Post-combustion carbon capture control method and system based on disturbance observation

Abstract

The invention discloses a post-combustion carbon capture control method and system based on disturbance observation. The post-combustion carbon capture control method and system are applied to a carbon capture system comprising an absorption tower and a reboiler. Constructing a system state space model and calibrating a multi-source disturbance boundary and a weight; a state observer is designed, a disturbance observer containing an LSTM prediction module is constructed, a gain matrix of the disturbance observer meets Hurwitz constraints, and the LSTM takes real-time disturbance estimation as input, outputs a future disturbance prediction value and feeds back and corrects dynamic logic of the observer; on the basis of the disturbance estimation and prediction value, combining an energy consumption-capture rate collaborative objective function and a DQN dynamic weight mechanism, calculating a feed-forward compensation amount with process hard constraints; meanwhile, designing a state feedback control law to generate a feedback quantity; and finally, the feedforward quantity and the feedback quantity are fused and input into a PID controller, so that smooth and asymptotic tracking control is realized. The problems of wide fluctuation of flue gas flow / CO2 concentration, multi-source disturbance coupling, contradiction between traditional control capture rate and energy consumption optimization, weak anti-interference, response oscillation and insufficient precision of a carbon capture system are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of carbon emission reduction, and particularly relates to a post-combustion carbon capture control method and system based on disturbance observation. BACKGROUND

[0002] Fossil fuel power plants are the largest stationary CO2 emission source, and in the future, fossil fuel power plants will still be the main power source. Under this background, carbon capture technology is still a key solution to rapidly reduce CO2 emissions. Among them, the post-combustion CO2 capture system based on chemical adsorption is the mainstream of current coal-fired power plant CO2 capture technology due to its low working pressure (atmospheric pressure), small investment cost, mature technology and other advantages.

[0003] Traditional fixed PID controllers are difficult to meet the precise adjustment requirements of the post-combustion CO2 capture system, so there is an urgent need to develop new control algorithms to optimize system performance. Considering the strong coupling characteristics between the coal-fired unit and the carbon capture system, such as the influence of flue gas flow and composition fluctuations on capture efficiency, and the possible disturbance of steam turbine extraction operation on the stability of the main machine, it is difficult for the independent control mode of the main machine or the capture subsystem to effectively cope with the challenges brought by multi-variable interaction.

[0004] Most of the current research on the design of post-combustion CO2 capture system controllers uses PI / PID controllers based on feedback action, which is difficult to handle the large inertia characteristics of the coal-fired power plant (CFPP)-PCC system and the coupling characteristics between the systems. Experiments show that when the PID controller faces the dramatic changes of flue gas flow, capture rate and other variables, the adjustment speed is slow, the overshoot is large, and the stability of the system cannot be guaranteed. In recent years, many researchers have used model predictive control (MPC) theory to design control systems for the capture process. MPC theory is suitable for handling multi-variable, slow dynamic, limited system characteristics, etc. Therefore, compared with PI control structure, MPC has better performance, but the performance of MPC depends largely on the quality of the prediction model. For linear MPC, the prediction model is always obtained by linearization of the mathematical model or identification at a given operating point. Since the PCC system faces flexible operation and the requirement of changing capture rate within a certain range, the PCC system must adapt to the changing flue gas, and the temperature of the reboiler will also change when the unit load demand changes. When these important variables deviate from the design point of the model, the dynamic behavior of the system will change greatly. This modeling mismatch will reduce the quality of predictive control, and in severe cases, it will even destroy the closed-loop control system. The current linear MPC only shows their performance near the design point.

[0005] Another problem of MPC is that the calculation of control law is of high complexity. To solve this problem, a stable prediction control algorithm based on extended state can be proposed based on the MPC controller in terms of algorithm, and a machine learning algorithm is used to establish and solve the economic index of the CFPP-PCC system, so that the flexible operation of the CFPP-PCC system is realized. The control purpose of the CFPP-PCC system is complex, and in actual operation, the carbon capture system control should consider environmental protection, energy utilization, system stability and other aspects. In the process of CO2 capture after combustion, the change of flue gas flow and CO2 capture rate will have a significant impact on the dynamic characteristics of the system, and this impact is not linear. However, the thermal power unit and the carbon capture system are tightly coupled, for example, the steam turbine steam absorption solvent regeneration will affect the power generation of the thermal power unit, and the change of the thermal power unit load will lead to the change of the flue gas flow, thereby affecting the carbon capture system, which leads to the fact that the traditional PID controller cannot achieve good results. SUMMARY

[0006] The purpose of the present application is to provide a post-combustion carbon capture control method and system based on disturbance observation, which can monitor external disturbances and internal uncertainties of the system in real time and generate corresponding compensation signals to enhance the robustness and accuracy of the system. In the carbon capture system of a coal-fired power plant, a disturbance observation technology is introduced, i.e., a state observation-based stable prediction control method, which regards the model deviation caused by the working condition change as a disturbance, estimates it using a state observer, and performs feedforward compensation, thereby realizing unbiased control of the system and solving the problems of wide fluctuation of flue gas flow / CO2 concentration, coupling of multiple source disturbances, contradiction between capture rate and energy consumption optimization in traditional control, weak anti-disturbance, response oscillation and insufficient accuracy of the carbon capture system.

[0007] The technical scheme of the present application is as follows:

[0008] A post-combustion carbon capture control method based on disturbance observation is applied to a carbon capture system comprising an absorption tower and a reboiler, comprising the following contents:

[0009] S1: system modeling and disturbance parameter calibration: constructing a state space model of the carbon capture system comprising the absorption tower and the reboiler, and calibrating the disturbance parameter boundary and priority weight in the system operation;

[0010] S2: design state observer, use gain matrix to ensure that the state estimation error converges quickly;

[0011] S3: construct a robust disturbance observer containing LSTM prediction: first, based on the system state, input and output, construct a disturbance observer, and the gain matrix of the disturbance observer Satisfy the Hurwitz matrix constraint, estimate the disturbance in real time, directly as the input data of the LSTM disturbance prediction model, the LSTM disturbance prediction model outputs the future disturbance prediction value, which is fed back to the next time operation of the disturbance observer as the pre-judgment reference item of the disturbance observer, and the observer gain matrix is dynamically adjusted The real-time disturbance estimation value of the disturbance observer and the future disturbance prediction value of the LSTM model are jointly input into the feedforward compensation control model:

[0012] S4: Design a disturbance feedforward compensation control model: define an energy capture rate collaborative objective function, set a reward function to guide the system to approach high precision and low energy consumption; dynamically adjust the energy consumption and capture rate weights by using the DQN algorithm; calculate the feedforward control amount based on the disturbance observation value and the gradient of the objective function; embed the process hard constraint to modify and limit the steam flow boundary;

[0013] S5: State feedback control model: use the estimated state output by the state observer to design a state feedback control law to output a feedback control amount;

[0014] S6: The feedforward control amount and the feedback control amount are input into the PID controller to smoothly fuse the feedforward and feedback amounts, so as to realize the asymptotic tracking control goal of the system under the disturbance condition.

[0015] Further, in the S1, comprising:

[0016] State space model construction: based on Aspen Plus, build an absorption tower-reboiler coupled model, and the model expression is:

[0017]

[0018]

[0019] Among them, is the first derivative of the system state vector, is the system state vector ( , CO2 molar concentration of the absorption tower, absorption tower temperature, absorption tower pressure, reboiler temperature, reboiler steam pressure), is the state matrix, is the input matrix, is the control input vector ( , lean liquid circulation flow, reboiler steam flow), is the disturbance matrix, is the disturbance vector ( , flue gas flow fluctuation, CO2 concentration fluctuation, steam pressure fluctuation, unmodeled dynamic disturbance, SOx / NOx pollutant disturbance);

[0020]

[0021] wherein, y is the system output vector (y = [CO2 capture rate, reboiler temperature]), respectively, CO2 capture rate, reboiler temperature, is the output matrix;

[0022] Disturbance parameter calibration: flue gas flow fluctuation Rated value, CO2 concentration (dry basis), steam pressure MPa, SOx / NOx≤10 mg / m³; Priority: flue gas flow (weight 0.4) > CO2 concentration (weight 0.3) > steam pressure (weight 0.2) > pollutants (weight 0.1).

[0023] Further, in the S2, comprising:

[0024] State observer model:

[0025]

[0026] wherein, is the first derivative of the state estimate, is the system state estimate, is the state observer gain matrix (3x3 dimension);

[0027] Gain setting: the target pole is 3-5 times the original pole of the system, and the setting formula is:

[0028]

[0029] wherein, is the target pole matrix;

[0030] Ensure that the state estimation error ( is the state estimation error) converges in time ≤2s, and the steady-state error is ≤±1%.

[0031] Further, in the S3, the robust disturbance observer with LSTM prediction is constructed, comprising:

[0032] Disturbance observer construction:

[0033]

[0034] wherein, is the disturbance observer gain matrix (5x5 dimension);

[0035]

[0036] in, This is the disturbance estimate. for The second row vector (1×5 dimensions). for The second line scalar;

[0037] Solving using LMI (Linear Matrix Inequality) Satisfying the Hurwitz constraint ( Real parts of all eigenvalues );

[0038] LSTM prediction module integration: Input history (Training with time series data of disturbance estimation from the past 10 seconds) Output 5-second disturbance prediction values ( (Predicted disturbance value for the next 5 seconds).

[0039] Collaborative logic: As input to LSTM, Feedback and correction :

[0040]

[0041] in, The gain matrix of the perturbation observer at the next time step. For the current moment , It is a symbolic function;

[0042] Ensure the disturbance estimation error (the true value of the disturbance and the actual value of the disturbance) The convergence time (difference) is ≤3s.

[0043] Furthermore, S4 includes: the disturbance feedforward compensation control model includes:

[0044] Objective function and reward function:

[0045]

[0046] in, To collaboratively optimize the objective function, The capture rate priority weight is used. As energy consumption priority weight, The target capture rate is 90%. CO2 capture rate, Real-time energy consumption (unit: GJ / t CO2). Rated energy consumption (2.0 GJ / t CO2). For the reward function;

[0047] ;

[0048] Dynamic weight adjustment based on the DQN algorithm: when the perturbation amplitude is ≥ ±20%, , When the disturbance amplitude is ≤ ±5%, , ;

[0049] Calculation of feedforward control quantity:

[0050]

[0051] in, This is the feedforward control variable. For the input matrix The inverse matrix, The feedforward differential gain is (0.08-0.12). To estimate the rate of change of the disturbance, To reinforce learning, control gain (0.09-0.11) was applied. The gradient of the objective function with respect to the control input;

[0052] in, ( (This is the estimated value of the disturbance at the previous time step). Solved using the backpropagation algorithm;

[0053] Process constraints: ( When the reboiler temperature is reached, ( (This refers to the steam flow component in feedforward control). Rated value ( (for low fluid flow rate) Rated value ( (Steam flow rate).

[0054] Furthermore, in S5, the state feedback control model includes:

[0055] State feedback control law: based on (State estimation) design, feedback gain Optimize using LQR ( , This is the state weight matrix; , (Input weight matrix)

[0056]

[0057] in, For feedback control, The state feedback gain matrix;

[0058] Ensure the damping ratio of the closed-loop system ( (This refers to the damping ratio of the closed-loop system).

[0059] Furthermore, S6 specifically includes:

[0060] PID parameter adaptive tuning based on the DQN algorithm:

[0061]

[0062] in, For PID proportional coefficient, It is the hyperbolic tangent function. For the reward function;

[0063]

[0064] in, For PID integral coefficients, It is an exponential function;

[0065]

[0066] in, These are the PID differential coefficients. This represents the absolute value of the estimated rate of change of the disturbance; the other parameters are defined as above.

[0067] Smooth composite control law:

[0068]

[0069] in, For the final composite control quantity, For composite control weighting coefficients, Transfer function for PID controller ( ), This refers to the deviation between the target and actual capture rate.

[0070]

[0071] in, The absolute value of the disturbance estimate. perturbation vector The boundary maximum value;

[0072] The control quantity adjustment rate is ≤±3% / min, enabling the system output to adjust under disturbances. Asymptotic tracking of (capture rate target value).

[0073] A post-combustion carbon capture control method based on perturbation observation also includes S7: closed-loop optimization: (System output vector) and (System state vector) is fed back to S1-S6 for correction. (System Matrix) (Observer gain matrix) (Objective function weights) and (PID parameters) to continuously optimize control performance and ensure that core objectives (capture rate ≥85%, energy consumption ≤2.0 GJ / t CO2, disturbance recovery time ≤50s, reboiler temperature ≤125℃) are achieved.

[0074] A post-combustion carbon capture control system based on perturbation observation includes: a sensing layer, a control layer, an execution layer, and a feedback layer, wherein the control layer includes:

[0075] System modeling and disturbance calibration module: used to construct the state-space model of the absorber-reboiler coupling and calibrate the boundary and priority weights of disturbance parameters;

[0076] State observation module: used to estimate the unmeasurable states of the system and ensure that the state estimation error converges quickly;

[0077] The robust perturbation observation and LSTM prediction module is used to estimate perturbations in real time, predict future perturbation trends, and dynamically adjust the observer gain.

[0078] The disturbance feedforward compensation control module is used to calculate the feedforward control quantity that combines disturbance rejection, optimization, and safety.

[0079] The state feedback control module is used to output feedback control quantities based on the optimal gain;

[0080] The PID fusion control module is used to adaptively tune PID parameters and smoothly fuse feedforward and feedback control quantities;

[0081] The perception layer collects system operation data and transmits it to the control layer. The execution layer receives the final control quantity output by the control layer to drive device regulation. The feedback layer monitors the system status in real time and feeds it back to the control layer to achieve closed-loop optimization.

[0082] Includes auxiliary modules:

[0083] The data storage module stores model parameters, disturbance data, control variables, observations, etc., and supports backtracking of historical data for ≥1 year;

[0084] The feedback monitoring module collects system output data in real time. With state Feedback is sent to each control unit for deviation correction;

[0085] The human-machine interface visually displays key indicators such as disturbance observations, predictions, control variables, energy consumption, and capture rate, and supports manual start / stop, parameter modification, and fault alarms.

[0086] The basic modeling and disturbance calibration of this invention: By using Aspen Plus coupled simulation and industrial data identification, a state-space model that accurately reflects the dynamic characteristics of the absorber-reboiler is established. At the same time, the boundaries and priorities of multi-source disturbances are calibrated, providing model support and disturbance benchmarks for subsequent control.

[0087] Dual-observer collaborative sensing: The state observer tunes its gain using the pole placement method to quickly converge and obtain a high-precision system state; the robust disturbance observer ensures stability based on Hurwitz matrix constraints, estimates multi-source disturbances in real time, and then combines the learning of the disturbance data from the past 10 seconds using the LSTM model to predict the disturbance trend in the next 5 seconds, forming a dual disturbance sensing of "real-time estimation + trend prediction". At the same time, the observer gain is dynamically adjusted through the predicted value to further improve the disturbance sensing accuracy.

[0088] Feedforward compensation active disturbance rejection: Based on reinforcement learning, an energy consumption-capture rate collaborative objective function is constructed. The weights of the two objectives are dynamically adjusted according to the disturbance intensity through the DQN algorithm. The feedforward control quantity is calculated by combining the disturbance observation, the rate of change and the gradient of the objective function to offset the disturbance in advance. At the same time, hard constraints of the process are embedded to ensure control safety.

[0089] Feedback control deviation correction: The LQR-optimized state feedback control law is designed using high-precision state estimates, and the output feedback control quantity is used to correct the system operation deviation and ensure the steady-state accuracy of the system.

[0090] PID smooth fusion and closed-loop optimization: The PID parameters are adaptively tuned using the DQN algorithm, and the fusion weights of the feedforward and feedback control quantities are dynamically adjusted based on the disturbance intensity to avoid abrupt control changes and achieve a smooth response. Finally, the system output and state are fed back to the entire process to dynamically correct the model, observer gain, control weights and PID parameters, forming a closed-loop optimization to ensure that the control objective is continuously achieved.

[0091] The innovation of this invention lies in the following aspects:

[0092] Robust perturbation observation and LSTM prediction co-design: For the first time, Hurwitz matrix constraints are introduced into the design of perturbation observers for carbon capture systems to ensure observation stability under multi-source coupled perturbations. At the same time, an LSTM time series prediction module is integrated to achieve dual perception of perturbation "real-time estimation + future prediction", which solves the defects of traditional observers such as slow convergence and weak perturbation resistance.

[0093] DQN-driven dual-objective dynamic optimization: Constructing a collaborative objective function of energy consumption and capture rate, and dynamically adjusting the objective weights according to the perturbation intensity through the DQN algorithm, to achieve intelligent switching between "maintaining capture rate under severe perturbation and reducing energy consumption during steady-state operation", thus resolving the contradiction of dual-objective optimization in traditional methods;

[0094] Adaptive weighted fusion control strategy: Based on the disturbance intensity, an adaptive fusion weight is designed to smoothly integrate feedforward compensation and feedback control quantities. At the same time, the PID parameters are dynamically tuned through the DQN algorithm to adapt to a wide range of disturbance conditions and avoid system oscillations caused by traditional fusion methods.

[0095] Full-process closed-loop optimization architecture: Establish a full-process closed loop of "modeling-observation-control-feedback-correction" to realize dynamic updates of model matrix, observer gain, control weight and PID parameters, and ensure the control accuracy and robustness of the system in long-term operation.

[0096] Advantages of this invention:

[0097] Strong anti-interference capability: It can effectively resist fluctuations of ±30% flue gas flow, CO2 concentration and multiple sources of disturbance such as steam pressure and pollutants. The convergence time of disturbance estimation error is ≤3s and the system recovery time is ≤50s, which is far superior to traditional methods.

[0098] High control precision: steady-state capture rate ≥90%, dynamic capture rate ≥85%, capture rate fluctuation deviation ≤±3%, solving the problem of insufficient precision of traditional PID control;

[0099] Significant energy consumption optimization: energy consumption per unit CO2 capture ≤2.0 GJ / t CO2, which is 10%-15% lower than traditional methods, achieving a synergy of high precision and low energy consumption;

[0100] Safe and reliable operation: Through hard constraint design such as reboiler temperature and flow boundary, the reboiler temperature is ensured to be ≤125℃, avoiding thermal degradation of absorbent, reducing equipment wear and extending system life;

[0101] Wide adaptability: It can be flexibly adapted to carbon capture systems of different sizes in industries such as thermal power, steel, and chemical industry. It has strong adaptability to fluctuations in operating conditions and high engineering application value. Detailed Implementation

[0102] A post-combustion carbon capture control method based on perturbation observation, applied to a carbon capture system including an absorber and a reboiler, includes the following:

[0103] S1: System Modeling and Disturbance Parameter Calibration: Construct a state-space model of the carbon capture system including the absorber and reboiler, and calibrate the boundaries and priority weights of the disturbance parameters during system operation;

[0104] S2: Design a state observer using a gain matrix to ensure that the state estimation error converges quickly;

[0105] S3: Constructing a robust perturbation observer with LSTM predictions: First, based on the system state, inputs, and outputs, construct the perturbation observer, including its gain matrix. Satisfying the Hurwitz matrix constraints, the disturbance is estimated in real time and directly used as input data for the LSTM disturbance prediction model. The LSTM disturbance prediction model outputs the predicted future disturbance value, which is fed back into the disturbance observer's calculation at the next time step, serving as a prediction reference term for the disturbance observer and correcting the observer gain matrix. The dynamic adjustment logic involves inputting the real-time disturbance estimate from the disturbance observer and the future disturbance prediction from the LSTM model into the feedforward compensation control model.

[0106] S4: Design a disturbance feedforward compensation control model: Define an energy consumption-capture rate collaborative objective function, set a reward function to guide the system toward high precision and low energy consumption; use the DQN algorithm to dynamically adjust the weights of energy consumption and capture rate; calculate the feedforward control quantity based on the disturbance observation value and the gradient of the objective function; embed process hard constraints and correct steam flow boundary limits;

[0107] S5: State Feedback Control Model: Using the estimated state output by the state observer, design the state feedback control law and output the feedback control quantity;

[0108] S6: The feedforward and feedback control quantities are input to the PID controller, smoothly merging the feedforward and feedback quantities to achieve the asymptotic tracking control objective of the system under disturbance conditions.

[0109] S1 includes:

[0110] State-space model construction: A coupled absorption tower-reboiler model was built based on Aspen Plus. The model expression is as follows:

[0111]

[0112]

[0113] in, The first derivative of the system state vector. The system state vector ( These are, respectively, the CO2 molar concentration in the absorber, the absorber temperature, the absorber pressure, the reboiler temperature, and the reboiler steam pressure. The state matrix, For the input matrix, To control the input vector ( (These are the lean liquor circulation flow rate and the reboiler steam flow rate, respectively). Here is the perturbation matrix. For the perturbation vector ( These are respectively flue gas flow fluctuations, CO2 concentration fluctuations, steam pressure fluctuations, unmodeled dynamic disturbances, and SOx / NOx pollutant disturbances.

[0114]

[0115] in, For the system output vector ( (represented by CO2 capture rate and reboiler temperature, respectively). This is the output matrix;

[0116] Disturbance parameter calibration: flue gas flow rate fluctuation Rated value, CO2 concentration (Dry basis), steam pressure MPa, SOx / NOx ≤ 10 mg / m³; Priority: flue gas flow rate (weight 0.4) > CO2 concentration (weight 0.3) > steam pressure (weight 0.2) > pollutants (weight 0.1).

[0117] The state-space model is constructed to provide accurate model support for subsequent control algorithms. Disturbance parameter calibration: The quantification boundaries and priority weights of multi-source disturbances such as flue gas flow rate and CO2 concentration are clearly defined, which not only provides an "observation benchmark" for the disturbance observer, but also provides a priority basis for the optimization of control objectives (such as feedforward weight allocation), ensuring that control resources are tilted towards the core disturbance.

[0118] Furthermore, S2 includes:

[0119] State observer model:

[0120]

[0121] in, The first derivative of the state estimate. This is the system state estimate. The state observer gain matrix (3×3 dimension);

[0122] Gain tuning: The target pole is 3-5 times the original pole of the system. The tuning formula is:

[0123]

[0124] in, The target pole matrix;

[0125] Ensure state estimation error ( (Convergence time ≤ 2s, steady-state error ≤ ±1%, for state estimation error)

[0126] State observer based on system input Output Based on the established state-space model, a closed-loop estimation mechanism is constructed: through... The actual output With observation output The deviation via the gain matrix After magnification, the feedback corrects the state estimate. This enables dynamic tracking of states that cannot be directly measured.

[0127] To improve the estimation response speed and accuracy, the target pole is set at 3-5 times the original pole of the system. Tuning Strengthen the bias correction function and ensure the state estimation error It converges quickly (≤2s) and has a steady-state error of ≤±1%, providing reliable state data support for subsequent control.

[0128] Furthermore, in S3, constructing a robust perturbation observer with LSTM predictions includes:

[0129] Perturbation observer construction:

[0130]

[0131] in, The perturbation observer gain matrix (5×5 dimensions);

[0132]

[0133] in, This is the disturbance estimate. for The second row vector (1×5 dimensions). for The second line scalar;

[0134] Solving using LMI (Linear Matrix Inequality) Satisfying the Hurwitz constraint ( Real parts of all eigenvalues );

[0135] LSTM prediction module integration: Input history (Training with time series data of disturbance estimation from the past 10 seconds) Output 5-second disturbance prediction values ( (Predicted disturbance value for the next 5 seconds).

[0136] Collaborative logic: As input to LSTM, Feedback and correction :

[0137]

[0138] in, The gain matrix of the perturbation observer at the next time step. For the current moment , It is a symbolic function;

[0139] Ensure the disturbance estimation error (the true value of the disturbance and the actual value of the disturbance) The convergence time (difference) is ≤3s.

[0140] Real-time disturbance estimation: based on system state ,enter and output ,pass Construct a closed-loop observation framework, using output bias to drive state estimation correction; then through Extract the gain matrix By using specific row vectors and scalars, a mapping relationship between state, output, and disturbance is established, enabling real-time decoupled estimation of multi-source disturbances.

[0141] Robustness Guarantee: Solved via LMI Force it to satisfy the Hurwitz constraint (real part of all eigenvalues) This ensures that the observer is asymptotically stable under multi-source coupled disturbances such as flue gas flow and pollutants, and avoids estimation divergence.

[0142] Time series trend prediction: The LSTM prediction model estimates time series data based on perturbations over the past 10 seconds. Using this as input, and leveraging its ability to learn temporal dependencies through deep learning, it outputs a predicted value of the perturbation for the next 5 seconds. To detect changes in disturbance trends in advance.

[0143] Dynamic gain correction: based on The deviation sign, through Dynamic adjustment This allows the observer gain to adapt to the changing trend of the disturbance, enhancing the ability to track rapid fluctuations.

[0144] Accuracy convergence guarantee: Through a closed-loop collaboration of "real-time estimation + trend prediction + gain correction", the accuracy of disturbance estimation error (between the actual disturbance and the actual disturbance) is ensured. The difference between the values ​​converges rapidly within 3 seconds, providing high-precision and forward-looking disturbance data support for feedforward compensation control.

[0145] Furthermore, S4 includes: the disturbance feedforward compensation control model includes:

[0146] Objective function and reward function:

[0147]

[0148] in, To collaboratively optimize the objective function, The capture rate priority weight is used. As energy consumption priority weight, The target capture rate is 90%. CO2 capture rate, Real-time energy consumption (unit: GJ / t CO2). Rated energy consumption (2.0 GJ / t CO2). For the reward function;

[0149] ;

[0150] Dynamic weight adjustment based on the DQN algorithm: when the perturbation amplitude is ≥ ±20%, , When the disturbance amplitude is ≤ ±5%, , ;

[0151] Calculation of feedforward control quantity:

[0152]

[0153] in, This is the feedforward control variable. For the input matrix The inverse matrix, The feedforward differential gain is (0.08-0.12). To estimate the rate of change of the disturbance, To reinforce learning, control gain (0.09-0.11) was applied. The gradient of the objective function with respect to the control input;

[0154] in, ( (This is the estimated value of the disturbance at the previous time step). Solved using the backpropagation algorithm;

[0155] Process constraints: ( When the reboiler temperature is reached, ( (This refers to the steam flow component in feedforward control). Rated value ( (for low fluid flow rate) Rated value ( (Steam flow rate).

[0156] Dual-objective collaborative optimization: By coupling the capture rate and energy consumption objectives through the objective function, the reward value is maximized when the capture rate approaches 90% and the energy consumption is close to 2.0 GJ / t CO2, guiding the system to converge toward the optimal state.

[0157] Dynamic weight adaptation: Based on the DQN algorithm, it senses the perturbation intensity and dynamically adjusts the weights. and When the disturbance is severe (≥±20%), the focus is on the stability of the capture rate. To avoid large fluctuations; when the system is in steady state (≤±5%), the focus is on energy consumption optimization. This achieves target equilibrium under different operating conditions.

[0158] Calculation of active disturbance rejection control quantity: Feedforward control quantity The interference resistance is achieved in three layers: firstly, through... To offset the direct impact of the current disturbance; secondly, through ( ) Predict the trend of disturbance changes and apply compensation in advance; 3. Through (The gradient is solved through backpropagation) The optimization objective is incorporated into the control quantity to ensure both disturbance rejection and dual objectives.

[0159] Process safety constraints: Embedded hard constraint logic ensures that when the reboiler temperature is ≥123℃, [the following occurs]. Suppressing steam flow prevents thermal degradation of the absorbent; simultaneously limiting the boundary between lean liquid and steam flow prevents equipment from operating beyond its range, ensuring system safety and stability, and ultimately outputting a feedforward control quantity that combines disturbance rejection, optimization, and safety.

[0160] Furthermore, in S5, the state feedback control model includes:

[0161] State feedback control law: based on (State estimation) design, feedback gain Optimize using LQR ( , This is the state weight matrix; , (Input weight matrix)

[0162]

[0163] in, For feedback control, The state feedback gain matrix;

[0164] Ensure the damping ratio of the closed-loop system ( (This refers to the damping ratio of the closed-loop system).

[0165] State data support: High-precision state estimation based on the output of the S2 state observer. (Covering key states such as CO2 concentration in the absorption tower and reboiler temperature), providing reliable state input for feedback control and ensuring that control decisions are based on the actual operating state of the system.

[0166] Optimal feedback gain optimization: The feedback gain matrix is ​​tuned using the LQR (Linear Quadratic Regulator) algorithm. Through the state weight matrix Assign differentiated weights to different state variables (e.g., reboiler temperature weight 8, CO2 concentration weight 10, highlighting the control priority of the core state), and input the weight matrix. By constraining the energy consumption of the control input, the optimal solution is found by minimizing the quadratic performance index of "state deviation cost + control input cost". .

[0167] Deviation correction control law: through Construct linear state feedback logic to estimate the state. With optimal gain Multiply and then invert to form the feedback control quantity—when the system state deviates from the target value. This will produce a reverse adjustment effect, offsetting the state deviation and driving the system to converge toward the target steady state.

[0168] Closed-loop stability assurance: LQR optimization ensures the damping ratio of the closed-loop system. This approach avoids system oscillations caused by insufficient damping while preventing slow response caused by excessive damping, achieving a balance between "rapid convergence and stable operation," thus laying a stable foundation for subsequent integration with feedforward control.

[0169] Furthermore, S6 specifically includes:

[0170] PID parameter adaptive tuning based on the DQN algorithm:

[0171]

[0172] in, For PID proportional coefficient, It is the hyperbolic tangent function. For the reward function;

[0173]

[0174] in, For PID integral coefficients, It is an exponential function;

[0175]

[0176] in, These are the PID differential coefficients. This represents the absolute value of the estimated rate of change of the disturbance; the other parameters are defined as above.

[0177] Smooth composite control law:

[0178]

[0179] in, For the final composite control quantity, For composite control weighting coefficients, Transfer function for PID controller ( ), This refers to the deviation between the target and actual capture rate.

[0180]

[0181] in, The absolute value of the disturbance estimate. perturbation vector The boundary maximum value;

[0182] The control quantity adjustment rate is ≤±3% / min, enabling the system output to adjust under disturbances. Asymptotic tracking of (capture rate target value).

[0183] Adaptive PID parameter tuning: Based on the DQN algorithm, the system's operating status is sensed, and the three main PID coefficients are dynamically optimized to achieve precise matching between operating conditions and parameters.

[0184] proportionality coefficient : Through reward function The synergistic effect of the feedback "capture rate - energy consumption" makes the system approach its optimal state. It stabilizes at around 2.8, but when the state deviation is large... Dynamically increased, enhancing the ability to make immediate corrections;

[0185] Integral coefficient Based on real-time energy consumption Adaptive adjustment enhances the ability to eliminate steady-state deviations and avoids integral overshoot when energy consumption is too high;

[0186] Differential coefficients Positively correlated with the estimated rate of change of the disturbance, the more drastic the disturbance fluctuation, the better. The larger the value, the earlier it can suppress state changes and improve the system's anti-disturbance response speed.

[0187] Smooth composite control quantity fusion: through adaptive weights Integrating feedforward, feedback, and PID deviation adjustment to avoid abrupt control changes:

[0188] Weight The intensity of the disturbance changes dynamically, and the closer the disturbance is to the boundary ( The larger ( The closer it gets to 0.8, the more the feedforward control quantity increases. Higher weights emphasize active disturbance rejection; when the disturbance is weak... Approaching 0.5, feedback control value The weight of the PID deviation adjustment is increased to enhance steady-state accuracy;

[0189] In a composite control law, the feedforward variable is responsible for "disturbance prediction," and the feedback variable is responsible for "state correction." The PID controller... Handling capture rate deviation The three Weighted fusion achieves a dynamic balance between "disturbance resistance and steady state".

[0190] Asymptotic tracking and safety constraints: By imposing a hard constraint that the control quantity adjustment rate is ≤±3% / min, drastic actions of actuators such as valves and pumps are avoided, reducing equipment wear. Simultaneously, a triple closed loop is formed by combining feedforward disturbance prediction, feedback state correction, and PID deviation compensation to ensure that the system output continuously moves towards the target capture rate value under any disturbance condition. Convergence ultimately achieves the asymptotic tracking control objective.

[0191] A post-combustion carbon capture control method based on perturbation observation also includes S7: closed-loop optimization: (System output vector) and (System state vector) is fed back to S1-S6 for correction. (System Matrix) (Observer gain matrix) (Objective function weights) and (PID parameters) to continuously optimize control performance and ensure that core objectives (capture rate ≥85%, energy consumption ≤2.0 GJ / t CO2, disturbance recovery time ≤50s, reboiler temperature ≤125℃) are achieved.

[0192] Example 2:

[0193] A post-combustion carbon capture control system based on perturbation observation includes: a sensing layer, a control layer, an execution layer, and a feedback layer, wherein the control layer includes:

[0194] System modeling and disturbance calibration module: used to construct the state-space model of the absorber-reboiler coupling and calibrate the boundary and priority weights of disturbance parameters;

[0195] State observation module: used to estimate the unmeasurable states of the system and ensure that the state estimation error converges quickly;

[0196] The robust perturbation observation and LSTM prediction module is used to estimate perturbations in real time, predict future perturbation trends, and dynamically adjust the observer gain.

[0197] The disturbance feedforward compensation control module is used to calculate the feedforward control quantity that combines disturbance rejection, optimization, and safety.

[0198] The state feedback control module is used to output feedback control quantities based on the optimal gain;

[0199] The PID fusion control module is used to adaptively tune PID parameters and smoothly fuse feedforward and feedback control quantities;

[0200] The perception layer collects system operation data and transmits it to the control layer. The execution layer receives the final control quantity output by the control layer to drive device regulation. The feedback layer monitors the system status in real time and feeds it back to the control layer to achieve closed-loop optimization.

[0201] Includes auxiliary modules:

[0202] The data storage module stores model parameters, disturbance data, control variables, observations, etc., and supports backtracking of historical data for ≥1 year;

[0203] The feedback monitoring module collects system output data in real time. With state Feedback is sent to each control unit for deviation correction;

[0204] The human-machine interface visually displays key indicators such as disturbance observations, predictions, control variables, energy consumption, and capture rate, and supports manual start / stop, parameter modification, and fault alarms.

[0205] Application Example Data

[0206] I. Basic Operating Parameters

[0207] Carbon capture system scale: 1 million Nm³ / h of flue gas to be processed, and a CO2 capture capacity of 300,000 tons / year;

[0208] Absorption tower parameters: diameter 14m, packing height 28m, operating pressure 0.12MPa, operating temperature 40-60℃;

[0209] Reboiler parameters: rated steam pressure 0.8MPa, rated heating power 22MW, operating temperature 120-123℃;

[0210] Control targets: capture rate ≥ 85% (steady-state ≥ 90%), energy consumption per unit of CO2 capture ≤ 2.0 GJ / t CO2.

[0211] II. Disturbance Conditions

[0212] Flue gas flow disturbance: Rated value 1 million Nm³ / h, fluctuation range 700,000-1,300,000 Nm³ / h (±30%), fluctuation period 30 min;

[0213] CO2 concentration perturbation: baseline 12% (dry basis), fluctuation range 10%-25%, random mutation amplitude ≤5% / min;

[0214] Steam pressure disturbance: rated value 0.8MPa, fluctuation range 0.7-0.9MPa (±0.1MPa);

[0215] Pollutant interference: SOx concentration 8 mg / m³, NOx concentration 6 mg / m³.

[0216] III. Control Effect Data

[0217] Monitoring index Conventional PID control The method of the invention Lifting effect Steady-state capture rate 86.2% 92.5% Lifting 6.3 percentage points Dynamic capture rate (when disturbed) 78.5%-88.3% 85.2%-93.1% Lifting 6.7 percentage points at the minimum Unit CO2 capture energy consumption 2.38 GJ / t CO2 1.89 GJ / t CO2 Reduced by 20.6% Disturbance recovery time (±30% flow mutation) 72s 41s Shortened by 43.1% Reboiler temperature fluctuation range 118-127℃ 120-123℃ Fluctuation amplitude reduced by 60% Lean liquid flow regulation fluctuation ±8.5% ±3.2% Fluctuation amplitude reduced by 62.4% Steam flow regulation fluctuation ±10.2% ±2.8% Fluctuation amplitude reduced by 72.5%

Claims

1. A post-combustion carbon capture control method based on perturbation observation, applied to a carbon capture system including an absorber and a reboiler, characterized in that, Includes the following: S1: System Modeling and Disturbance Parameter Calibration: Construct a state-space model of the carbon capture system including the absorber and reboiler, and calibrate the boundaries and priority weights of the disturbance parameters during system operation; S2: Design a state observer using a gain matrix to ensure that the state estimation error converges quickly; S3: Constructing a robust perturbation observer with LSTM predictions: First, based on the system state, inputs, and outputs, construct the perturbation observer, including its gain matrix. Satisfying the Hurwitz matrix constraints, the disturbance is estimated in real time and directly used as input data for the LSTM disturbance prediction model. The LSTM disturbance prediction model outputs the predicted future disturbance value, which is fed back into the disturbance observer's calculation at the next time step, serving as a prediction reference term for the disturbance observer and correcting the observer gain matrix. The dynamic adjustment logic involves inputting the real-time disturbance estimate from the disturbance observer and the future disturbance prediction from the LSTM model into the feedforward compensation control model. S4: Design a disturbance feedforward compensation control model: Define an energy consumption-capture rate collaborative objective function, set a reward function to guide the system toward high precision and low energy consumption; use the DQN algorithm to dynamically adjust the weights of energy consumption and capture rate; calculate the feedforward control quantity based on the disturbance observation value and the gradient of the objective function; embed process hard constraints and correct steam flow boundary limits; S5: State Feedback Control Model: Using the estimated state output by the state observer, design the state feedback control law and output the feedback control quantity; S6: The feedforward and feedback control quantities are input to the PID controller, smoothly merging the feedforward and feedback quantities to achieve the asymptotic tracking control objective of the system under disturbance conditions.

2. The post-combustion carbon capture control method based on perturbation observation according to claim 1, characterized in that, S1 includes: State-space model construction: Building an absorption tower-reboiler coupling model based on Aspen Plus: ， ， Disturbance parameter calibration: flue gas flow rate fluctuation Rated value, CO2 concentration Steam pressure MPa, SOx / NOx≤10 mg / m³; Priority: flue gas flow rate (0.4) > CO2 concentration (0.3) > steam pressure (0.2) > pollutants (0.1).

3. The post-combustion carbon capture control method based on perturbation observation according to claim 1, characterized in that, S2 includes: State observer model: ， Gain tuning: The target pole is 3-5 times the original pole. Tuning formula: ， Ensure state estimation error Convergence time ≤ 2s, steady-state error ≤ ±1%.

4. The post-combustion carbon capture control method based on perturbation observation according to claim 1, characterized in that, In S3, constructing a robust perturbation observer with LSTM predictions includes: Perturbation observer construction: ， ， in for The second row vector (1×5) for The second line is a scalar; solved using LMI. It satisfies the Hurwitz constraint; LSTM prediction module integration: Enter history Training, output 5 seconds ) Collaborative logic: As input to LSTM, Feedback and correction : ； Ensure that the convergence time of the disturbance estimation error is ≤3s.

5. The post-combustion carbon capture control method based on perturbation observation according to claim 1, characterized in that, The S4 mentioned above includes: the disturbance feedforward compensation control model includes: Objective function and reward function: ， ， Dynamic weight adjustment based on the DQN algorithm: Disturbance ≥ ±20% Disturbance ≤ ±5% ; Calculation of feedforward control quantity: ， in , Solve using backpropagation; Process constraints: hour, ; Rated value, Rated value.

6. The post-combustion carbon capture control method based on perturbation observation according to claim 1, characterized in that, In S5, the state feedback control model includes: State feedback control law: based on Design, feedback gain Optimize using LQR ( , ): ， Ensure the damping ratio of the closed-loop system .

7. The post-combustion carbon capture control method based on perturbation observation according to claim 1, characterized in that, Specifically, S6 includes: PID parameter adaptive tuning based on the DQN algorithm: ， ， ， Smooth composite control law: ， ， The control quantity adjustment rate is ≤±3% / min, enabling the system output to adjust under disturbances. Asymptotic tracking.

8. The post-combustion carbon capture control method based on perturbation observation according to claim 1, characterized in that, Also includes S7: Closed-loop optimization: and Feedback sent to S1-S6 for correction. , , and We will continue to optimize control performance to ensure the achievement of core objectives.

9. A post-combustion carbon capture control system based on disturbance observation, characterized in that, include: The system comprises a perception layer, a control layer, an execution layer, and a feedback layer. The control layer includes: System modeling and disturbance calibration module: used to construct the state-space model of the absorber-reboiler coupling and calibrate the boundary and priority weights of disturbance parameters; State observation module: used to estimate the unmeasurable states of the system and ensure that the state estimation error converges quickly; The robust perturbation observation and LSTM prediction module is used to estimate perturbations in real time, predict future perturbation trends, and dynamically adjust the observer gain. The disturbance feedforward compensation control module is used to calculate the feedforward control quantity that combines disturbance rejection, optimization, and safety. The state feedback control module is used to output feedback control quantities based on the optimal gain; The PID fusion control module is used to adaptively tune PID parameters and smoothly fuse feedforward and feedback control quantities; The perception layer collects system operation data and transmits it to the control layer. The execution layer receives the final control quantity output by the control layer to drive device regulation. The feedback layer monitors the system status in real time and feeds it back to the control layer to achieve closed-loop optimization.

10. The post-combustion carbon capture control system based on disturbance observation according to claim 9, characterized in that, Includes auxiliary modules: The data storage module stores model parameters, disturbance data, control variables, observations, etc., and supports backtracking of historical data for ≥1 year; The feedback monitoring module collects system output data in real time. With state Feedback is sent to each control unit for deviation correction; The human-machine interface visually displays key indicators such as disturbance observations, predictions, control variables, energy consumption, and capture rate, and supports manual start / stop, parameter modification, and fault alarms.

Images