time control cannot be met, and great difficulty is brought to production quality monitoring and control. The quality control of the industrial site mainly depends on the experience of operators, so that the problems of unstable production process, low product quality, first economic benefit and the like are caused. The international advanced control optimizing software for olefin polymerization device has extremely high price, the later service is difficult to ensure, and the accuracy of melt index forecast is difficult to meet the requirements of 'safe, stable, long, full and excellent' high-quality production of industrial process. Therefore, the research of the online melt index forecasting method with high precision and high robustness in the propylene polymerization process has important significance. In the related art, soft measurement modeling methods can be classified into two types, namely process mechanism modeling and data-driven modeling. The process mechanism modeling needs to be deeply known on the internal reaction mechanism of the industrial process, and needs to be modeled by detailed high-order differential equations such as heat balance, energy balance, material balance and the like. However, in the actual process, due to the fact that raw materials are available, raw material components are large in fluctuation, key technological parameters are large, and a detailed mechanism formula is difficult to obtain. Along with the rapid development of data storage technology and sensing technology, mass data are collected and stored in an industrial field, and the data-driven modeling method becomes a mainstream soft measurement modeling method, and the method only depends on industrial process data, learns the nonlinear relation between auxiliary variables and difficult-to-measure target variables through algorithms, and does not need complex mechanism equations or expert knowledge. However, in the related art, the data-driven modeling method generally assumes that the production process is under a stable working condition, and the polypropylene production process needs to frequently switch product grades to meet market demands, so that the process presents non-static and nonlinear dynamic characteristics, and further when the working condition is adjusted, the model is difficult to adapt to complex dynamic changes, the generalization performance of the model is reduced, and improvement is needed. Disclosure of Invention The application provides a characteristic decoupling self-coding polypropylene production prediction model based on a residual network, which aims to solve the problems that in the related art, a data driving modeling method assumes that a production process is in a stable working condition, and the polypropylene production process is a non-static and nonlinear complex process, so that the model is difficult to adapt to complex dynamic changes, the generalization performance of the model is reduced and the like. The embodiment of the first aspect of the application provides a polypropylene production prediction model based on characteristic decoupling self-coding of a residual network, which comprises the following steps of obtaining original data of polypropylene production, extracting trend characteristics and period characteristics of the original data through a trend period long-short-term memory neural network, extracting spatial characteristics of the original data through a dynamic self-attention convolution neural network, splicing the trend characteristics, the period characteristics and the spatial characteristics by utilizing characteristic decoupling self-coding to obtain decoupled multi-source characteristics, establishing a relation between the decoupled multi-source characteristics and output by utilizing the residual network, and constructing the polypropylene production model according to the relation between the deco