A method, equipment and medium for constructing a large-scale chemical production model

By constructing a multi-scale graph network and fusing multi-source time-series data into a large-scale chemical production model, the problem of balancing physical interpretability and dynamic modeling in chemical production processes was solved, thereby improving the physical reliability and generalization ability of the model.

CN123116885APending Publication Date: 2026-11-17SHANDONG JINGBO POLYOLEFIN NEW MATERIALS CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611620474.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-20
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing chemical production process modeling methods struggle to balance physical interpretability with complex dynamic modeling capabilities, resulting in unreliable and non-transferable model predictions.

Method used

A multi-scale graph network is constructed for structured representation, multi-source time series data is integrated, and physical mechanism constraints of soft and hard constraints are introduced to form a large model of chemical production.

Benefits of technology

This approach achieves a unified representation of the multi-scale spatial characteristics and dynamic modeling of chemical production models, ensuring that the model prediction results follow physical laws and improving the physical reliability and generalization ability of the models.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123116885A_ABST
    Figure CN123116885A_ABST
Patent Text Reader

Abstract

This application discloses a method, equipment, and medium for constructing a large-scale chemical production model, relating to the field of electronic digital data processing technology. The method includes: performing structured representation on a multi-scale graph network to obtain corresponding multi-scale structure embedding vectors; performing time-series modeling on multi-source time-series data generated during the chemical production process to obtain time-series embedding vectors; fusing the multi-scale structure embedding vectors and the time-series embedding vectors to obtain corresponding multi-scale fusion vectors; constructing a large-scale chemical production model embedded with physical mechanism constraints; wherein the physical mechanism constraints include soft constraint embedding and hard constraint embedding, where soft constraint embedding involves introducing a physical mechanism constraint term into the loss function, and hard constraint embedding involves using a graph neural network that satisfies physical conservation laws in the network structure of the large-scale chemical production model; and inputting the multi-scale fusion vectors into the large-scale chemical production model for training to obtain a large-scale chemical production model that integrates the multi-scale graph network and mechanism constraints.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of electronic digital data processing technology, specifically to a method, equipment, and medium for constructing a large-scale model of chemical production. Background Technology

[0002] Currently, chemical production processes are rapidly transforming towards intelligent and low-carbon technologies, placing higher demands on the accuracy, generalization ability, and physical reliability of process modeling. However, existing chemical process modeling methods still primarily rely on pure mechanistic modeling or pure data-driven modeling, making it difficult to balance physical interpretability with the ability to model complex dynamics. This results in model predictions that are neither reliable nor transferable. Summary of the Invention

[0003] To address the aforementioned issues, this application proposes a method for constructing a large-scale chemical production model, comprising: Based on the collected chemical production data, a corresponding multi-scale graph network is constructed, and the multi-scale graph network is structurally represented to obtain the corresponding multi-scale structure embedding vector. Time series modeling is performed on multi-source time series data generated in chemical production processes to obtain time series embedding vectors. The multi-scale structure embedding vectors are then fused with the time series embedding vectors to obtain the corresponding multi-scale fusion vectors. A large-scale chemical production model is constructed with embedded physical mechanism constraints. The physical mechanism constraints include soft constraint embedding and hard constraint embedding. The soft constraint embedding is to introduce physical mechanism constraint terms into the loss function, and the hard constraint embedding is to use a graph neural network that satisfies the physical conservation law in the network structure of the large-scale chemical production model. The multi-scale fusion vector is input into the large-scale chemical production model for training, resulting in a large-scale chemical production model that integrates multi-scale graph networks and mechanistic constraints.

[0004] This application provides a large-scale chemical production model construction device, characterized in that the device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform a chemical production large-scale model construction method as described in any of the preceding claims.

[0005] This application provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: A method for constructing a large-scale chemical production model, as described in any of the preceding items.

[0006] The method for constructing a large-scale chemical production model proposed in this application can bring the following beneficial effects: By employing multi-scale graph networks, the multi-scale spatial features of chemical production processes are structurally represented, and multi-source time-series data are fused to obtain multi-scale fusion vectors. This enables the model to simultaneously possess spatial structure representation and dynamic modeling capabilities, resolving the insufficient generalization ability caused by single-scale modeling. Furthermore, by introducing physical mechanism constraints including soft and hard constraints, the model's prediction results are ensured to strictly adhere to physical laws such as mass conservation and energy conservation. This makes the prediction results of the large-scale chemical production model physically reliable, providing technical support for intelligent optimization control, fault diagnosis, and safety early warning in chemical processes. Attached Figure Description

[0007] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a method for constructing a large-scale chemical production model, provided in an embodiment of this application; Figure 2 This is a schematic diagram of a large-scale chemical production model construction device provided in an embodiment of this application. Detailed Implementation

[0008] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0009] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0010] like Figure 1 As shown in the embodiments of this application, a method for constructing a large-scale chemical production model includes: S101: Based on the collected chemical production data, construct the corresponding multi-scale graph network, and perform structured representation of the multi-scale graph network to obtain the corresponding multi-scale structure embedding vector.

[0011] Chemical production processes are simultaneously influenced by reaction kinetics, transport behavior, and operating conditions. These different scales exhibit both physical coupling and significantly different spatiotemporal characteristics. To model chemical production processes uniformly, it is first necessary to collect chemical production data reflecting the operational status at each scale, and then construct a multi-scale graph network based on this data. Chemical production data refers to multi-source heterogeneous data that characterizes the operational status of the production process, specifically including process flow topology information, equipment connection relationships and logistics flow information, material chemical composition information, and process operating condition information. This data can originate from multiple data sources such as Distributed Control Systems (DCS), Laboratory Information Management Systems (LIMS), Enterprise Resource Planning (ERP) systems, and process design documents, and is collected and aggregated through a unified data interface. A multi-scale graph network is a data structure that uses a graph structure to uniformly represent objects at different scales and their interrelationships in the chemical production process. Its core idea is to encode information originally scattered across different scales and data sources in the chemical production process into nodes and edges, thereby enabling information at different scales to propagate and merge within the same framework. Specifically, the multi-scale graph network in this application includes a graph structure with three levels: microscopic molecular graph, mesoscopic particle graph, and macroscopic device graph. The three levels of graph structure correspond to modeling objects of different scales in the chemical production process, and the information is transmitted and expressed in a unified manner through a cross-scale graph alignment mechanism.

[0012] When constructing multi-scale graph networks, it is first necessary to construct corresponding graph structures for different scales. For microscopic molecular graphs, molecules participating in chemical reactions are used as modeling objects, atoms within the molecule are treated as nodes, and chemical bonds between atoms are treated as edges. Node features include information characterizing atomic chemical properties such as atom type, atomic charge, and hybridization state, while edge features include information characterizing interatomic interactions such as bond type, bond length, and bond angle. Since chemical reaction behavior at the microscopic scale has a decisive influence on reaction paths and reaction rates, graph attention networks (GAT) are used to encode microscopic molecular graphs. This allows attention weights to be adaptively allocated based on the importance of adjacent atoms during message passing, thereby more accurately capturing key structural features and reactive sites within the molecule. For mesoscopic particle graphs, particles, bubbles, or droplets in reactors or separation equipment are used as modeling objects, particles or bubbles are treated as nodes, and contact, collision, or interphase interactions between particles are treated as edges. Node features include information characterizing the particle physical state such as particle size, density, and temperature, while edge features include information characterizing interphase interactions such as contact force, mass transfer coefficient, and interphase interaction strength. Since aggregation, fragmentation, and mass transfer behaviors at the mesoscale have a significant impact on equipment-level macroscopic performance, Graph Convolutional Networks (GCNs) are used to encode mesoscopic particle graphs, effectively capturing local aggregation patterns and transfer behaviors between particles. For macroscopic equipment graphs, equipment in the chemical production process is used as the modeling object, with equipment as nodes and material flow and energy flow connections between equipment as edges. Node features include information characterizing equipment operating status such as equipment type, operating temperature, operating pressure, and throughput, while edge features include information characterizing material or energy flow characteristics such as flow rate, composition, and heat load. The macroscopic equipment graph is constructed based on a process flow diagram (PFD), which can realistically reflect the topology and equipment coupling relationships of the chemical production process, providing a structural foundation for equipment-level information dissemination and collaborative modeling.

[0013] After constructing the microscopic molecular graph, mesoscopic particle graph, and macroscopic device graph respectively, the multi-scale graph network needs to be further structurally represented. This involves encoding the graph structures at each of the three levels using a graph neural network, mapping each node and its neighborhood information into a low-dimensional, dense embedding vector. Specifically, for the microscopic molecular graph, a graph attention network is used for encoding. Through a multi-layer message passing mechanism, each atomic node can aggregate the feature information of its neighboring atoms, and the neighborhood information is weighted and fused according to attention weights, ultimately outputting the embedding vector of each atomic node. For the mesoscopic particle graph, a graph convolutional network is used for encoding. Through layer-by-layer convolution operations, each particle node can aggregate the feature information of its neighboring particles, capturing particle aggregation and mass transfer behavior, ultimately outputting the embedding vector of each particle node. For the macroscopic device graph, a graph neural network is used for encoding. Through a message passing mechanism, each device node can aggregate the state information of its upstream and downstream devices, ultimately outputting the embedding vector of each device node.

[0014] Since the three levels of graph structures reside in different scale spaces, their node and edge features differ significantly in physical meaning and numerical dimensions. Directly concatenating or simply fusing the embedding vectors of the three levels makes it difficult to achieve consistent information representation across scales. Therefore, this application introduces a cross-scale graph alignment mechanism, mapping the embedding vectors encoded from the microscopic molecular graph, mesoscopic particle graph, and macroscopic device graph to a unified embedding space. The core of the cross-scale graph alignment mechanism lies in establishing correspondences between graph structures of different scales. Specifically, there is a component affiliation relationship between molecular nodes in the microscopic molecular graph and particle nodes in the mesoscopic particle graph, and a spatial inclusion relationship between particle nodes in the mesoscopic particle graph and device nodes in the macroscopic device graph. Cross-scale alignment edges are constructed through these correspondences, and alignment constraints are applied to the embedding vectors of each scale in the unified embedding space, ensuring semantic consistency of the embedding representation of the same material at different scales within the unified space. Through the cross-scale graph alignment mechanism, atomic-level molecular structure information can be progressively transferred to the particle and device levels, ultimately forming multi-scale structural embedding vectors containing information across the entire "atom → particle → device" chain.

[0015] The multi-scale structural embedding vector obtained through the above method encompasses chemical structural information at the microscopic molecular level, transport behavior information at the mesoscopic particle level, and topological coupling information at the macroscopic device level, achieving a unified structured expression of multi-scale information in chemical production processes. This multi-scale structural embedding vector provides the basic representation of structural domains for subsequent temporal modeling and fusion, enabling the subsequently constructed large-scale chemical production model to simultaneously process structural and dynamic information in a unified high-dimensional semantic space, thereby overcoming the shortcomings of existing technologies where multi-scale information is fragmented and difficult to model uniformly.

[0016] S102: Perform time series modeling on multi-source time series data generated by chemical production processes to obtain time series embedding vectors. Then, fuse the multi-scale structural embedding vectors with the time series embedding vectors to obtain the corresponding multi-scale fusion vectors.

[0017] Chemical production processes continuously generate a large amount of operational data that changes over time. This data records the evolution of key information such as operating conditions, material properties, and equipment status over time, serving as an important source of information reflecting the dynamic characteristics of chemical processes. Multi-source time-series data refers to data sequences collected from different data sources during chemical production processes, characterized by timestamps and arranged chronologically. These sources include, but are not limited to: time-series data on operating parameters such as temperature, pressure, flow rate, and liquid level collected by distributed control systems; time-series data on raw material composition and product properties collected by laboratory information management systems; time-series data on component concentrations collected by chromatographs; and time-series data on real-time quality indicators collected by online analytical instruments. While these multi-source time-series data reflect the operational status of chemical production processes from different perspectives, the data sources differ in sampling frequency, time base, and data format. Directly using them for modeling makes it difficult to form a unified time-aligned representation. Therefore, it is necessary to first perform time-series modeling on the multi-source time-series data, mapping it to a unified time-series embedding vector. Mapping multi-source time-series data into a unified time-series embedding vector involves three steps: variable-level tokenization, temporal location encoding, and variable attention pooling. (1) Variable-level tokenization: For the i-th measurement point at time step t, normalize its scalar value. The vector is linearly projected to increase its dimension to d, and then superimposed with the identity embedding, data source type embedding, and time location encoding of the measurement point:

[0018] in, Let d be the linear projection matrix from scalar to vector, and take d=256; The learnable identity embedding for the i-th test point is given by the learnable matrix. Obtained by looking up the table; Embedded for data source type, by a learnable matrix According to the table, K=4 corresponds to four types of data sources: DCS operating parameters, LIMS laboratory analysis, chromatographic components, and online quality instruments. Time position encoding, given in sine / cosine form:

[0019]

[0020] (2) Variable attention pooling: Aggregate all M variable tokens at the same time step t into a single vector for that time step:

[0021]

[0022] in, , These are learnable parameters; (3) Representation of temporal embedding vectors: Stack the vectors of each time step in chronological order to obtain the input representation matrix: ,Right now After the following multi-layer self-attention encoding, the result is obtained. Among them, the t-th row That is, the temporal embedding vector at time step t, and matrix H is the temporal embedding vector representation of the entire time series.

[0023] Before performing time series modeling on multi-source time series data, preprocessing is required. Specifically, since different data sources have different sampling frequencies—for example, DCS data is typically collected at second or minute intervals, while LIMS data is typically collected at hourly or batch-level intervals—time alignment and resampling are necessary for the time series data from each source. This resampling uses a unified time reference to ensure consistent sampling intervals across different sources on the same time axis. Simultaneously, for data gaps caused by instrument malfunctions or communication interruptions, interpolation or imputation based on nearby data is used. For data with different dimensions and numerical ranges, normalization is applied to map them to a unified numerical interval, preventing deviations in subsequent modeling. After this preprocessing, a time-aligned, uniformly formatted multi-source time series data sequence is obtained.

[0024] After obtaining the preprocessed multi-source time-series data, a self-attention mechanism is employed for time-series modeling. The self-attention mechanism is a modeling method capable of capturing long-distance dependencies within a sequence. Its core idea is that for each time step in the sequence, by calculating the attention weights between that time step and all other time steps, the representation of that time step can aggregate relevant information from the entire sequence, thereby capturing long-term dependencies and dynamic collaborative relationships across units. Specifically, the preprocessed multi-source time-series data is input into the self-attention encoder in chronological order. For each time step, the attention weights between it and other time steps in the sequence are calculated, and the features of other time steps are weighted and fused based on these attention weights, thus outputting the temporal embedding vector for that time step. Through multi-layer self-attention encoding, the embedding vector of each time step can aggregate contextual information from the entire time series, effectively capturing long-term dependencies between operating parameters and dynamic collaborative relationships between different units in the chemical production process. For example, when the operating conditions of a certain piece of equipment change, the change may be transmitted to downstream equipment after a certain time delay. The self-attention mechanism can capture this dynamic correlation across time steps and across units through attention weights, so that the time-series embedding vector can accurately reflect the dynamic evolution characteristics of the chemical production process.

[0025] Multi-layer self-attention coding consists of stacked coding layers with the same L_enc layer structure. The calculation of each layer is as follows: (1) Multi-head self-attention: for input Calculate the query, key, and value matrix. , , .in, , , ; The output of the m-th attention head is: ,

[0026] in, , , The m-th slices of Q, K, and V are respectively. ; Multi-head outputs are spliced ​​and linearly transformed: , ; (2) Residual connectivity and layer normalization:

[0027] (3) Location feedforward network:

[0028] in, , ,Pick , and Indicates the bias term; (4) Second residual and layer normalization:

[0029] Given the following parameters: number of encoding layers L_enc=4, number of attention heads h=8, model dimension d=256, and dimension per head... =32, with both attention and hidden layer dropout rates set to 0.1; by Let be the input, and let be the output of the l-th layer. ,Pick This is the final temporal embedding vector.

[0030] (5) Implementation mechanism of long-term dependency and cross-unit collaboration: time step t and time step The attention weights between them are:

[0031] This weight directly measures the state at time step t in relation to time step t. The correlation between states, and its computational path length is... Not with time interval Growth; therefore, when upstream equipment is Time-based operation adjustment When downstream equipment is affected after a certain time step, the downstream time step The query vector can be directly compared with The key vectors at time points are associated without needing to go through... The intermediate propagation of the steps avoids gradient decay caused by long path propagation. In the multi-head mechanism, different heads can focus on the correlation of different time scales (for example, some heads focus on the short-term dynamics of adjacent time steps, while others focus on slow time-varying trends on the order of several hours).

[0032] After obtaining the multi-scale structural embedding vector and the temporal embedding vector, they need to be fused to obtain a multi-scale fused vector. The multi-scale structural embedding vector represents the spatial structure and topological coupling relationships between objects of different scales, such as molecules, particles, and equipment, in the chemical production process, belonging to the structural domain information. The temporal embedding vector represents the dynamic evolution characteristics of operating parameters and material properties over time in the chemical production process, belonging to the dynamic domain information. Both describe different aspects of the chemical production process from spatial and temporal dimensions, respectively. If only one type of information is used, the model cannot simultaneously possess structural interpretability and dynamic modeling capabilities. Therefore, it is necessary to deeply fuse the structural domain information and the dynamic domain information to form a unified high-dimensional semantic vector.

[0033] Specifically, this application employs a cross-layer attention mechanism and a feature-gated fusion strategy to fuse structural domain information and dynamic domain information. The core idea of ​​the cross-layer attention mechanism is to use a temporal embedding vector as the query and a multi-scale structural embedding vector as the key and value. By calculating the attention weights between temporal and structural information, it determines which structural information is most relevant to the current dynamic state at the current time step, and then aggregates the structural information based on these attention weights. Through this mechanism, the model can dynamically select the structural information most relevant to the current state, avoiding information redundancy and interference caused by simply concatenating structural and temporal information. The core idea of ​​the feature-gated fusion strategy is to dynamically adjust the fusion ratio of structural and dynamic domain information through a gating mechanism. The gating value is adaptively calculated based on the current input. When structural information is more important for prediction at a certain moment, the gating mechanism automatically increases the fusion weight of structural information; conversely, it increases the fusion weight of temporal information. Through this strategy, the model can adaptively balance the roles of structural and dynamic information under different conditions, thereby generating more accurate and robust multi-scale fusion vectors.

[0034] In practical implementation, the multi-scale structural embedding vector and the temporal embedding vector are linearly transformed and then input into a cross-layer attention module to calculate the interaction weights between structural and temporal information. Based on these interaction weights, the structural information is weighted and aggregated to obtain a structure-aware temporal representation. Subsequently, the structure-aware temporal representation is concatenated with the original temporal embedding vector, and a gating network is used to calculate the gating value. The concatenated features are then weighted and fused based on the gating value to finally obtain a multi-scale fusion vector. This multi-scale fusion vector is a unified high-dimensional semantic vector that simultaneously contains multi-scale structural information and multi-source temporal dynamic information from the chemical production process. It combines physical structure interpretability with complex dynamic modeling capabilities, providing a generalizable high-dimensional representation foundation for subsequent mechanism-constrained embedding and multi-objective collaborative optimization.

[0035] The cross-layer attention module is calculated as follows: (1) Linear transformation: Using temporal embedding as the source of query and structural embedding as the source of key and value, both are mapped to the same subspace through linear transformation: ,

[0036] ,

[0037] ,

[0038] Pick ; (2) Interaction weight calculation: The interaction weight between the temporal representation of time step t and the j-th structural node is:

[0039] in, for The OK, for The OK, Interaction weight matrix satisfy This weight is the interaction weight between structural information and temporal information. Its physical meaning is the degree of correlation between the j-th structural node (atom, particle, or device) and the working condition represented by the t-th time step. (3) Weighted aggregation: The structural information is weighted and summed according to the interaction weights to obtain the temporal representation of the structure perception: ,Right now ,

[0040] (4) Multi-head splicing and residual fusion: adopting Each attention point is computed in parallel and then concatenated, followed by a linear transformation. Map back to 256 dimensions and perform residual connections and layer normalization:

[0041] in, The t-th row is the temporal representation of the structure-aware time step t.

[0042] when When the product of T and T is too large, causing the size of the interaction weight matrix to exceed the memory, a sparsification scheme can be adopted. The interaction weight is calculated only for the structural nodes that have a topological association with the device to which the current time step belongs, and the weights of the other nodes are set to zero.

[0043] S103: Construct a large-scale chemical production model embedded with physical mechanism constraints; wherein, physical mechanism constraints include soft constraint embedding and hard constraint embedding. Soft constraint embedding is to introduce physical mechanism constraint terms into the loss function, and hard constraint embedding is to use a graph neural network that satisfies the physical conservation law in the network structure of the large-scale chemical production model.

[0044] Chemical production processes strictly adhere to fundamental physical laws such as the conservation of mass, energy, and reaction kinetics. However, when training a data-driven deep learning model, the model parameters are entirely determined by the data fitting results. While this can achieve high accuracy on the training data, it may violate basic physical laws, such as predicting negative component concentrations, energy imbalances, or reaction rates that do not conform to kinetic laws. This physical inconsistency leads to insufficient generalization ability of the model under scenarios such as operating condition drift, production of new grades, or deployment of new plants, and may even produce unreliable predictions, failing to meet the safety and reliability requirements of chemical production. To address this issue, this application embeds the fundamental physical mechanisms of chemical processes into the model as constraints when constructing a large-scale chemical production model. This ensures that the model satisfies physical conservation laws during both training and inference, thus balancing the high accuracy of data-driven models with the physical reliability of mechanistic models. Physical mechanism constraints refer to the restrictions imposed on the model prediction results based on the basic physical laws of chemical processes. These constraints include mass conservation constraints, energy conservation constraints, and reaction kinetic constraints. Embedding refers to integrating the above-mentioned physical mechanism constraints into the construction process of a large-scale chemical production model in the form of soft or hard constraints, so that the model naturally satisfies physical laws during training and inference.

[0045] The physical mechanism constraints in this application include two forms: soft constraint embedding and hard constraint embedding. These two methods embed the physical mechanisms at the loss function and network structure levels, respectively. Soft constraint embedding refers to introducing physical mechanism constraint terms into the loss function of the large-scale chemical production model, guiding the model's predictions to conform to physical laws through the optimization process during training. Specifically, when constructing the loss function of the large-scale chemical production model, in addition to the conventional data fitting loss term, residual terms based on the mass conservation equation, energy conservation equation, and reaction kinetic equation are introduced as physical mechanism constraint terms. The mass conservation equation residual is used to characterize the deviation of the model's predicted flow rate or concentration of each component in the material balance; the energy conservation equation residual is used to characterize the deviation of the model's predicted energy of each stream in the energy balance; and the reaction kinetic equation residual is used to characterize the deviation between the model's predicted reaction rate and the calculated value from the kinetic equation. These residual terms are assigned appropriate weights and then added to the total loss function, forming a total loss function that includes both the data fitting loss term and the physical mechanism constraint term. During model training, by minimizing the total loss function, the model not only needs to fit the training data but also needs to minimize the physical mechanism constraint term. This ensures that the model's prediction results satisfy the data fitting accuracy while naturally adhering to fundamental physical laws such as mass conservation, energy conservation, and reaction kinetics. Since soft constraint embedding is achieved through a penalty term in the loss function, the physical mechanism constraint term plays a guiding role during training, but the model output does not structurally enforce the conservation laws; hence, it is called soft constraint.

[0046] Hard constraint embedding refers to the use of graph neural networks that satisfy physical conservation laws in the network structure of large-scale chemical production models, making the model structurally inherently satisfy these laws. Specifically, hard constraint embedding is achieved by designing conservation-type graph neural networks. Conservation-type graph neural networks, through specific structural designs, ensure that the updates of node and edge features satisfy mass or energy conservation laws during message passing and node updates. For example, during message passing, for each node, the sum of messages received from its neighbors is balanced with the sum of messages it sends, thus ensuring that the updated features of the node structurally satisfy mass conservation; alternatively, an energy balancing layer can be introduced into the network structure to ensure that energy transfer between nodes structurally satisfies energy conservation. Furthermore, hard constraint embedding can also be implemented by constructing graph structures based on Hamiltonian or Lagrange forms. The Hamiltonian form uses the total energy of the system as the state function, describing the evolution of the system state over time through the Hamiltonian equation. It represents the node and edge states in the graph structure as generalized coordinates and generalized momentum in the Hamiltonian, ensuring that the model's state updates structurally satisfy energy balance and conservation constraints. The Lagrange form uses the difference between the system's kinetic and potential energy as the Lagrange function, describing the system dynamics through the Lagrange equation. Similarly, it naturally ensures that the model structurally satisfies conservation laws. By using the above-mentioned conservation-type graph neural networks or graph structures based on the Hamiltonian / Lagrange form as the basic network structure of a large-scale chemical production model, the model's internal state updates and feature transfers satisfy physical conservation laws during forward and backward propagation, thus structurally guaranteeing the physical consistency of the model output. Since hard constraint embedding is implemented through the network structure itself, the model output is structurally forced to satisfy conservation laws; therefore, it is called a hard constraint.

[0047] Soft-constraint embedding and hard-constraint embedding complement and reinforce each other in large-scale chemical production models. Soft-constraint embedding introduces physical mechanism constraints into the loss function, allowing the model to continuously approximate physical laws during training. It can flexibly handle various forms of physical constraints, including mass conservation, energy conservation, and reaction kinetics, without altering the network structure. It is simple to implement and highly applicable. Hard-constraint embedding, through the design of conservation-type graph neural networks or graph structures based on Hamiltonian / Lagrange forms, ensures that the model structurally satisfies physical conservation laws. It guarantees physical consistency without relying on penalty terms in the loss function, resulting in higher reliability and generalization ability. The combination of these two approaches allows the model to learn physical laws through soft constraints during training, while hard constraints in the network structure ensure the physical consistency of the model's output, thus constructing a large-scale chemical production model embedded with physical mechanism constraints. When this large-scale chemical production model is subsequently trained using multi-scale fusion vectors, it can achieve a balance between data fitting and physical constraints. This allows the model's prediction results to have both high accuracy driven by data and to meet the basic physical laws of chemical processes. It significantly improves the model's physical consistency, interpretability, and generalization ability, providing a reliable modeling foundation for the intelligent optimization and safe operation of complex chemical processes.

[0048] S104: Input the multi-scale fusion vector into the large-scale chemical production model for training to obtain a large-scale chemical production model that integrates multi-scale graph network and mechanism constraints.

[0049] Multi-scale fusion vectors are used as training samples to train a large-scale chemical production model. Through iterative optimization, the model parameters gradually converge, ultimately yielding a large-scale chemical production model that integrates multi-scale graph networks and mechanistic constraints. Multi-scale fusion vectors are high-dimensional semantic vectors that simultaneously contain multi-scale structural information and multi-source temporal dynamic information of the chemical production process. They represent both the spatial structure and topological coupling relationships between objects of different scales, such as molecules, particles, and equipment, and the dynamic evolution characteristics of operating parameters and material properties over time. Therefore, they can serve as a unified input representation for the large-scale chemical production model, enabling the model to simultaneously process structural domain information and dynamic domain information in the same high-dimensional semantic space.

[0050] During training, the multi-scale fusion vector is first input into a large-scale chemical production model. The model then performs forward propagation on this vector, outputting the corresponding prediction results. Specifically, after the multi-scale fusion vector is input into the model, it sequentially passes through the message passing layer, feature aggregation layer, and output layer. The model transforms and aggregates the input features layer by layer based on the current parameters, ultimately outputting the predicted results of the chemical production process, such as predicted product properties, equipment operating status, or key process indicators. After obtaining the model's prediction results, the loss value corresponding to the current prediction result is calculated based on a pre-constructed total loss function. This total loss function consists of a weighted sum of a data fitting loss term and a physical mechanism constraint term. The data fitting loss term measures the deviation between the model's prediction results and the true labels of the training data, while the physical mechanism constraint term measures the deviation of the model's prediction results from physical laws in terms of mass conservation, energy conservation, and reaction kinetics. Through the total loss function, the model needs to both fit the training data and satisfy the physical mechanism constraints during training, thus achieving a balance between data fitting accuracy and physical consistency in the model's prediction results.

[0051] After calculating the total loss function, backpropagation and gradient updates are performed on the parameters of the large-scale chemical production model based on the total loss function, and the model parameters are gradually optimized through iterative training. During the iterative training process, the weights of each residual term in the physical mechanism constraint term are dynamically adjusted with the training process. Specifically, in the early stage of training, the model parameters are not yet stable and the data fitting loss term is large. At this time, the weight of the physical mechanism constraint term is appropriately reduced so that the model learns the basic laws and feature representations in the data first. As the training progresses, the data fitting loss term gradually decreases and the model parameters tend to stabilize. At this time, the weight of the physical mechanism constraint term is gradually increased so that the model further satisfies the physical conservation laws on the basis of fitting the data. Through the above-mentioned dynamic adjustment strategy of residual weights, the problem of model convergence difficulty caused by excessively strong physical constraints in the early stage of training can be avoided, while ensuring that the model output has high physical consistency in the later stage of training. It should be noted that the dynamic adjustment method of residual weights can be set according to actual training needs, such as linear increment, exponential increment, or piecewise adjustment based on training rounds. This application does not make specific limitations in this regard.

[0052] During iterative training, the changes in the total loss function and the magnitudes of the residual terms in the physical mechanism constraint terms are continuously monitored. When the total loss function converges and the residuals in the physical mechanism constraint terms are less than the preset residual threshold, it indicates that the deviation between the model's prediction results and the training data has reached an acceptable range. Simultaneously, the deviations of the model's prediction results from physical laws in terms of mass conservation, energy conservation, and reaction kinetics also meet the preset requirements. At this point, training is stopped, resulting in a large-scale chemical production model integrating multi-scale graphical networks and mechanistic constraints. The preset residual threshold is a threshold value used to determine whether the physical mechanism constraint terms meet the physical consistency requirements. It can be preset according to the physical consistency requirements of the chemical production process. By setting the preset residual threshold, it is ensured that the final large-scale chemical production model not only meets the requirements in terms of data fitting accuracy but also achieves the preset standard in terms of physical consistency, thereby avoiding the model violating basic physical laws due to overfitting the data.

[0053] In one embodiment, the optimization of a chemical production process is typically not a single-objective problem, but rather involves multiple coupled and even conflicting optimization objectives, such as output, energy consumption, quality, cost, and safety. For example, increasing output often leads to increased energy consumption, improving quality may increase costs, and any operational adjustments must be made while ensuring safety. Therefore, it is necessary to first determine the optimization objectives in the chemical production process and, using chemical operating parameters as decision variables, construct a multi-objective optimization function to uniformly describe the quantitative relationship between each optimization objective and the decision variables. Optimization objectives refer to multiple performance indicators that need to be considered simultaneously in chemical production processes, including but not limited to at least two of the following: output, energy consumption, quality, cost, and safety. Chemical operating parameters refer to process parameters that directly affect the operating status of the chemical production process and can be adjusted by operators or control systems, including but not limited to at least one of hydrogen feed rate, reflux ratio, temperature setpoint, pressure setpoint, feed flow rate, and catalyst addition. A multi-objective optimization function refers to a mathematical function relationship constructed with chemical operating parameters as independent variables and multiple optimization objectives as dependent variables. It can calculate the values ​​of each optimization objective under a given combination of operating parameters based on the predicted output of a large-scale chemical production model.

[0054] After constructing the multi-objective optimization function, it needs to be transformed into an optimization problem within the reinforcement learning framework. This allows the reinforcement learning strategy to search for the optimal or near-optimal combination of operational parameters in a continuous operation space. Specifically, the multi-scale fusion vector obtained in the preceding steps is used as the state space in reinforcement learning. The state space refers to the set of variables used in reinforcement learning to describe the current state of the environment, and the agent selects actions based on the current state. Since the multi-scale fusion vector simultaneously contains multi-scale structural information and multi-source temporal dynamic information of the chemical production process, it can comprehensively and accurately reflect the current operating state of the chemical production process. Therefore, using it as the state space allows the reinforcement learning strategy to fully perceive the structural coupling relationship and dynamic evolution trend of the process during decision-making, avoiding decision bias due to incomplete state information. The chemical operation parameters are used as the action space in reinforcement learning, that is, the set of actions that the agent can choose at each moment. Each action in the action space corresponds to a specific set of chemical operation parameter values. A reward function is constructed based on the multi-objective optimization function. The reward function is used to evaluate the immediate reward obtained by the agent after taking a certain action in the current state. Its core idea is to transform the multiple optimization objectives described by the multi-objective optimization function into a scalar reward signal that reinforcement learning can optimize. Specifically, the optimization objectives in the multi-objective optimization function can be weighted and combined according to their importance or actual production needs, and the weighted combination result can be used as a component of the reward function. In an optional implementation, an attention mechanism can be used to dynamically allocate the weights of each optimization objective based on the current operating conditions, enabling the reward function to adaptively balance objectives such as output, energy consumption, quality, cost, and safety. Guided by the reward function, the reinforcement learning strategy continuously adjusts its action selection strategy during training to maximize long-term cumulative rewards, thereby gradually approaching the optimal solution of the multi-objective optimization problem.

[0055] After constructing the state space, action space, and reward function, a reinforcement learning strategy is employed to explore and learn policies within the reinforcement learning environment. A reinforcement learning strategy refers to a decision rule by which an agent selects actions based on its current state. It continuously updates its policy parameters through interaction with the environment to maximize long-term cumulative rewards. Reinforcement learning strategies can be implemented using any reinforcement learning algorithm capable of handling continuous action spaces, including but not limited to deep deterministic policy gradient, proximal policy optimization, or soft-actor commentator strategies. During policy exploration, the agent outputs a set of candidate chemical operation parameters based on the state represented by the current multi-scale fusion vector. It then predicts the process response under these parameters using a large-scale chemical production model and calculates the corresponding reward value based on the reward function. The agent updates its policy parameters based on the reward feedback and continues exploration at the next time step. To prevent the reinforcement learning strategy from entering hazardous conditions during exploration, safety constraints are set. These constraints are pre-defined boundary conditions that limit the range of chemical operation parameters, including but not limited to equipment operating limits, safety interlock conditions, upper and lower limits of process parameters, and hazardous area boundaries. During strategy exploration, safety constraints are applied to the action space or strategy output, ensuring that the agent selects actions only within the range that meets these constraints, prohibiting the exploration of dangerous states such as overheating, overpressure, and flash explosion. Ultimately, the reinforcement learning strategy outputs target chemical operation parameters while satisfying the safety constraints. These target chemical operation parameters refer to a combination of chemical operation parameters obtained after strategy optimization that simultaneously considers multiple optimization objectives and meets the safety constraints. These parameters can be directly used for the optimized control of chemical production processes or issued as operational guidance to the production site.

[0056] After obtaining the target chemical operating parameters, the operating conditions of the chemical production process have been optimized on a short-term timescale. However, chemical production typically involves continuous switching between multiple batches and grades, rather than operating under a single condition or with a single grade. Furthermore, catalysts gradually degrade over time, and equipment performance slowly changes due to factors such as scaling and impurity accumulation. Therefore, in addition to optimizing operating parameters in real-time or short cycles, it is necessary to optimize production scheduling on a longer timescale to determine scheduling parameters such as catalyst batch switching timing and grade switching sequence. Scheduling parameters are decision variables that determine production plans, batch arrangements, and switchover actions over a relatively long timescale. They do not directly correspond to continuous process setpoints at a specific moment, but rather to the timing and execution sequence of discrete events in the production process. Specifically, catalyst batch switching timing refers to determining when to switch the current catalyst batch to a new one based on catalyst activity decay patterns, product quality requirements, and the production plan, to avoid catalyst deactivation leading to decreased product quality or reduced reaction efficiency. Grade switching sequence refers to determining the production order of multiple product grades when the production plan includes them, to minimize material loss, increased energy consumption, and equipment disturbance during the switchover process while meeting order demands. Since scheduling parameters typically take values ​​that are discrete or mixed discrete variables, and each scheduling decision evaluation requires long-term simulations using large-scale chemical production models, the evaluation cost is high and the available sample size is limited. Therefore, optimization methods suitable for small samples and with high evaluation costs are needed to optimize scheduling parameters. In one alternative implementation, Bayesian optimization can be used to optimize the scheduling parameters. A surrogate model is constructed to approximate the mapping relationship between the scheduling parameters and the optimization objective. A balance between exploration and utilization is achieved based on the acquisition function, thereby efficiently searching for a better combination of scheduling parameters with a limited number of evaluations. The optimization objective can be consistent with the aforementioned multi-objective optimization function, simultaneously considering factors such as output, energy consumption, quality, cost, and safety. Alternatively, it can be tailored to the actual needs of the scheduling level, focusing on scheduling-related indicators such as switching losses, catalyst lifetime, and grade switching time.

[0057] After optimizing the scheduling parameters and obtaining the optimal ones, the target chemical operation parameters output from the previous steps need to be integrated with the scheduling parameters to generate operation instructions and issue them to the corresponding chemical production equipment. Integration refers to unifying the operation parameters on the continuous time scale and the scheduling parameters on the discrete time scale into the same time frame, so that the two are interconnected in time sequence and coordinated in execution to form a complete sequence of operation instructions. Specifically, the target chemical operation parameters correspond to the set values ​​of various process parameters under the current operating conditions, such as hydrogen feed rate, reflux ratio, and temperature set value, and their objects are specific equipment such as reactors, distillation columns, and heat exchangers; the scheduling parameters correspond to the occurrence timing and execution sequence of discrete events such as catalyst switching and grade switching, and their objects are production plans and batch arrangements. During integration, the production period of each grade, the usage range of catalyst batches, and the trigger time of switching actions are first determined according to the scheduling parameters. Then, within each production period, the corresponding target chemical operation parameters are used as the process set values ​​for that period, and corresponding switching instructions are inserted at the switching time, such as catalyst unloading and loading instructions, grade switching transition instructions, etc. The integrated operation instructions include at least the process parameter settings for each piece of equipment, the trigger time and execution sequence of switching actions, and the adjustment strategy for the transition process. Finally, the integrated operation instructions are issued to the corresponding chemical production equipment. Specifically, this can be achieved through distributed control systems, programmable logic controllers (PLCs), or manufacturing execution systems (MAS) to transmit the operation instructions to the field equipment. The field equipment then executes the corresponding adjustment and switching actions according to the instructions, thereby achieving full-cycle coordination from real-time operation optimization to long-term scheduling optimization.

[0058] In one embodiment, during the actual operation of a chemical production process, the properties of raw materials, equipment status, environmental conditions, and production plans may all change over time. For example, sudden changes in raw material sulfur content, catalyst activity decay, equipment scaling, or grade switching may occur. These changes can lead to discrepancies between the data distribution upon which the previously trained large-scale chemical production model relies and the current actual operating conditions, resulting in decreased model prediction accuracy and inaccurate optimization decisions. Therefore, after obtaining a large-scale chemical production model that integrates multi-scale graph networks and mechanistic constraints, it is necessary to continuously monitor the model's performance and optimize it promptly when a performance decline is detected, ensuring that the model can continuously adapt to changes in operating conditions and maintain long-term stable operation. Model performance parameters refer to quantitative indicators used to evaluate the performance of the large-scale chemical production model during operation, specifically including at least one of model prediction error, stability of target chemical operating parameters, and safety boundary satisfaction rate. Among them, model prediction error is used to measure the deviation between the predicted results output by the large-scale chemical production model and the actual measured values ​​of the chemical production process. It can be measured using indicators such as mean absolute error, root mean square error, or mean absolute percentage error. Target chemical operation parameter stability measures the fluctuation of the target chemical operation parameters output by the reinforcement learning strategy at consecutive time steps. If the output operation parameters change too drastically between adjacent time steps, it indicates insufficient strategy stability, which may lead to frequent adjustments by the actuators and affect production stability. Safety boundary satisfaction rate measures the proportion of operation parameters output by the reinforcement learning strategy that meet safety constraints during the strategy exploration process. If the safety boundary satisfaction rate decreases, it indicates that there are many operation parameters in the strategy output that are close to or exceed the safety boundary, requiring timely intervention. The above model performance parameters can be collected and calculated in real time through an online performance monitoring system. The online performance monitoring system can be deployed in the control system at the chemical production site or in a cloud server. It communicates with the large-scale chemical production model and the on-site data acquisition system through an interface, periodically acquiring model prediction results, actual measured values, and strategy outputs, and obtaining the performance parameters of each model according to preset calculation rules.

[0059] After monitoring the model performance parameters, the current model performance parameters are compared with the corresponding preset values ​​to determine whether the model performance of the large-scale chemical production model has declined. Preset values ​​are thresholds used to determine whether the model performance meets operational requirements. These values ​​are pre-set based on the actual requirements of the chemical production process for model accuracy, stability, and safety. For example, a preset error threshold can be set for model prediction error; when the model prediction error exceeds this threshold, model performance is considered to have declined. Similarly, a preset stability threshold can be set for the stability of target chemical operating parameters; when the fluctuation of operating parameters exceeds this threshold, model performance is considered to have declined. A preset satisfaction rate threshold can be set for the safety boundary satisfaction rate; when the safety boundary satisfaction rate is lower than this threshold, model performance is considered to have declined. When any one or more of the above model performance parameters are detected as not meeting the preset values, it indicates that the performance of the large-scale chemical production model is no longer suitable for the current operating conditions, and model optimization is required.

[0060] In one optional implementation, optimizing the large-scale chemical production model specifically includes: incremental fine-tuning the large-scale chemical production model based on newly added chemical production data. The newly added chemical production data refers to operational data generated during the chemical production process after the last training of the large-scale chemical production model, including new operating parameters, laboratory analysis results, product quality data, and information on changes in operating conditions. Incremental fine-tuning refers to using the newly added data as training samples to update the model parameters in small steps while maintaining the original parameters, enabling the model to quickly learn new characteristics and patterns under the current operating conditions without retraining from scratch. Because incremental fine-tuning only uses newly added data for local adjustments, it has low computational cost and fast update speed, making it suitable for online updates of the model when operating conditions experience slow drift. After completing the incremental fine-tuning, the reinforcement learning strategy is re-executed based on the incrementally fine-tuned large-scale chemical production model to output updated target chemical operating parameters, ensuring that the optimization decision matches the predictive ability of the updated model. If the performance parameters of the large-scale chemical production model after incremental fine-tuning still do not meet the preset requirements, it indicates that the current operating conditions have changed too drastically, and incremental fine-tuning alone is insufficient to restore the model to a usable state. In this case, the large-scale chemical production model is rolled back to a historical version. The historical version refers to a version of the model whose performance parameters meet the preset requirements and which is saved during the model version management process. Version rollback can quickly restore the model to a known usable state, avoiding the impact on production safety and optimization results due to continuous decline in model performance. Through the combination of monitoring, incremental fine-tuning, strategy re-execution, and version rollback, the large-scale chemical production model can continuously adapt to operating condition drift and performance degradation during chemical production, forming a closed-loop self-evolution capability, thereby maintaining high-precision, high-stability, and high-safety modeling and optimization performance in long-term operation.

[0061] The above are embodiments of the method proposed in this invention. Based on the same idea, some embodiments of this invention also provide devices and non-volatile computer storage media corresponding to the above methods.

[0062] Figure 2 This is a schematic diagram of a large-scale chemical production model construction device provided in an embodiment of the present invention. Figure 2 As shown, it includes: At least one processor; and, At least one processor-communication-connected memory; wherein, The memory stores instructions that can be executed by at least one processor, such that the at least one processor is able to execute a method for constructing a large model of chemical production as described in any of the preceding items.

[0063] This invention provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: A method for constructing a large-scale chemical production model, as described in any of the preceding items.

[0064] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0065] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for constructing a large-scale chemical production model, characterized in that, The method includes: Based on the collected chemical production data, a corresponding multi-scale graph network is constructed, and the multi-scale graph network is structurally represented to obtain the corresponding multi-scale structure embedding vector. Time series modeling is performed on multi-source time series data generated in chemical production processes to obtain time series embedding vectors. The multi-scale structure embedding vectors are then fused with the time series embedding vectors to obtain the corresponding multi-scale fusion vectors. A large-scale chemical production model is constructed with embedded physical mechanism constraints. The physical mechanism constraints include soft constraint embedding and hard constraint embedding. The soft constraint embedding is to introduce physical mechanism constraint terms into the loss function, and the hard constraint embedding is to use a graph neural network that satisfies the physical conservation law in the network structure of the large-scale chemical production model. The multi-scale fusion vector is input into the large-scale chemical production model for training, resulting in a large-scale chemical production model that integrates multi-scale graph networks and mechanistic constraints.

2. The method for constructing a large-scale chemical production model according to claim 1, characterized in that, Constructing a large-scale chemical production model embedded with physical mechanism constraints, specifically including: Based on a pre-defined graph neural network, a basic network for a large-scale chemical production model is constructed; wherein the input and output of the graph neural network satisfy the laws of physical conservation. Based on the aforementioned basic network, the residuals of the mass conservation equation, energy conservation equation, and reaction kinetic equation are used as physical mechanism constraints and added to the loss function of the large-scale chemical production model. The physical mechanism constraint term and the data fitting loss term are weighted and summed to construct the total loss function for training the large chemical production model, so that the model prediction results satisfy the physical conservation law during the training process.

3. The method for constructing a large-scale chemical production model according to claim 2, characterized in that, The multi-scale fusion vector is input into the large-scale chemical production model for training, resulting in a large-scale chemical production model that integrates multi-scale graph networks and mechanistic constraints, specifically including: The multi-scale fusion vector is input into the large-scale chemical production model, and the large-scale chemical production model is iteratively trained with the loss function as the optimization objective; wherein, the weight of the residual is dynamically adjusted during the training process. When the loss function converges and the residual is less than a preset residual threshold, a large-scale chemical production model integrating multi-scale graph networks and mechanistic constraints is obtained.

4. The method for constructing a large-scale chemical production model according to claim 1, characterized in that, After obtaining a large-scale chemical production model that integrates multi-scale graph networks and mechanistic constraints, the method further includes: The optimization objectives of the chemical production process are determined, and the chemical operation parameters in the chemical production process are used as decision variables to construct a multi-objective optimization function; The multi-scale fusion vector is used as the state space, the chemical operation parameters are used as the action space, and a reward function is constructed based on the multi-objective optimization function. Based on the reward function, a reinforcement learning strategy is adopted to output the target chemical operation parameters that meet the safety constraints.

5. The method for constructing a large-scale chemical production model according to claim 4, characterized in that, A reinforcement learning strategy is employed to output target chemical operation parameters that satisfy safety constraints, specifically including: A reinforcement learning strategy is employed to output candidate operation parameters and their corresponding prediction distributions. Based on the predicted distribution, the risk boundary of the candidate operating parameters is determined; The risk boundary is compared with the safety constraints to screen target chemical operation parameters that satisfy the safety constraints, and the target chemical operation parameters are output.

6. The method for constructing a large-scale chemical production model according to claim 5, characterized in that, After outputting the target chemical operation parameters, the method further includes: The scheduling parameters are optimized; wherein, the scheduling parameters include catalyst batch switching timing and grade switching order; The chemical operation parameters are integrated with the scheduling parameters to generate operation instructions and send them to the corresponding chemical production equipment.

7. The method for constructing a large-scale chemical production model according to claim 1, characterized in that, After obtaining a large-scale chemical production model that integrates multi-scale graph networks and mechanistic constraints, the method further includes: Monitor the model performance parameters of the large-scale chemical production model; wherein, the model performance parameters include at least one or more of the following: model prediction error, stability of target chemical operating parameters, and safety boundary satisfaction rate; Based on the model performance parameters, when a preset value of performance degradation of the large-scale chemical production model is detected, the large-scale chemical production model is optimized.

8. The method for constructing a large-scale chemical production model according to claim 7, characterized in that, The large-scale chemical production model is optimized, specifically including... Based on the newly added chemical production data, the large-scale chemical production model is incrementally fine-tuned. Based on the incrementally fine-tuned large-scale chemical production model, the reinforcement learning strategy is re-executed to output updated target chemical operation parameters. If the performance parameters of the incrementally fine-tuned chemical production model do not meet the preset requirements, the chemical production model will be rolled back to a historical version.

9. A large-scale chemical production model construction device, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform a method for constructing a large-scale chemical production model as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are set as follows: A method for constructing a large-scale chemical production model as described in any one of claims 1-8.