A data-driven based dual-temperature real-time optimization control method for distillation column

Abstract

This invention discloses a data-driven real-time dual-temperature optimization control method for distillation columns, belonging to the field of chemical process control and energy conservation. First, a global optimization of the steady-state model of the two-column series system is performed based on a genetic algorithm, and a dual-temperature dynamic control architecture with cross-coordination of the reboiler and reflux ratio is constructed by combining sensitivity analysis. Second, full operating space data is extracted through Latin hypercube sampling, and a high-precision artificial neural network (ANN) surrogate model is trained to replace the mechanistic equations. Finally, when facing operating disturbances, the GA-ANN is used for real-time rapid optimization to find the optimal temperature setpoint that minimizes total energy consumption, and the result is sent to the closed-loop execution of the dual-temperature PID architecture. This invention completely overcomes computational latency, limits product purity fluctuations to within ±0.002 when facing severe disturbances, effectively reduces the total system energy consumption by 1.82%, and achieves efficient, stable, and low-carbon operation of the distillation process.

Description

Technical Field

[0001] This invention belongs to the field of energy integration and advanced control of chemical processes, and specifically relates to a data-driven real-time optimization control method for dual-temperature distillation columns. Background Technology

[0002] Distillation, as the most widely used separation technology in the chemical industry, transfers the mass and heat of mixtures through a multi-stage distillation process. However, its high energy consumption, especially during high-load operation and the separation of complex multi-component materials (such as the separation of BTX aromatic mixtures like benzene, toluene, and xylene), has long constrained the green and low-carbon development of the industry. To address the high energy consumption of distillation processes, existing technologies typically optimize both process structure and operating parameters. Regarding process structure optimization, for example, the literature (Villegas-Uribe, CA, et al., Optimal design and control of three simplified sargent four-product dividing-wall columns, Chemical Engineering and Processing, 2022) improves separation efficiency by designing complex partitioned-wall column structures. While this effectively reduces steady-state energy consumption, its operational flexibility and anti-interference capabilities are severely insufficient when facing frequent fluctuations in feed composition in actual industrial applications, and the energy-saving state established by the steady-state process is easily disrupted. To address the problem of increased energy consumption under varying operating conditions, real-time optimization (RTO) strategies based on operating parameters have emerged. However, existing technologies still have significant shortcomings in handling the rapid response of dynamic models and multi-objective constraints. For example, the literature (Kumar, V., et al., Real-time optimization of a reactor–separator–recycle process II: dynamic evaluation, Ind.Eng.Chem.Res.58, 2019) proposes a real-time optimization framework based on fitting a steady-state mechanism model. However, its practice shows that traditional RTO strategies, when faced with rapid dynamic disturbances, suffer from extremely high online computational loads due to their reliance on rigorous mechanism models, often resulting in severe convergence delays. To overcome the problem of time-consuming online computation of mechanism models, some scholars have attempted to introduce data-driven models. For example, the literature (Li, H., et al., Dynamic real-time energy saving control of pressure-swing distillation based on artificial neural networks, Chem.Eng.Sci.282, 2023) explores energy-saving control of distillation based on artificial neural networks.However, when the upper-level RTO quickly calculates new operating setpoints and sends them down to the lower-level basic control system (such as a conventional single-loop PID controller), existing technologies generally face the most critical challenge: as pointed out in the literature (Kaymak, DB, & Luyben, WL, Evaluation of a two-temperature control structure..., Chem. Eng. Sci. 61, 2006), traditional single-temperature PID control systems (i.e., maintaining the temperature of a single sensitive plate solely by adjusting the reboiler heat load) have certain control performance when faced with feed flow disturbances, but when faced with disturbances in feed composition, the product concentration fluctuates drastically (the change can even exceed 0.02), completely failing to provide effective control assurance for the process. This makes the process extremely prone to loss of control during the transition phase, not only failing to achieve the energy-saving targets set by the RTO, but also causing a sharp rebound in energy consumption and an increase in the amount of defective products. Therefore, developing a novel RTO-dual-temperature control architecture for complex multi-component distillation processes that can completely overcome the delay in mechanism calculations and solve the problem of weak resistance to component interference in lower-level control is an extremely urgent and unsolved technical challenge in this field. Summary of the Invention

[0003] To address the shortcomings of existing technologies, such as slow calculation of rigorous mechanistic models and the inability of traditional single-temperature PID control to smoothly track component disturbances and optimization instructions, this invention provides a data-driven dual-temperature real-time optimization control method for distillation columns. This method is an online real-time optimization (RTO) method for complex multi-component distillation processes that combines genetic algorithm optimization, artificial neural network (ANN) surrogate model prediction, and a multi-loop dual-temperature PID dynamic control architecture.

[0004] The method of this invention comprises three progressive stages: Phase 1: Construct a steady-state separation model for a dual-tower series connection and use a genetic algorithm for global optimization of economic efficiency (TAC). Based on this, the Luyben sensitivity analysis method is used to accurately locate the two temperature-sensitive plates in each tower and construct a dual-temperature collaborative dynamic control architecture (i.e., in the single distillation tower, the reboiler heat load is used to control the lower sensitive plate, and the reflux ratio is used to control the upper sensitive plate).

[0005] The second stage: Using the steady-state model from the first stage, sample data from the entire operating space is obtained through Latin hypercube sampling (LHS), and a high-precision artificial neural network (ANN) model is trained to replace the computationally intensive mechanistic model as the prediction layer of the RTO layer.

[0006] Phase 3: Constructing a collaborative execution architecture between upper-level GA-ANN real-time optimization decision-making and lower-level dual-temperature PID control. When a disturbance in the operating condition is detected, the upper level calculates the optimal combination of temperature setpoints that minimizes the global heat load; the lower-level dual-temperature PID control system, with its strong robustness, quickly eliminates component and flow disturbances, smoothly tracks and maintains the new temperature setpoint.

[0007] The technical solution of the present invention is as follows: A data-driven real-time optimization control method for a distillation column with dual temperatures, comprising the following steps: [hM1] Step 1: Construct a steady-state separation model for multi-component distillation and a dual-temperature dynamic control architecture Step (1.1) Constructing a steady-state separation model: Establish a steady-state separation model for the dual-tower series structure, and use a genetic algorithm with a total annual cost ( The objective is to perform global optimization of process parameters with the goal of minimization, thereby obtaining the initial optimal equipment and operating parameters; the specific mathematical description of the objective function for steady-state optimization is as follows: in, This describes the functional relationship between each operating parameter and the total annual cost. , These represent the total number of trays in the first and second distillation columns, respectively. , These are the locations of the feed trays; These are the reflux ratios for the first and second distillation columns, respectively. These are the distillate rates of the first and second distillation columns, respectively. , , These are the target product concentrations produced at the top and bottom of the column, respectively. The cost of distillation column vessels, stages, and heat exchanger equipment. Operating costs for steam, cooling water, and electricity; The investment payback period for the equipment is set at 3 years. The heat load of the reboiler; These are the investment costs for the reboiler, condenser, tower shell, and tower tray, respectively. , These are the heat exchange areas of the reboiler and condenser, respectively. It should be noted that in the above economic evaluation mathematical model, the equipment size parameters of the distillation column (including the heat exchange area) are... , Tower shell height Tower diameter ) and energy consumption parameters (such as reboiler heat load) All of these are not independent variables. During the global optimization process, whenever the genetic algorithm generates a new set of independent process parameters ( When combining, input it into the pre-built strict steady-state material and energy balance model for solution, so that the corresponding size parameters and energy consumption parameters mentioned above can be naturally mapped and extracted, and then substituted into formulas (3) to (9) to complete the closed-loop calculation of the annual total cost TAC.

[0008] Step (1.2) Construct a dual-temperature collaborative dynamic control architecture: Based on steady-state design parameters, at least two independent temperature PID controllers are configured in each distillation column; the first distillation column controls the temperature of the first lower temperature-sensitive plate by adjusting the heat input of the reboiler, and controls the temperature of the first upper temperature-sensitive plate by adjusting the reflux ratio; the second distillation column controls the temperature of the second lower temperature-sensitive plate by adjusting the heat input of the reboiler, and controls the temperature of the second upper temperature-sensitive plate by adjusting the reflux ratio.

[0009] Step 2: Construct and train an artificial neural network agent model Step (2.1) Steady-state data sampling: Under the premise of meeting the preset product purity constraints, the Latin hypercube sampling (LHS) strategy is used to perform full-space perturbation sampling on the feeding conditions and operating variable conditions to extract the steady-state dataset that meets the requirements. Step (2.2) Data normalization: The extracted sample data is normalized using the deviation standardization method. The specific mathematical description is as follows: in, This represents the normalized sample data; This represents the original extracted sample data; and These represent the maximum and minimum values ​​of the corresponding variables in the original sample dataset, respectively. and The upper and lower limits of the normalized mapping interval are not specified.

[0010] Step (2.3) Proxy model training: A multilayer feedforward artificial neural network (ANN) is used. The input variables of the network include: feed flow rate, mole fraction of each component in the feed, and temperature of the four temperature-sensitive plates; the output variables of the network include: purity of the target product and reboiler heat load of the first and second distillation columns; the backpropagation algorithm is used to train the network until the prediction accuracy meets the threshold, and a fast nonlinear mapping between feed disturbance and optimal operating conditions is established.

[0011] Step 3: Real-time optimization decision-making based on GA-ANN and collaborative execution of dual-temperature PID Real-time data on feed flow rate and component disturbances from the industrial site are acquired and fixed as known variables, then input into a trained artificial neural network surrogate model. A coupled genetic algorithm (GA) is used, with a preset product purity as a hard constraint and minimizing the total reboiler heat load of the two towers as the optimization objective, to perform rapid iterative optimization. The specific mathematical description of the objective function and constraints of the real-time optimization model is as follows: in, This is the sum of the heat loads of the reboilers in the two distillation columns; These are the optimal temperature settings for each temperature-sensitive plate (it should be noted that, under the specific material system of this embodiment, they specifically correspond to...). (In other systems, this corresponds to the actual set value of the sensitivity board). This represents the minimum product molar concentration threshold. After calculating and obtaining the optimal combination of temperature setpoints that minimizes total energy consumption, the setpoints are updated in real time and sent to each temperature PID controller in the dual-temperature collaborative dynamic control architecture as new setpoints for closed-loop control.

[0012] Furthermore, the dual-temperature collaborative dynamic control architecture described in step (1.2) also includes a basic stable control loop, the control logic of which is as follows: The feed flow rate is controlled in a closed loop by manipulating the opening of the feed valve. The liquid level in the reflux tanks of the first and second distillation columns is controlled in a closed loop by the corresponding distillate flow rates. The bottom liquid level in the first and second distillation columns is controlled in a closed loop by the corresponding bottom product discharge flow rate. The above level control loop needs to specify the volume of the reflux tank and the bottom tank to provide a 5-minute material residence time when the container is 50% full; The control parameters of the temperature PID controller are tuned based on the Tyreus-Luyben adjustment rules.

[0013] Furthermore, in step (2.1), the disturbance boundary range of the Latin hypercube sampling is set as follows: the total feed flow rate and the composition of each component are sampled and combined in steps of 2% within ±20% of the initial operating conditions; the reflux ratio of the first distillation column... The sampling range is 1.4 to 6, and the reflux ratio of the second distillation column is... The sampling range is 2.4 to 8; the distillate flow rate of the first distillation column. The sampling range is 30 to 68, and the distillate flow rate of the second distillation column is... The sampling range is 25 to 59; the extraction criterion for the qualified steady-state dataset is that the molar concentration of all target products is greater than or equal to 99.4%.

[0014] Furthermore, the location of the temperature-sensitive plate mentioned in step (2.3) is determined by combining open-loop steady-state simulation experiments with Luyben sensitivity analysis. The specific steps are as follows: apply an independent step disturbance with an amplitude of ±0.1% to the reflux ratio or reboiler heat load of the distillation column, monitor and record the temperature change ΔT of each plate, and select the plate with the largest absolute value of temperature change as the temperature-sensitive plate for dual-temperature control.

[0015] Further, in step (2.3), the multilayer feedforward artificial neural network comprises one input layer, four hidden layers, and one output layer; wherein the hidden layers use sigmoid activation function neurons, with each layer containing 20 neurons; the output layer uses linear output neurons; the network training process divides the sample data into 80% training set, 10% validation set, and 10% test set, and uses the Levenberg-Marquardt backpropagation algorithm executed through the Trainlm function; the mean squared error (MSE) is used. ) and coefficient of determination ( The mathematical evaluation formula for assessing model accuracy is as follows: in, For the sample size, It is the first The true value of each. No. The predicted value of each.

[0016] Furthermore, when evaluating the dynamic response performance and energy-saving effect of the aforementioned dual-temperature collaborative dynamic control architecture when executing real-time optimization commands, the product concentration absolute error integral (…) is used. ) and total energy consumption ( As an evaluation indicator, its mathematical description is as follows: in, This represents the actual product concentration. To set the product concentration; This represents the start time of the disturbance. The end date of the observation period. This is the total heat load of the dual-tower reboiler.

[0017] The beneficial effects of this invention are as follows: (1) Excellent control layer resistance to component disturbances and tracking stability: It completely eliminates the defect of conventional single-temperature control that is prone to loss of control when faced with component disturbances. The dual-temperature control system implemented in this invention uses dual PID feedback of reboiler and reflux ratio for cross-coordinated regulation. Dynamic tests show that when faced with severe disturbances of up to ±20% in feed flow rate or feed composition, the product concentration fluctuation can be strictly limited to ±0.002, and the system stabilization time is shortened to less than 4 hours, giving the lower-level system excellent robustness.

[0018] (2) Extremely fast online optimization response capability: The ANN surrogate model trained based on LHS sampling and deviation standardization perfectly replaces the rigorous solution process of the mechanistic equation (for the prediction determination coefficient of high purity products). (MSE < 0.005). When faced with sudden changes in operating conditions, the optimized instruction generation time combined with the GA algorithm is shortened from several hours to tens of seconds, fully meeting the hard real-time requirements of continuous production chemical plants.

[0019] (3) This invention does not rely on the robustness of the controller to "hard-fight" disturbances, but achieves energy savings by dynamically updating the setpoints. When dealing with simultaneous combined disturbances of flow rate and composition, this architecture can automatically and reasonably allocate the reboiler heat load of the two towers, reducing the total energy consumption by 1.82% compared to the unoptimized state. At the same time, the dynamic absolute error integral (IAE) of product purity is significantly reduced by more than 50%, achieving a high degree of unity between economic benefits and product quality. Attached Figure Description

[0020] Figure 1 Abstract and accompanying figures Figure 2 is a schematic diagram of the dual-tower series steady-state distillation process used in an embodiment of the present invention.

[0021] Figure 3 is a diagram of the iterative convergence process of the genetic algorithm for global economic optimization in an embodiment of the present invention.

[0022] Figure 4 This is a schematic diagram of the flow path and signal loop structure of the dual-temperature collaborative dynamic control system constructed in this invention.

[0023] Figure 5The graph shows the selection result of the temperature-sensitive plate of the first distillation column (C1) determined by the Luyben sensitivity analysis method in the example.

[0024] Figure 6 This is a graph showing the selection results of the temperature-sensitive plate for the second distillation column (C2) determined by the Luyben sensitivity analysis method in the example.

[0025] Figure 7 This is a comparison curve of the dynamic anti-interference performance of a single temperature control system and the dual temperature control system of the present invention when faced with feeding interference.

[0026] Figure 8 This is a graph showing the prediction performance of the trained artificial neural network on the validation set in an embodiment of the present invention.

[0027] Figure 9 This is a dynamic comparison curve of product concentration between the real-time optimization (RTO+PID) system and the unoptimized system (PID) when dealing with complex operating conditions.

[0028] Figure 10 This is a dynamic comparison curve of the total reboiler heat load between the real-time optimization (RTO+PID) system and the unoptimized system (PID) when dealing with combined disturbances in operating conditions. Detailed Implementation

[0029] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0030] This embodiment focuses on the separation process of a mixture of benzene, toluene, and m-xylene (BTX), a core basic raw material in the chemical industry. Due to the high energy consumption and stringent purity requirements (typically >99.5%) of this process, traditional control strategies easily lead to energy waste. In this embodiment, the feed mixture is a saturated liquid containing 35.3509% benzene, 29.9693% toluene, and 34.6797% m-xylene at 310 K, with an initial flow rate of 135.8011 kmol / h. Thermodynamic properties are estimated using the NRTL-RK model, assuming a pressure drop of 0.0068 atm across the entire tray.

[0031] Step 1: Steady-state process design and GA-based economic optimization like Figure 2As shown, the system employs a two-tower direct-sequence cascade structure. The mixed feed enters the first distillation column C1, from which high-purity benzene is collected at the top. The mixture from the bottom is pumped into the second distillation column C2, where toluene is collected at the top and m-xylene at the bottom. In the steady-state design, a genetic algorithm (GA) is used in MATLAB to optimize global parameters with the objective of minimizing the total annual cost (TAC). The optimization constraint is that the molar purity of all three products is 99.5%. The core parameters of the GA are set as follows: population size 400, crossover probability 0.8, mutation probability 0.01, and maximum number of iterations 500. Figure 3 As shown in the convergence curve, the iteration stopped at generation 183. The optimal TAC value significantly decreased from the initial design of 488291.80 to 460521.52. The optimized core parameters are as follows: the total number of C1 trays remains at 30, the feed tray position is optimized from 15 to 13, and the reflux ratio... The distillate rate decreased from 2.027 to 1.791. The value was 48.169; the total number of trays in C2 increased from 28 to 29, the feed tray position was optimized from 19 to 20, and the reflux ratio was... The distillate rate decreased from 3.2 to 2.895. The value is 40.424. The optimization directly reduces the amount of liquid returning from the top of the tower, lowers the gas-liquid contact and equipment investment costs, and lays a high-efficiency benchmark for dynamic control.

[0032] Step 2: Establishment and parameter tuning of the dual-temperature coordinated dynamic PID control system Based on the above optimal steady-state parameters, a dynamic closed-loop control system is established, as shown in the schematic diagram below. Figure 4 As shown in Figure 4, the meanings of each control element and symbol are as follows: C1 and C2 represent the first and second distillation columns, respectively; FC is the feed flow controller; PC1 and PC2 are the top pressure controllers of the first and second distillation columns, respectively; LC1 to LC4 are the level controllers of the reflux tanks and bottom tanks of the two columns, respectively; TC1 and TC2 are the tray temperature controllers; RR1 and RR2 are the reflux ratio controllers; the numbers on the column (such as 15, 30, 19, 28) represent the position number of the tray; ΔT represents the temperature difference measurement link.

[0033] A heuristic method was used to specify the volumes of the reflux tank and the bottom tank, providing a 5-minute buffer holding time when the tank is 50% full. The empirical controller parameters for the basic loop were set as follows: feed flow control Kc=0.3, Ti=0.5min; level control Kc=2, Ti=9999min; pressure control Kc=20, Ti=12min, where Kc is the proportional gain and Ti is the integral time constant.

[0034] Precise search for temperature-sensitive plates: Abandoning empirical specifications, this embodiment employs the rigorous Luyben sensitivity analysis method. Under open-loop steady-state conditions, independent step perturbations of ±0.1% are applied to either the reflux ratio (RR) or the reboiler heat load (Q), and the temperature change ΔT of each tray is monitored and recorded. Figure 5 , Figure 6 As shown, plot the absolute value curve of temperature change.

[0035] Dual-temperature control architecture construction and comparison: To avoid severe coupling between the two temperature loops, this invention introduces the "region-dominated and spatial decoupling constraint principle" when selecting the sensitive board: 1. If the maximum ΔT peak value caused by Q and RR disturbances points to the same tray, then find the locally sensitive peak plate affected by RR in the rectification section above the feed tray and find the locally sensitive peak plate affected by Q in the stripping section below the feed tray. 2. If the physical distance between the selected upper and lower sensitive plates is too close, in order to avoid strong dynamic coupling of the control loop, a suboptimal peak plate with sufficient physical buffer space (e.g., at least 3 trays apart) will be selected as the final sensitive plate along the sensitivity curve.

[0036] Based on the above analysis and constraints, after applying a disturbance, this embodiment ultimately and precisely locates the 7th and 17th trays of the first distillation column C1, and the 15th and 24th trays of the second distillation column C2, as temperature-sensitive actuators in the dual-temperature control architecture. It should be noted that the specific tray numbers determined above are only preferred parameters for the specific material system of this embodiment; the specific positions of the temperature-sensitive trays will differ for different distillation processes and material systems. Those skilled in the art can use conventional testing methods (such as Luyben sensitivity analysis) and combine them with the "region-dominant and spatial decoupling constraint principle" proposed in this invention to re-search and determine the optimal temperature-sensitive actuator positions for the corresponding system.

[0037] This invention constructs as follows Figure 4 The dual temperature control architecture shown: each column is equipped with a second temperature controller. In the first distillation column C1, tray 17 ( The temperature of tray 7 is controlled by the heat input of the reboiler. The temperature of the column is controlled by adjusting the reflux ratio (using a proportional controller); in the second distillation column C2, tray 24 ( ) and tray 15 ( These are respectively controlled by the reboiler and reflux ratio. The Tyreus-Luyben adjustment rule is adopted.

[0038] Anti-interference test verification: such as Figure 7 As shown in the comparison, the dual temperature control system of the present invention can limit the product concentration fluctuation within the range of ±0.002 when faced with a wide range of disturbances of ±20% in the feed flow rate, and can completely stabilize the product concentration within 4 hours, effectively solving the problem of weak anti-interference ability of the prior art.

[0039] Step 3: Extraction and Construction of ANN Proxy Model To reduce the dimensionality of complex, rigorous mechanistic models for real-time optimization, this embodiment uses Latin hypercube sampling (LHS) to obtain a full-space dataset. Upper and lower boundaries are defined for the perturbation domain and the domain of operated variables: feed flow rate. Values ​​[108.64, 162.96]; Benzene component ∈[0.28,0.42]; Toluene component [0.24, 0.36]; Xylene components [0.22, 0.47]; First tower reflux ratio [1.4,6], Distillate flow rate ∈[30,68]; Second tower reflux ratio [2.4,8], Distillate flow rate [25,59]. Within the aforementioned space, random sampling and process execution were performed. After removing low-purity dead zones, 7912 sets of high-quality steady-state data with a product molar concentration ≥99.4% were extracted. The data were linearly mapped to the [0,1] interval using deviation normalization. The preprocessed data was input into a feedforward neural network containing four hidden layers (each with 20 sigmoid activation function neurons, and the output layer being a linear neuron). The data was divided into an 80% training set, a 10% validation set, and a 10% test set. Training was performed using the Levenberg-Marquardt backpropagation algorithm via the Trainlm function.

[0040] like Figure 8 Images (a) through (e) respectively demonstrate the effect of the artificial neural network (ANN) surrogate model constructed in this invention on the purity of benzene at the top of the first distillation column. ), purity of toluene at the top of the second distillation column ( ), purity of xylene in the bottom of the second distillation column ( ), and the heat load of the reboiler in the first distillation column ( ) and the heat load of the reboiler in the second distillation column ( The regression fitting results for the predictive performance of five core control variables are presented. In the figure, the horizontal axis represents the target value of the true steady-state data obtained from the rigorous mechanistic model, the vertical axis represents the predicted value rapidly output by the ANN surrogate model, and the solid line running diagonally is the ideal baseline for completely accurate prediction (i.e., Y=X). The scatter distribution of the subplots in the figure clearly shows that the massive number of test sample points in the validation set converge very closely to and fit the ideal baseline; the mathematical evaluation indicators further confirm that the prediction determination coefficient for high-purity products... All values ​​are greater than 0.999 and the mean square error (MSE) is as low as 0.003, for predicting reboiler heat load. Greater than 0.997. This extremely high regression fitting accuracy objectively and fully demonstrates that the ANN surrogate model used in this invention perfectly overcomes the fatal flaw of large latency in online solution of traditional mechanistic equations. It can accurately map the nonlinear physical behavior of complex distillation systems while ensuring highly accurate calculation results, thus providing a solid data-driven engine for the implementation of ultra-fast and reliable online optimization in the upper-level RTO architecture.

[0041] Step 4: Evaluation of the Effectiveness of Online Collaborative Execution of Operating Conditions and RTO To evaluate... The performance of a real-time optimization (RTO) system based on GA-ANN and dual-temperature PID collaborative optimization was tested, and a long-term dynamic simulation was performed in an Aspen Dynamics module and Simulink-connected environment. The initial settings for the four sensitive boards were as follows: =84.507℃ =109.193℃ =117.958℃ =134.826℃. A step disturbance signal is added at time 0, and the disturbance is adjusted every 15 hours.

[0042] (1) Dynamic optimization execution test: Test conditions (dual abrupt changes in flow rate and composition perturbation): For example, when the feed flow rate jumps to 162 kmol / h, and the benzene / toluene / xylene composition changes abruptly to 0.31 / 0.34 / 0.35: the upper-level RTO fixes the current perturbation input parameters and rapidly iterates to calculate the total energy consumption ( The minimum new temperature setpoint sequence is then issued: =84.73℃ =107.90℃ =117.84℃ =133.19℃.

[0043] (2) Comparative evaluation of effects: (2) Comparative evaluation of effects: like Figure 9 and Figure 10 As shown, dynamic response comparison data of product concentration and total reboiler heat load were extracted throughout the entire 90-hour assessment cycle.

[0044] against Figure 9 Product concentration dynamic comparison: In the graph, the solid curve labeled (PID) represents the unoptimized system, and the curve labeled (RTO+PID) represents the state after implementing the optimization strategy of this invention. It can be clearly seen that conventional (PID) control, in order to prevent product defects, can only set conservative fixed values, resulting in excessively high product purity during stable operation (e.g., ...). The temperature setpoint remained consistently around 0.997, resulting in significant energy waste. However, by adopting the (RTO+PID) strategy, the system proactively shifted its operating point towards a lower energy consumption state. Furthermore, when faced with severe combined disturbances at nodes such as 15h and 30h, the RTO system dynamically adjusted the temperature setpoint, effectively suppressing drastic concentration overshoot. Concentration fluctuations of all products were forcibly confined within a safe range, and the dynamic absolute error integral (IAE) dropped dramatically to 0.029, 0.064, and 0.068 respectively (all reductions exceeding 50%).

[0045] Combination Figure 9 Energy consumption comparison data: Throughout the entire assessment period, considering the combined disturbances in flow rate and composition, the total reboiler energy consumption of the unoptimized system reached as high as 299,321.99 kW·h; while after adopting RTO dynamic adjustment, the severe fluctuations in heat load were effectively "peak-shaving," and the total energy consumption was reduced to 293,868.77 kW·h, resulting in direct energy savings of 1.82%. The above detailed data fully verify that the present invention possesses excellent tracking stability and extremely high energy-saving and low-carbon value when facing complex disturbances.

Claims

1. A data-driven real-time optimization control method for a distillation column with dual temperatures, characterized in that, The steps are as follows: Step 1: Construct a steady-state separation model for multi-component distillation and a dual-temperature dynamic control architecture Step (1.1) Constructing a steady-state separation model: Establish a steady-state separation model for the dual-tower series structure, and use a genetic algorithm with a total annual cost. The process parameters are globally optimized with the goal of minimization to obtain the initial optimal equipment and operating parameters; Step (1.2) Construct a dual-temperature collaborative dynamic control architecture: Based on steady-state design parameters, at least two independent temperature PID controllers are configured in each distillation column; the temperature of the first lower temperature-sensitive plate in the first distillation column is controlled by adjusting the heat input of the reboiler, and the temperature of the first upper temperature-sensitive plate is controlled by adjusting the reflux ratio. The temperature of the second lower temperature-sensitive plate in the second distillation column is controlled by adjusting the heat input of the reboiler, and the temperature of the second upper temperature-sensitive plate is controlled by adjusting the reflux ratio. Step 2: Construct and train an artificial neural network agent model Step (2.1) Steady-state data sampling: Under the premise of meeting the preset product purity constraints, the Latin hypercube sampling (LHS) strategy is used to perform full-space perturbation sampling of the feeding conditions and operating variable conditions, and extract the steady-state dataset that meets the requirements. Step (2.2) Data normalization: The extracted sample data is normalized and mapped using the deviation normalization method; Step (2.3) Proxy model training: A multilayer feedforward artificial neural network (ANN) is used. The input variables of the network include: feed flow rate, mole fraction of each component in the feed, and temperature of the four temperature-sensitive plates; the output variables of the network include: purity of the target product and reboiler heat load of the first and second distillation columns; the backpropagation algorithm is used to train the network until the prediction accuracy meets the threshold, and a fast nonlinear mapping between feed disturbance and optimal operating conditions is established. Step 3: Real-time optimization decision-making based on GA-ANN and collaborative execution of dual-temperature PID Real-time data on feed flow and component disturbances from the industrial site are acquired and fixed as known variables, then input into a trained artificial neural network surrogate model. A genetic algorithm (GA) is coupled in, using a preset product purity as a hard constraint and minimizing the total reboiler heat load of the two towers as the optimization objective, to perform rapid iterative optimization. After calculating the optimal temperature setpoint combination that minimizes total energy consumption, the setpoint is updated in real-time and sent to each temperature PID controller in the dual-temperature collaborative dynamic control architecture as a new setpoint for closed-loop control.

2. The data-driven real-time dual-temperature optimization control method for a distillation column according to claim 1, characterized in that, In step (1.1), the objective function for steady-state optimization is specifically mathematically described as follows: in, This describes the functional relationship between each operating parameter and the total annual cost. , These represent the total number of trays in the first and second distillation columns, respectively. , These are the locations of the feed trays; These are the reflux ratios for the first and second distillation columns, respectively. These are the distillate rates of the first and second distillation columns, respectively. , , These are the target product concentrations produced at the top and bottom of the column, respectively. The cost of distillation column vessels, stages, and heat exchanger equipment. Operating costs for steam, cooling water, and electricity; The payback period for the equipment can be set to 3 years; The heat load of the reboiler; These are the investment costs for the reboiler, condenser, tower shell, and tower tray, respectively. , These are the heat exchange areas of the reboiler and the condenser, respectively.

3. The data-driven real-time dual-temperature optimization control method for a distillation column according to claim 1, characterized in that, In step (3), the specific mathematical description of the objective function and constraints of the real-time optimization model is as follows: in, This is the sum of the heat loads of the reboilers in the two distillation columns; These are the optimal temperature settings for each temperature-sensitive plate; This represents the minimum product molar concentration threshold.

4. The data-driven real-time dual-temperature optimization control method for a distillation column according to claim 1, characterized in that, The dual-temperature collaborative dynamic control architecture described in step (1.2) also includes a basic stable control loop, the control logic of which is as follows: The feed flow rate is controlled in a closed loop by manipulating the opening of the feed valve. The liquid level in the reflux tanks of the first and second distillation columns is controlled in a closed loop by the corresponding distillate flow rates. The bottom liquid level in the first and second distillation columns is controlled in a closed loop by the corresponding bottom product discharge flow rate. The above level control loop needs to specify the volume of the reflux tank and the bottom tank to provide a 5-minute material residence time when the container is 50% full; The control parameters of the temperature PID controller are tuned based on the Tyreus-Luyben adjustment rules.

5. The data-driven real-time dual-temperature optimization control method for a distillation column according to claim 1, characterized in that, In step (2.1), the disturbance boundary range of the Latin hypercube sampling is set as follows: the total feed flow rate and the composition of each component are sampled in steps of 2% within ±20% of the initial operating conditions; the reflux ratio of the first distillation column... The sampling range is 1.4 to 6, and the reflux ratio of the second distillation column is... The sampling range is 2.4 to 8; the distillate flow rate of the first distillation column. The sampling range is 30 to 68, and the distillate flow rate of the second distillation column is... The sampling range is 25 to 59; the extraction criterion for the qualified steady-state dataset is that the molar concentration of all target products is greater than or equal to 99.4%.

6. The data-driven real-time dual-temperature optimization control method for a distillation column according to claim 1, characterized in that, The location of the temperature-sensitive plate mentioned in step (2.3) is determined by combining open-loop steady-state simulation experiments with Luyben sensitivity analysis. The specific steps are as follows: apply an independent step disturbance with an amplitude of ±0.1% to the reflux ratio or reboiler heat load of the distillation column, monitor and record the temperature change ΔT of each plate, and select the plate with the largest absolute value of temperature change as the temperature-sensitive plate for dual-temperature control.

7. The data-driven real-time dual-temperature optimization control method for a distillation column according to claim 1, characterized in that, In step (2.3), the multilayer feedforward artificial neural network comprises one input layer, four hidden layers, and one output layer; the hidden layers use sigmoid activation function neurons, with each layer containing 20 neurons; the output layer uses linear output neurons; the network training process divides the sample data into 80% training set, 10% validation set, and 10% test set, and uses the Levenberg-Marquardt backpropagation algorithm executed through the Trainlm function; mean squared error is used. and coefficient of determination The mathematical evaluation formula for assessing model accuracy is as follows: in, For the sample size, It is the first The true value of each. No. The predicted value of each.

8. The data-driven real-time dual-temperature optimization control method for a distillation column according to claim 1, characterized in that, When evaluating the dynamic response performance and energy-saving effect of the aforementioned dual-temperature collaborative dynamic control architecture when executing real-time optimization commands, the product concentration absolute error integral is used. and total energy consumption As an evaluation metric, its mathematical description is as follows: in, This represents the actual product concentration. To set the product concentration; This represents the start time of the disturbance. The end date of the observation period. The total reboiler heat load is for the dual towers.

Images