Chemical process parameter prediction method based on space time sequence block network
By using a spatial temporal block network approach, and leveraging a fast-slow dual-stream architecture and a learnable frequency domain filter, the problem of weak prediction accuracy and generalization of chemical process parameters was solved, enabling accurate prediction of chemical processes and intelligent early warning of safety risks.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA ACAD OF SAFETY SCI & TECH
- Filing Date
- 2026-01-16
- Publication Date
- 2026-05-08
AI Technical Summary
Existing chemical process parameter prediction models suffer from poor prediction accuracy and weak generalization in the field of chemical production. In particular, when faced with complex chemical processes with high dynamics, strong noise, and non-stationary characteristics, they are unable to capture frequency domain features, lose key safety boundary information, and disrupt global continuity.
A spatial temporal block network-based approach is adopted, which uses a fast and slow dual-stream architecture, a learnable frequency domain filter, and a statistical injection module to process the fast and slow channel data streams respectively. Combined with multi-scale spatial feature fusion, it can achieve accurate prediction of chemical processes.
It improves the prediction accuracy of chemical process parameters, retains key safety boundary information, reduces overfitting, and enhances the intelligent early warning capability for dynamic safety risks.
Smart Images
Figure CN121996958A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial artificial intelligence and multivariate time series analysis technology, and in particular to a method for predicting chemical process parameters based on spatial temporal block networks. Background Technology
[0002] As a fundamental pillar of modern industry, the chemical industry's production processes are generally characterized by strong multivariate coupling, high nonlinearity, large inertial time delays, and complex reaction mechanisms. Furthermore, they often operate under extreme conditions such as high temperature and high pressure, placing stringent requirements on safe operation. Traditional experience-driven control models are no longer sufficient to meet the demands of modern factory management. With the development of artificial intelligence technology and the continuous advancement of intelligent manufacturing strategies, chemical production is undergoing a transformation from the passive response of traditional distributed control systems (DCS) to proactive intelligent decision-making systems based on data-driven approaches. In this process, high-precision time-series prediction of key process parameters (such as temperature, pressure, and flow rate) has become a core driving technology for achieving optimized production control, early warning of faults and dynamic safety risks, and improved energy efficiency.
[0003] Existing long-term prediction techniques for chemical process parameters mainly include various implementation methods such as recurrent neural networks, full attention mechanisms, sparse attention variants, and block processing architectures. Although existing deep learning models have established a leading position in general long-term prediction tasks, their practical application still faces three significant technical bottlenecks when dealing with complex chemical processes with high dynamics, strong noise, and non-stationary characteristics. First, in terms of feature extraction, existing technologies are mostly limited to the time domain perspective, making it difficult to capture signals with significant characteristics in the frequency domain, such as equipment vibration or fluid pulsation. This results in the model being unable to effectively distinguish between high-frequency random noise and critical high-frequency operating condition changes. Second, in terms of data distribution processing, while the normalization processing (such as RevIN) commonly used in existing technologies to address data non-stationarity stabilizes training, it erases the mean and variance of the data, leading to the loss of absolute physical quantities containing key safety boundary information (such as over-temperature and over-pressure warnings). Finally, in terms of temporal evolution modeling, while the mainstream patching strategy in existing technologies improves the efficiency of capturing local transient fluctuations, the segmentation operation to some extent disrupts the global continuity of the sequence, making it difficult to capture long-spanning slow-changing trends such as the drift of raw material component properties. Summary of the Invention
[0004] Based on the above analysis, the present invention aims to provide a chemical process parameter prediction method based on spatial temporal block networks, in order to solve the problems of poor accuracy and weak generalization of existing time-series prediction models in the field of chemical production.
[0005] This invention provides a method for predicting chemical process parameters based on spatial temporal block networks. The method includes the following steps: Acquire multivariate time-series data collected by the distributed control system of a chemical enterprise, and perform fast and slow dual-stream segmentation on the multivariate time-series data to obtain a slow channel data stream and a fast channel data stream; Statistics are extracted from the fast channel data stream, and statistical context embedding vectors, normalized slow channel data streams, and normalized fast channel data streams are generated based on the statistics. Trend and seasonal terms are extracted from the normalized fast channel data stream to generate a fast channel context embedding vector; a slow channel time-frequency context embedding vector is generated based on the normalized slow channel data stream. The statistical context embedding vector, the fast channel context embedding vector, and the slow channel time-frequency context embedding vector are fused to obtain the final encoded output; Based on the trend term and the final encoded output, the final prediction result of the chemical process parameters is obtained.
[0006] Based on the above solution, the present invention also makes the following improvements: Further, the step of generating a statistical context embedding vector, a normalized slow-channel data stream, and a normalized fast-channel data stream based on the statistical measure includes: Extract statistics from the fast channel data stream, and generate a statistics context embedding vector based on the statistics; Based on the statistics, the fast channel data stream is subjected to reversible instance normalization to obtain a normalized fast channel data stream. The slow channel data stream is aligned and normalized based on the statistics to obtain a normalized slow channel data stream.
[0007] Furthermore, the generation of the slow channel time-frequency context embedding vector includes: Time-domain feature extraction and frequency-domain feature extraction are performed on the normalized slow channel data stream, respectively; The time-domain features and the frequency-domain features are fused to generate a slow-channel time-frequency context embedding vector.
[0008] Further, the fusion of the statistical context embedding vector, the fast-channel context embedding vector, and the slow-channel time-frequency context embedding vector to obtain the final encoded output includes: The statistical context embedding vector, the fast channel context embedding vector, and the slow channel time-frequency context embedding vector are fused to obtain the fused result; The fused result is used as the input to the decoupled spatiotemporal multi-head attention network to obtain the final encoded output.
[0009] Furthermore, the decoupled spatiotemporal multi-head attention network is composed of L layers of decoupled spatial-temporal modules stacked sequentially, and each layer of decoupled spatial-temporal modules includes a temporal multi-head self-attention network unit, a spatial self-attention network unit, and a feedforward network unit. The fused result is input into a temporal multi-head self-attention network unit for temporal attention feature extraction, and the output of the temporal multi-head self-attention network unit is obtained. The output of the temporal multi-head self-attention network unit is input into the spatial multi-head self-attention network unit to extract spatial attention features, and the output of the spatial multi-head self-attention network unit is obtained. The output of the spatial multi-head self-attention network unit is input into the feedforward network unit for feedforward network feature extraction to obtain the final encoded output.
[0010] Furthermore, obtaining the final chemical process parameter prediction result based on the trend term and the final encoded output includes: The prediction result of the trend term is obtained based on the trend term; the prediction result of the trend term is fused with the final encoded output to obtain a normalized prediction output; The normalized prediction output is subjected to an inverse transform of reversible instance normalization to obtain the final prediction results of chemical process parameters.
[0011] Furthermore, the statistical context embedding vector is calculated using the following formula:
[0012] In the formula, and This represents the mean and standard deviation sequences obtained by extracting statistics from the fast-channel data stream. Indicates the operation of combining. The feature dimension of the encoded vector. This represents a multilayer perceptron, and M represents the bit number.
[0013] Furthermore, the slow channel time-frequency context embedding vector is calculated using the following formula:
[0014] In the formula, The weight of the gating unit, This represents the time-domain characteristics of a normalized slow-channel data stream. The frequency domain characteristics of the normalized slow channel data stream. This indicates an element-wise multiplication operation.
[0015] Furthermore, the final encoded output is calculated using the following formula:
[0016] In the formula, Indicates the first The output of a multi-head self-attention network unit in a layer space. The layer normalization function is represented. This represents a feedforward network.
[0017] Furthermore, the predicted chemical process parameters are calculated using the following formula: .
[0018] In the formula, This represents the normalized prediction output. and Represents the learnable affine parameters. It is a very small positive number.
[0019] The beneficial effects of this technical solution are: This invention provides a chemical process parameter prediction method based on spatial temporal block networks, which addresses the problems of poor accuracy and weak generalization in existing time-series prediction models for process parameters in the chemical production field. First, this invention introduces a learnable frequency domain filter based on RFFT, which can automatically filter key frequency components in the frequency domain, achieving accurate capture of periodic process changes and effective noise suppression. This overcomes the shortcomings of existing time-series prediction models, which focus on feature extraction from a time-domain perspective and struggle to identify frequency domain features such as equipment vibration and fluid pulsation in chemical processes, and distinguish between high-frequency random noise and effective high-frequency abrupt changes in operating conditions. Second, this invention designs a statistical injection module, which remaps the distribution statistical features of the original sequence back to a high-dimensional feature space. This allows the model to maintain training stability while preserving absolute physical quantity context information, addressing the problem of existing time-series prediction models' inability to accurately predict multiple process parameters. Normalization methods for non-stationary data remove the mean and variance of the data, thus losing key safety boundary information (such as over-temperature warnings) represented by absolute physical quantities. Finally, this invention proposes a fast-slow dual-flow architecture, which retains the "fast channel" for local processing while adding a "slow channel" specifically for handling long-scale global evolution features. Through the fusion of multi-scale spatial features, it achieves simultaneous and accurate prediction of transient changes and steady-state drifts in chemical processes. This solves the problem that the block operation used in existing time series prediction models destroys the global continuity of the sequence, making it difficult for the model to capture long-span slow trends (such as drift of raw material component properties, fouling effects of heat exchange systems, etc.).
[0020] In summary, this invention improves the prediction accuracy of chemical process parameters and reduces overfitting through the aforementioned RFFT-based learnable frequency domain filter, statistical injection module, and fast-slow dual-stream architecture. It provides a foundation model for enhancing the intelligent early warning capability of dynamic safety risks and solves the problems of poor prediction accuracy and weak generalization of existing time series prediction models in the field of chemical production.
[0021] In this invention, the above-described technical solutions can be combined with each other to achieve more preferred combinations. Other features and advantages of this invention will be set forth in the following description, and some advantages may become apparent from the description or be learned by practicing the invention. The objects and other advantages of this invention can be realized and obtained from what is particularly pointed out in the description and drawings. Attached Figure Description
[0022] The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Throughout the drawings, the same reference numerals denote the same parts.
[0023] Figure 1 The flowchart illustrates a chemical process parameter prediction method based on spatial temporal block networks, as provided in an embodiment of the present invention. Detailed Implementation
[0024] Preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, which form part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not intended to limit the scope of the present invention.
[0025] A specific embodiment of the present invention discloses a method for predicting chemical process parameters based on spatial temporal block networks. The method flow is as follows: Figure 1 As shown, the specific steps include: Step S1: Acquire multivariate time-series data collected by the distributed control system of the chemical enterprise, and perform fast and slow dual-stream segmentation on the multivariate time-series data to obtain slow channel data stream and fast channel data stream.
[0026] Specifically, by splitting the fast and slow dual-channel data streams, the temporal receptive field scale of the model is expanded without significantly increasing computational overhead.
[0027] In practice, chemical process parameters are a set of physical and chemical quantities that need to be monitored and controlled during chemical production, defining the state of production. Examples of chemical process parameters include reaction temperature, reactor pressure, feed flow rate, product concentration, and catalyst activity. Multivariate time-series data is a standardized data format formed after the above chemical process parameters are continuously measured, sampled, and recorded in the chemical enterprise's distributed control system (DCS). The future prediction result of any chemical process parameter is not merely a function of its own history, but rather the result of the combined effect of all relevant variables.
[0028] Acquire multivariate time-series data collected by the DCS system, wherein the format of the multivariate time-series data is an M×L matrix. Here, M represents the number of tag numbers, and L represents the time step. Each chemical process parameter in the DCS system corresponds to at least one sensor and is assigned a unique tag number; each tag number continuously generates a time-varying numerical sequence at a fixed sampling frequency. The time-series data of all tag numbers, after being aligned by time, constitute a multivariate time-series data matrix. Each row of the matrix corresponds to a specific chemical process parameter; each column of the matrix corresponds to a specific sampling time, containing the instantaneous state of all M chemical process parameters at that time.
[0029] The multivariate time-series data is segmented along the time dimension by performing fast and slow dual-stream splitting on the matrix, that is, dividing the matrix into segments of size [missing information] along the time dimension. Slow channel data stream to be downsampled and compressed and size are Fast channel data stream ,in .
[0030] Furthermore, for slow channel data streams subjected to downsampling compression Perform equidistant downsampling with equal step size to obtain the compressed slow channel data stream. Finally, the slow channel data stream is obtained. and fast channel data stream This downsampling mechanism extends the model's receptive field to an extremely long timescale without significantly increasing computational overhead.
[0031] Step S2: Extract statistics from the fast channel data stream, and generate a statistics context embedding vector, a normalized slow channel data stream, and a normalized fast channel data stream based on the statistics.
[0032] Specifically, by introducing statistical alignment and reversible normalization with injected residuals in data preprocessing, the non-stationarity of the data is addressed while ensuring that the fast and slow channel features are fused within the same distribution space. For the non-stationarity characteristics of industrial data, a reversible instance normalization mechanism is integrated and improved.
[0033] Step S21: Extract statistics from the fast channel data stream and generate a statistics context embedding vector based on the statistics.
[0034] In the specific implementation process, from the fast channel data stream The statistics for the time series data of each bit number are extracted, including the mean and standard deviation of the time series data for each bit number, forming the mean sequence of the fast channel data stream. and standard deviation series The mean sequence is injected using the statistics injection module. and standard deviation series Mapped to a high-dimensional latent space, the statistics injection module is implemented using a multilayer perceptron (MLP) to generate statistics context embedding vectors. The calculation formula is as follows:
[0035] In the formula, Indicates the operation of combining. This represents the feature dimension of the encoded vector.
[0036] Step S22: Perform reversible instance normalization on the fast channel data stream based on the statistics to obtain a normalized fast channel data stream.
[0037] In the specific implementation process, the mean sequence is used. and standard deviation series The fast channel data stream is subjected to reversible instance normalization to obtain normalized fast channel data. The calculation formula is as follows:
[0038] In the formula, and Represents the learnable affine parameters. This indicates an element-wise multiplication operation.
[0039] Step S23: Perform alignment and normalization processing on the slow channel data stream based on the statistics to obtain a normalized slow channel data stream.
[0040] In the specific implementation process, the mean sequence is used. and standard deviation series The slow channel data stream is aligned and normalized to obtain a normalized slow channel data stream, calculated using the following formula:
[0041] In the formula, It should be a very small positive number to prevent the denominator from being 0.
[0042] Step S3: Extract trend and seasonal terms from the normalized fast channel data stream to generate a fast channel context embedding vector; generate a slow channel time-frequency context embedding vector based on the normalized slow channel data stream.
[0043] Specifically, by fusing fast-channel trends and block-based data, the processing time series length is effectively shortened and the model's ability to represent local patterns is enhanced. At the same time, it can effectively capture the changing trends of chemical process parameters with large time inertia. By fusing slow-channel time-frequency context information, the problem of difficulty in extracting and fusing time-domain and frequency-domain features of medium- and long-term data with short windows is solved.
[0044] Step S31: Extract the trend term and seasonal term from the normalized fast channel data stream.
[0045] In the specific implementation process, the moving average operator is used to extract the trend term and seasonal term from the normalized fast channel data stream.
[0046] Preferably, a window size of [size missing] is used. The smoothed trend is extracted using 1D average pooling. First, the beginning and end of the time-series data for each bit in the normalized fast channel data stream are padded with the first element, and the end is padded with the last element. The padded length depends on the window size and stride to offset the length reduction caused by convolution. Then, a sliding window average is calculated for the time-series data of each bit in the padded and normalized fast channel data stream. The average value within each window replaces the original data within that window, resulting in the smoothed trend term. Finally, subtracting the trend term from the normalized fast-channel data stream yields the seasonal term, which includes high-frequency fluctuations and patterns. .
[0047] Step S32: Segment the seasonal item in the time dimension, and generate a fast-channel context embedding vector by performing matrix multiplication on the segmented fragments.
[0048] In practice, the seasonal item is divided into N segments along the time dimension with a fixed step size. Each segment is processed by dividing the seasonal item into sub-matrices with an equal number of columns. The sub-matrices are then multiplied sequentially using matrix multiplication to generate a fast-channel context embedding vector. .
[0049] Step S33: Perform time-domain feature extraction and frequency-domain feature extraction on the normalized slow channel data stream.
[0050] In the specific implementation process, the normalized slow channel data stream is sent to the time-domain trend encoder and the frequency-domain phase encoder for time-domain feature extraction and frequency-domain feature extraction, respectively.
[0051] By using a time-domain trend encoder and superimposing tanh operations, the time-domain features of long-term time series trends can be captured. The calculation formula is as follows:
[0052] In the formula, and This indicates the parameter weights for the tanh operation and the preset parameter weights.
[0053] Using a frequency-domain phase encoder, the calculation is first performed by combining a Fast Fourier Transform (RFFT) function with a complex filter, as shown in the following formula:
[0054] In the formula, This represents the complex filter weight matrix. It is used to filter the frequency domain complex sequence obtained by RFFT calculation, resulting in the filtered sequence. .
[0055] Reuse Will Projecting onto the frequency domain dimension captures the frequency domain characteristics of long-term frequency domain fluctuations. The calculation formula is as follows:
[0056] In the formula, Indicates taking the real part, This indicates taking the imaginary part.
[0057] Step S34: Fuse the time-domain features with the frequency-domain features to generate a slow-channel time-frequency context embedding vector.
[0058] In the specific implementation process, the time-domain features and frequency-domain features are finally fused through a gating unit to generate a slow-channel time-frequency context embedding vector. The calculation formula is as follows: ;
[0059] In the formula, The weight of the gating unit, This represents a trainable weight matrix. This represents the normalized weight allocation function.
[0060] Step S4: Fuse the statistical context embedding vector, the fast channel context embedding vector, and the slow channel time-frequency context embedding vector to obtain the final encoded output.
[0061] Specifically, by fusing the statistical context embedding vector, the fast-channel context embedding vector, and the slow-channel time-frequency context embedding vector, and inputting the fused result into a decoupled spatiotemporal multi-head attention network, the computational efficiency can be significantly optimized by simultaneously capturing multivariate spatial and temporal relationships through a self-attention mechanism.
[0062] Step S41: Fuse the statistical context embedding vector, the fast channel context embedding vector, and the slow channel time-frequency context embedding vector to obtain the fused result.
[0063] In the specific implementation process, the statistical context is embedded into the vector. Fast-channel context embedding vector and slow channel time-frequency context embedding vector The fusion occurs in the latent space, yielding the fused result. The calculation formula is as follows:
[0064] Step S42: Use the fused result as input to the decoupled spatiotemporal multi-head attention network to obtain the final encoded output.
[0065] Specifically, the decoupled spatiotemporal multi-head attention network is composed of The decoupled spatial-temporal modules are stacked sequentially, and each layer of decoupled spatial-temporal modules includes a temporal multi-head self-attention network unit, a spatial self-attention network unit, and a feedforward network unit.
[0066] Step S421: Input the fused result into the temporal multi-head self-attention network unit to extract temporal attention features and obtain the output of the temporal multi-head self-attention network unit.
[0067] In practice, the temporal multi-head self-attention network unit aims to capture inter-fragment dependencies within a single variable, specifically for the first... Layer output Reconstruct its vector dimension mixed batch and number of position variables The resulting mixed vector :
[0068] Will As a residual term and a time-multi-head self-attention network unit The calculation terms are summed and then passed through... Perform layer normalization to obtain the first layer. Multi-head self-attention output in layer time :
[0069] Preferably, a temporal multi-head self-attention network unit The calculation process is expressed as follows: Weight matrix of temporal multi-head self-attention network unit , and The query, key, and value matrix of a multi-head self-attention network unit is calculated using the following formula: ; ;
[0070] The attention score of a temporal multi-head self-attention network unit is obtained by calculating the correlation between the query and key matrix of the network unit. :
[0071] In the formula, Indicates the scaling factor. is the number of heads. The scaling factor is used to prevent gradient vanishing.
[0072] Furthermore, regarding the first Normalization and weighted summation are performed at each end of the layer. The normalization and summation results are then multiplied by the value matrix of the temporal multi-head self-attention network unit to obtain the output matrix of each end of the temporal multi-head self-attention network unit. The calculation formula is as follows:
[0073] Will The outputs of each head are concatenated and projected again to obtain a temporal multi-head self-attention network unit. The calculation term is as follows:
[0074] In the formula, This represents the output projection matrix of a temporal multi-head self-attention network unit.
[0075] Step S422: Input the output of the temporal multi-head self-attention network unit into the spatial multi-head self-attention network unit to extract spatial attention features, and obtain the output of the spatial multi-head self-attention network unit.
[0076] In the specific implementation, the output of the temporal multi-head self-attention network unit undergoes a tensor dimension transformation Permute operation, swapping the dimension of the number of bit variables and the dimension of the number of segments. The calculation formula is as follows:
[0077] Will As a residual term and a spatial multi-head self-attention network unit The calculation terms are summed and then passed through... Perform layer normalization to obtain the first layer. Output of multi-head self-attention network unit in layer space :
[0078] Preferably, a spatial multi-head self-attention network unit The calculation process is expressed as follows: Weight matrix of spatial multi-head self-attention network unit , and The query, key, and value matrix of a multi-head self-attention network unit in the computational space: ; ;
[0079] The attention score of a spatial multi-head self-attention network unit is obtained by calculating the correlation between the query and key matrix of the spatial multi-head self-attention network unit. :
[0080] Furthermore, regarding the first Normalization and weighted summation are performed at each end of the layer. The normalization and summation results are then multiplied by the value matrix of the spatial multi-head self-attention network unit to obtain the output matrix of each end of the spatial multi-head self-attention network unit. The calculation formula is as follows:
[0081] Will The outputs of each unit are spliced together and projected again to obtain a spatial multi-head self-attention network unit. The calculation term is as follows:
[0082] In the formula, This represents the output projection matrix of a spatial multi-head self-attention network unit.
[0083] Step S423: Input the output of the spatial multi-head self-attention network unit into the feedforward network unit to extract feedforward network features and obtain the final encoded output.
[0084] In the specific implementation process, the output of the spatial multi-head self-attention network unit will be used. The input is fed into the feedforward network unit, that is, the output of the spatial multi-head self-attention network unit is used as the computational term and residual term of the feedforward network, and the result is obtained by layer normalization. The final encoded output of the layer :
[0085] Preferably, a feedforward network The calculation process is expressed as follows: enter Through the first linear layer, the feature dimension is reduced from... Expand to ,get In the formula The expansion factor is a positive integer. A linear layer is represented using a weight matrix. and bias matrix Linear transformation performed:
[0086] Furthermore, the GELU activation function is used on... By performing a nonlinear transformation, we obtain :
[0087] Compared to ReLU, GELU has a smooth gradient in the negative region, which is more conducive to training complex patterns.
[0088] Furthermore, regarding After performing the Dropout operation, the feature dimensions are remapped back to... In order to perform residual connections and use weight matrices and bias matrix Perform a linear transformation of the second linear layer:
[0089] Step S5: Based on the trend term and the final encoded output, the final chemical process parameter prediction result is obtained.
[0090] Specifically, by fusing the calculation results of trend and seasonal terms, it is possible to integrate the changing trend and the prediction signal to obtain the final output prediction of the fast and slow dual-stream spatial temporal block network.
[0091] Step S51: Obtain the prediction result of the trend term based on the trend term; fuse the prediction result of the trend term with the final encoded output to obtain a normalized prediction output.
[0092] In the specific implementation process, the prediction result of the trend term is obtained through a multilayer perceptron (MLP) based on the trend term. Then, the output of the last layer of the decoupled spatiotemporal multi-head attention network is fused with the prediction result of the trend term to obtain a normalized prediction output. The calculation formula is as follows:
[0093] Step S52: Perform an inverse transform of reversible instance normalization on the normalized prediction output to obtain the final prediction results of the chemical process parameters.
[0094] In practice, the normalized prediction output is subjected to an inverse transform of reversible instance normalization to obtain the final prediction results of the chemical process parameters. The calculation formula is as follows:
[0095] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware, and the program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.
[0096] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.
Claims
1. A method for predicting chemical process parameters based on spatial temporal block networks, characterized in that, The method includes the following steps: Acquire multivariate time-series data collected by the distributed control system of a chemical enterprise, and perform fast and slow dual-stream segmentation on the multivariate time-series data to obtain a slow channel data stream and a fast channel data stream; Statistics are extracted from the fast channel data stream, and statistical context embedding vectors, normalized slow channel data streams, and normalized fast channel data streams are generated based on the statistics. Trend and seasonal terms are extracted from the normalized fast channel data stream to generate a fast channel context embedding vector; a slow channel time-frequency context embedding vector is generated based on the normalized slow channel data stream. The statistical context embedding vector, the fast channel context embedding vector, and the slow channel time-frequency context embedding vector are fused to obtain the final encoded output; Based on the trend term and the final encoded output, the final prediction result of the chemical process parameters is obtained.
2. The method for predicting chemical process parameters based on spatial temporal block networks according to claim 1, characterized in that, The process of generating a statistical context embedding vector, a normalized slow-channel data stream, and a normalized fast-channel data stream based on the statistical measures includes: Extract statistics from the fast channel data stream, and generate a statistics context embedding vector based on the statistics; Based on the statistics, the fast channel data stream is subjected to reversible instance normalization to obtain a normalized fast channel data stream. The slow channel data stream is aligned and normalized based on the statistics to obtain a normalized slow channel data stream.
3. The method for predicting chemical process parameters based on spatial temporal block networks according to claim 2, characterized in that, The generated slow channel time-frequency context embedding vector includes: Time-domain feature extraction and frequency-domain feature extraction are performed on the normalized slow channel data stream, respectively; The time-domain features and the frequency-domain features are fused to generate a slow-channel time-frequency context embedding vector.
4. The method for predicting chemical process parameters based on spatial temporal block networks according to claim 3, characterized in that, The process of fusing the statistical context embedding vector, the fast channel context embedding vector, and the slow channel time-frequency context embedding vector to obtain the final encoded output includes: The statistical context embedding vector, the fast channel context embedding vector, and the slow channel time-frequency context embedding vector are fused to obtain the fused result; The fused result is used as the input to the decoupled spatiotemporal multi-head attention network to obtain the final encoded output.
5. The method for predicting chemical process parameters based on spatial temporal block networks according to claim 4, characterized in that, The decoupled spatiotemporal multi-head attention network is composed of L layers of decoupled spatial-temporal modules stacked sequentially. Each layer of decoupled spatial-temporal modules includes a temporal multi-head self-attention network unit, a spatial self-attention network unit, and a feedforward network unit. The fused result is input into a temporal multi-head self-attention network unit for temporal attention feature extraction, and the output of the temporal multi-head self-attention network unit is obtained. The output of the temporal multi-head self-attention network unit is input into the spatial multi-head self-attention network unit to extract spatial attention features, and the output of the spatial multi-head self-attention network unit is obtained. The output of the spatial multi-head self-attention network unit is input into the feedforward network unit for feedforward network feature extraction to obtain the final encoded output.
6. The method for predicting chemical process parameters based on spatial temporal block networks according to claim 5, characterized in that, The process of obtaining the final chemical process parameter prediction result based on the trend term and the final encoded output includes: The prediction result of the trend term is obtained based on the trend term; the prediction result of the trend term is fused with the final encoded output to obtain a normalized prediction output; The normalized prediction output is subjected to an inverse transform of reversible instance normalization to obtain the final prediction results of chemical process parameters.
7. The method for predicting chemical process parameters based on spatial temporal block networks according to any one of claims 1-6, characterized in that, The statistical context embedding vector is calculated using the following formula: In the formula, and This represents the mean and standard deviation sequences obtained by extracting statistics from the fast-channel data stream. Indicates the operation of combining. The feature dimension of the encoded vector. This represents a multilayer perceptron, and M represents the number of bit numbers.
8. The method for predicting chemical process parameters based on spatial temporal block networks according to claim 7, characterized in that, The slow channel time-frequency context embedding vector is calculated using the following formula: In the formula, The weight of the gating unit, This represents the time-domain characteristics of a normalized slow-channel data stream. The frequency domain characteristics of the normalized slow channel data stream. This indicates an element-wise multiplication operation.
9. The method for predicting chemical process parameters based on spatial temporal block networks according to claim 8, characterized in that, The final encoded output is calculated using the following formula: In the formula, Indicates the first The output of a multi-head self-attention network unit in a layer space. The layer normalization function is represented. This represents a feedforward network.
10. The method for predicting chemical process parameters based on spatial temporal block networks according to claim 9, characterized in that, The predicted chemical process parameters are calculated using the following formula: 。 In the formula, This represents the normalized prediction output. and Represents the learnable affine parameters. It is a very small positive number.