Model-based multivariable pre-estimation control method for pentafluoroethane production process
By constructing a digital twin hybrid model, the catalyst performance is estimated in real time and the target is dynamically optimized, solving the model mismatch and safety constraints in the pentafluoroethane production process, and realizing efficient and sustainable pentafluoroethane production.
Patent Information
- Application Number
- CN202511818320.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-04
- Publication Date
- 2026-03-17
AI Technical Summary
The existing pentafluoroethane production process suffers from model mismatch, static optimization objectives, and rigid safety constraints, making it difficult to achieve multi-dimensional sustainable optimization.
A digital twin hybrid model is constructed to estimate catalyst performance in real time, dynamically reconstruct economic optimization objectives, calculate dynamic risk boundaries, implement multi-objective rolling optimization of economic and safety that integrates dynamic risk boundaries, and integrate process energy efficiency and environmental emission monitoring.
It achieves high-precision adaptive control, continuously optimizes economic benefits, flexibly manages safety risks, constructs an economic-safety-environmental collaborative optimization framework, and improves the system's intelligence level and stability.
Smart Images
Figure CN121680307A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of pentafluoroethane production technology, and in particular to a model-based multivariate predictive control method for pentafluoroethane production processes. Background Technology
[0002] Pentafluoroethane (HFC-125), as an important refrigerant and fire extinguishing agent, typically involves complex catalytic chemical reactions and distillation separation in its production process. This process is characterized by strong nonlinearity, large time delays, and strong coupling of multiple variables. Furthermore, the catalyst activity in its core reactor gradually declines over the production cycle, leading to time-varying process kinetics. In addition, the production process must be carried out under high temperature and high pressure conditions, imposing strict constraints on key safety parameters such as reactor hotspot temperature and system pressure.
[0003] Currently, industrial control of this process mostly employs traditional PID control or model predictive control (MPC) based on linear models. However, these methods have significant limitations:
[0004] Model mismatch problem: Traditional PID control cannot effectively handle the coupling between multiple variables, while the linear model on which conventional MPC relies is difficult to accurately describe the nonlinear dynamics and time-varying characteristics of the process. Especially during the period of catalyst performance degradation, the model prediction accuracy will decrease significantly, leading to the deterioration of control performance.
[0005] Static optimization objective: The upper-level economic optimization layer usually uses the operating point calculation based on the steady-state model, which cannot adjust the economic objective in real time and adaptively according to the dynamic decay of catalyst performance, making it difficult to achieve continuous economic optimization throughout the entire production cycle.
[0006] Rigid safety constraints: Existing methods mostly treat safety parameters as fixed hard constraints, lacking the foresight and flexible management of dynamic risks. When operating conditions fluctuate drastically or the feed load changes, fixed constraint boundaries may be too conservative (sacrificing economic benefits) or too aggressive (causing safety hazards).
[0007] Isolated objective function: Traditional optimization control often considers economic, safety, and environmental goals in isolation, or only as ex-post constraints, failing to achieve deep integration and synergistic optimization of multi-dimensional sustainability goals. Summary of the Invention
[0008] The purpose of this invention is to address the shortcomings of the prior art by proposing a model-based multivariate predictive control method for pentafluoroethane production processes.
[0009] To achieve the above objectives, the present invention adopts the following technical solution:
[0010] A model-based multivariate predictive control method for pentafluoroethane production process includes:
[0011] S1: Construct a digital twin hybrid model of the pentafluoroethane production process;
[0012] S2: Based on the aforementioned digital twin hybrid model, estimate the catalyst performance index online;
[0013] S3: Dynamically reconstruct the economic optimization objective function based on the catalyst performance index;
[0014] S4: Calculate the dynamic risk boundary of key safety parameters;
[0015] S5: Perform economic-safety multi-objective rolling optimization with integrated dynamic risk boundaries;
[0016] S6: Implement control and update the digital twin model online;
[0017] S7: Monitors and controls system performance indicators in real time, and adaptively adjusts control parameters based on performance evaluation;
[0018] S8: Integrate process energy efficiency and environmental emission monitoring to build multi-dimensional sustainability optimization goals.
[0019] Preferably, in step S1, a dynamic mechanism model including core units such as reactors and distillation columns is established based on process mechanisms and historical data; simultaneously, real-time operation data is collected, and a data-driven neural network is trained as a compensation model; the mechanism model and the data-driven model are run in parallel, and their outputs are fused through a weighting factor that is adaptively adjusted based on recent prediction errors to form the final output of the digital twin hybrid model.
[0020] Preferably, in step S2, the mixing model is used in conjunction with real-time reactor inlet and outlet component and temperature data, and a nonlinear state observer or moving window parameter estimation algorithm is used to calculate in real time the comprehensive performance index reflecting the catalyst activity and selectivity.
[0021] Preferably, in step S3, a multi-parameter objective function library with economic benefits as the core is established. This function library contains optimization focuses for different catalyst performance stages. Based on the performance index obtained in step S2, the current optimal economic objective function is selected or interpolated from the objective function library for use in upper-level optimization.
[0022] Preferably, in step S4, a functional relationship is established between the reactor hot spot temperature, the system pressure safety valve setpoint, and the operating variables, feed load, and catalyst performance; based on the current operating conditions and the future dynamic prediction of the digital twin hybrid model, a risk probability model or safety margin calculation method is used to calculate the time-varying and flexible dynamic risk boundary in the future time domain.
[0023] Preferably, in S5, in each control cycle, the economic objective reconstructed in S3 is used as the main optimization objective, while a penalty term is introduced for the predicted value exceeding the dynamic risk boundary described in S4, to construct a new multi-objective optimization problem; a numerical optimization algorithm is used to solve for a series of operational variable increments in the future control time domain, so as to minimize the multi-objective optimization function.
[0024] Preferably, in step S6, the first control increment obtained from step S5 is applied to the actual production process; a new round of process measurement data is collected, which is used for feedback correction of MPC on the one hand, and for updating the parameters of the data-driven compensation model in step S1 on the other hand, and the weighting factors of the hybrid model are readjusted to realize the online collaborative evolution of the digital twin model.
[0025] Preferably, in step S7, key performance indicators of the control system are defined, including but not limited to setpoint tracking error, fluctuation range of operating variables, and achievement of economic targets; based on the performance indicators within the rolling time window, the weight coefficients, prediction time domain, and control time domain of the predictive controller are dynamically adjusted to balance the response speed and stability of the control system.
[0026] Preferably, in step S8, the process energy efficiency index and environmental emission parameters are calculated in real time; the energy efficiency and emission index are integrated as soft constraints or additional penalty terms into the multi-objective rolling optimization in step S5 to form an economic-safety-environmental comprehensive optimization framework.
[0027] Preferably, in step S8, the process energy efficiency and environmental emission monitoring are integrated to construct a multi-dimensional sustainability optimization target. Based on the cumulative deviation between historical data and real-time data, a model confidence index is calculated. When the confidence level is lower than a threshold, the model retraining process is automatically initiated to update the mechanistic model parameters and the data-driven compensation model, and the weighting factor of the digital twin hybrid model is recalibrated.
[0028] The beneficial effects of the model-based multivariate predictive control method for pentafluoroethane production process described in this invention are as follows:
[0029] High-precision and self-evolving process control has been achieved: By constructing a digital twin hybrid model that combines a mechanistic model and a data-driven model, and utilizing real-time data for online collaborative updates, the mismatch problem of traditional linear models is effectively overcome. This model can adaptively track changes in process dynamic characteristics caused by catalyst decay and other factors, significantly improving model prediction accuracy and laying a solid foundation for advanced optimization control.
[0030] This method achieves full-cycle, adaptive economic optimization: by estimating catalyst performance indices online and dynamically reconstructing the economic objective function, it breaks through the limitations of traditional static optimization models. It can intelligently adjust the optimization focus based on the actual activity stage of the catalyst (e.g., pursuing high throughput in the fresh catalyst stage and high selectivity in the degradation stage), thereby continuously tapping the economic potential of the process throughout the entire production cycle and maximizing economic benefits.
[0031] This approach achieves a leap from "rigid constraints" to "flexible risk management": by calculating dynamic risk boundaries, fixed hard safety constraints are transformed into time-varying, flexible safety boundaries. This method can proactively assess risks based on predictions of future operating conditions, avoiding the sacrifice of economic benefits due to overly tight constraints under safe operating conditions, and enabling the implementation of mitigation measures in advance when risks accumulate, significantly improving the safety and operational flexibility of process operation.
[0032] A multi-objective synergistic optimization framework encompassing economy, safety, and environmental protection was constructed. This framework deeply integrates economic objectives, dynamic safety boundaries, and energy efficiency and emission targets into a single rolling optimization problem, resolving the issue of objective fragmentation in traditional control methods. This approach enables the simultaneous balancing of economic benefits, safety risks, and environmental impacts in every control decision, thereby achieving truly sustainable, green, and intelligent production.
[0033] The system's intelligence and long-term stability have been enhanced: By monitoring control system performance in real time and adaptively adjusting control parameters, the system can self-tune based on actual control effects (such as tracking error and fluctuation amplitude), maintaining the controller's performance at its optimal state. Combined with model confidence assessment and automatic retraining mechanisms, the long-term effectiveness and reliability of the entire digital twin and control system are ensured, reducing maintenance costs. Attached Figure Description
[0034] Figure 1 This is a flowchart of a model-based multivariate predictive control method for pentafluoroethane production process proposed in this invention. Detailed Implementation
[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0036] Example 1
[0037] Reference Figure 1 A model-based multivariate predictive control method for pentafluoroethane production process, comprising:
[0038] S1: Construct a digital twin hybrid model of the pentafluoroethane production process;
[0039] S2: Online estimation of catalyst performance index based on digital twin hybrid model;
[0040] S3: Dynamically reconstruct the economic optimization objective function based on the catalyst performance index;
[0041] S4: Calculate the dynamic risk boundary of key safety parameters;
[0042] S5: Perform economic-safety multi-objective rolling optimization with integrated dynamic risk boundaries;
[0043] S6: Implement control and update the digital twin model online;
[0044] S7: Monitors and controls system performance indicators in real time, and adaptively adjusts control parameters based on performance evaluation;
[0045] S8: Integrate process energy efficiency and environmental emission monitoring to build multi-dimensional sustainability optimization goals.
[0046] In this embodiment, in S1, a dynamic mechanism model containing core units such as reactors and distillation columns is established based on process mechanism and historical data; at the same time, real-time operation data is collected, and a data-driven neural network is trained as a compensation model; the mechanism model and the data-driven model are run in parallel, and their outputs are fused through a weighting factor that is adaptively adjusted based on recent prediction errors to form the final output of the digital twin hybrid model.
[0047] In this embodiment, in S2, a hybrid model is used, combined with real-time reactor inlet and outlet component and temperature data, and a nonlinear state observer or moving window parameter estimation algorithm is used to calculate in real time the comprehensive performance index reflecting the catalyst activity and selectivity.
[0048] In this embodiment, in S3, a multi-parameter objective function library with economic benefits as the core is established. This function library contains optimization focuses for different catalyst performance stages. Based on the performance index obtained in S2, the current optimal economic objective function is selected or interpolated from the objective function library for use in upper-level optimization.
[0049] In this embodiment, in S4, a functional relationship is established between the reactor hot spot temperature, the system pressure safety valve setpoint, and the operating variables, feed load, and catalyst performance. Based on the current operating conditions and the future dynamic prediction of the digital twin hybrid model, a risk probability model or safety margin calculation method is used to calculate the time-varying and flexible dynamic risk boundary in the future time domain.
[0050] In this embodiment, in S5, in each control cycle, the economic objective reconstructed in S3 is taken as the main optimization objective, and a penalty term for the predicted value exceeding the dynamic risk boundary in S4 is introduced to construct a new multi-objective optimization problem. A numerical optimization algorithm is used to solve for a series of operational variable increments in the future control time domain, so as to minimize the multi-objective optimization function.
[0051] In this embodiment, in S6, the first control increment obtained from S5 is applied to the actual production process; a new round of process measurement data is collected, which is used for feedback correction of MPC on the one hand, and for updating the parameters of the data-driven compensation model in S1 on the other hand, and the weighting factor of the hybrid model is readjusted to realize the online collaborative evolution of the digital twin model.
[0052] In this embodiment, in S7, key performance indicators of the control system are defined, including but not limited to setpoint tracking error, fluctuation range of operating variables, and achievement of economic goals. Based on the performance indicators within the rolling time window, the weight coefficients, prediction time domain, and control time domain of the predictive controller are dynamically adjusted to balance the response speed and stability of the control system.
[0053] In this embodiment, in S8, the process energy efficiency index and environmental emission parameters are calculated in real time; the energy efficiency and emission index are integrated as soft constraints or additional penalty terms into the multi-objective rolling optimization in S5 to form an economic-safety-environmental comprehensive optimization framework.
[0054] In this embodiment, in step S8, the process energy efficiency and environmental emission monitoring are integrated to construct a multi-dimensional sustainability optimization target. Based on the cumulative deviation between historical and real-time data, the model confidence index is calculated. When the confidence level is lower than the threshold, the model retraining process is automatically initiated to update the mechanistic model parameters and the data-driven compensation model, and the weighting factor of the digital twin hybrid model is recalibrated.
[0055] Example 2: A Multivariate Prediction and Control Method Based on Deep Reinforcement Learning
[0056] This embodiment uses deep reinforcement learning (DRL) to replace the hybrid model and optimization framework of Embodiment 1, aiming to achieve more efficient control through autonomous learning.
[0057] Steps Overview:
[0058] S1: Build a deep neural network (DNN) model, train it using historical process data, and use it to predict process states and outputs.
[0059] S2: The catalyst performance index is directly estimated through DNN, without the need for an explicit state observer.
[0060] S3: The economic objective function is dynamically adjusted based on the DRL reward function, which includes economic benefit indicators.
[0061] S4: The security risk boundary is predicted by DNN and incorporated into the DRL strategy as a constraint.
[0062] S5: Multi-objective optimization is performed through DRL algorithms (such as near-end policy optimization) to balance economy and security.
[0063] S6: Control actions are generated by the DRL policy, and the model is updated online through experience replay and periodic retraining.
[0064] S7: Performance monitoring automatically adjusts DRL hyperparameters (such as learning rate) through reward signals.
[0065] S8: Energy efficiency and emissions indicators are incorporated as part of the reward function to form comprehensive sustainability goals.
[0066] Example 3: Adaptive Control Method Based on Fuzzy Logic
[0067] This embodiment employs fuzzy logic control, combined with expert knowledge, to achieve simple and reliable adaptive control, suitable for scenarios with high model uncertainty.
[0068] step:
[0069] S1: Construct a fuzzy logic model and define the membership functions of process variables (such as temperature and pressure) based on expert rules.
[0070] S2: The catalyst performance index is estimated by a fuzzy inference system, with the reactor inlet and outlet data as inputs.
[0071] S3: The economic objective function is adjusted through fuzzy rules, and the optimization focus is switched according to different catalyst states.
[0072] S4: The safety risk boundary is defined as a fuzzy set and is dynamically adjusted based on the current operating conditions.
[0073] S5: Multi-objective optimization is performed through fuzzy decision-making, and the output is the adjustment amount of the operational variables.
[0074] S6: The control action is output by the fuzzy controller, and the membership function is adjusted online based on error feedback.
[0075] S7: Performance monitoring adjusts control parameters (such as quantization factors) through fuzzy evaluation.
[0076] S8: Energy efficiency and emissions are incorporated as fuzzy variables into the optimization, and the weighted average method is used to synthesize the objectives.
[0077] Comparative Example 1: Traditional PID Control Method
[0078] This comparative example uses a traditional PID control strategy without advanced optimization or model prediction, and is used to compare baseline performance.
[0079] step:
[0080] Multiple single-loop PID controllers are used to control key variables (such as reactor temperature and pressure).
[0081] Economic goals are fixed, and operational points are set based on experience.
[0082] Safety is ensured through fixed high-limit alarms and emergency shutdown procedures, with no dynamic risk boundaries.
[0083] There are no online model updates or adaptive performance adjustments.
[0084] No energy efficiency and emissions integration optimization.
[0085] Comparative Example 2: Basic MPC Control Method
[0086] This comparative example uses basic model predictive control (MPC), but lacks the advanced features of Example 1, such as digital twin and dynamic optimization.
[0087] step:
[0088] S1: Prediction is made using a linear state-space model based on offline identification.
[0089] S2: Catalyst-free performance estimation, with fixed model parameters.
[0090] S3: The economic objective function is fixed and there is no dynamic reconstruction.
[0091] S4: Safety constraints are fixed values, with no dynamic risk boundaries.
[0092] S5: Single-objective economic optimization with no safety penalty.
[0093] S6: Control is implemented, but the model is not updated online.
[0094] S7: No performance adaptive adjustment.
[0095] S8: No energy efficiency or emissions optimization.
[0096] The data from Examples 1, 2, and 3 are compared with those from Comparative Examples 1 and 2 in the table below:
[0097] index Example 1 Example 2 Example 3 Comparative Example 1 Comparative Example 2 Control error (°C) ±0.55 ±0.50 ±1.00 ±2.50 ±1.20 Production increase (%) +5% +8% +5% -8% -3% Energy consumption reduction (%) -5% -9% -8% +10% +4% Number of events exceeding the limit 3 2 4 15 8 Emissions reduction (%) -10% -15% -12% +12% +5% Calculation time (ms) 40 50 20 5 30
[0098] Example 1 is used as a benchmark, and others are compared with it.
[0099] Experimental data show that Examples 1 and 2 have the best overall performance, while Comparative Examples 1 and 2 are inferior in terms of control accuracy and economy. Example 3 has an advantage in computational efficiency, but its control accuracy is slightly inferior.
[0100] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A model-based multivariable predictive control method for a pentafluoroethane production process, characterized by, Comprise: S1: Construct a digital twin hybrid model of the pentafluoroethane production process; S2: Based on the digital twin hybrid model, estimate the catalyst performance index online; S3: According to the catalyst performance index, dynamically reconstruct the economic optimization objective function; S4: Calculate the dynamic risk boundary of the key safety parameters; S5: Perform economic-safety multi-objective rolling optimization fused with dynamic risk boundary; S6: Implement control and update the digital twin model online; S7: Real-time monitor the control system performance indicators, and adaptively adjust the control parameters based on performance evaluation; S8: Integrate process energy efficiency and environmental emission monitoring to build a multi-dimensional sustainability optimization target.
2. The method according to claim 1, wherein In S1, based on process mechanism and historical data, a dynamic mechanism model of the core unit including reactor and rectifying column is established; At the same time, real-time operation data is collected to train a data-driven neural network as a compensation model; The mechanism model and the data-driven model are run in parallel, and the output is fused through a weighting factor based on the recent prediction error adaptive adjustment to form the final output of the digital twin hybrid model.
3. The method according to claim 2, wherein the model-based multivariable predictive control method for a process of producing pentafluoroethane is characterized by, In S2, using the hybrid model, combining real-time reactor inlet and outlet components and temperature data, through a nonlinear state observer or a moving window parameter estimation algorithm, the comprehensive performance index reflecting the catalyst activity and selectivity is calculated in real time.
4. The method according to claim 3, wherein the model-based multivariable predictive control method for a process of producing pentafluoroethane is characterized by, In S3, a multi-parameter objective function library with economic benefit as the core is established, which includes optimization focus for different catalyst performance stages; According to the performance index obtained in S2, the current optimal economic objective function is selected or interpolated from the target function library for upper optimization.
5. The model-based multivariable predictive control method for a process of pentafluoroethane production according to claim 4, wherein, In S4, for the reactor hot spot temperature and the system pressure safety valve set value, the function relationship between its operating variables, feed load and catalyst performance is established; Based on the current working condition and the future dynamic prediction of the digital twin hybrid model, the risk probability model or the safety margin calculation method is used to roll out the time-varying and flexible dynamic risk boundary in the future time domain.
6. The model-based multivariable predictive control method for a process of pentafluoroethane production according to claim 5, wherein, In S5, in each control period, the economic target reconstructed by S3 is used as the main optimization target, and a penalty term for the predicted value exceeding the dynamic risk boundary in S4 is introduced to construct a new multi-objective optimization problem; Use numerical optimization algorithm to solve a series of operating variable increments in the future control time domain, so that the multi-objective optimization function is minimized.
7. The model-based multivariable predictive control method for a process of pentafluoroethane production according to claim 6, wherein, In S6, the first control increment obtained by S5 is applied to the actual production process; Collect new round of process measurement data, which is used for feedback correction of MPC on one hand, and for updating parameters of data-driven compensation model in S1 and readjusting the weighting factor of hybrid model on the other hand, to realize online co-evolution of digital twin model.
8. The model-based multivariable predictive control method for a process of producing pentafluoroethane according to claim 7, wherein In S7, define the key performance indicators of the control system, including but not limited to set value tracking error, operating variable fluctuation amplitude, economic target achievement degree; Based on the performance index calculation in the rolling time window, dynamically adjust the weight coefficient, prediction time domain and control time domain of the predictive controller to balance the response speed and stability of the control system.
9. The model-based multivariable predictive control method for a process of pentafluoroethane production according to claim 8, wherein, In the S8, real-time process energy efficiency indicators and environmental emissions parameters are calculated; energy efficiency and emissions indicators are integrated into the multi-objective rolling optimization of S5 as soft constraints or additional penalty terms, forming an economic-safety-environmental protection comprehensive optimization framework.
10. The model-based multivariable predictive control method for a process of pentafluoroethane production according to claim 9, wherein, In the S8, process energy efficiency and environmental emissions monitoring are integrated, multi-dimensional sustainability optimization objectives are constructed, model confidence indicators are calculated based on the cumulative deviation of historical data and real-time data; when the confidence is lower than the threshold, the model retraining process is automatically started, the mechanism model parameters and the data-driven compensation model are updated, and the weighting factor of the digital twin hybrid model is recalibrated; the model confidence indicators are calculated based on the cumulative deviation of historical data and real-time data; when the confidence is lower than the threshold, the model retraining process is automatically started, the mechanism model parameters and the data-driven compensation model are updated, and the weighting factor of the digital twin hybrid model is recalibrated.