A real-time optimization and model predictive control method for multi-component separation processes

Abstract

The present application belongs to the field of energy integration and advanced control of chemical process, and relates to a real-time optimization and model predictive control method for multi-component separation process, comprising: constructing a real-time optimization decision layer (RTO) based on data driving; constructing a multivariable model predictive control execution layer (MPC); and constructing a real-time optimization decision layer-model predictive control execution layer layered collaborative execution architecture; the present application is a double-layer collaborative architecture constructed by "upper-layer data-driven RTO economic optimization + lower-layer MPC multivariable execution", which utilizes low-order linear time-invariant models, limited time-domain objective functions and rolling optimization to coordinate the execution of the optimal temperature set value issued by the RTO layer.

Description

Technical Field

[0001] This invention belongs to the field of energy integration and advanced control of chemical processes, and relates to a real-time optimization and model predictive control method for multi-component separation processes. Background Technology

[0002] Multi-component separation processes are widely used in the production of petrochemicals, fine chemicals, and basic chemicals. For multi-component materials with similar boiling points, such as benzene, toluene, and m-xylene, multiple separation units are typically connected in series to achieve stepwise separation. These separation units interact through material and energy flows, resulting in processes characterized by multivariable coupling, nonlinearity, and dynamic lag. When the feed flow rate or composition changes, the original material balance, energy balance, and economic operating point also change. Maintaining a fixed operating target can easily lead to product quality deviations and increased energy consumption during transition phases.

[0003] For the dynamic control of multi-component separation processes, existing technologies typically employ proportional-integral-derivative (PID) controllers to regulate process variables such as temperature, flow rate, pressure, and level. This type of control method offers advantages such as simple structure, mature parameter tuning, and low engineering implementation costs. However, it primarily relies on feedback regulation based on current and historical deviations, lacking the ability to predict the future dynamic behavior of the system. When multiple control loops act simultaneously on a separation system with strong coupling characteristics, the regulatory action of one loop may become a disturbance to other loops, leading to mutual interference between loops. Furthermore, PID controllers struggle to directly handle operational constraints such as reflux ratio, reboiler heat load, and product quality range during the control process. When feed conditions or operating setpoints change significantly, problems such as response lag, increased overshoot, and prolonged transient processes can easily occur.

[0004] To improve the multivariate coordinated control capability of complex separation processes, Model Predictive Control (MPC) has been introduced into the field of chemical process control. MPC utilizes a process dynamic model to predict the system's output changes over a finite time domain, calculates the control sequence by solving a constrained optimization problem, and continuously updates the control action using rolling optimization and feedback correction methods. It can simultaneously consider variable coupling relationships and process constraints. For example, one paper established an MPC control structure suitable for multi-tower azeotropic separation processes based on a linear time-invariant model and used a third-generation non-dominated sorting genetic algorithm to optimize the weight parameters of the MPC controller, showing that its control performance is superior to traditional PID control. Another paper established a linear state-space model for an internally thermally coupled separation system and designed an MPC controller to coordinate the interaction between different control loops and achieve stable control of product composition. Yet another paper combined linear MPC with traditional PI control to construct a cascade control structure, improving the system's disturbance rejection capability under fluctuating feed flow rate and feed composition conditions. These techniques demonstrate that MPC can improve multivariate coordinated control; however, MPC is essentially still a setpoint tracking control, and its control objective depends on a pre-given operating setpoint. When feed conditions and production load change, even if the MPC controller can stabilize the system near its original setpoint, that setpoint may no longer correspond to the economically optimal operating state under the current conditions. Therefore, Real-Time Optimization (RTO) technology has been introduced into multi-level process control systems. One paper proposes an RTO framework based on a steady-state model for a "reactor-separator-circulation" process, adjusting the system operating point according to changes in operating conditions by setting economic objectives such as minimum energy consumption or maximum yield. To reduce the online computational burden of rigorous mechanistic models, another paper uses artificial neural networks to approximate the entire natural gas-to-oil process model, shortening the originally time-consuming optimization process to the second level. Another paper generates samples through process simulation and optimization algorithms, training an artificial neural network to predict optimal operating parameters under different feed composition conditions, reducing system energy consumption while meeting product purity requirements.

[0005] Building upon this, patent application number ZL202610314218.7 discloses a data-driven RTO-PID control method, which solves the problems of computational delay in traditional mechanism RTO and weak resistance to component disturbances in single-temperature PID. However, this method still uses independent dual-temperature PID loops to execute the optimized setpoint. PID is essentially a post-event feedback adjustment based on deviation, which cannot explicitly describe the dynamic coupling relationship between the two towers in a unified model, nor can it predict the combined impact of disturbances and manipulated variables on future outputs. When executing the changed setpoints issued by RTO, it suffers from response lag, large overshoot, and high energy consumption during the transition process, failing to fully unleash the energy-saving potential of data-driven RTO. Therefore, how to construct a hierarchical collaborative architecture of data-driven RTO and MPC to achieve a closed-loop unity of "economic optimization decision-making - multi-variable precise execution" and further explore the dynamic energy-saving potential of the separation process is a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0006] To address the issues of insufficient multivariate coordination, dynamic tracking lag, and long transition time in existing RTO-PID control methods that employ dual-temperature PID controllers to execute the optimal temperature setpoint, this invention provides a real-time optimization and model predictive control method for multi-component separation processes. This method employs a two-layer collaborative architecture of "upper-layer data-driven RTO economic optimization + lower-layer MPC multivariate execution," utilizing a low-order linear time-invariant model, a finite-time objective function, and rolling optimization to coordinate the execution of the optimal temperature setpoint issued by the RTO layer.

[0007] The method of this invention comprises three progressive stages: Phase 1: Constructing an RTO decision layer based on genetic algorithms and artificial neural networks (GA-ANN).

[0008] Phase 2: Constructing a multivariate model predictive control execution layer. A linear state-space model near the operating point is extracted from the Aspen rigorous mechanism model. After variable scaling, model order reduction, and dynamic response verification, a low-order linear time-invariant (LTI) model for online multi-step prediction is obtained. Based on this, a multivariate MPC controller is established, incorporating output tracking error, manipulated variable increment, and manipulated variable constraints into finite-time domain optimization. The output tracking error weights and manipulated variable increment suppression weights are then tuned offline using a GA (Gaussian General Aspect) method.

[0009] Phase 3: Constructing a multi-timescale collaborative architecture between the upper-layer RTO and the lower-layer MPC. The upper-layer RTO generates the optimal temperature setpoint; the lower-layer MPC coordinates the execution of the optimal temperature setpoint through rolling optimization and feedback correction.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows: A real-time optimization and model predictive control method for multi-component separation processes, comprising the following steps: Step 1: Construct a data-driven real-time optimization decision layer (RTO). For a multi-component separation system of benzene-toluene-m-xylene, a steady-state model and a dynamic model of a dual-tower series connection are first established to determine the four temperature-sensitive plates and their manipulated variables. Then, steady-state samples of the entire operating space are obtained through Latin hypercube sampling, and an artificial neural network (ANN) surrogate model is trained. Finally, the real-time feed conditions are input into the ANN surrogate model and coupled with a genetic algorithm (GA) to obtain the optimal temperature setpoint under the conditions of minimizing the total reboiler heat load and product molar purity constraints. The details are as follows: A steady-state model for multi-component separation in a dual-tower series configuration was established using Aspen Plus, i.e., the Aspen rigorous mechanism model. A dual-temperature coordinated dynamic control architecture for the dual-tower series configuration was established using Aspen Dynamics, i.e., the Aspen dynamic model. An open-loop steady-state sensitivity analysis method was used to determine the temperature-sensitive plate. Two temperature controllers were configured in each distillation column. The temperature of the upper sensitive plate in both columns was determined by the reflux ratio (…). The temperature of the sensitive plates at the bottom of the two towers is controlled by the heat load of the reboiler at the bottom of the tower. The control is achieved using Latin hypercube sampling. Steady-state samples of the entire operating space, including total feed flow, are obtained from the Aspen rigorous mechanism model. Feed composition of benzene, toluene, and m-xylene The temperatures of the four temperature-sensitive plates Molar purity of the three products And the reboiler heat load of the two distillation columns An ANN surrogate model is trained using steady-state samples. The input variables of the ANN surrogate model include... , Output variables include , The total feed flow rate and the composition of the three components are acquired in real time and fixed as known variables. These are then input into the ANN surrogate model and coupled with a genetic algorithm (GA). The temperatures of the four temperature-sensitive plates are used as the variables to be optimized. The objective is to minimize the sum of the heat loads of the two reboilers, and the molar purity of the three products is constrained to be no less than the target purity. Rapid iterative optimization is performed. The objective function and constraints are as follows: (1) (2) in, The objective function is... This is the sum of the heat loads of the two reboilers. ; This represents the minimum product molar purity. After the genetic algorithm converges, it yields the optimal temperature setpoint vector. .

[0011] Step 2: Construct the multivariate model predictive control execution layer (MPC); First, a low-order linear time-invariant (LTI) model suitable for online prediction is extracted and validated from the Aspen dynamic model. Then, based on the LTI model, an MPC finite-time domain optimization problem is constructed, incorporating output tracking error, manipulated variable increments, and manipulated variable constraints. Rolling optimization and feedback correction are used to implement closed-loop control. Finally, the four output tracking error weights and four manipulated variable increment suppression weights are tuned offline using GA. Details are as follows: Step 2.1: Construct a low-order linear time-invariant model; To address the requirements of a dual-tower, dual-temperature process, a highly coupled multivariable control network is established. Manipulated variable vectors are used. Controlled variable vector and measurable perturbation vector They are as follows: (3) (4) (5) in, For the first Each sampling time; , The first The reflux ratio and reboiler heat load of the first distillation column at each sampling time; , The first The reflux ratio and reboiler heat load of the second distillation column at each sampling time; , , , The first Temperatures of the four temperature-sensitive plates at each sampling time; For the first The total feed flow rate at each sampling time. , , The first The mole fraction of each component in the feed at each sampling time.

[0012] The Aspen dynamic model is run to its optimal steady-state operating point, and a linear state-space model near this point is extracted. However, this model still has a high order, and directly using it for online predictive control would significantly increase the computational cost of rolling optimization and weaken the controller's real-time performance. Therefore, the manipulated variables, controlled variables, and measurable disturbance variables in the linear state-space model are normalized to reduce the impact of different dimensions and orders of magnitude on the controller's computation. Subsequently, while preserving as much of the system's dominant low-frequency dynamic characteristics as possible, the order of the linear state-space model is reduced to obtain an LTI model suitable for online solution. z (6) (7) in, For discrete state matrices, They are respectively The input matrix; This is the output matrix. For the first The internal state vector at each sampling time is composed of the dominant state components formed by the order reduction transformation of state variables such as the temperature of each tray of the two distillation columns, the gas-liquid phase composition and the liquid holdup. Each component is usually a linear combination of the above-mentioned multiple physical states, rather than a single process variable that can be directly measured.

[0013] Subsequently, under the same initial state and input conditions, a consistent step test signal was applied to both the Aspen strict mechanism model and the LTI model, and the consistency of their dynamic response results on the key output variables—the temperatures of the four temperature-sensitive plates—was compared. Within a certain deviation range, when the LTI model could reproduce the main dynamic behavior of the Aspen strict mechanism model, it was adopted as the internal predictive model (LTI) of the MPC controller.

[0014] Step 2.2: Construct the MPC objective function and implement rolling optimization; At each sampling time, the internal prediction model LTI is based on the vector of controlled variables. Measurable perturbation vector Optimal temperature setpoint After predicting the future output temperature trajectory, the MPC controller needs to operate within the set prediction time domain. and control time domain Inside( Find the quadratic performance index function that minimizes the sum of the output tracking error and the cost of the manipulated variable increment: (8) in, It is a quadratic performance index function used to achieve a mathematical trade-off between control accuracy and execution cost. To output the tracking error penalty term, To suppress the penalty term for the increment of the manipulated variable, This represents the weighted L2 norm of the output tracking error. This represents the weighted L2 norm of the increment of the manipulated variable. For the first The information at the sampling time is obtained for the first... The predicted output temperature is given by step, where r(k+i) is the optimal temperature setpoint vector given by the upper-layer RTO at the (k+i)th sampling time. To manipulate the increment of the variable, For the first The increment of the manipulated variable at the j-th future step is obtained from the information at each sampling time. To output the tracking error weight matrix, express It is a symmetric positive semi-definite matrix. To suppress the increment of the manipulated variable weight matrix, >0 indicates It is a symmetric positive definite matrix, thus ensuring that both penalties are non-negative.

[0015] For the constructed Aspen dynamic model, the output tracking error weight matrix is ​​defined as follows: (9) in, , , , These are the four output tracking error weights, corresponding to... , , , .

[0016] definition The output tracking error penalty term is then expanded as follows: (10) in, Represents the output tracking error. , , , express , , , The tracking error of the predicted output relative to the corresponding reference setpoint.

[0017] Similarly, the expansion of the manipulator's increment suppression penalty term is: (11) in, , , , These are the incremental suppression weights for the four manipulated variables, respectively, corresponding to... ; , , , They are respectively The increment.

[0018] After solving the MPC finite-time optimization problem shown in equation (8), the optimal manipulator increment sequence is obtained. : (12) in, This is the first step optimal manipulator increment obtained based on the information at the k-th sampling time. This represents the increment of the optimal manipulated variable at the M-th step within the same control time domain.

[0019] According to the rolling optimization principle, only the first term in the optimal manipulator increment sequence is selected. Acting on multi-component separation processes; in the first At each sampling time point, the actual output is reread, the deviation between it and the predicted output at the previous time point is calculated, and the internal state and the predicted future temperature trajectory are corrected using a state estimator. Then, the finite-time domain optimization problem is solved again. This rolling closed-loop mechanism effectively overcomes control drift caused by Aspen model mismatch and external disturbances. Upper and lower bound constraints on the manipulated variables are also incorporated into the solution process to ensure that the control actions always remain within the safe operating range of the equipment. The manipulated variable constraints are RR... 1,min ≤RR1≤RR 1,max Q 1,min ≤Q1≤Q 1,max RR 2,min ≤RR2≤RR 2,max and Q 2,min ≤Q2≤Q 2,max Each lower limit is the minimum reflux ratio or minimum reboiler heat load allowed to maintain stable separation in the corresponding distillation column, and each upper limit is the maximum value allowed by the equipment processing capacity and safe heat load. The values ​​are jointly determined by the Aspen steady-state feasible region, the equipment design capacity and the on-site safety interlock limits.

[0020] Step 2.3: Use GA to tune the MPC controller weights offline; In an MPC controller, the selection of parameters for the output tracking error weight matrix and the manipulated variable increment suppression weight matrix directly affects the system's dynamic decoupling capability, anti-interference performance, and control smoothness. and The eight diagonal elements form the parameter vector to be optimized, with the goal of achieving the optimal combination of control performance and motion smoothness under multiple disturbance conditions, based on the integral absolute error (...). The evaluation approach uses a weighted sum of the output tracking error term and the manipulation variable increment suppression term as the fitness function of the genetic algorithm. (13) (14) (15) (16) in, The fitness function of the genetic algorithm is... To output the tracking error term, This is the increment suppression term for the manipulated variable; To balance the weighting coefficients for tracking accuracy and control motion smoothness, The eight weight parameters to be optimized are: The number of sampling points in the evaluation time domain is determined by the weighting. This is the reference setpoint vector, which is the reference vector for the four temperature-sensitive plates in the weight tuning scenario; For the first The incremental vector of the manipulated variables at each sampling time. When the fitness function satisfies the preset convergence condition or reaches the maximum number of iterations, the resulting eight weights are written to the MPC controller.

[0021] Step 3: Construct a hierarchical collaborative execution architecture of real-time optimization decision layer - model prediction control execution layer (RTO-MPC). First, configure the update cycles of RTO and MPC to form a time-scale stratification with a slower upper scale and a faster lower scale. Then, the RTO calculates and issues four optimal temperature setpoints based on real-time feed conditions, and the MPC coordinates the rolling tracking of four manipulated variables within operational constraints. Finally, the synergistic control effect is evaluated by the integral of the total reboiler heat load and the absolute error of the product molar purity. Details are as follows: Step 3.1: Configure time-scale hierarchical structure; The upper layer, the Real-Time Optimization (RTO) layer, operates on a slow timescale, periodically updating four optimal temperature setpoints based on real-time feeding conditions. The lower layer, the Model Predictive Control (MPC) layer, operates on a fast timescale. The update cycle for the slow timescale can be 15 hours, while the cycle for the fast timescale can be 1 minute. After each update of the optimal temperature setpoint by the RTO, the MPC can complete multiple rolling optimizations and feedback corrections before the next update by the upper layer.

[0022] Step 3.2, Collaborative Closed-Loop Execution; The system collects data on total feed flow rate and feed composition in real time, inputs it into the upper-level RTO to calculate the optimal temperature setpoint under the current operating conditions, and sends it to the MPC controller. The MPC controller coordinates the actions of each manipulated variable based on the predictive model. Under the premise of meeting the operating constraints, it coordinates the execution of the optimal temperature setpoint through rolling optimization and feedback correction to achieve the economical and optimal operation of the process.

[0023] To ensure that the improvement effect reflects only the differences in the lower-level execution controller, the RTO-MPC cooperative control system of this invention uses the same external disturbance amplitude as the RTO-PID and the optimal temperature setpoint calculated by the upper-level RTO under the corresponding operating conditions. Therefore, the differences in dynamic quality and economic performance can more realistically reflect the difference in control capabilities between MPC and PID when executing the same optimization instructions under multivariable coupling conditions.

[0024] Furthermore, the total energy consumption of the reboiler heat load using two separation units is... The integral of the absolute error of the product molar purity The two indicators are used to evaluate economic performance and dynamic control performance. They are used to compare the energy-saving effect and dynamic tracking improvement of different control schemes under the same disturbance and the same optimal temperature setpoint.

[0025] (17) (18) in, The start time of the disturbance. To evaluate the termination time, For the occurrence of disturbance The sum of the real-time heat loads of the total reboilers of the two distillation columns at any given time. For the occurrence of disturbance The actual molar purity of the product at any given time. To set the molar purity of the product.

[0026] The beneficial effects of this invention are as follows: (1) Significantly improves dynamic tracking performance and shortens system settling time: By replacing single-loop PID with multivariable MPC, the variable coupling relationship of the separation process is explicitly handled. When tracking the dynamic setpoint of RTO, the average settling time of key products is shortened by 29%~47%, and the product purity dynamic absolute error integral ( The efficiency has been further reduced, effectively solving the problems of slow response and large overshoot when traditional PID is executed.

[0027] (2) Reduce energy consumption during the transition phase and fully unleash the energy-saving potential of RTO: MPC significantly reduces the ineffective heat load fluctuations during the transition phase through predictive regulation and smooth control actions; under disturbance conditions, the total energy consumption of RTO-MPC is reduced by 2.5% compared to the fixed setpoint MPC; compared to RTO-PID, the total energy consumption of RTO-MPC is further reduced by 0.095%, achieving the unity of steady-state energy saving and dynamic energy saving.

[0028] (3) Explicitly handle operational constraints to improve the safety of operating conditions: MPC directly incorporates operational constraints such as the upper and lower limits of reflux ratio and the upper and lower limits of reboiler heat load into the rolling optimization solution, avoiding the risk of equipment exceeding limits when operating conditions fluctuate drastically, and improving the robustness of system operation while ensuring economy. Attached Figure Description

[0029] Figure 1 This is a basic flowchart of the present invention; Figure 2 This is a comparison chart of the responses of the LTI prediction model and the Aspen strict mechanism model in this embodiment of the invention; Figure 3 This is a convergence curve of the GA offline tuning of the MPC weight parameters in an embodiment of the present invention; Figure 4 This is a comparison chart of the product purity response of MPC and RTO-MPC under combined disturbance conditions in the embodiments of the present invention; Figure 5 This is a comparison diagram of the reboiler thermal load response of MPC and RTO-MPC under feed flow disturbance conditions in the embodiments of the present invention; Figure 6 This is a comparison chart of the product purity response of RTO-PID and RTO-MPC under feed composition disturbance conditions in the embodiments of the present invention; Figure 7 This is a comparison diagram of the reboiler thermal load response of RTO-PID and RTO-MPC under combined disturbance conditions in an embodiment of the present invention. Detailed Implementation

[0030] The present invention will be described in detail below with reference to specific embodiments and accompanying drawings.

[0031] This embodiment focuses on the separation process of a typical multi-component mixture of benzene, toluene, and m-xylene (BTX) in the petrochemical industry. For example... Figure 1 As shown, the feed is a saturated liquid at 310K, and the initial total molar flow rate is... The molar fractions of benzene, toluene, and m-xylene were 35.3509%, 29.9693%, and 34.6797%, respectively. The thermodynamic properties were calculated using the NRTL-RK model, and the pressure drop across the entire tray was set to 0.0068 atm. Benzene was collected from the top of the first distillation column, and the bottom stream from the first distillation column flowed into the second distillation column. Toluene was collected from the top of the second distillation column, and m-xylene was collected from the bottom.

[0032] Step 1: Building a data-driven RTO decision-making layer; 7912 valid steady-state samples were obtained using LHS (Local Hysteresis Halo), satisfying the requirement that the molar purity of the three products be no less than 0.994. These samples were then used to train an ANN (Application Not Resistant) model with four hidden layers and 20 neurons per layer. The ANN surrogate model was coupled with a GA (Gas Optimization) model, aiming to minimize the total reboiler heat load and constraining the molar purity of the three products to ≥99.4%. Online optimization was achieved within seconds. After the GA converged, the optimal temperature setpoint vector was obtained. (19) in, Vector of optimal temperature setpoints; These are the optimal setpoints for the 7th and 17th temperature-sensitive plates of the first distillation column, respectively. These are the optimal setpoints for the 15th and 24th temperature-sensitive plates of the second distillation column, respectively. Under initial operating conditions, the four optimal temperature setpoints are... , , , When the total feed flow rate or feed composition changes, the current feed conditions are fixed as known variables, and equations (1) and (2) are solved using GA to output the result. The minimum molar purity of the three products satisfies the constraints. This is used as the reference setting trajectory for the lower-level MPC controller.

[0033] Step 2: Extract, reduce, and validate the MPC internal prediction model; In the Aspen Dynamics and Simulink co-simulation environment, the Aspen rigorous mechanism model is run to the initial steady-state operating point, and the control boundary is determined according to equations (3) to (5): As four manipulated variables, As four controlled variables, As four measurable disturbance variables, a linear state-space model is extracted through the Aspen Control Design Interface (CDI).

[0034] Considering the differences in the dimensions and orders of magnitude of different variables, a scaling transformation is first performed on each variable of the linear state-space model. Then, the model order is reduced while preserving the dominant low-frequency dynamics and main coupling relationships of the system, resulting in the LTI models shown in equations (6) and (7). Subsequently, the same test signal is applied to both the LTI model and the Aspen strict mechanism model, and four temperature outputs are recorded. Figure 2 As shown, the two types of models maintain good consistency in response direction, dynamic trend and main transition characteristics. Therefore, the obtained LTI model is loaded into the MPC controller as the internal prediction model used for online multi-step prediction.

[0035] Step 3: Configure the multivariable MPC controller and tune the weights offline; After the internal prediction model was validated, the MPC controller was configured according to equations (8) to (12). To achieve the best balance between tracking accuracy and operational stability, feed flow rate, feed composition, and composite disturbances were used as offline tuning scenarios. GA search optimization was performed on the four output tracking error weights and the four manipulated variable increment suppression weights according to equations (13) to (16). The search range for the output tracking error weights was 10–50, and the search range for the manipulated variable increment suppression weights was 0.01–10. Figure 3 As shown, the algorithm converges after approximately 60 generations, yielding the optimal weight parameters: , , , as well as , , , The four output tracking error weights correspond sequentially to... The increment suppression weights of the four manipulated variables correspond sequentially. The output tracking error weights range from 22.76 to 29.09, while the manipulated variable increment suppression weights range from 0.10 to 8.54. The relatively close weights of the four output tracking errors are used to maintain relatively coordinated tracking performance across the temperature loops. The significant differences in the weights of the four manipulated variable increment suppression reflect the different dynamic sensitivities and action costs of the different manipulated variables, necessitating the use of non-uniform action penalties.

[0036] Step 4: Set up a perturbation control experiment and evaluate the synergistic effect; A complete RTO-MPC dual-layer collaborative control architecture was constructed on the Simulink-Aspen co-simulation platform. The upper-layer RTO was set to update the optimal temperature setpoint every 15 hours based on the disturbance conditions, and the lower-layer MPC had a sampling period of 1 minute for rolling optimization control. Tests were conducted under three scenarios: feed flow disturbance, feed composition disturbance, and combined flow and composition disturbance. Two sets of comparisons were used: the first set compared the fixed setpoint MPC with RTO-MPC to identify the effect of the upper-layer RTO updating the optimal temperature setpoint; the second set compared RTO-PID with RTO-MPC to evaluate the differences of the lower-layer execution controller separately. The economic and dynamic performance were evaluated using Equations (17) and (18), respectively.

[0037] Figure 4 and Figure 5 Product purity and reboiler thermal load response at fixed setpoints MPC and RTO-MPC are presented. Regarding dynamic performance, under combined disturbances, the purity of the three products... The values ​​decreased from 0.1130, 0.0913, and 0.1666 to 0.0290, 0.0559, and 0.0657 respectively, indicating that the upper-level RTO can promptly correct the control target based on the current operating conditions, thereby reducing the cumulative deviation during the system transition process. In terms of economic performance, under feed flow disturbances, the total energy consumption of the fixed setpoint MPC is... The total energy consumption of RTO-MPC is Reduced energy consumption The energy saving reached 2.5%. This indicates that the cooperative control strategy not only improved the quality of disturbance response but also enhanced the system's ability to track the economically optimal operating point under new operating conditions. Overall, the introduction of RTO (Real-Time Objective) allows the MPC controller to evolve from a simple setpoint tracking controller into a cooperative optimization control structure that balances dynamic control and operational economy.

[0038] Figure 6 and Figure 7 Product purity and reboiler thermal load response for RTO-PID and RTO-MPC are presented. Regarding dynamic performance, under feed composition disturbances, the three products... The values ​​decreased from 0.0150, 0.0370, and 0.1160 to 0.0136, 0.0345, and 0.1151, respectively. In terms of economic performance, under combined disturbances, the total energy consumption of RTO-PID... The total energy consumption of RTO-MPC is Energy consumption is further reduced. The energy saving reached 0.095%. The above difference only comes from the lower-level controller, indicating that MPC not only improves the quality of the system's transient process, but also reduces the additional energy consumption during the disturbance transient phase to a certain extent.

[0039] The improvement in settling time is even more significant. Taking feed flow disturbance as an example, the settling times for the three products under RTO-PID were approximately 3.41 h, 5.81 h, and 4.41 h, respectively, while under RTO-MPC, these times were shortened to 2.43 h, 3.07 h, and 2.60 h, representing a reduction of 29% to 47%. This indicates that MPC can more quickly pull the system back to a new optimized operating range, thereby reducing the system's prolonged stay in a transitional state with high deviation and high heat load fluctuations. The reduction in settling time not only improves control quality but also directly reduces ineffective heat consumption during the transition phase.

Claims

1. A real-time optimization and model predictive control method for multi-component separation processes, characterized in that, The specific steps are as follows: Step 1: Construct a data-driven real-time optimization decision layer (RTO); For the multi-component separation system of benzene-toluene-m-xylene, a steady-state model and a dynamic model of the two towers in series are first established, and the four temperature-sensitive plates and their manipulated variables are determined. Subsequently, steady-state samples of the entire operating space were obtained through Latin hypercube sampling and an artificial neural network (ANN) surrogate model was trained. Finally, the real-time feed conditions were input into the ANN surrogate model and coupled with the genetic algorithm (GA) to obtain the optimal temperature setpoint under the conditions of minimum total reboiler heat load and product molar purity constraint. Step 2: Construct the Multivariate Model Predictive Control Execution Layer (MPC); First, a low-order linear time-invariant (LTI) model suitable for online prediction is extracted and verified from the Aspen dynamic model. Then, based on the LTI model, an MPC finite-time domain optimization problem containing output tracking error, manipulator increment, and manipulator constraint is constructed, and closed-loop control is implemented using rolling optimization and feedback correction. Finally, the four output tracking error weights and four manipulator increment suppression weights are tuned offline using GA. Step 3: Construct a hierarchical collaborative execution architecture RTO-MPC consisting of a real-time optimization decision layer, a model prediction control execution layer, and a control layer. First, configure the update cycles of RTO and MPC to form a time-scale layering with a slower top and faster bottom; The RTO then calculates and issues four optimal temperature setpoints based on the real-time feed conditions, and the MPC coordinates the rolling tracking of the four manipulated variables within the operating constraints; finally, the effect of the coordinated control is evaluated by the integral of the total reboiler heat load and the absolute error of the product molar purity.

2. The real-time optimization and model predictive control method for multi-component separation processes according to claim 1, characterized in that, Step 1 is as follows: A steady-state model for multi-component separation in a dual-column series configuration was established using Aspen Plus, i.e., the Aspen rigorous mechanism model. A dual-temperature coordinated dynamic control architecture for the dual-column series configuration was established using Aspen Dynamics, i.e., the Aspen dynamic model. An open-loop steady-state sensitivity analysis method was employed to determine the temperature-sensitive plate. Two temperature controllers were configured in each distillation column, and the temperature of the upper sensitive plate in both columns was determined by the reflux ratio. The temperature of the sensitive plates at the bottom of the two towers is controlled by the heat load of the reboiler at the bottom of the tower. To control this; Latin hypercube sampling is used to obtain steady-state samples of the entire operating space from the Aspen rigorous mechanism model, including the total feed flow rate. Feed composition of benzene, toluene, and m-xylene The temperatures of the four temperature-sensitive plates Molar purity of the three products And the reboiler heat load of the two distillation columns ; An ANN surrogate model is trained using steady-state samples. The input variables of the ANN surrogate model include... , Output variables include , The total feed flow rate and the feed composition of the three components are obtained in real time and fixed as known variables. They are then input into the ANN surrogate model and coupled with the genetic algorithm GA. The temperature of the four temperature-sensitive plates is used as the variable to be optimized, the sum of the heat loads of the two reboilers is minimized as the objective, and the molar purity of the three products is not lower than the target purity as a constraint. The model is then used for rapid iterative optimization.

3. The real-time optimization and model predictive control method for multi-component separation processes according to claim 2, characterized in that, The objective function and constraints are as follows: (1) (2) in, The objective function is... This is the sum of the heat loads of the two reboilers. ; To achieve the lowest product molar purity; after the genetic algorithm converges, the optimal temperature setpoint vector is obtained. .

4. The real-time optimization and model predictive control method for multi-component separation processes according to claim 1, characterized in that, Step 2 is as follows: Step 2.1: Construct a low-order linear time-invariant model; To address the requirements of a dual-tower, dual-temperature process, a highly coupled multivariable control network is established; the variable vector is manipulated. Controlled variable vector and measurable perturbation vector They are as follows: (3) (4) (5) in, For the first Each sampling time; , The first The reflux ratio and reboiler heat load of the first distillation column at each sampling time; , The first The reflux ratio and reboiler heat load of the second distillation column at each sampling time; , , , The first Temperatures of the four temperature-sensitive plates at each sampling time; For the first The total feed flow rate at each sampling time. , , The first Mole fraction of each component in the feed at each sampling time; The Aspen dynamic model is run to its optimal steady-state operating point. A linear state-space model near the optimal steady-state operating point is extracted. The manipulated variables, controlled variables, and measurable disturbance variables in the linear state-space model are normalized. The order of the linear state-space model is reduced to obtain an LTI model suitable for online solution. z (6) (7) in, For discrete state matrices, They are respectively The input matrix; This is the output matrix; For the first The internal state vector at each sampling moment consists of state components formed by the reduced-order transformation of the temperature of each tray, the gas-liquid phase composition, and the liquid holdup of the two distillation columns. Subsequently, under the same initial state and input conditions, a consistent step test signal was applied to the Aspen strict mechanism model and the LTI model respectively, and the consistency of the dynamic response results of the Aspen strict mechanism model and the LTI model on the key output variables was compared. The key output variables were the temperatures of the four temperature-sensitive plates. Within a certain deviation range, when the LTI model could reproduce the main dynamic behavior of the Aspen strict mechanism model, it was used as the internal predictive model LTI of the MPC controller. Step 2.2: Construct the MPC objective function and implement rolling optimization; At each sampling time, the internal prediction model LTI is based on the vector of controlled variables. Measurable perturbation vector Optimal temperature setpoint After predicting the future output temperature trajectory, the MPC controller needs to operate within the set prediction time domain. and control time domain Inside, Find the quadratic performance index function that minimizes the sum of the output tracking error and the cost of the manipulated variable increment: (8) in, It is a quadratic performance index function used to achieve a mathematical trade-off between control accuracy and execution cost; To output the tracking error penalty term, To suppress the penalty term for the increment of the manipulated variable, This represents the weighted L2 norm of the output tracking error. The weighted L2 norm represents the increment of the manipulated variable; For the first The information at the sampling time is obtained for the first time. The predicted output temperature is given by step, where r(k+i) is the optimal temperature setpoint vector given by the upper-layer RTO at the (k+i)th sampling time. To manipulate the increment of the variable, For the first The increment of the manipulated variable at the j-th future step is obtained from the information at each sampling time. To output the tracking error weight matrix, express It is a symmetric positive semi-definite matrix. To suppress the increment of the manipulated variable weight matrix, >0 indicates It is a symmetric positive definite matrix, thus ensuring that both penalties are non-negative; After solving the MPC finite-time optimization problem shown in equation (8), the optimal manipulator increment sequence is obtained. : (12) in, This is the first step optimal manipulator increment obtained based on the information at the k-th sampling time. This represents the increment of the optimal manipulated variable at the M-th step within the same control time domain; According to the rolling optimization principle, only the first term in the optimal manipulated variable increment sequence is selected. Acting on multi-component separation processes; in the first The actual output is reread at each sampling time, the deviation between it and the predicted output at the previous time is calculated, and the internal state and the future predicted temperature trajectory are corrected using the state estimator. Then the finite time domain optimization problem is solved again. The upper and lower limits of the manipulated variables are also included in the solution process to ensure that the control action is always within the safe operating range of the equipment. Step 2.3: Use GA to tune the MPC controller weights offline; Will and The eight diagonal elements form the parameter vector to be optimized, aiming to achieve the optimal combination of control performance and motion smoothness under multiple disturbance conditions, based on the integral absolute error. The evaluation approach uses the weighted sum of the output tracking error term and the manipulation variable increment suppression term as the fitness function of the genetic algorithm: (13) (14) (15) (16) in, The fitness function of the genetic algorithm is... To output the tracking error term, This is the increment suppression term for the manipulated variable; To balance the weighting coefficients for tracking accuracy and control motion smoothness, The eight weight parameters to be optimized are: The number of sampling points in the evaluation time domain is determined by the weighting. This is the reference setpoint vector, which is the reference vector for the four temperature-sensitive plates in the weight tuning scenario; For the first The incremental vector of the manipulated variables at each sampling time; when the fitness function satisfies the preset convergence condition or reaches the maximum number of iterations, the resulting eight weights are written into the MPC controller.

5. The real-time optimization and model predictive control method for multi-component separation processes according to claim 4, characterized in that, For the constructed Aspen dynamic model, the output tracking error weight matrix is ​​defined as follows: (9) in, , , , These are the four output tracking error weights, corresponding to... , , , ; definition The output tracking error penalty term is then expanded as follows: (10) in, Represents the output tracking error. , , , express , , , The tracking error of the predicted output relative to the corresponding reference setpoint; Similarly, the expansion of the manipulator's increment suppression penalty term is: (11) in, , , , These are the incremental suppression weights for the four manipulated variables, respectively, corresponding to... ; , , , They are respectively The increment.

6. The real-time optimization and model predictive control method for multi-component separation processes according to claim 4, characterized in that, Total energy consumption of reboiler with two separation units The integral of the absolute error of the product molar purity To evaluate economic performance and dynamic control performance, these two indicators are used to compare the energy-saving effect and dynamic tracking improvement of different control schemes under the same disturbance and the same optimal temperature setpoint. (17) (18) in, The start time of the disturbance. To evaluate the termination time, For the occurrence of disturbance The sum of the real-time heat loads of the total reboilers of the two distillation columns at any given time; For the occurrence of disturbance The actual molar purity of the product at any given time. To set the molar purity of the product.

7. The real-time optimization and model predictive control method for multi-component separation processes according to claim 1, characterized in that, Step 3 is as follows: Step 3.1: Configure time-scale hierarchical structure; The upper layer, the Real-Time Optimization Decision Layer (RTO), operates on a slow time scale and periodically updates four optimal temperature setpoints based on real-time feeding conditions. The lower layer, the Model Predictive Control Execution Layer (MPC), operates on a fast time scale. After each update of the optimal temperature setpoint of the RTO, the MPC can complete multiple rolling optimizations and feedback corrections before the next update of the upper layer. Step 3.2, Collaborative Closed-Loop Execution; The system collects data on total feed flow rate and feed composition in real time, inputs it into the upper-level RTO to calculate the optimal temperature setpoint under the current operating conditions, and sends it to the MPC controller. The MPC controller coordinates the actions of each manipulated variable based on the predictive model. Under the premise of meeting the operating constraints, it coordinates the execution of the optimal temperature setpoint through rolling optimization and feedback correction to achieve the economical and optimal operation of the process.

Images