CN-121996958-A - Chemical process parameter prediction method based on space time sequence blocking network

CN121996958ACN 121996958 ACN121996958 ACN 121996958ACN-121996958-A

Abstract

A chemical process parameter prediction method based on a space time sequence block network belongs to the technical field of industrial artificial intelligence and multivariate time sequence analysis and is used for solving the problems of poor process parameter prediction precision and weak generalization of the existing time sequence prediction model in the chemical production field. The method comprises the steps of obtaining multivariable time sequence data, carrying out fast-slow double-flow segmentation to obtain a slow-channel data stream and a fast-channel data stream, extracting statistics from the fast-channel data stream to respectively generate a statistics context embedding vector, a normalized slow-channel data stream and a normalized fast-channel data stream, extracting trend items and season items from the normalized fast-channel data stream to generate a fast-channel context embedding vector, generating a slow-channel time-frequency context embedding vector based on the normalized slow-channel data stream, and fusing the statistics context embedding vector, the fast-channel context embedding vector and the slow-channel time-frequency context embedding vector to obtain a final chemical process parameter prediction result.

Inventors

GUAN LEI
LI PEIZHANG
ZONG KAI
WEI LIANGXIAO
XU XUERUI

Assignees

中国安全生产科学研究院

Dates

Publication Date: 20260508
Application Date: 20260116

Claims (10)

1. The chemical process parameter prediction method based on the space time sequence blocking network is characterized by comprising the following steps of: The method comprises the steps of obtaining multivariable time sequence data collected by a chemical enterprise distributed control system, and performing fast and slow double-flow segmentation on the multivariable time sequence data to obtain slow channel data flow and fast channel data flow; Extracting statistics from the fast channel data stream, and generating a statistics context embedding vector, a normalized slow channel data stream and a normalized fast channel data stream based on the statistics, respectively; extracting trend items and season items from the normalized fast channel data stream to generate a fast channel context embedding vector; generating a slow channel time-frequency context embedding vector based on the normalized slow channel data stream; Fusing the statistic context embedded vector, the fast channel context embedded vector and the slow channel time-frequency context embedded vector to obtain a final coding output; And obtaining a final chemical process parameter prediction result based on the trend item and the final code output.
2. The method of claim 1, wherein generating a statistic context embedding vector, a normalized slow channel data stream, and a normalized fast channel data stream based on the statistic comprises: Extracting statistics from the fast channel data stream, generating a statistics context embedding vector based on the statistics; performing reversible instance normalization processing on the fast channel data stream based on the statistic to obtain a normalized fast channel data stream; and carrying out alignment normalization processing on the slow channel data stream based on the statistic to obtain a normalized slow channel data stream.
3. The method for predicting chemical process parameters based on a space-time-sequence block network according to claim 2, wherein generating the slow-channel time-frequency context embedding vector comprises: respectively carrying out time domain feature extraction and frequency domain feature extraction on the normalized slow channel data stream; And fusing the time domain features and the frequency domain features to generate a slow channel time-frequency context embedding vector.
4. The method for predicting chemical process parameters based on a space-time-sequence block network according to claim 3, wherein the fusing the statistic context embedding vector, the fast channel context embedding vector and the slow channel time-frequency context embedding vector to obtain a final encoded output comprises: Fusing the statistic context embedded vector, the fast channel context embedded vector and the slow channel time-frequency context embedded vector to obtain a fused result; And taking the fused result as the input of the decoupling space-time multi-head attention network to obtain the final coding output.
5. The method for predicting chemical process parameters based on a space-time-sequence block network according to claim 4, wherein the decoupling space-time multi-head attention network is formed by sequentially stacking L layers of decoupling space-time sequence modules, and each layer of decoupling space-time sequence modules comprises a time multi-head self-attention network unit, a space self-attention network unit and a feedforward network unit; Inputting the fused result into a time multi-head self-attention network unit for time attention feature extraction, and obtaining the output of the time multi-head self-attention network unit; The output of the time multi-head self-attention network unit is input to the space multi-head self-attention network unit to extract the space attention characteristics, so that the output of the space multi-head self-attention network unit is obtained; and inputting the output of the spatial multi-head self-attention network unit into a feedforward network unit to extract feedforward network characteristics, and obtaining final coding output.
6. The method for predicting chemical process parameters based on a space-time-series block network according to claim 5, wherein the obtaining a final chemical process parameter prediction result based on the trend term and the final code output comprises: Fusion is carried out on the predicted result of the trend item and the final coding output to obtain normalized predicted output; And carrying out inverse transformation of reversible instance normalization on the normalized prediction output to obtain a final chemical process parameter prediction result.
7. The method of predicting chemical process parameters based on a spatially-sequential blocking network according to any one of claims 1-6, wherein the statistic context embedding vector is calculated by the following formula: in the formula, And Representing a sequence of mean and standard deviation obtained by extracting statistics from the fast channel data stream, The operation of the spelling is indicated, Representing the characteristic dimension of the encoded vector, Representing a multi-layer perceptron, M representing the number of bit numbers.
8. The method for predicting chemical process parameters based on a space-time-sequence block network according to claim 7, wherein the slow-channel time-frequency context embedding vector is calculated by the following formula: in the formula, The weight of the gating cell is determined, Representing the time domain characteristics of the normalized slow channel data stream, Representing the frequency domain characteristics of the normalized slow-channel data stream, Representing an element-wise multiplication operation.
9. The method for predicting chemical process parameters based on a space-time-series blocking network according to claim 8, wherein the final encoded output is calculated by the following formula: in the formula, Represent the first The output of the layer space multi-headed self-focusing network element, The layer normalization function is represented as a function of the layer, Representing a feed forward network.
10. The method for predicting chemical process parameters based on a space-time-sequence block network according to claim 9, wherein the prediction result of the chemical process parameters is calculated by the following formula: 。 in the formula, Representing the normalized predicted output of the device, And Representing the affine parameters that can be learned, Is a very small positive number.

Description

Chemical process parameter prediction method based on space time sequence blocking network Technical Field The invention relates to the technical field of industrial artificial intelligence and multivariate time sequence analysis, in particular to a chemical process parameter prediction method based on a space time sequence block network. Background The chemical industry is taken as a basic support of modern industry, the production process of the chemical industry generally has the remarkable characteristics of multivariable strong coupling, high nonlinearity, large inertia time lag, complex reaction mechanism and the like, is always in extreme working conditions of high temperature, high pressure and the like, and has strict requirements on safe operation. The traditional experience-driven control mode has hardly met the requirement of modern factory fine management, and with the development of artificial intelligence technology and the continuous advancement of industrial intelligent manufacturing strategies, chemical production is undergoing a transition from passive response of a traditional Distributed Control System (DCS) to a data-driven active intelligent decision system. In the process, high-precision time sequence prediction for key process parameters (such as temperature, pressure, flow and the like) has become a core driving technology for realizing production optimization control, early warning of faults and dynamic safety risks and energy efficiency improvement. The technical scheme of the existing chemical process parameter long time sequence prediction mainly comprises a plurality of implementation modes such as a cyclic neural network, a full attention mechanism, a sparse attention variant and a block processing architecture. Although the existing deep learning model has established a leading position in the general long-time-sequence prediction task, its practical application still faces three significant technical bottlenecks in the face of complex chemical processes with high dynamics, strong noise and non-stationary characteristics. Firstly, in a feature extraction dimension, the prior art is limited to a time domain visual angle, signals with remarkable features in a frequency domain such as equipment vibration or fluid pulsation are difficult to capture, so that a model cannot effectively distinguish high-frequency random noise from critical high-frequency working condition mutation, secondly, in a data distribution processing dimension, the prior art stabilizes training for normalization processing (such as RevIN) which is generally adopted for dealing with data non-stationarity, but erases the mean value and variance of data, so that absolute physical magnitude containing critical safety boundary information (such as overtemperature and overpressure early warning) is lost, and finally, in a time sequence evolution modeling dimension, a partitioning (Patch) strategy of a main stream of the prior art improves the capture efficiency of local transient fluctuation, but the global continuity of a sequence is damaged to a certain extent by the partitioning operation, so that long-span slow variation trend such as difficulty in capturing raw material component attribute drift is caused. Disclosure of Invention In view of the analysis, the invention aims to provide a chemical process parameter prediction method based on a space time sequence block network, which is used for solving the problems of poor process parameter prediction precision and weak generalization of a time sequence prediction model in the chemical production field in the prior art. The embodiment of the invention provides a chemical process parameter prediction method based on a space time sequence block network, which comprises the following steps: The method comprises the steps of obtaining multivariable time sequence data collected by a chemical enterprise distributed control system, and performing fast and slow double-flow segmentation on the multivariable time sequence data to obtain slow channel data flow and fast channel data flow; Extracting statistics from the fast channel data stream, and generating a statistics context embedding vector, a normalized slow channel data stream and a normalized fast channel data stream based on the statistics, respectively; extracting trend items and season items from the normalized fast channel data stream to generate a fast channel context embedding vector; generating a slow channel time-frequency context embedding vector based on the normalized slow channel data stream; Fusing the statistic context embedded vector, the fast channel context embedded vector and the slow channel time-frequency context embedded vector to obtain a final coding output; And obtaining a final chemical process parameter prediction result based on the trend item and the final code output. Based on the scheme, the invention also makes the following improvements: Further, the generating a statistic context