A dynamic closed-loop control method, equipment and medium based on chemical production
By constructing a large-scale model of chemical production and embedding process mechanism constraints, and deploying it to a digital twin platform for dynamic mapping and alignment with the DCS system, the problems of mechanism constraints and data adaptability, low-latency control and multi-device collaborative optimization in chemical production are solved, and efficient and safe dynamic closed-loop control is achieved.
Patent Information
- Application Number
- CN202611588166.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-10-14
- Publication Date
- 2026-11-17
AI Technical Summary
The deep coupling between mechanistic constraints and data-driven models in existing chemical production has not yet been resolved, making it difficult to guarantee physical consistency and data adaptability; there is a contradiction between low-latency closed-loop control and large model inference capabilities, making it difficult to meet the real-time control requirements of industrial processes; model updates under multi-device coupling lack a collaborative mechanism, making it difficult to guarantee global consistency, continuous evolution, and collaborative optimization.
By acquiring multi-source data from chemical production plants, preprocessing and uniformly characterizing it, a large-scale chemical production model is constructed and embedded with process mechanism constraints and production constraints. This model is then deployed to a digital twin platform for dynamic mapping and alignment with the DCS system. Multi-scale graph neural networks and self-attention mechanisms are used for collaborative optimization to achieve step-by-step incremental control and incremental updates, forming a dynamic closed-loop control.
While ensuring physical consistency, it improves data adaptability, achieves low-latency closed-loop control, enhances the collaborative optimization capability of multiple devices, improves production efficiency, energy utilization and operational safety, and solves the problems of robustness and generalization ability under complex time-varying operating conditions.
Smart Images

Figure CN123113397A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of chemical production technology, and in particular to a dynamic closed-loop control method, equipment and medium based on chemical production. Background Technology
[0002] Chemical production processes generate massive amounts of highly heterogeneous data. Intelligent technologies combining data-driven approaches with mechanistic models are profoundly transforming the chemical industry, driving its shift from traditional digital simulation to a more efficient, low-carbon, and safe intelligent production model. Currently, generative artificial intelligence (AIGC) technologies, represented by large-scale models, are developing rapidly, bringing new historical opportunities for the transformation of industrial production models. Digital twins, by combining mechanistic models, reduced-order models, and data-driven proxy models, can achieve soft measurement, anomaly diagnosis, what-if simulation, and closed-loop control, and have become an important technological path for the intelligentization of chemical processes.
[0003] However, existing solutions still have significant shortcomings in practical industrial applications: First, the deep coupling between mechanistic constraints and data-driven models remains unresolved, making it difficult to simultaneously guarantee physical consistency and data adaptability. Second, there is an inherent contradiction between low-latency closed-loop control and the inference capabilities of large models; the high inference overhead of large models makes it difficult to meet the real-time control requirements of industrial processes. Third, model updates under multi-device coupling lack a collaborative mechanism, making it difficult to guarantee global consistency, continuous evolution, and collaborative optimization. These problems restrict the practical application of AIGC, digital twins, and the integration of chemical process mechanistic knowledge, necessitating the development of modeling and control methods that balance physical consistency, real-time performance, and multi-device collaboration. Summary of the Invention
[0004] To address the aforementioned issues, this application proposes a dynamic closed-loop control method for chemical production, comprising: acquiring multi-source data from a chemical production unit; preprocessing and uniformly characterizing the multi-source data to obtain a digital representation of the process; constructing and training a model based on the digital representation of the process to obtain a large-scale chemical production model; embedding pre-set process mechanism constraints and production constraints into the large-scale chemical production model, and outputting multi-objective collaborative optimization process parameters and operating condition predictions; deploying the trained large-scale chemical production model to a pre-set digital twin platform to determine the dynamic mapping between the virtual twin and the physical entity, and aligning the input / output protocol of the large-scale chemical production model with the DCS system of the chemical plant; sending the process parameters and the operating condition predictions to the DCS system, and controlling the process parameters adaptively according to a pre-set step-by-step incremental method through the DCS system, and collecting field data after execution; transmitting the field data back to the digital twin platform to update the large-scale chemical production model, and performing repeated predictions through the updated large-scale chemical production model to form dynamic closed-loop control.
[0005] In one example, a model is constructed and trained based on the digital representation of the process to obtain a large-scale chemical production model. Specifically, this includes: using reaction units, heat exchange units, and separation units within the chemical production unit as graph nodes, material transport pipelines and energy heat exchange links as graph edges, and material flow rate and heat exchange coefficient as edge weights; determining a multi-scale graph neural network based on the graph nodes, edges, and edge weights; dividing the chemical production unit into multiple scales according to pre-set scale rules, where the scale rules include fine-scale, medium-scale, and coarse-scale. The fine-scale represents partial reaction characteristics within the reaction unit, the medium-scale represents material and energy transfer characteristics between chemical production units, and the coarse-scale represents the global production condition characteristics of the chemical production unit; performing graph convolution operations according to the scale rules to aggregate and update the graph nodes and edge weights to obtain a multi-scale fused latent feature vector; determining the time sequence of process measurement points in the chemical production unit, and mining the coupling correlation between different process measurement point time sequences through a self-attention mechanism to obtain a time-series fused latent feature vector; and generating the output of the large-scale chemical production model based on the multi-scale fused latent feature vector or the time-series fused latent feature vector.
[0006] In one example, embedding pre-set process mechanism constraints and production constraints into the large-scale chemical production model specifically includes: embedding pre-set hard constraint loss terms into the loss function of the large-scale chemical production model, and superimposing pre-set production constraints into the output verification of the large-scale chemical production model; during model training, jointly constraining the parameter update direction of the large-scale chemical production model based on the hard constraint loss terms and the production constraints.
[0007] In one example, the output of multi-objective collaborative optimization of process parameters and operating condition predictions specifically includes: using the real-time operating status of the chemical production unit as input and the setpoint fine-tuning operation corresponding to the process parameters as output to construct a reinforcement learning agent; constructing a production benefit reward item consisting of improved product yield and reduced unit energy consumption, an equipment life protection reward item consisting of equipment operating load stability and temperature and pressure fluctuation amplitude, and a safety constraint reward item consisting of equipment safety boundary margin and no over-threshold operating conditions; and constructing a multi-objective fusion reward function based on the production benefit reward item, the equipment life protection reward item, and the safety constraint reward item; and iteratively optimizing the reinforcement learning agent based on the multi-objective fusion reward function to output the optimal reaction temperature, system pressure, feed ratio, reflux ratio, and future operating condition predictions.
[0008] In one example, the trained large-scale chemical production model is deployed to a pre-configured digital twin platform to determine the dynamic mapping between the virtual twin and the physical entity. The input / output protocol of the large-scale chemical production model is then aligned with the DCS system at the chemical plant. Specifically, this includes: packaging the trained large-scale chemical production model into a containerized inference service and deploying the packaged model to the service cluster of the digital twin platform; configuring the inference interface and input / output data protocol of the DCS system to align the input / output data protocol with the data point table of the DCS system; and determining pre-configured inference call triggering conditions, including periodic triggering and event triggering.
[0009] In one example, the DCS system controls the process parameters adaptively using a pre-set step-by-step incremental method, and collects field data after execution. Specifically, this includes: calculating the difference between the optimal target parameters output by the large-scale chemical production model and the actual operating parameters of the DCS system to obtain the adjustment deviation; processing the adjustment deviation based on the response characteristics of the equipment actuators to set the adjustment step size; determining the adjustment step size through the DCS system; processing the adjustment step size using a step-by-step incremental method to output control commands; and gradually adjusting the process parameters according to the control commands. After adjustment, field data is collected, and the field data undergoes mechanism verification and safety compliance checks. If the checks fail, the issuance of control commands is stopped and the target parameters are corrected; if the checks pass, adjustments continue until the optimal target parameters are reached.
[0010] In one example, the field data is transmitted back to the digital twin platform to update the large-scale chemical production model. The updated model is then used for repeated predictions to form dynamic closed-loop control. Specifically, this includes: collecting operating data of the chemical production unit in the digital twin platform; performing frame-by-frame calculations on the operating data to obtain the mapping residual between the predicted operating value and the operating data; comparing the mapping residual with a pre-set residual threshold; if the mapping residual is greater than the residual threshold, recording the corresponding number of sampling periods; if the number of records exceeds a pre-set threshold, triggering an incremental update mechanism; freezing the underlying features of the large-scale chemical production model under the update mechanism; extracting basic weights based on the underlying features; and locally adjusting the mechanism constraint coefficients, optimization hyperparameters, and simulation correction coefficients based on the basic weights to determine new operating condition samples; iteratively updating the model output strategy using the new operating condition samples to dynamically correct the output target parameters; and repeatedly executing the step of sending the process adjustment parameters and the predicted operating values to the DCS system to form dynamic closed-loop control.
[0011] In one example, multi-source data from a chemical production plant is acquired, and the multi-source data is preprocessed and uniformly characterized to obtain a digital representation of the process. Specifically, this includes: collecting historical datasets of the chemical production plant, which include feed flow rates, feed components, reaction temperatures, pressures, recirculation ratios, discharge components, equipment safety valve thresholds, and the plant's maximum processing load for each reactor; normalizing the historical datasets and filtering the normalized historical datasets to directly remove sampling points that exceed pre-set process alarm limits; and slicing the filtered historical datasets according to pre-set fixed time windows to obtain the digital representation of the process.
[0012] On the other hand, this application also proposes a dynamic closed-loop control device based on chemical production, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the dynamic closed-loop control device based on chemical production to perform: the method described in any of the examples above.
[0013] On the other hand, this application also proposes a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to be the method described in any of the examples above.
[0014] This application utilizes a unified representation of multi-source data and the construction of a large-scale chemical production model. It embeds process mechanisms such as matter conservation, energy conservation, and reaction kinetics as hard constraints into the model's loss function, and superimposes production constraints, thereby improving data adaptability while ensuring physical consistency. The model employs a multi-scale graph neural network or Transformer with a self-attention mechanism, combined with a reinforcement learning multi-objective reward function and Bayesian optimization. This enables collaborative optimization of output, energy consumption, quality, and safety boundaries, outputting optimal process parameters and predicted operating conditions. By leveraging containerized deployment on a digital twin platform and alignment with DCS data point tables, dynamic mapping between the virtual twin and the physical entity and low-latency online inference are achieved. Through step-by-step incremental adaptive adjustment of the DCS, mechanism verification, and safety compliance checks, sudden parameter changes are effectively avoided, ensuring safe plant operation. After field data is transmitted back, an incremental update mechanism triggered by the mapping residual is used to fine-tune the mechanism constraint coefficients and optimize hyperparameters locally. This continuously reduces the virtual-to-real deviation without requiring downtime for retraining, resolving operating condition drift and model aging issues. The system forms a dynamic closed loop of "model prediction - actual execution - data feedback - model update", which significantly improves robustness, generalization ability and multi-device collaborative optimization under complex time-varying conditions, and ultimately improves production efficiency, energy utilization and operational safety. Attached Figure Description
[0015] 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 This is a schematic flowchart of a dynamic closed-loop control method based on chemical production in an embodiment of this application; Figure 2 This is a schematic diagram of a dynamic closed-loop control device based on chemical production in an embodiment of this application. Detailed Implementation
[0016] 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.
[0017] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.
[0018] like Figure 1 As shown, in order to solve the above problems, this application provides a dynamic closed-loop control method based on chemical production, the method comprising: S101. Acquire multi-source data from chemical production equipment, preprocess and uniformly characterize the multi-source data to obtain a digital representation of the process.
[0019] Collect multi-source data from chemical production plants. This multi-source data can include real-time historical datasets and online real-time datasets from the physical units. Specific raw data may include: feed flow rate, feed composition, reaction temperature, pressure, recirculation ratio, discharge composition, equipment safety valve thresholds, and maximum processing load of each reactor. This data can come from on-site DCS systems, sensors, laboratory test data, equipment condition monitoring systems, and production management systems.
[0020] The collected historical or real-time datasets are normalized. As an optional method, the normalization parameter uses min-max scaling, with a value range of [0,1]. The normalized data is then filtered to directly remove sampling points exceeding pre-set process alarm limits, thus filtering outlier samples. Subsequently, the filtered historical dataset is sliced according to a pre-set fixed time window to obtain a digital representation of the process. This digital representation can uniformly express process measurement points, equipment units, material flows, energy flows, and their temporal relationships, providing a unified input for the subsequent construction of a large-scale chemical production model.
[0021] In one embodiment, the multi-source data may also include slowly varying operating condition data such as equipment operating load, hourly energy consumption, product purity, equipment micro-deterioration status, raw material moisture content fluctuations, and heat exchange efficiency decay, as well as transient disturbance data such as sudden changes in feed, local pipeline pressure fluctuations, and short-term temperature jumps. Through the above preprocessing and unified characterization, standardized samples can be formed for large-scale model training and online inference.
[0022] S102. Based on the digital representation of the process, a model is constructed and trained to obtain a large-scale chemical production model. Pre-set process mechanism constraints and production constraints are embedded in the large-scale chemical production model, and process parameters and operating condition predictions for multi-objective collaborative optimization are output.
[0023] A large-scale chemical production model is obtained by constructing and training a model based on the digital representation of the process. This model can employ Transformer or multi-scale graph neural networks as its core. The model structure may include: a multi-scale graph neural network module / Transformer encoding module, a mechanism constraint embedding layer, a reinforcement learning agent module, and a Bayesian optimization solution module. The running hardware can be an industrial server CPU / GPU computing unit. The model embeds process mechanism constraints and production constraints, and integrates reinforcement learning and Bayesian optimization techniques to learn collaborative optimization strategies for multiple objectives, such as output, energy consumption, quality, and safety boundaries.
[0024] In one embodiment, reaction units, heat exchange units, and separation units within a chemical production plant are used as graph nodes; material conveying pipelines and energy heat exchange links are used as graph edges; and material flow rate and heat exchange coefficient are used as edge weights. A multi-scale graph neural network is determined based on the graph nodes, graph edges, and edge weights.
[0025] The chemical production unit is divided into multiple scales according to pre-set scaling rules. These scaling rules include fine-scale, medium-scale, and coarse-scale. The fine-scale represents the reaction characteristics within a single reaction unit; the medium-scale represents the material and energy transfer characteristics between chemical production units; and the coarse-scale represents the overall production conditions of the chemical production unit. Graph convolution operations are performed on each of the three scales. The convolution operations aggregate and update node features and edge weights, outputting a multi-scale fused latent feature vector.
[0026] In the scenario of equipment optimization, the reactor can contain n reaction units, and there are material flow and energy flow coupling relationships between the units. Through the multi-scale graph neural network mentioned above, the coupling correlation characteristics between various equipment and operating parameters can be extracted.
[0027] In one embodiment, as an alternative or replacement for multi-scale graph neural networks, the large-scale chemical production model can use a Transformer as its core. Specifically, the time series sequences of each process measurement point are used as token sequences, and the time window length T, sampling period, and number of multi-head attention heads are set. In one embodiment, the time window length T is 60 minutes, the sampling period is 1 minute, and the number of multi-head attention heads is set to 8. The coupling correlation between time series variables of different measurement points is explored through a self-attention mechanism, and a time series fusion latent feature vector is output. The output of the large-scale chemical production model is generated based on the multi-scale fusion latent feature vector or the time series fusion latent feature vector.
[0028] In one embodiment, pre-set process mechanism constraints and production constraints are embedded into the large-scale chemical production model. Specifically, the mass conservation equation, energy conservation equation, and reaction kinetic formula are embedded as hard constraint loss terms into the model loss function, rather than being used merely as business rules for post-processing validation. Simultaneously, pre-set production constraints are superimposed on the output validation of the large-scale chemical production model. During model training, the parameter update direction of the large-scale chemical production model is jointly constrained based on the hard constraint loss terms and production constraints.
[0029] The mechanism constraint embedding layer is a key technical means, and its execution is mainly carried out by the mechanism constraint embedding layer. Its execution process includes: embedding hard constraints of process mechanisms such as conservation of reactant substances, conservation of heat transfer energy, and reaction kinetics into the model loss function, and superimposing equipment pressure and temperature safety boundary conditions, so as to ensure that the model output is physically consistent and safe.
[0030] In one embodiment, the process parameters and predicted operating conditions are output through multi-objective collaborative optimization. Specifically, a reinforcement learning agent is constructed using the real-time operating status of the chemical production unit as input and the setpoint fine-tuning operations corresponding to the process parameters as output. Product yield improvement and unit energy consumption reduction constitute production benefit rewards; equipment operating load stability and temperature and pressure fluctuation amplitude constitute equipment lifespan protection rewards; and equipment safety boundary margin and absence of threshold exceedance conditions constitute safety constraint rewards. A multi-objective fusion reward function is constructed based on the production benefit rewards, equipment lifespan protection rewards, and safety constraint rewards. The reinforcement learning agent is iteratively optimized based on the multi-objective fusion reward function to output the optimal reaction temperature, system pressure, feed ratio, reflux ratio, and predicted future operating conditions.
[0031] By combining a Bayesian optimization solution module, the reinforcement learning optimization process is accelerated and global optimization is achieved, thereby realizing multi-objective collaborative optimization of output, energy consumption, quality, and safety boundaries. Model output can include process adjustment parameters such as optimal reaction temperature, system pressure, and feed ratio, as well as predicted values for short-term operating conditions. After mechanistic verification and safety compliance checks, the prediction results can be sent to the on-site DCS system, providing accurate and actionable pre-decision inputs for the complete closed loop of subsequent equipment execution, data feedback, and online model correction and updates.
[0032] S103. Deploy the trained large-scale chemical production model to a pre-set digital twin platform to determine the dynamic mapping between the virtual twin and the physical entity, and align the input / output protocol of the large-scale chemical production model with the DCS system of the chemical site.
[0033] In one embodiment, the trained large-scale chemical production model is packaged into a containerized inference service, and the packaged model is deployed to the service cluster of the digital twin platform. The digital twin platform server can provide a container environment, and the deployed model achieves dynamic mapping between the "virtual twin and the physical entity".
[0034] The inference interface and input / output data protocol of the DCS system are configured to align the input / output data protocol with the data point table of the DCS system. Pre-set inference call triggering conditions are determined, including periodic triggering and event triggering. In one embodiment, the periodic triggering period is 1 minute; event triggering occurs when a physical entity experiences process disturbances, alarm events, etc. Through the above methods, a real-time data access interface, online inference and low-latency decision-making modules, data synchronization and feedback mechanisms, and model update and transfer learning strategies can be implemented.
[0035] In one embodiment, the digital twin platform may also include a visual interface and interactive functions to enable real-time monitoring and dynamic optimization of critical equipment, energy consumption, and safety risks, thereby improving operational efficiency (OEE), energy utilization, and operational safety.
[0036] S104. The process parameters and the predicted operating conditions are sent to the DCS system. The DCS system controls the process parameters in a pre-set step-by-step incremental manner to adaptively adjust the process parameters and collect the field data after execution.
[0037] The system uses a chemical reaction unit as its specific application, continuously collecting key production data online, including real-time reactor temperature, feed flow rate, equipment operating load, hourly energy consumption, and product purity. After anomaly removal, normalization, and fixed-time-window slicing, the data is analyzed using a multi-scale graph neural network to uncover the coupling and correlation features of various equipment and operating parameters. The model is forcibly embedded with hard constraints of process mechanisms such as reactant conservation, heat transfer energy conservation, and reaction kinetics, and superimposed with equipment pressure and temperature safety boundary conditions. Relying on reinforcement learning combined with Bayesian optimization iterative optimization, the system outputs optimal process control parameters such as reaction temperature, system pressure, and feed ratio, as well as predicted values for short-term operating conditions. The prediction results are then verified by mechanism calculations and safety compliance checks before being sent to the on-site DCS system.
[0038] Specifically, the difference between the optimal target parameters output by the large-scale chemical production model and the actual operating parameters of the DCS system is calculated to obtain the adjustment deviation. The adjustment deviation is processed based on the response characteristics of the equipment actuators to set the adjustment step size. In one embodiment, the single temperature adjustment step size is ≤2℃, and the single pressure adjustment step size is ≤0.03MPa to avoid sudden parameter changes that could disrupt the device's operation. The DCS system determines the adjustment step size and processes it in a step-by-step incremental manner to output control commands. The DCS outputs control commands to actuators such as heating valves and feed regulating valves in a step-by-step incremental manner to gradually adjust key process parameters such as reaction temperature, feed flow rate, system pressure, and reflux ratio.
[0039] After adjustments, field data is collected and implemented, and the data undergoes mechanism calculations and safety compliance checks. If the check fails, the process is stopped and the target parameters are corrected; if the check passes, adjustments continue until the optimal target parameters are achieved. Through this method, adaptive adjustment and dynamic optimization control of key process parameters are realized.
[0040] S105. The field data is transmitted back to the digital twin platform to update the chemical production model, and repeated predictions are made using the updated chemical production model to form dynamic closed-loop control.
[0041] After the adjustment is executed, the sensors collect real-time on-site process, energy consumption, and quality data and transmit them back to the digital twin platform. The measured values are compared with the model's predicted target values to obtain the residuals. These residuals are then used to update the model weights and constraint coefficients online, dynamically correcting the target parameters for the next round of output. By repeatedly executing the process of "parameter issuance—on-site execution—feedback comparison—model correction," key parameters such as temperature and pressure are continuously and automatically adjusted in response to raw material fluctuations and load changes, achieving dynamic optimization control of the unit.
[0042] Specifically, operating condition data of the chemical production unit is collected in the digital twin platform. The operating condition data is calculated frame by frame to obtain the mapping residual between the predicted operating condition value and the actual operating condition data. The mapping residual is compared with a pre-set residual threshold. If the mapping residual is greater than the residual threshold, the corresponding number of sampling periods is recorded; if the number of records exceeds a pre-set threshold, an incremental update mechanism is triggered. In one embodiment, the incremental update mechanism is triggered when the average residual of 10 consecutive sampling periods (1 min / period) exceeds a set threshold. The set thresholds include temperature deviation > 1.5℃ and pressure deviation > 0.02 MPa.
[0043] Under the update mechanism, the underlying features of the large-scale chemical production model are frozen. Basic weights are extracted based on these features, and the mechanistic constraint coefficients, optimization hyperparameters, and simulation correction coefficients are locally adjusted according to these weights to determine the new operating condition samples. The model output strategy is iteratively updated using these new operating condition samples to dynamically correct the output target parameters. The process adjustment parameters and predicted operating conditions are repeatedly sent to the DCS system to form a dynamic closed-loop control.
[0044] In one embodiment, for complex time-varying operating conditions such as fluctuations in raw material composition, load switching, minor equipment degradation, and transient disturbances in chemical reaction units, methods for improving the accuracy of virtual-real mapping and model robustness correction may also be included. Specifically, these methods include online model updates based on incremental learning, abnormal operating condition detection and adaptive adjustment, and multi-scenario simulation verification.
[0045] When updating online models based on incremental learning, traditional offline training models only adapt to steady-state conditions. Faced with slowly changing and complex conditions such as fluctuations in raw material moisture content and decline in equipment heat exchange efficiency, prediction bias is prone to occur. This embodiment uses an incremental learning algorithm to achieve online iterative updates of model parameters, eliminating the need for downtime retraining and reliance on historical full data. The twin platform collects physical device operating data in real time, calculating the mapping residual between virtual twin predictions and physical measured values frame by frame. When the average residual over 10 consecutive sampling periods (1 min / period) exceeds a set threshold (temperature deviation > 1.5℃, pressure deviation > 0.02 MPa), the incremental update mechanism is triggered. The model freezes the basic weights of the bottom feature extraction layer, making only local fine-tuning to the mechanism constraint coefficients, optimization hyperparameters, and simulation correction coefficients. The model output strategy is iteratively updated by adding new real-time operating condition samples. When the device experiences catalyst activity decay and heat exchange fouling leading to operating condition drift during long-term operation, incremental learning can adapt to changes in equipment status in real time, correct model kinetic calculation parameters, continuously reduce the virtual-to-real mapping deviation, and avoid accuracy degradation caused by model solidification.
[0046] During abnormal operating condition detection and adaptive adjustment, a mechanism for identifying abnormal operating conditions and an adaptive control logic are constructed to address sudden and complex abnormal operating conditions such as sudden changes in feed, local pipeline pressure fluctuations, and short-term temperature jumps, enabling rapid identification and accurate correction of disturbances. Using historical operating condition features such as normal steady-state, variable load, and small-amplitude disturbances as a benchmark library, the twin platform compares the similarity distance between the real-time operating condition feature vector and the benchmark feature library in real time. When a feature distance exceeding the limit is detected, it is determined as an abnormal operating condition and classified into three categories: abnormal feed composition, critical conditions of over-temperature and over-pressure, and micro-disturbances in equipment operation. The model automatically locks the abnormal correlation parameters, pauses the conventional iterative optimization logic, and initiates an adaptive constraint adjustment strategy. For example, when a sudden increase in feed flow is detected, the reaction pressure safety constraint weight is automatically tightened, the temperature adjustment step size is reduced (from the usual 2℃ / time to 1℃ / time), and the virtual-real mapping simulation boundary conditions of the twin are simultaneously corrected to suppress prediction oscillations and mapping distortions under abnormal operating conditions, ensuring system operational stability.
[0047] During multi-scenario simulation verification, to ensure the model's adaptability under all operating conditions, various complex operating condition simulation environments were constructed based on digital twin virtual scenarios to complete closed-loop verification of model robustness and mapping accuracy. Implementation examples verified scenarios covering: high-low load switching conditions, conditions with significant fluctuations in raw material composition, conditions with minor equipment fault disturbances, boundary parameter extreme conditions, and conditions with multi-parameter coupled disturbances. Various extreme and complex production scenarios were reproduced in the virtual twin environment, and the model's virtual-real mapping error, parameter adjustment stability, and multi-objective optimization adaptability were tested. By comparing the mapping accuracy, operating condition fluctuation amplitude, and parameter adjustment response speed before and after correction, the improvement in model robustness was quantified, and defects in weakly adaptable scenarios were eliminated. Ultimately, this ensured that the model could adapt to all-dimensional complex and changing operating conditions on-site and possessed extremely strong engineering generalization capabilities.
[0048] The digital twin platform can also include a process optimization layer and a safety simulation layer. The process optimization layer is used to empower production, while the safety simulation layer is used to empower safety. Based on a unified digital representation system for processes, deep reinforcement learning intelligent optimization and physical-data fusion causal inference technology are respectively employed to achieve intelligent closed-loop optimization of production and dynamic prediction and simulation of accident risks.
[0049] In one embodiment, the process optimization layer is used for multi-objective closed-loop process optimization based on deep reinforcement learning. This layer relies on the fine-grained digital unified representation of the process, using deep reinforcement learning as the core technology. It abandons the traditional fixed parameter control mode, constructs a multi-dimensional reward function, and drives the agent to autonomously iteratively learn in a twin virtual environment, adaptively outputting the optimal process parameter adjustment strategy to achieve dynamic closed-loop optimization of the production process. Specific implementation steps include: First, constructing a twin virtual training environment. Based on a fine-grained model of a chemical reaction unit, the entire unit structure and material / energy transfer mechanism of the reactor, heat exchange equipment, and pipelines are replicated. All adjustable process parameters such as temperature, pressure, feed flow rate, and reflux ratio are mapped, building a virtual training environment synchronized 1:1 with the physical production scenario, providing a simulation carrier for the agent's iterative learning. Second, constructing a multi-objective fusion reward function. A quantitative reward system is constructed with production efficiency, equipment lifespan, and operational safety as core dimensions. Specifically, this includes: rewarding increased product yield and reduced unit energy consumption as production efficiency benefits; rewarding stable equipment operation load and temperature and pressure fluctuations as equipment lifespan protection benefits; and rewarding adequate safety margins and absence of threshold-exceeding conditions as safety constraints, achieving a balanced and quantitative evaluation of multiple objectives. The third step involves the agent's autonomous iterative learning and optimization. A reinforcement learning agent is deployed in a twin virtual environment, taking the real-time operating status of the equipment as input and fine-tuning of various process parameter setpoints as output. The agent continuously interacts with the virtual operating environment, accumulating high-reward operating condition samples, eliminating ineffective control strategies, and iteratively learning the optimal parameter adjustment strategy to adapt to complex conditions such as raw material fluctuations and load switching. For example, adaptively optimizing reactor temperature setpoints, feed ratios, and system pressure parameters. The fourth step involves closed-loop optimization execution between the virtual and real environments. The optimal process parameter strategy learned by the intelligent agent is sent to the on-site DCS system for actual execution. Simultaneously, the production data after execution is collected and sent back to the twin platform to continuously update the virtual environment operating condition characteristics, completing the intelligent closed-loop control of "virtual learning - on-site execution - data feedback - iterative optimization", and steadily improving production economy and operational safety.
[0050] In one embodiment, the safety simulation layer is used for dynamic risk simulation and consequence prediction driven by physical-data fusion. This layer adopts a technical architecture driven by the fusion of physical mechanism models and data causal inference, breaking through the limitations of traditional single early warning systems that can only identify anomalies but cannot predict consequences. By mining risk causal chains and constructing dynamic risk maps, it achieves high-precision dynamic simulation of accidents such as chemical leaks, fires, and explosions, as well as domino disaster chains, completing the upgrade from passive early warning to proactive consequence prediction. Specific implementation steps include: First, embedding physical mechanism models to build a fusion simulation foundation. The core physical mechanism equations of chemical material diffusion, combustion, and explosion are solidified in the digital twin model, binding the equipment layout, medium properties, and operating boundary conditions to construct a physical simulation foundation that fits actual production laws, ensuring the authenticity and professionalism of accident simulation. Second, causal inference to mine risk links and construct a dynamic risk map. Historical fault data, abnormal operating condition data, and real-time operating data of the equipment are collected. Causal inference algorithms are used to eliminate false data correlations, accurately mining the risk causal chains between abnormal process parameters, equipment defects, environmental disturbances, and safety accidents. The system updates the associated weights of each risk node in real time, dynamically generating a comprehensive risk map adapted to the current operating conditions, and clearly defining the propagation path and triggering conditions of potential hazards. The third step involves triggering early warning signals and initiating a specialized simulation mechanism. When the platform detects early warning signals such as sudden pressure drops, abnormal medium concentrations, or excessive temperatures, it immediately identifies the source of the risk, calls the corresponding accident simulation model, and initiates high-fidelity dynamic simulations for typical accident scenarios such as material leaks, open flame fires, and pressure explosions. The fourth step involves quantifying and extrapolating the uncertainty of the disaster chain and outputting its consequences. This simulates the spread range, intensity changes, and temporal evolution of the accident after it occurs, while quantifying the uncertainties caused by disturbances in operating conditions and deviations in equipment status. It accurately extrapolates the domino effect of the accident, predicts the risk of cascading failures in adjacent equipment and pipelines, and ultimately outputs quantitative results such as the accident's impact range, hazard level, and spread duration, providing precise data support for emergency response.
[0051] Through the aforementioned process optimization and safety simulation layers, combined with the visual interface and interactive functions of the production digital twin platform, real-time monitoring and dynamic optimization of key equipment, energy consumption, and safety risks can be achieved, thereby significantly improving operational efficiency (OEE), energy utilization, and operational safety. The integration technology route between the large-scale chemical production model and the production digital twin platform is as follows: Figure 1 As shown.
[0052] This application's embodiments acquire multi-source data from chemical production plants, preprocess and uniformly characterize it to obtain a digital representation of the process; construct and train a large-scale chemical production model based on this digital representation, embedding process mechanism constraints and production constraints, and outputting multi-objective collaborative optimization process parameters and operating condition predictions; deploy the large-scale chemical production model to a digital twin platform to achieve dynamic mapping between the virtual twin and the physical entity, and align the data with the DCS system; adaptively adjust the DCS system in a step-by-step incremental manner, and collect and transmit field data; update the large-scale chemical production model with field data and repeat predictions to form dynamic closed-loop control. Therefore, this application can balance physical consistency and data adaptability, improve low-latency closed-loop control capabilities, enhance model update collaboration under multi-device coupling, and continuously reduce the virtual-real mapping deviation under complex time-varying operating conditions, thereby improving production efficiency, energy utilization, and operational safety.
[0053] like Figure 2 As shown in the illustration, this application also provides a dynamic closed-loop control device based on chemical production, comprising: At least one processor; and, A memory that is communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by at least one processor to enable a dynamic closed-loop control device based on chemical production to perform the method as described in any of the embodiments above.
[0054] This application also provides a non-volatile computer storage medium storing computer-executable instructions, which are configured as described in any of the above embodiments.
[0055] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to hardware circuit structures. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program and "integrate" a digital system onto a PLD themselves, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Furthermore, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software. Similar to the software compiler used in program development, the original code before compilation must also be written in a specific programming language, called a Hardware Description Language (HDL). There are many HDLs, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also understand that by simply performing some logic programming on the method flow using one of these hardware description languages and programming it into an integrated circuit, the hardware circuit implementing the logical method flow can be easily obtained.
[0056] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.
[0057] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.
[0058] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.
[0059] 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.
[0060] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.
[0061] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0062] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0063] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0064] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0065] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.
[0066] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.
[0067] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.
[0068] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0069] The above are merely embodiments of this application and are 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 dynamic closed loop control method based on chemical production, characterized by, include: Acquire multi-source data from chemical production facilities, preprocess and uniformly characterize the multi-source data to obtain a digital representation of the process; The model is constructed and trained based on the digital representation of the process to obtain a large chemical production model. Pre-set process mechanism constraints and production constraints are embedded in the large chemical production model, and process parameters and operating condition predictions for multi-objective collaborative optimization are output. The trained large-scale chemical production model is deployed to a pre-set digital twin platform to determine the dynamic mapping between the virtual twin and the physical entity, and the input-output protocol of the large-scale chemical production model is aligned with the DCS system of the chemical site. The process parameters and the predicted operating conditions are sent to the DCS system, which then controls the process parameters in a pre-set step-by-step incremental manner to adaptively adjust them and collect field data after execution. The field data is transmitted back to the digital twin platform to update the large-scale chemical production model, and repeated predictions are made using the updated large-scale chemical production model to form a dynamic closed-loop control.
2. The method of claim 1, wherein, Based on the digital representation of the process, a model is constructed and trained to obtain a large-scale chemical production model, specifically including: The reaction unit, heat exchange unit, and separation unit within the chemical production device are used as graph nodes, material conveying pipelines and energy heat exchange links are used as graph edges, and material flow rate and heat exchange coefficient are used as edge weights. A multi-scale graph neural network is determined based on the graph nodes, graph edges, and edge weights. The chemical production unit is divided into multiple scales according to the pre-set scale rules. The scale rules include fine scale, medium scale and coarse scale. The fine scale is the reaction characteristics within the reaction unit, the medium scale is the material and energy transfer characteristics between chemical production units, and the coarse scale is the global production condition characteristics of the chemical production unit. Graph convolution operations are performed according to the scale rules to aggregate and update the graph nodes and edge weights to obtain a multi-scale fused latent feature vector. The timing sequence of process measurement points in the chemical production unit is determined, and the coupling correlation between different process measurement point timing sequences is mined through a self-attention mechanism to obtain a timing fusion latent feature vector. The output of the large-scale chemical production model is generated based on the multi-scale fusion latent feature vector or the temporal fusion latent feature vector.
3. The method of claim 1, wherein, The large-scale chemical production model is embedded with pre-set process mechanism constraints and production constraints, specifically including: The pre-set hard constraint loss term is embedded into the loss function of the large chemical production model, and the pre-set production constraints are superimposed on the output verification of the large chemical production model. During model training, the parameter update direction of the large chemical production model is jointly constrained based on the hard constraint loss term and the production constraints.
4. The method according to claim 1, characterized in that, Output multi-objective collaborative optimization process parameters and operating condition predictions, specifically including: Using the real-time operating status of the chemical production unit as input and the fine-tuning operation of the setpoint corresponding to the process parameters as output, a reinforcement learning agent is constructed. The system includes a production benefit reward item consisting of improved product yield and reduced unit energy consumption, an equipment life protection reward item consisting of stable equipment operating load and temperature and pressure fluctuations, and a safety constraint reward item consisting of device safety boundary margin and no over-threshold operating conditions. A multi-objective fusion reward function is constructed based on the production benefit reward item, the equipment life protection reward item, and the safety constraint reward item. The reinforcement learning agent is iteratively optimized according to the multi-objective fusion reward function to output the optimal reaction temperature, system pressure, feed ratio, reflux ratio, and future operating condition prediction.
5. The method according to claim 1, characterized in that, The trained large-scale chemical production model is deployed to a pre-set digital twin platform to determine the dynamic mapping between the virtual twin and the physical entity. The input / output protocol of the large-scale chemical production model is then aligned with the DCS system at the chemical plant site. Specifically, this includes: The trained large-scale chemical production model is packaged into a containerized inference service, and the packaged large-scale chemical production model is deployed to the service cluster of the digital twin platform. Configure the inference interface and input / output data protocol of the DCS system to align the input / output data protocol with the data point table of the DCS system, and determine the pre-set inference call triggering conditions, including periodic triggering and event triggering.
6. The method according to claim 1, characterized in that, The DCS system controls the process parameters adaptively according to a pre-set step-by-step incremental method, and collects field data after execution, specifically including: The difference between the optimal target parameters output by the large chemical production model and the actual operating parameters of the DCS system is calculated to obtain the adjustment deviation. The adjustment deviation is then processed according to the response characteristics of the equipment actuator to set the adjustment step size. The adjustment step size is determined by the DCS system, and the adjustment step size is processed in a step-by-step incremental manner to output control commands. The process parameters are then adjusted step by step according to the control commands. After adjustment, the field data collected and executed is used to perform mechanism verification and safety compliance verification. If the verification fails, the distribution is stopped and the target parameters are corrected. If the verification passes, the adjustment continues until the optimal target parameters are reached.
7. The method according to claim 1, characterized in that, The field data is transmitted back to the digital twin platform to update the large-scale chemical production model. Repeated predictions are then performed using the updated model to form a dynamic closed-loop control. Specifically, this includes: The operating condition data of the chemical production unit is collected in the digital twin platform, and the operating condition data is calculated frame by frame to obtain the mapping residual between the predicted operating condition value and the operating condition data. The mapping residual is compared with a preset residual threshold. If the mapping residual is greater than the residual threshold, the number of corresponding sampling periods is recorded. If the number of records exceeds a preset number threshold, an incremental update mechanism is triggered. Under the update mechanism, the underlying features of the large chemical production model are frozen, basic weights are extracted based on the underlying features, and the mechanism constraint coefficients, optimization hyperparameters and simulation correction coefficients are locally adjusted based on the basic weights to determine the new working condition samples. The model output strategy is iteratively updated using the newly added operating condition samples to dynamically correct the output target parameters, and the step of sending the process adjustment parameters and the operating condition prediction values to the DCS system is repeated to form dynamic closed-loop control.
8. The method according to claim 1, characterized in that, Acquire multi-source data from chemical production plants, preprocess and uniformly characterize the multi-source data to obtain a digital representation of the process, specifically including: Collect historical data of the chemical production unit, which includes feed flow rate, feed composition, reaction temperature, pressure, recirculation ratio, discharge composition, equipment safety valve threshold and maximum processing load of each reactor; The historical dataset is normalized, and the normalized historical dataset is filtered to directly remove sampling points that exceed the preset upper and lower limits of process alarms. The filtered historical dataset is sliced according to a pre-set fixed time window to obtain a digital representation of the process.
9. A dynamic closed-loop control device based on chemical production, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the dynamic closed-loop control device based on chemical production to perform the method 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 configured to be the method as described in any one of claims 1-8.
