Benzene, toluene and xylene yield prediction method based on multi-scale space-time attention gating long short-term memory network

Abstract

The invention relates to the technical field of industrial control, in particular to a benzene, toluene and xylene yield prediction method based on a multi-scale space-time attention gating long-short term memory network. According to the method, process variables in the continuous catalytic reforming production process are collected and normalized to obtain normalized process variables, and then a spatial dimension graph representing the spatial dependency relationship of the normalized process variables is obtained on the basis of a graph attention network to extract spatial features; and obtaining a time dimension graph representing normalized process variable time characteristics by combining a pyramid attention mechanism so as to extract time characteristics, fusing the time-space characteristics, and inputting the fused characteristics into a long short-term memory network model so as to obtain predicted values of the yields of benzene, toluene and xylene. Therefore, the problems of unstable prediction precision and weak generalization ability in the face of multivariable coupling and long-distance dependence due to lack of a spatial modeling mechanism and only capability of sensing time information in a single scale in the prior art are solved.

Description

Technical Field

[0001] This application relates to the field of industrial control technology, and in particular to a method for predicting the yields of benzene, toluene and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network. Background Technology

[0002] In modern chemical industry, benzene, toluene, and xylene, as basic chemical raw materials, are widely used in fine chemicals, pharmaceuticals, coatings, and polymer materials, and are key components in industrial production processes. With the continuous development of my country's economy and the acceleration of industrialization, the market demand for benzene, toluene, and xylene products is showing a steady upward trend. Effectively predicting the production of benzene, toluene, and xylene is an important technical means to improve production efficiency, reduce operating costs, and optimize process management.

[0003] In related technologies, methods for predicting the production of benzene, toluene, and xylene include empirical formula methods, statistical modeling methods, and data-driven methods based on machine learning. Empirical formula methods rely on the knowledge accumulated by process experts and specific operating condition assumptions to build predictive models, making them difficult to adapt to complex and changing production environments and exhibiting poor robustness. Statistical modeling methods, such as multiple regression and time series modeling, often approximate industrial processes with linear or low-order nonlinear functions, making it difficult to effectively characterize the complex nonlinear coupling relationships between variables and failing to meet the requirements of actual production for modeling accuracy and stability. Data-driven methods based on machine learning, such as deep neural networks, possess strong nonlinear fitting and feature extraction capabilities, making them suitable for high-dimensional and complexly coupled modeling tasks.

[0004] However, due to insufficient feature fusion capabilities, related data-driven methods have limitations when dealing with complex data that has high-dimensional redundancy, strong variable coupling, and spatiotemporal dynamic changes. At the same time, due to their strong dependence on massive samples and computing resources, the overall cost of engineering deployment is high, making it difficult to implement and scale up in actual industrial scenarios. These issues urgently need to be addressed. Summary of the Invention

[0005] This application provides a method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network. This method addresses the problem that related technologies, due to the lack of spatial modeling mechanisms, can only perceive temporal information at a single scale, resulting in unstable prediction accuracy and weak generalization ability when faced with multivariate coupling and long-distance dependence.

[0006] The first aspect of this application provides a method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network, comprising the following steps: collecting process variables in a continuous catalytic reforming process and normalizing the process variables to obtain normalized process variables; establishing a spatial dimension graph reflecting the spatial dependencies of the normalized process variables based on the normalized process variables and a graph attention network model, and establishing a temporal dimension graph reflecting the temporal characteristics of the normalized process variables based on the normalized process variables and a pyramid attention mechanism module; extracting the spatial features of the normalized process variables based on the spatial dimension graph, and extracting the temporal features of the normalized process variables based on the temporal dimension graph; performing weighted fusion based on the spatial features and the temporal features to generate fused features, and inputting the fused features into a pre-constructed long short-term memory network model to output predicted yields of benzene, toluene, and xylene.

[0007] Based on the above technical means, the embodiments of this application break through the limitations of traditional single time series modeling methods, and innovatively integrate graph attention network model, pyramid attention mechanism module and long short-term memory network model to construct a deep learning architecture with spatiotemporal feature fusion capability. The graph attention network model extracts the complex spatial interaction relationship between process variables, the pyramid attention mechanism module captures multi-scale time-dependent structure, and the long short-term memory network model strengthens the modeling of long-term time series laws. The three work together to achieve accurate prediction of benzene, toluene and xylene yields under complex operating conditions. This not only significantly improves prediction performance and stability, but is also easy to deploy in existing industrial systems, and has good versatility and promotion value.

[0008] Optionally, in one embodiment of this application, the step of extracting the spatial features of the normalized process variables based on the spatial dimension graph includes: performing a weighted summation on each node and its adjacent nodes of the spatial dimension graph according to a first attention weight to obtain a first weighted summation result; extracting multiple spatial features of the normalized process variables based on the first weighted summation result, and generating a spatial feature matrix based on the multiple spatial features.

[0009] Based on the above technical means, the embodiments of this application extract multi-view features through multi-head attention mechanism, dynamically quantifies the spatial dependence strength between process variables, and generates a spatial feature matrix that completely preserves the nonlinear spatial correlation information between process parameters, providing key spatial dimension feature support for subsequent benzene, toluene and xylene yield prediction, effectively improving the fitting ability of the prediction model to complex spatial interaction relationships, and thus significantly improving the accuracy and reliability of yield prediction.

[0010] Optionally, in one embodiment of this application, the formula for calculating the spatial feature matrix may be, but is not limited to: , in, For the first Layer nodes No. The spatial feature matrix output by each attention head For the first The output of each attention head is spliced ​​together. For activation function, For nodes The set of adjacent nodes, For the first Layer nodes For adjacent nodes Attention weights For the first Layer Adjacent Nodes Its characteristics.

[0011] Based on the above technical means, the embodiments of this application effectively integrate multi-perspective spatial correlation information by using the splicing operation of the multi-head attention mechanism and the weighted aggregation of adjacent node features through the calculation formula of the spatial feature matrix. This enhances the distinguishability of spatial dependencies between nodes, enables the spatial feature matrix to accurately depict the nonlinear spatial interaction details between nodes, provides more hierarchical spatial feature support for subsequent yield prediction models, and improves the model's fitting accuracy for the spatial correlation of process variables.

[0012] Optionally, in one embodiment of this application, the step of extracting the time features of the normalized process variables based on the time dimension graph includes: performing a weighted summation on each node and adjacent nodes of the time dimension graph according to a second attention weight to obtain a second weighted summation result; extracting multiple time features of the normalized process variables according to the second weighted summation result, and generating a time feature vector according to the multiple time features.

[0013] Based on the above technical means, the embodiments of this application fully preserve the multi-level dynamic laws of process variables evolving over time by extracting multi-scale time features and generating time feature vectors. This allows subsequent models to make full use of the temporal correlation of time series data, enhance the fitting ability of dynamic changes in the continuous catalytic reforming process, provide more accurate time dimension support for the prediction of benzene, toluene and xylene yields, and improve the reliability of prediction results.

[0014] Optionally, in one embodiment of this application, the formula for calculating the time feature vector may be, but is not limited to: , in, For nodes Time feature vector, For scale The next node The set of adjacent nodes, For nodes The query vector, Adjacent nodes The key vector, Adjacent nodes The value vector, is the dimension of the key vector.

[0015] Based on the above technical means, the embodiments of this application accurately quantify the temporal correlation strength between nodes and adjacent nodes through the calculation formula of time feature vectors, realize dynamic weighted aggregation of adjacent node value vectors, effectively enhance the distinguishability and correlation of multi-scale time features, accurately capture the deep time dependency details across scales and within the same scale in time series data, inject more targeted time correlation information into time feature vectors, and improve the utilization efficiency and prediction accuracy of subsequent yield prediction models for time dimension features.

[0016] Optionally, in one embodiment of this application, the process variables are normalized to obtain normalized process variables. The calculation formula for the normalized process variables may be, but is not limited to, the following: , in, For the sample index of process variables, The total sample size for process variables. For indexing process variables, To normalize process variables, For the first The first process variable One sample, For the first The maximum value of each process variable among all samples. For the first The minimum value among all samples of a process variable.

[0017] Based on the above technical means, the embodiments of this application, through normalization processing, make various process variables comparable in numerical scale. This ensures that when constructing spatial dimension graphs based on graph attention network models, constructing temporal dimension graphs based on pyramid attention mechanism modules, and using long short-term memory network models for yield prediction, each variable can be fairly and effectively learned by the model to contribute its features. This avoids the model from paying excessive attention to variables with large numerical ranges due to inconsistent dimensions, thereby improving the accuracy and stability of benzene, toluene, and xylene yield predictions and providing reliable data support for process optimization and production decisions.

[0018] A second aspect of this application provides a benzene, toluene, and xylene yield prediction device based on a multi-scale spatiotemporal attention-gated long short-term memory network, comprising: a data acquisition module for acquiring process variables during a continuous catalytic reforming process and normalizing the process variables to obtain normalized process variables; a data establishment module for establishing a spatial dimension map reflecting the spatial dependencies of the normalized process variables based on the normalized process variables and a graph attention network model, and establishing a temporal dimension map reflecting the temporal characteristics of the normalized process variables based on the normalized process variables and a pyramid attention mechanism module; an extraction module for extracting the spatial features of the normalized process variables based on the spatial dimension map, and extracting the temporal features of the normalized process variables based on the temporal dimension map; and a prediction module for performing weighted fusion based on the spatial features and the temporal features to generate fused features, and inputting the fused features into a pre-constructed long short-term memory network model to output predicted benzene, toluene, and xylene yields.

[0019] Based on the above technical means, the embodiments of this application break through the limitations of traditional single time series modeling methods, and innovatively integrate graph attention network model, pyramid attention mechanism module and long short-term memory network model to construct a deep learning architecture with spatiotemporal feature fusion capability. The graph attention network model extracts the complex spatial interaction relationship between process variables, the pyramid attention mechanism module captures multi-scale time-dependent structure, and the long short-term memory network model strengthens the modeling of long-term time series laws. The three work together to achieve accurate prediction of benzene, toluene and xylene yields under complex operating conditions. This not only significantly improves prediction performance and stability, but is also easy to deploy in existing industrial systems, and has good versatility and promotion value.

[0020] Optionally, in one embodiment of this application, the extraction module includes: a first calculation unit, configured to perform a weighted summation on each node and adjacent nodes of the spatial dimension graph according to a first attention weight, to obtain a first weighted summation result; and a first generation unit, configured to extract multiple spatial features of the normalized process variables based on the first weighted summation result, and generate a spatial feature matrix based on the multiple spatial features.

[0021] Based on the above technical means, the embodiments of this application extract multi-view features through multi-head attention mechanism, dynamically quantifies the spatial dependence strength between process variables, and generates a spatial feature matrix that completely preserves the nonlinear spatial correlation information between process parameters, providing key spatial dimension feature support for subsequent benzene, toluene and xylene yield prediction, effectively improving the fitting ability of the prediction model to complex spatial interaction relationships, and thus significantly improving the accuracy and reliability of yield prediction.

[0022] Optionally, in one embodiment of this application, the formula for calculating the spatial feature matrix may be, but is not limited to: , in, For the first Layer nodes No. The spatial feature matrix output by each attention head For the first The output of each attention head is spliced ​​together. For activation function, For nodes The set of adjacent nodes, For the first Layer nodes For adjacent nodes Attention weights For the first Layer Adjacent Nodes Its characteristics.

[0023] Based on the above technical means, the embodiments of this application effectively integrate multi-perspective spatial correlation information by using the splicing operation of the multi-head attention mechanism and the weighted aggregation of adjacent node features through the calculation formula of the spatial feature matrix. This enhances the distinguishability of spatial dependencies between nodes, enables the spatial feature matrix to accurately depict the nonlinear spatial interaction details between nodes, provides more hierarchical spatial feature support for subsequent yield prediction models, and improves the model's fitting accuracy for the spatial correlation of process variables.

[0024] Optionally, in one embodiment of this application, the extraction module includes: a second calculation unit, configured to perform weighted summation on each node and adjacent nodes of the time dimension graph according to a second attention weight, to obtain a second weighted summation result; and a second generation unit, configured to extract multiple time features of the normalized process variable based on the second weighted summation result, and generate a time feature vector based on the multiple time features.

[0025] Based on the above technical means, the embodiments of this application fully preserve the multi-level dynamic laws of process variables evolving over time by extracting multi-scale time features and generating time feature vectors. This allows subsequent models to make full use of the temporal correlation of time series data, enhance the fitting ability of dynamic changes in the continuous catalytic reforming process, provide more accurate time dimension support for the prediction of benzene, toluene and xylene yields, and improve the reliability of prediction results.

[0026] Optionally, in one embodiment of this application, the formula for calculating the time feature vector may be, but is not limited to: , in, For nodes Time feature vector, For scale The next node The set of adjacent nodes, For nodes The query vector, Adjacent nodes The key vector, Adjacent nodes The value vector, is the dimension of the key vector.

[0027] Based on the above technical means, the embodiments of this application accurately quantify the temporal correlation strength between nodes and adjacent nodes through the calculation formula of time feature vectors, realize dynamic weighted aggregation of adjacent node value vectors, effectively enhance the distinguishability and correlation of multi-scale time features, accurately capture the deep time dependency details across scales and within the same scale in time series data, inject more targeted time correlation information into time feature vectors, and improve the utilization efficiency and prediction accuracy of subsequent yield prediction models for time dimension features.

[0028] Optionally, in one embodiment of this application, the calculation formula for the normalized process variable in the acquisition module may be, but is not limited to, the following: , in, For the sample index of process variables, The total sample size for process variables. For indexing process variables, To normalize process variables, For the first The first process variable One sample, For the first The maximum value of each process variable among all samples. For the first The minimum value among all samples of a process variable.

[0029] Based on the above technical means, the embodiments of this application, through normalization processing, make various process variables comparable in numerical scale. This ensures that when constructing spatial dimension graphs based on graph attention network models, constructing temporal dimension graphs based on pyramid attention mechanism modules, and using long short-term memory network models for yield prediction, each variable can be fairly and effectively learned by the model to contribute its features. This avoids the model from paying excessive attention to variables with large numerical ranges due to inconsistent dimensions, thereby improving the accuracy and stability of benzene, toluene, and xylene yield predictions and providing reliable data support for process optimization and production decisions.

[0030] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the benzene, toluene, and xylene yield prediction method based on a multi-scale spatiotemporal attention-gated long short-term memory network as described in the above embodiments.

[0031] A fourth aspect of this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network.

[0032] A fifth aspect of this application provides a computer program product, including a computer program that, when executed, implements the above-described method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network.

[0033] This application's embodiments break through the limitations of traditional single-time-series modeling methods by innovatively integrating a graph attention network model, a pyramid attention mechanism module, and a long short-term memory network model to construct a deep learning architecture with spatiotemporal feature fusion capabilities. The graph attention network model extracts complex spatial interactions between process variables, the pyramid attention mechanism module captures multi-scale temporal dependency structures, and the long short-term memory network model strengthens the modeling of long-term temporal patterns. The three work synergistically to achieve accurate prediction of benzene, toluene, and xylene yields under complex operating conditions. This not only significantly improves prediction performance and stability but is also easy to deploy in existing industrial systems, possessing good versatility and promotional value. Therefore, it solves the problem that related technologies, lacking a spatial modeling mechanism, can only perceive temporal information at a single scale, leading to unstable prediction accuracy and weak generalization ability when facing multi-variable coupling and long-distance dependencies.

[0034] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0035] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a schematic diagram of a continuous catalytic reforming production process according to an embodiment of this application; Figure 2 This is a flowchart of a method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network according to an embodiment of this application; Figure 3This is a schematic diagram of the yield characteristics of benzene, toluene, and xylene provided according to an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a multi-scale spatiotemporal attentive–gated long short-term memory (MST-AGL) network model provided according to an embodiment of this application; Figure 5 This is a block diagram of a benzene, toluene, and xylene yield prediction device based on a multi-scale spatiotemporal attention-gated long short-term memory network according to an embodiment of this application. Figure 6 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application.

[0036] Figure label: 50 - A device for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network; 100 - Acquisition module, 200 - Establishment module, 300 - Extraction module, 400 - Prediction module; 601 - Memory, 602 - Processor, 603 - Communication interface. Detailed Implementation

[0037] The embodiments of this application are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0038] The following description, with reference to the accompanying drawings, describes a method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network (LSTMN) according to embodiments of this application. Addressing the issues mentioned in the background art, where the lack of spatial modeling mechanisms and the ability to perceive temporal information only at a single scale leads to unstable prediction accuracy and weak generalization ability when faced with multivariate coupling and long-distance dependencies, this application provides a method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated LTMN. This method overcomes the limitations of traditional single-temporal modeling methods by innovatively integrating a graph attention network model, a pyramid attention mechanism module, and an LTMN model to construct a deep learning architecture with spatiotemporal feature fusion capabilities. The graph attention network model extracts complex spatial interactions between process variables, the pyramid attention mechanism module captures multi-scale temporal dependency structures, and the LTMN model strengthens the modeling of long-term temporal patterns. The three work synergistically to achieve accurate prediction of benzene, toluene, and xylene yields under complex operating conditions. This not only significantly improves prediction performance and stability but is also easy to deploy in existing industrial systems, possessing good versatility and promotional value. This solves the problem that related technologies, lacking spatial modeling mechanisms, can only perceive temporal information at a single scale, resulting in unstable prediction accuracy and weak generalization ability when faced with multivariate coupling and long-distance dependence.

[0039] Before introducing the method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network proposed in the embodiments of this application, we will first briefly introduce the continuous catalytic reforming production process.

[0040] like Figure 1 As shown, mixed naphtha is mixed with hydrogen and enters a pre-hydrogenation reactor. Impurities such as sulfur and nitrogen are removed through a hydrogenation reaction, completing the raw material refining. Then, it enters a stripping tower to remove light components from the pre-hydrogenated material. Next, it enters a pre-fractionation tower to perform preliminary fractionation of the material, adjust the fraction range, and obtain refined naphtha.

[0041] Refined naphtha enters an improved feed heat exchanger for preheating to improve energy efficiency. It then enters reactors 1-4, which consist of a four-in-one heating furnace, where it undergoes reactions such as cycloalkanes dehydrogenation and alkane aromatization to produce aromatics (benzene, toluene, xylene, etc.). The naphtha then enters an improved gas-liquid separator, where it is separated into gas and liquid phases. The gas phase consists of recycled hydrogen, which is pressurized by a recycled hydrogen compressor and recycled back to the reaction system to continue participating in the reaction. The liquid phase consists of a mixture of hydrocarbons produced by the reaction.

[0042] The liquid phase enters the deasphalting tower to separate the heavy components in the liquid phase product, and then enters the reforming oil separator to separate the material into two paths. One path enters the extraction unit to separate non-aromatic hydrocarbons, benzene, and toluene products, while the other path enters the xylene tower to separate xylene products. Heavy aromatic hydrocarbons are obtained at the bottom of the tower.

[0043] Specifically, Figure 2 This is a flowchart of a method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network, according to an embodiment of this application.

[0044] like Figure 2 As shown, the method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network includes the following steps: In step S201, process variables in the continuous catalytic reforming production process are collected and normalized to obtain normalized process variables.

[0045] In the embodiments of this application, process variables can be understood as core parameters in chemical production processes (such as the production of benzene, toluene, and xylene) that can be directly measured, controlled, or indirectly reflect the production status. These parameters may include, but are not limited to, reaction temperature, operating pressure, raw material flow rate, and raw material component ratio. They are key factors used to characterize the production process and drive changes in product yield.

[0046] Furthermore, normalized process variables can be understood as parameters obtained after standardizing the original process variables. For example, by using certain mathematical methods (such as linear mapping, Z-score standardization, etc.), the dimensional differences (such as temperature unit ℃ and flow rate unit m³ / h) and the differences in numerical range between different process variables are eliminated, so that each variable is on a uniform numerical scale and has comparability, which is used to avoid the imbalance of feature weights caused by inconsistent scales of the original variables.

[0047] In actual implementation, this application embodiment collects process variables from the continuous catalytic reforming production process to serve as model inputs. Specifically, the process variables include 24 categories, categorized into raw material properties, feed operation, reactor process, and separation and control.

[0048] Among them, the feed properties include CCR2_YL_ZGWT (Naphtha feed n-alkanes content, naphtha feed n-alkanes content), CCR2_YL_YGWT (Naphtha feed iso-paraffins content, naphtha feed iso-alkanes content), CCR2_YL_HWT (Naphtha feed naphthenes content, naphtha feed cycloalkanes content), CCR2_YL_FT (Naphtha feed aromatics content, naphtha feed aromatics content), CCR2_YL_CLD (Naphtha feed initial boiling point, naphtha feed initial boiling point), CCR2_YL_10 (Naphtha feed 10% distillation temperature, naphtha feed 10% distillation temperature), CCR2_YL_50 (Naphtha feed 50% distillation temperature, naphtha feed 50% distillation temperature), and CCR2_YL_ZLD (Naphtha feed final boiling point). temperature, naphtha feed final boiling point temperature).

[0049] Feed operation categories include CCR2_FT4050T_PV (Feed (hot naphtha) temperature), CCR2_FQ4082_PV (Reforming naphtha feed flow control), CCR2_FT4050D_PV (Feed (hot naphtha) density), and CCR2_TI4002_PV (First reactor inlet temperature).

[0050] Reactor process classes include CCR2_TC4001_PV (F1401 outlet temperature), CCR2_TC4003_PV (F1402 outlet temperature), CCR2_TC4005_PV (F1403 outlet temperature), CCR2_TC4007_PV (F1404 outlet temperature), CCR2_H2HC_C_CPV (Reforming reaction hydrogen-to-oil ratio), CCR2_TC4033_PV (C1401 tray 8 temperature), CCR2_TC4044_PV (C1402 tray 22 temperature), and CCR2_TDC4057_PV (C1403 temperature). (C1403 tower temperature difference), AHU_TI4220-3_PV (C2501 tray 31 temperature).

[0051] The separation and control categories include AHU_TDC4257_PV (C-2531 plate 5-20 temperature difference indication control), AHU_TC4231-4_PV (Solvent recovery column bottom temperature), and AHU_TDC4268_5_PV (Toluene column temperature difference control).

[0052] Furthermore, in one embodiment of this application, the process variables are normalized to obtain normalized process variables. The calculation formula for the normalized process variables may be, but is not limited to, the following: , in, For the sample index of process variables, ; The total sample size for process variables; This serves as an index for process variables, used to distinguish different process variables, such as temperature, pressure, and hydrogen-to-oil ratio. To normalize process variables; For the first The first process variable A sample, such as a specific measurement of temperature; For the first The maximum value of each process variable among all samples. For example, the highest value among all measurements of temperature; For the first The minimum value among all samples of each process variable. For example, the lowest value among all temperature measurements. It should be noted that... .

[0053] In actual implementation, the embodiments of this application are based on a linear transformation mechanism of max-min normalization, with the first... Based on the maximum and minimum values ​​of the sample for the nth process variable, the nth process variable is used as the benchmark. Each sample is mapped to the [0,1] interval to achieve standardized transformation of process variables with different dimensions and numerical ranges, thereby eliminating the interference of dimensional differences on the subsequent modeling process.

[0054] This application's embodiments, through normalization processing, ensure that various process variables are comparable on a numerical scale. This guarantees that when constructing spatial dimension graphs based on graph attention network models, constructing temporal dimension graphs based on pyramid attention mechanism modules, and using long short-term memory network models for yield prediction, each variable can be fairly and effectively learned by the model to contribute its features. This avoids the model's excessive focus on variables with large numerical ranges due to inconsistent dimensions, thereby improving the accuracy and stability of benzene, toluene, and xylene yield predictions and providing reliable data support for process optimization and production decisions.

[0055] Furthermore, the model outputs predicted yields of benzene, toluene, and xylene. Although the outputs (benzene, toluene, and xylene yields) are all yield indicators, their numerical ranges may differ due to the influence of process conditions (e.g., benzene yields are concentrated in the 10%-20% range, and toluene yields are concentrated in the 20%-30% range). If the raw yield values ​​are directly used as the model output target, the loss function will be more sensitive to yield variables with larger numerical ranges, causing the model to prioritize fitting these types of variables during training and ignore variables with smaller numerical ranges, resulting in learning bias.

[0056] Therefore, in one embodiment of this application, the predicted yield values ​​of benzene, toluene, and xylene are normalized in the form of an output mode vector to obtain a normalized output mode vector. The formula for calculating the normalized output mode vector can be, but is not limited to, the following: , in, For sample index, To normalize the output pattern vector, To output the pattern vector, To output the minimum value among all samples of the pattern vector, This is the maximum value of the output variable across all samples.

[0057] It can be understood from the description of the above embodiments that, as Figure 3 The diagram illustrates the process variables and their actual yields (i.e., benzene, toluene, and xylene yields) during continuous catalytic reforming production (data derived from historical operating data). The process variables serve as the initial input to the model, while the actual yields serve as the model's labeled data. Together, these variables and yields constitute the model's initial dataset, which can be divided into training and testing sets to provide data support for subsequent model training and performance validation.

[0058] In step S202, a spatial dimension graph reflecting the spatial dependencies of the normalized process variables is established based on the normalized process variables and the graph attention network model, and a temporal dimension graph reflecting the temporal characteristics of the normalized process variables is established based on the normalized process variables and the pyramid attention mechanism module.

[0059] like Figure 4 As shown, the structure of the MST-AGL model can include, but is not limited to, graph attention network structure, pyramid attention mechanism structure, and long short-term memory network structure. Specifically, the graph attention network structure corresponds to the graph attention network model, the pyramid attention mechanism structure corresponds to the pyramid attention mechanism module, and the long short-term memory network structure corresponds to the long short-term memory network model.

[0060] In the embodiments of this application, the graph attention network model can be understood as a model that can effectively model the complex spatial relationship structure between parameters by introducing a self-attention mechanism to dynamically weight the relationship between variables, and has the ability to adapt to high-dimensional, non-Euclidean spatial data.

[0061] Furthermore, the pyramid attention module can be understood as a model that captures deep dependencies in the time dimension through multi-scale modeling, possessing the ability to model remote dependencies and local changes, as well as generalization performance in complex chemical environments.

[0062] In actual implementation, the embodiments of this application use a normalized process variable input graph attention network model, which can take normalized process variables (such as temperature, pressure, and hydrogen-to-oil ratio) as nodes and physical relationships (such as temperature increase leading to pressure change) or statistical relationships (such as aromatic yield fluctuating with hydrogen-to-oil ratio) between variables as edges. Based on the nodes and edges, an adjacency matrix is ​​constructed to quantify the relationship between nodes and edges, thereby establishing a spatial dimension graph that reflects the spatial dependence of normalized process variables.

[0063] In addition, since process variables change over time, this application embodiment continuously collects process variables to form time-series data (such as recording temperature and flow rate once every minute). The time-series data is then input into the pyramid attention mechanism module, and features of different time granularities are extracted through multi-scale convolution (such as fine granularity, which can be the temperature fluctuation of the most recent minute; medium granularity, which can be the temperature trend of the past hour; and coarse granularity, which can be the temperature cycle of the past day). The features of different time granularities are then arranged in a hierarchy from fine to coarse over time (fine granularity at the bottom layer and coarse granularity at the top layer) to establish a time dimension diagram that reflects the time characteristics of normalized process variables.

[0064] In step S203, the spatial features of the normalized process variables are extracted based on the spatial dimension map, and the temporal features of the normalized process variables are extracted based on the time dimension map.

[0065] In the embodiments of this application, spatial features can be understood as feature representations extracted from normalized process variables through graph attention networks, reflecting the complex spatial relationships between various process variables in the chemical production process, covering the coupling effects, dependencies and interaction patterns between variables, and this application does not impose specific limitations.

[0066] Furthermore, the time features can be understood as multi-scale features extracted from the time-series data of normalized process variables through the pyramid attention mechanism, reflecting the evolution of process variables and the yields of benzene, toluene and xylene over time. These features cover the short-term fluctuations, medium-term trends, long-term dependence and cross-scale and same-scale time correlation patterns of the variables. This application does not impose any specific limitations.

[0067] The following details how embodiments of this application extract spatial features of normalized process variables based on a spatial dimension map.

[0068] Specifically, in one embodiment of this application, the spatial features of the normalized process variables are extracted based on the spatial dimension graph, including: weighting and summing each node and its adjacent nodes in the spatial dimension graph according to a first attention weight to obtain a first weighted summation result; extracting multiple spatial features of the normalized process variables based on the first weighted summation result, and generating a spatial feature matrix based on the multiple spatial features.

[0069] In the embodiments of this application, the nodes of the spatial dimension graph can be understood as the basic feature units of the spatial dimension graph constructed based on the graph attention network, corresponding to a single normalized single process variable.

[0070] Furthermore, the adjacent nodes of the spatial dimension diagram can be understood as feature units that have process coupling and feature interaction relationships with the target node in the spatial dimension diagram. They can be defined based on the inherent mechanism of the chemical process and the variable correlation characteristics reflected by the actual operating data. This application does not impose specific restrictions.

[0071] In addition, the first attention weight can be understood as the spatial correlation between a node and its neighboring nodes in the spatial dimension graph obtained by using the multi-head attention mechanism. For example, in the spatial dimension graph, when the target node is the feed (hot naphtha) temperature, and the neighboring nodes of the target node are reforming naphtha feed flow control, feed (hot naphtha) density, and first reactor inlet temperature, the first attention weights obtained after calculation by the multi-head attention mechanism are 0.6, 0.3, and 0.1 respectively. This indicates that the spatial correlation between reforming naphtha feed flow control and feed (hot naphtha) temperature is the strongest.

[0072] Furthermore, the first weighted summation result can be understood as an intermediate quantization result obtained by weighting and summing each node and its adjacent nodes in the spatial dimension graph according to the first attention weight.

[0073] In actual implementation, this embodiment uses each normalized process variable in the spatial dimension graph as a node. It utilizes a multi-head attention mechanism to quantify the first attention weight of each node and its neighboring nodes. Then, it performs a weighted summation on the node and its neighboring nodes based on the first attention weight, aggregates the surrounding association information of the node, and obtains the first weighted summation result. Based on the first weighted summation result, it extracts multiple spatial features for spatial interaction relationships (such as direct association and indirect association) in different dimensions, and integrates multiple spatial features into a spatial feature matrix. This allows the system to capture the nonlinear spatial dependency relationship between process variables, providing a structured spatial information carrier for subsequent feature fusion.

[0074] This application embodiment extracts features from multiple perspectives through a multi-head attention mechanism, dynamically quantifies the spatial dependence strength between various process variables, and generates a spatial feature matrix that fully preserves the nonlinear spatial correlation information between process parameters. This provides key spatial dimension feature support for subsequent yield prediction of benzene, toluene, and xylene, effectively improving the ability of the prediction model to fit complex spatial interaction relationships, thereby significantly improving the accuracy and reliability of yield prediction.

[0075] Optionally, in one embodiment of this application, the formula for calculating the spatial feature matrix may be, but is not limited to, the following: , in, For the first Layer nodes No. The spatial feature matrix output by each attention head For the first The output of each attention head is spliced ​​together. For activation function, For nodes The set of adjacent nodes, For the first Layer nodes For adjacent nodes Attention weights For the first Layer Adjacent Nodes Its characteristics.

[0076] In actual implementation, after constructing the spatial dimension graph, the embodiments of this application use each process variable as a node and employ a multi-head attention mechanism to process the set of adjacent nodes. Middle Adjacent Nodes Features By attention weight Perform weighted summation, followed by activation function After introducing nonlinearity, for The output of each attention point is spliced ​​together. This allows us to learn the nonlinear spatial interaction relationships between process variables and generate a spatial feature matrix. This enables the modeling of the spatial dependencies of process parameters.

[0077] This application embodiment effectively integrates multi-perspective spatial correlation information by using the calculation formula of the spatial feature matrix, the splicing operation of the multi-head attention mechanism, and the weighted aggregation of the features of adjacent nodes. This enhances the distinguishability of spatial dependencies between nodes, enabling the spatial feature matrix to accurately depict the details of nonlinear spatial interactions between nodes. This provides more hierarchical spatial feature support for subsequent yield prediction models and improves the model's fitting accuracy to the spatial correlation of process variables.

[0078] Optionally, in one embodiment of this application, the formula for calculating the attention weight may be, but is not limited to: , in, For nodes For adjacent nodes Attention weights It is an exponential function. For a leaky linear rectified activation function, For nodes The input feature vector, For nodes The input feature vector, For nodes The input feature vector, It is a learnable linear transformation matrix. This is a learnable attention weight vector. For splicing operations, For nodes The set of adjacent nodes.

[0079] The following details how embodiments of this application extract the time features of normalized process variables based on a time dimension graph.

[0080] Specifically, in one embodiment of this application, extracting time features of normalized process variables based on a time dimension graph includes: performing weighted summation on each node and adjacent nodes of the time dimension graph according to a second attention weight to obtain a second weighted summation result; extracting multiple time features of normalized process variables based on the second weighted summation result, and generating a time feature vector based on the multiple time features.

[0081] In the embodiments of this application, the nodes of the time dimension graph can be understood as the basic feature units of the time dimension graph constructed based on the pyramid attention mechanism, corresponding to different time granularity features extracted through multi-scale convolution operations (such as fine-grained short-term fluctuations, medium-grained time-period trends, and coarse-grained long-term changes).

[0082] Furthermore, the adjacent nodes of the time dimension graph can be understood as feature units that have a direct temporal relationship with the target node in the time dimension graph. They can reflect the hierarchical dependence and intra-scale association between time granularities through inter-scale and intra-scale connections. This application does not impose any specific limitations.

[0083] In addition, the second attention weight can be understood as using the scaled dot product attention mechanism to quantify the temporal correlation between a node and its neighboring nodes in the obtained time dimension graph.

[0084] Furthermore, the second weighted summation result can be understood as an intermediate quantization result obtained by weighting and summing each node and its adjacent nodes in the time dimension graph according to the second attention weight.

[0085] In actual implementation, after constructing the time dimension graph, this embodiment calculates the second attention weight of each node and its neighboring nodes through a scaling dot product attention mechanism. Based on the second attention weight, the nodes and their neighboring nodes are weighted and summed to aggregate the time dimension association information and obtain the second weighted sum result. Then, based on the second weighted sum result, multiple time features are extracted for dynamic patterns at different time scales (such as short-term mutations and long-term trends), and multiple time features are integrated into a time feature vector. Thus, the system captures the multi-scale time dependencies in time series data and provides a structured time information carrier for subsequent feature fusion.

[0086] This application's embodiments, through cross-scale and same-scale attention weight aggregation, can specifically capture the correlation between features at different time granularities, dynamically focus on key time segments for yield prediction, and effectively filter out irrelevant temporal noise. In addition, through the extraction of multi-scale time features and the generation of time feature vectors, the multi-level dynamic laws of process variables evolving over time are fully preserved, enabling subsequent models to fully utilize the temporal correlation of time series data, enhance the fitting ability to dynamic changes in the continuous catalytic reforming process, provide more accurate time dimension support for benzene, toluene, and xylene yield prediction, and improve the reliability of prediction results.

[0087] Optionally, in one embodiment of this application, the formula for calculating the time feature vector may be, but is not limited to, the following: , in, For nodes Time feature vector, For scale The next node The set of adjacent nodes, For nodes The query vector, Adjacent nodes The key vector, Adjacent nodes The value vector, is the dimension of the key vector.

[0088] In actual implementation, the embodiments of this application are based on a scaled dot product attention mechanism, using a time-dimensional graph as a foundation and a scale... The next node The set of adjacent nodes To process objects, through nodes query vector With adjacent nodes key vector Perform a dot product operation, along the key vector dimension. After scaling by the square root, attention weights are generated using exponential function normalization to quantify the temporal correlation strength between neighboring nodes and the current node. These normalized weights are then used to assign attention weights to neighboring nodes. value vector Weighted summation to generate nodes Time feature vector This allows us to capture deep structural dependencies across multiple scales in time series data.

[0089] This application embodiment accurately quantifies the temporal correlation strength between nodes and adjacent nodes through the calculation formula of time feature vectors, realizes dynamic weighted aggregation of adjacent node value vectors, effectively enhances the distinguishability and correlation of multi-scale time features, accurately captures deep temporal dependency details across scales and within the same scale in time series data, injects more targeted temporal correlation information into time feature vectors, and improves the utilization efficiency and prediction accuracy of subsequent yield prediction models for time dimension features.

[0090] In step S204, a weighted fusion is performed based on spatial and temporal features to generate fused features. The fused features are then input into a pre-built long short-term memory network model to output predicted yields of benzene, toluene, and xylene.

[0091] In the embodiments of this application, the fusion feature can be understood as a comprehensive feature representation formed by weighting and fusing the spatial features extracted by the graph attention network model and the temporal features extracted by the pyramid attention mechanism module using a dynamic weight allocation strategy.

[0092] In addition, the Long Short-Term Memory (LSTM) network model can be understood as a recurrent neural network model that incorporates structures such as forget gates, input gates, and output gates to alleviate the gradient vanishing problem.

[0093] In actual implementation, the embodiments of this application can perform weighted fusion of spatial and temporal features based on spatial and temporal features through a dynamic weight allocation mechanism (the weights are adaptively adjusted according to the significance of the influence of spatial and temporal features on the yields of benzene, toluene, and xylene) to generate fused features containing comprehensive information on spatial interaction and temporal evolution. Then, the fused features are input into a pre-constructed long short-term memory network model, and its gating mechanism (forget gate, input gate, output gate) is used to model the long-term temporal dependence implied in the fused features. Through multi-layer nonlinear transformation, it is mapped to the yield prediction space, thereby outputting the predicted yield values ​​of benzene, toluene, and xylene, realizing comprehensive feature modeling and accurate prediction of complex processes.

[0094] In summary, after obtaining the predicted yields of benzene, toluene, and xylene, the embodiments of this application compare the model predictions with the actual yields and calculate evaluation indicators such as MAE (Mean Absolute Error), RMSE (Root Mean Squared Error), and MAPE (Mean Absolute Percentage Error) to comprehensively evaluate the prediction performance.

[0095] Meanwhile, the embodiments of this application compare the provided MST-AGL model with several mainstream models such as LTM (Long Short-Term Memory), GRU (Gated Recurrent Unit), Attention-LSTM (Attention-Based Long Short-Term Memory), and GAT-LSTM (Graph Attention Network - Long Short-Term Memory) to verify its accuracy and superiority in the yield prediction task. The specific results are shown in Table 1.

[0096] In addition, to verify the generalization ability of the model, this application embodiment conducted transfer experiments on two UCI open-source datasets (EnergyEfficiency and Concrete Strength). Through cross-domain testing, the applicability and stability of the proposed model in different task scenarios were evaluated. The relevant results are shown in Tables 2 and 3.

[0097] Experimental results show that traditional single models (such as LSTM, GRU, Attention-LSTM, etc.) exhibit significant performance fluctuations and unstable prediction errors during different training processes. In contrast, the MST-AGL model proposed in this application significantly improves model stability and generalization ability by integrating multi-scale temporal features and spatial interaction relationships. Experimental results demonstrate that the MST-AGL model outperforms other comparative models on multiple evaluation metrics, including MAE, RMSE, and MAPE, exhibiting the smallest error and minimal fluctuations, demonstrating stronger prediction accuracy and robustness. This proves the practicality and superiority of the benzene, toluene, and xylene yield prediction method proposed in this application for complex industrial processes.

[0098] Table 1 Comparison of prediction models based on the MST-AGL model

[0099] Table 2 Energy Efficiency Dataset Validation

[0100] Table 3 Validation of Concrete Compressive Strength Dataset

[0101] The principle of the benzene, toluene, and xylene yield prediction method based on a multi-scale spatiotemporal attention-gated long short-term memory network proposed in this application is illustrated below with a specific embodiment.

[0102] In this embodiment, process variables in the continuous catalytic reforming production process are collected and used as model inputs. The process variables are then normalized to obtain normalized process variables. The normalized data is then input into a graph attention network model to establish a spatial dimension graph reflecting the spatial dependencies of the normalized variables. The normalized process variables are then input into a pyramid attention mechanism module to establish a temporal dimension graph reflecting the temporal characteristics of the normalized variables.

[0103] Subsequently, in this embodiment of the application, each process variable in the spatial dimension graph is used as a node. A multi-head attention mechanism is used to quantify the importance of the association between each node and its neighboring nodes to obtain a first attention weight. Then, the nodes and their neighboring nodes are weighted and summed according to the first attention weight to aggregate the spatial dimension association information and obtain a first weighted summation result. Then, based on the first weighted summation result, multiple spatial features are extracted for spatial interaction relationships in different dimensions, and the multiple spatial features are integrated into a spatial feature matrix according to a certain dimension.

[0104] Meanwhile, after constructing the time dimension graph, this embodiment calculates the importance of each node's association with its neighboring nodes through a scaling dot product attention mechanism to obtain a second attention weight. Then, it performs a weighted summation on the node and its neighboring nodes based on the second attention weight to aggregate the time dimension association information and obtain a second weighted summation result. Based on the second weighted summation result, it extracts multiple time features for the dynamic patterns of different time scales and integrates these features into a time feature vector.

[0105] Furthermore, in this embodiment, spatial and temporal features can be weighted and fused using a dynamic weight allocation mechanism to generate a fused feature containing comprehensive information on spatial interaction and temporal evolution. This fused feature is then input into a pre-constructed long short-term memory network model, which uses its gating mechanism to model the long-term temporal dependence implied in the fused feature. The model is then mapped to the yield prediction space through multi-layer nonlinear transformation, thereby outputting the yield prediction values ​​of benzene, toluene, and xylene.

[0106] The benzene, toluene, and xylene yield prediction method based on a multi-scale spatiotemporal attention-gated long short-term memory network proposed in this application overcomes the limitations of traditional single-time-series modeling methods. It innovatively integrates a graph attention network model, a pyramid attention mechanism module, and a long short-term memory network model to construct a deep learning architecture with spatiotemporal feature fusion capabilities. The graph attention network model extracts complex spatial interactions between process variables, the pyramid attention mechanism module captures multi-scale temporal dependency structures, and the long short-term memory network model strengthens the modeling of long-term temporal patterns. The three work synergistically to achieve accurate prediction of benzene, toluene, and xylene yields under complex operating conditions. This not only significantly improves prediction performance and stability but is also easy to deploy in existing industrial systems, possessing good versatility and promotional value. Therefore, it solves the problem that related technologies, lacking a spatial modeling mechanism, can only perceive temporal information at a single scale, leading to unstable prediction accuracy and weak generalization ability when facing multi-variable coupling and long-distance dependencies.

[0107] Next, referring to the accompanying drawings, a benzene, toluene, and xylene yield prediction device based on a multi-scale spatiotemporal attention-gated long short-term memory network, according to an embodiment of this application, is described.

[0108] Figure 5 This is a block diagram of a benzene, toluene, and xylene yield prediction device based on a multi-scale spatiotemporal attention-gated long short-term memory network provided in an embodiment of this application.

[0109] like Figure 5 As shown, the benzene, toluene and xylene yield prediction device 50 based on a multi-scale spatiotemporal attention-gated long short-term memory network includes: an acquisition module 100, an establishment module 200, an extraction module 300 and a prediction module 400.

[0110] The acquisition module 100 is used to acquire process variables in the continuous catalytic reforming production process and normalize the process variables to obtain normalized process variables.

[0111] Module 200 is established to create a spatial dimension graph reflecting the spatial dependencies of normalized process variables based on normalized process variables and a graph attention network model, and to create a temporal dimension graph reflecting the temporal characteristics of normalized process variables based on normalized process variables and a pyramid attention mechanism module.

[0112] The extraction module 300 is used to extract the spatial features of normalized process variables based on the spatial dimension map and the temporal features of normalized process variables based on the time dimension map.

[0113] The prediction module 400 is used to perform weighted fusion based on spatial and temporal features to generate fused features. The fused features are then input into a pre-built long short-term memory network model to output predicted yields of benzene, toluene, and xylene.

[0114] Optionally, in one embodiment of this application, the extraction module 300 includes a first calculation unit and a first generation unit.

[0115] The first calculation unit is used to perform a weighted summation on each node and its adjacent nodes in the spatial dimension graph according to a first attention weight, so as to obtain a first weighted summation result.

[0116] The first generation unit is used to extract multiple spatial features of the normalized process variables based on the first weighted summation result, and generate a spatial feature matrix based on the multiple spatial features.

[0117] Optionally, in one embodiment of this application, the formula for calculating the spatial feature matrix may be, but is not limited to, the following: , in, For the first Layer nodes No. The spatial feature matrix output by each attention head For the first The output of each attention head is spliced ​​together. For activation function, For nodes The set of adjacent nodes, For the first Layer nodes For adjacent nodes Attention weights For the first Layer Adjacent Nodes Its characteristics.

[0118] Optionally, in one embodiment of this application, the extraction module 300 includes: a second calculation unit and a second generation unit.

[0119] The second calculation unit is used to perform a weighted summation on each node and its adjacent nodes in the time dimension graph according to the second attention weight, so as to obtain the second weighted summation result.

[0120] The second generation unit is used to extract multiple time features of the normalized process variables based on the second weighted summation result, and generate a time feature vector based on the multiple time features.

[0121] Optionally, in one embodiment of this application, the formula for calculating the time feature vector may be, but is not limited to, the following: , in, For nodes Time feature vector, For scale The next node The set of adjacent nodes, For nodes The query vector, Adjacent nodes The key vector, Adjacent nodes The value vector, is the dimension of the key vector.

[0122] Optionally, in one embodiment of this application, the data acquisition module 100 may use, but is not limited to, the formula for calculating the normalized process variables as follows: , in, For the sample index of process variables, The total sample size for process variables. For indexing process variables, To normalize process variables, For the first The first process variable One sample, For the first The maximum value of each process variable among all samples. For the first The minimum value among all samples of a process variable.

[0123] It should be noted that the foregoing explanation of the embodiment of the benzene, toluene and xylene yield prediction method based on multi-scale spatiotemporal attention-gated long short-term memory network also applies to the benzene, toluene and xylene yield prediction device based on multi-scale spatiotemporal attention-gated long short-term memory network in this embodiment, and will not be repeated here.

[0124] The benzene, toluene, and xylene yield prediction device based on a multi-scale spatiotemporal attention-gated long short-term memory network proposed in this application overcomes the limitations of traditional single-time-series modeling methods. It innovatively integrates a graph attention network model, a pyramid attention mechanism module, and a long short-term memory network model to construct a deep learning architecture with spatiotemporal feature fusion capabilities. The graph attention network model extracts complex spatial interactions between process variables, the pyramid attention mechanism module captures multi-scale time-dependent structures, and the long short-term memory network model strengthens the modeling of long-term time-series patterns. The three work synergistically to achieve accurate prediction of benzene, toluene, and xylene yields under complex operating conditions. This not only significantly improves prediction performance and stability but is also easy to deploy in existing industrial systems, possessing good versatility and promotional value. Therefore, it solves the problem that related technologies, lacking a spatial modeling mechanism, can only perceive temporal information at a single scale, leading to unstable prediction accuracy and weak generalization ability when facing multi-variable coupling and long-distance dependencies.

[0125] Figure 6 This is a schematic diagram of the structure of an electronic device provided according to an embodiment of this application. The electronic device may include: The memory 601, the processor 602, and the computer program stored on the memory 601 and capable of running on the processor 602.

[0126] When the processor 602 executes the program, it implements the benzene, toluene and xylene yield prediction method based on a multi-scale spatiotemporal attention-gated long short-term memory network provided in the above embodiments.

[0127] Furthermore, electronic devices also include: Communication interface 603 is used for communication between memory 601 and processor 602.

[0128] The memory 601 is used to store computer programs that can run on the processor 602.

[0129] The memory 601 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0130] If the memory 601, processor 602, and communication interface 603 are implemented independently, then the communication interface 603, memory 601, and processor 602 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0131] Optionally, in a specific implementation, if the memory 601, processor 602, and communication interface 603 are integrated on a single chip, then the memory 601, processor 602, and communication interface 603 can communicate with each other through an internal interface.

[0132] The processor 602 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0133] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network.

[0134] This application also provides a computer program product, including a computer program that, when executed, implements the above-described method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network.

[0135] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0136] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0137] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0138] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other media on which the program can be printed, because the program can be obtained electronically by optically scanning the paper or other media, then editing, interpreting, or otherwise processing it as appropriate, and then stored in computer memory.

[0139] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0140] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0141] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0142] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network, characterized in that, Includes the following steps: Process variables in the continuous catalytic reforming production process are collected and normalized to obtain normalized process variables. Based on the normalized process variables and the graph attention network model, a spatial dimension graph reflecting the spatial dependencies of the normalized process variables is established, and based on the normalized process variables and the pyramid attention mechanism module, a temporal dimension graph reflecting the temporal characteristics of the normalized process variables is established. Based on the spatial dimension map, the spatial features of the normalized process variables are extracted, and based on the temporal dimension map, the temporal features of the normalized process variables are extracted. The spatial and temporal features are weighted and fused to generate fused features. The fused features are then input into a pre-constructed long short-term memory network model to output predicted yields of benzene, toluene, and xylene.

2. The method according to claim 1, characterized in that, The step of extracting the spatial features of the normalized process variables based on the spatial dimension map includes: Each node and its adjacent nodes in the spatial dimension graph are summed using a first attention weight to obtain a first weighted summation result; Based on the first weighted summation result, multiple spatial features of the normalized process variables are extracted, and a spatial feature matrix is ​​generated based on the multiple spatial features.

3. The method according to claim 2, characterized in that, The formula for calculating the spatial feature matrix is: ， in, For the first Layer nodes No. The spatial feature matrix output by each attention head For the first The output of each attention head is spliced ​​together. For activation function, For nodes The set of adjacent nodes, For the first Layer nodes For adjacent nodes Attention weights For the first Layer Adjacent Nodes Its characteristics.

4. The method according to claim 1, characterized in that, The step of extracting the time features of the normalized process variables based on the time dimension diagram includes: Each node and its adjacent nodes in the time dimension graph are weighted and summed according to the second attention weight to obtain the second weighted summation result; Based on the second weighted summation result, multiple time features of the normalized process variables are extracted, and a time feature vector is generated based on the multiple time features.

5. The method according to claim 4, characterized in that, The formula for calculating the time feature vector is: ， in, For nodes Time feature vector, For scale The next node The set of adjacent nodes, For nodes The query vector, Adjacent nodes The key vector, Adjacent nodes The value vector, is the dimension of the key vector.

6. The method according to claim 1, characterized in that, The process variables are then normalized to obtain normalized process variables, wherein the formula for calculating the normalized process variables is as follows: ， in, For the sample index of process variables, The total sample size for process variables. For indexing process variables, To normalize process variables, For the first The first process variable One sample, For the first The maximum value of each process variable among all samples. For the first The minimum value among all samples of a process variable.

7. A benzene, toluene, and xylene yield prediction device based on a multi-scale spatiotemporal attention-gated long short-term memory network, characterized in that, include: The data acquisition module is used to acquire process variables in the continuous catalytic reforming production process and normalize the process variables to obtain normalized process variables. A module is established to build a spatial dimension graph reflecting the spatial dependencies of the normalized process variables based on the normalized process variables and the graph attention network model, and to build a temporal dimension graph reflecting the temporal characteristics of the normalized process variables based on the normalized process variables and the pyramid attention mechanism module. The extraction module is used to extract the spatial features of the normalized process variables based on the spatial dimension map, and to extract the temporal features of the normalized process variables based on the time dimension map. The prediction module is used to perform weighted fusion based on the spatial features and the temporal features to generate fused features, and input the fused features into a pre-constructed long short-term memory network model to output predicted yields of benzene, toluene and xylene.

8. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and capable of running on the processor, wherein the processor executes the program to implement the method for predicting the yields of benzene, toluene, and xylene based on a multi-scale spatiotemporal attention-gated long short-term memory network as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, The program is executed by the processor to implement the benzene, toluene, and xylene yield prediction method based on a multi-scale spatiotemporal attention-gated long short-term memory network as described in any one of claims 1-6.

10. A computer program product, comprising a computer program, characterized in that, The computer program is executed to implement the benzene, toluene, and xylene yield prediction method based on a multi-scale spatiotemporal attention-gated long short-term memory network as described in any one of claims 1-6.

Images