Low-energy-consumption phase-change CO2 capture-methanol co-production process system for flue gas of thermal power plant

Abstract

The present invention relates to the technical field of thermal power plant flue gas carbon capture and resource utilization, and discloses a thermal power plant flue gas low energy consumption phase change CO2 capture-methanol co-production process system, which comprises: a CO2 absorption unit, which enables a non-aqueous phase change absorbent to absorb CO2 in thermal power plant flue gas to form a rich absorption liquid; the liquid-liquid phase separation unit is used for spontaneously separating the rich absorption liquid into a barren liquid phase and a rich liquid phase, and returning the barren liquid phase to the CO2 absorption unit; the rich liquid regeneration unit is used for heating and regenerating the rich liquid phase so as to desorb CO2 gas; the CO2 catalytic methanol synthesis unit is used for synthesizing methanol from CO2 gas and external hydrogen-rich gas; the heat pump waste heat recovery unit is used for recovering methanol synthesis reaction heat, upgrading the methanol synthesis reaction heat and supplying the methanol synthesis reaction heat to the pregnant solution regeneration unit; and the AI intelligent regulation and control unit is used for collecting process parameters of each unit and carrying out collaborative optimization control on an execution mechanism of each unit based on an improved model prediction control algorithm. Therefore, the energy consumption in the carbon capture process is greatly reduced, and the high-value utilization of carbon resources is realized.

Description

Technical Field

[0001] This invention relates to the field of carbon capture and resource utilization technology for flue gas from thermal power plants, specifically to a low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants. Background Technology

[0002] As a major source of CO2 emissions in my country, the capture and utilization of flue gas carbon from thermal power plants is a core pathway to achieving the low-carbon transformation of the power industry. Currently, CO2 capture from flue gas in thermal power plants mainly relies on the chemical amine method, with monoethanolamine (MEA) absorption being the most widely used. However, this process has several key drawbacks: firstly, the CO2 loading of the absorbent is low, only about 0.5 mol / mol of amine, resulting in a large circulation volume; secondly, the regeneration process requires heating the entire absorbent, with regeneration energy consumption exceeding 3.8 GJ / ton of CO2, accounting for more than 70% of the operating cost of the carbon capture system, severely hindering the large-scale promotion of the technology.

[0003] The emergence of phase change absorbents has provided a new direction for reducing regeneration energy consumption. They can spontaneously separate phases after absorbing CO2, requiring only the regeneration of the CO2-rich liquid phase, which can significantly reduce ineffective heat consumption. However, existing phase change absorbents still have problems such as insufficient CO2 loading capacity, unstable phase separation performance, and large cycle losses, making it difficult to adapt to the flue gas characteristics of thermal power plants with large flow rates and wide operating conditions.

[0004] Meanwhile, the high-value utilization of captured CO2 is key to improving the economics of carbon capture, utilization, and storage (CCUS) projects. Coupled CO2 capture with methanol synthesis, on-site conversion of carbon resources can be achieved, transforming environmental governance costs into chemical production benefits. However, the carbon capture-methanol cogeneration system is a typical complex system with large inertia, large lag, strong coupling, and multiple variables, presenting three major control challenges: First, traditional proportional-integral-derivative (PID) control lags behind in responding to fluctuations in operating conditions, failing to achieve global collaborative optimization of multiple variables; second, conventional model predictive control (MPC) suffers from the industry pain point of relying on engineering experience to tune multi-objective weight factors, making it difficult to balance steady-state accuracy and dynamic response; third, factors such as equipment aging, scaling, and catalyst activity decay can cause continuous drift in system parameters, and existing control methods lack robustness, making it difficult to ensure stable and efficient operation throughout the system's entire lifecycle.

[0005] In summary, existing technologies cannot simultaneously solve the core problems of high energy consumption in CO2 capture and regeneration of flue gas from thermal power plants, poor economic efficiency in carbon resource utilization, and difficulty in controlling cogeneration systems. There is an urgent need to develop a complete technical solution that integrates low-energy capture, high-value utilization, and intelligent regulation. Summary of the Invention

[0006] To address the aforementioned shortcomings of existing technologies, this invention aims to provide a low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants, achieving the following core objectives: constructing a process system deeply coupled with CO2 capture and methanol synthesis; achieving system energy self-balancing through waste heat recovery via heat pumps; converting captured CO2 into high-value-added methanol products; improving the overall economic efficiency of the CCUS project; proposing an improved model predictive control method adapted to the characteristics of chemical processes, solving the problems of difficulty in tuning weight factors and the inability to simultaneously balance dynamic response and steady-state accuracy in conventional model predictive control, while improving the system's robustness to parameter drift and operating condition fluctuations; and achieving coordinated control of dynamic optimization of absorbent performance and global system operation optimization, ensuring stable and efficient operation of the system within the 10%~100% rated load range, thereby improving the industrial feasibility, reproducibility, and engineering feasibility of the technical solution.

[0007] To achieve the above objectives, the following technical solution is adopted:

[0008] This invention provides a low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants, comprising a CO2 absorption unit, a liquid-liquid phase separation unit, a rich liquid regeneration unit, a CO2 catalytic synthesis methanol unit connected in sequence, a heat pump waste heat recovery unit connected to the CO2 catalytic synthesis methanol unit and the rich liquid regeneration unit respectively, and an AI intelligent control unit electrically connected to each of the aforementioned units respectively.

[0009] The CO2 absorption unit is used to absorb CO2 in the flue gas of thermal power plants using a non-aqueous phase change absorbent to form a rich absorbent liquid.

[0010] The liquid-liquid phase separation unit is used to spontaneously separate the rich absorbent into a lean liquid phase and a rich liquid phase, and return the lean liquid phase to the CO2 absorption unit.

[0011] The rich liquid regeneration unit is used to heat and regenerate the rich liquid phase to release CO2 gas.

[0012] The CO2 catalytic methanol synthesis unit is used to synthesize methanol from CO2 gas and external hydrogen-rich gas.

[0013] The heat pump waste heat recovery unit is used to recover the heat of methanol synthesis reaction and then supply it to the rich liquid regeneration unit after upgrading.

[0014] The AI ​​intelligent control unit is used to collect the process parameters of the aforementioned units and to perform coordinated optimization control of the actuators of each unit based on the improved model predictive control algorithm.

[0015] In a preferred embodiment of the present invention, the non-aqueous phase change absorbent is composed of a lipophilic tertiary amine, a sterically hindered amine activator, and a hydrophobic organic solvent.

[0016] The lipophilic tertiary amine is N-methylcyclohexylamine, accounting for 40% to 60% of the absorbent by mass; the hindered amine activator is 2-amino-2-methyl-1-propanol, accounting for 10% to 20% of the absorbent by mass.

[0017] The hydrophobic organic solvent is n-octanol, which accounts for 20% to 40% of the mass of the absorbent.

[0018] In a preferred embodiment of the present invention, the non-aqueous phase change absorbent is a homogeneous solution at 20~40℃. After absorbing CO2, liquid-liquid phase separation occurs to form a rich liquid phase and a poor liquid phase. The volume of the rich liquid phase accounts for 30%~40% of the total volume of the absorbent, and it enriches more than 95% of the CO2 loading in the system.

[0019] As a preferred embodiment of the present invention, the AI ​​intelligent control unit includes a state observation and delay compensation module, a multi-objective reference value conversion and normalization module, a dual-mode control quantity optimization and action time calculation module, and a rolling optimization and feedback correction module, which are set up to realize the improved model predictive control algorithm.

[0020] The state observation and delay compensation module is used to establish a system state space prediction model and compensate for system delay. The system state space prediction model includes a state vector, a control input vector, a measurable disturbance vector, and an output vector.

[0021] The multi-objective reference value conversion and normalization module is used to convert multiple control objectives into a single comprehensive state reference vector and to normalize all state variables.

[0022] The dual-mode control quantity optimization and action time calculation module is used to determine the first optimal control quantity and the second optimal control quantity within a control cycle, and to calculate the optimal action time of the first optimal control quantity.

[0023] The rolling optimization and feedback correction module is used to repeatedly execute the optimization process in each control cycle based on the real-time state estimate.

[0024] In a preferred embodiment of the present invention, the state observation and delay compensation module is specifically used for:

[0025] A hybrid modeling approach combining system identification and mechanism correction is used to establish continuous domain state-space equations, which serve as the basis for the system state-space prediction model.

[0026] The Heinrich prediction-correction method is used to discretize the continuous domain state space equations and achieve one-step delay compensation.

[0027] The continuous domain state space equation includes a state equation and an output equation. The state equation is used to describe the relationship between the state vector and time, and the output equation is used to establish the mapping relationship between the state vector and the output vector.

[0028] The state vector includes the CO2 concentration in the flue gas at the absorber outlet, the CO2 loading rate in the rich liquid, the temperature of the regeneration tower bottom, the bed temperature of the methanol synthesis reactor, the heating capacity of the heat pump system, the effective amine concentration in the absorbent, the system heat loss coefficient, and the relative activity of the catalyst.

[0029] The control input vector includes the opening degree of the absorbent feed regulating valve, the frequency of the rich liquid feed pump, the opening degree of the reboiler steam regulating valve, the frequency of the heat pump compressor, and the opening degree of the syngas feed regulating valve.

[0030] The measurable disturbance vector includes the inlet flue gas flow rate, the inlet flue gas CO2 concentration, and the hydrogen concentration of the hydrogen-rich source.

[0031] The output vector includes CO2 capture rate, system unit energy consumption, and methanol single-pass yield.

[0032] The coefficient matrices in the state equation and output equation are obtained by identifying data from the on-site operation of thermal power plants and correcting them in conjunction with the chemical process mechanism.

[0033] In a preferred embodiment of the present invention, the multi-target reference value conversion and normalization module is specifically used for:

[0034] The state variables are normalized by scaling all state variables to a predetermined numerical range through a linear transformation to eliminate dimensional differences.

[0035] The process constraint boundaries of the system are determined, including the lower limit of CO2 capture rate, the upper limit of system unit energy consumption, and the lower limit of methanol single-pass yield.

[0036] Based on the optimal parameters of the system under rated operating conditions, and combined with the process constraint boundaries, a normalized comprehensive state reference vector composed of the normalized optimal reference values ​​of each state variable is constructed as a unified reference benchmark for multi-objective control.

[0037] In a preferred embodiment of the present invention, the dual-mode control quantity optimization and action time calculation module is specifically used for:

[0038] Using the normalized composite state reference vector at a specified future time as the target, a cost function is constructed to measure the degree of deviation between the normalized predicted value of the system state and the normalized composite state reference vector.

[0039] By traversing the candidate control variables and calculating the cost function value, the combination of control variables that minimizes the cost function value is selected as the first optimal control variable.

[0040] The candidate set of the second optimal control quantity is determined based on the single-step adjustment principle, which means that only a single control quantity is allowed to switch gears once relative to the first optimal control quantity, while the other control quantities remain unchanged.

[0041] Assuming that the action time of the first optimal control variable within a single control cycle is the optimal action time to be solved, and the action time of the second candidate control variable is the remaining time of the control cycle, the normalized state prediction value at the specified future time is derived by combining the normalized state prediction model.

[0042] Substitute the normalized state prediction value at the specified future time into the cost function, solve for the partial derivative of the cost function with respect to the optimal action time to obtain the optimal action time analytically, and then perform amplitude limiting processing on the optimal action time to make its value within a single control cycle.

[0043] Iterate through all the second candidate control variables, calculate the corresponding optimal action time and cost function value, select the second candidate control variable that minimizes the cost function value as the second optimal control variable, and use the first optimal control variable and the second optimal control variable as the final dual-mode control variable combination output.

[0044] Within the current control cycle, the first optimal control quantity is applied and maintained for the optimal duration, and then the second optimal control quantity is applied and maintained for the remaining time of the control cycle.

[0045] In a preferred embodiment of the present invention, the rolling optimization and feedback correction module is specifically used for:

[0046] Within each control cycle, the real-time state estimate of the system is obtained through the full-order closed-loop state observer, and the state estimate is fed back and corrected in combination with the field measured output value to update the initial value of the state prediction model.

[0047] Based on the corrected initial state value, the multi-target reference value conversion and normalization module and the dual-mode control quantity optimization and action time calculation module are triggered to re-execute the corresponding operations to output the updated control quantity;

[0048] The feedback correction and rolling optimization process is executed cyclically to achieve real-time closed-loop optimization control of the system.

[0049] As a preferred embodiment of the present invention, the mathematical model of the full-order closed-loop state observer includes an observer state equation and an observer output equation. The observer state equation is used to describe the relationship between the state observation estimate and time, and the observer output equation is used to establish the mapping relationship between the state observation estimate and the output observation estimate.

[0050] The observer state equation includes an observer gain matrix, which is obtained by solving the pole placement method or the Riccati equation and is used to adjust the convergence speed and stability of the observer.

[0051] In a preferred embodiment of the present invention, the AI ​​intelligent control unit further includes an LSTM deep learning absorbent ratio optimization module that works in conjunction with the improved model predictive control algorithm;

[0052] The LSTM deep learning absorbent ratio optimization module adopts a three-layer LSTM network structure. Its input parameters include inlet flue gas CO2 concentration, flue gas flow rate, flue gas temperature, absorbent circulation volume and rich liquid CO2 loading rate. The output parameter is the optimal replenishment ratio of N-methylcyclohexylamine, 2-amino-2-methyl-1-propanol and n-octanol.

[0053] The training dataset for the LSTM deep learning absorbent ratio optimization module comes from on-site operation data of thermal power plants and laboratory absorbent performance test data. The model training and iterative optimization are completed through supervised learning.

[0054] The LSTM deep learning absorbent ratio optimization module outputs the optimal replenishment ratio command in real time to control the operation of the absorbent replenishment system.

[0055] Compared with the prior art, the present invention achieves the following beneficial effects:

[0056] 1. This invention deeply couples CO2 capture with methanol synthesis, using the captured high-purity CO2 directly as a carbon feedstock for methanol synthesis, thus realizing the on-site high-value conversion of carbon resources. At the same time, the low-grade heat energy released by the methanol synthesis reaction is recovered through a heat pump waste heat recovery system, and after upgrading, it is reused in the CO2 regeneration process, thus constructing an energy self-balancing mode of "chemical-driven capture and capture-chemical linkage". The system's energy utilization efficiency is improved by more than 30%, significantly reducing dependence on external high-quality steam and significantly improving the economic feasibility of the CCUS project.

[0057] 2. The non-aqueous phase change absorbent developed in this invention achieves a CO2 molar loading rate more than twice that of traditional MEA absorbents under typical operating conditions of flue gas in thermal power plants. Through spontaneous liquid-liquid phase separation, only 30% to 40% of the volume of the rich liquid phase needs to be heated for regeneration, eliminating the ineffective heat consumption of heating the entire absorbent in traditional processes from the thermodynamic source. The regeneration energy consumption is reduced by more than 50% compared to traditional MEA processes, and the CO2 capture rate is stable at ≥90%, fully meeting the national carbon capture standards for thermal power plants.

[0058] 3. The improved model predictive control algorithm proposed in this invention abandons the vector geometry concept and microsecond-level time scale in the field of motor control, and constructs a control architecture that is fully adapted to the large inertia and large lag characteristics of chemical industry: it solves the problem of incompatibility of multi-variable dimensions through normalization processing; it completely solves the industry problem of traditional MPC multi-objective weight factors relying on engineering experience tuning by converting multi-objectives to a single comprehensive state reference vector; through dual-mode control quantity combination optimization and action time analytical solution, it takes into account both ultra-fast dynamic response and high steady-state control accuracy. The dynamic response speed is improved by more than 50% compared with traditional PID control, and it can quickly respond to large fluctuations in flue gas load within 3-5 minutes, while avoiding frequent actuator operation and extending equipment service life.

[0059] 4. The full-order closed-loop state observer designed in this invention can estimate unmeasurable state variables in real time and has strong robustness to parameter drift caused by equipment aging and scaling, as well as external disturbances caused by flue gas composition and load fluctuations. Through rolling optimization and feedback correction, the model error is continuously corrected, and the system can operate stably within the range of 10% to 100% of the rated load, solving the problem of poor stability of traditional control methods under low load conditions. The algorithm has a moderate computational load and can be reliably executed within the interruption cycle of conventional industrial distributed control systems (DCS) / programmable logic controllers (PLCs), without the need for high-end hardware support, making it highly practical for engineering implementation.

[0060] 5. This invention deeply integrates the improved model predictive control algorithm with long short-term memory network (LSTM) deep learning, realizing both real-time global optimization of system energy flow and material flow, and long-term dynamic optimization of absorbent performance. It can continuously tap the system's energy-saving potential and maximize the economic benefit of methanol co-production. It solves the problems of slow response and inability to find global optimization in traditional control systems, ensuring efficient, economical and stable operation of the system throughout its entire life cycle, and providing core technical support for the large-scale promotion of CCUS technology in thermal power plants.

[0061] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0062] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0063] Figure 1This is a schematic diagram of the system architecture of a low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas in a thermal power plant, according to an embodiment of the present invention.

[0064] Figure 2 This is a schematic diagram of the internal module structure of the AI ​​intelligent control unit in an embodiment of the present invention;

[0065] Figure 3 This is a flowchart illustrating the improved model predictive control algorithm in an embodiment of the present invention;

[0066] Figure 4 This is a schematic diagram of the network structure of the LSTM deep learning absorber ratio optimization module in an embodiment of the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0068] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0069] This invention addresses the core pain points of existing carbon dioxide (CO2) capture technologies in thermal power plants, aiming to overcome the shortcomings of existing technologies: First, traditional amine-based CO2 capture and regeneration technologies have high energy consumption and low absorbent loading capacity, making them unsuitable for the high-flow-rate operating conditions of flue gas in thermal power plants; second, there is a lack of highly economical ways to dispose of captured CO2, resulting in poor overall profitability of carbon capture, utilization, and storage (CCUS) projects; third, the carbon capture-methanol cogeneration system has the characteristics of large inertia, large lag, strong coupling, and multiple variables, making traditional proportional-integral-derivative (PID) control slow in dynamic response and poor adaptability to operating conditions, while conventional model predictive control (MPC) faces industry challenges such as difficulty in tuning multi-objective weight factors and the inability to balance steady-state accuracy and dynamic response; fourth, chemical process operating conditions fluctuate greatly, and equipment parameters continue to drift with aging and scaling, making existing control methods insufficiently robust and difficult to achieve stable and optimized operation throughout the system's entire life cycle.

[0070] This invention provides a low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants, a non-aqueous phase change absorbent adapted to the system, and an AI intelligent control method based on improved model predictive control proposed in this invention. This achieves a significant reduction in energy consumption during the carbon capture process and high-value utilization of carbon resources. At the same time, it solves the core problem of multivariate optimization control in chemical co-production systems, ensuring efficient, stable, and economical operation of the system over a wide range of operating conditions, and improving the industrial feasibility, reproducibility, and engineering applicability of the technical solution.

[0071] Figure 1 This is a schematic diagram of the system architecture of a low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from a thermal power plant, according to an embodiment of the present invention. Figure 1 As shown, the present invention provides a low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas in thermal power plants, comprising a CO2 absorption unit 110, a liquid-liquid phase separation unit 120, a rich liquid regeneration unit 130, a CO2 catalytic synthesis methanol unit 140 connected in sequence, a heat pump waste heat recovery unit 150 connected to the CO2 catalytic synthesis methanol unit 140 and the rich liquid regeneration unit 130 respectively, and an AI intelligent control unit 160 electrically connected to each of the aforementioned units respectively.

[0072] CO2 absorption unit 110 is used to absorb CO2 in flue gas from thermal power plants using a non-aqueous phase change absorbent to form a rich absorbent liquid.

[0073] The liquid-liquid phase separation unit 120 is used to spontaneously separate the rich absorbent into a lean liquid phase and a rich liquid phase, and return the lean liquid phase to the CO2 absorption unit 110.

[0074] The rich liquid regeneration unit 130 is used to heat and regenerate the rich liquid phase to release CO2 gas.

[0075] The CO2 catalytic methanol synthesis unit 140 is used to synthesize methanol from CO2 gas and external hydrogen-rich gas.

[0076] Specifically, the CO2 catalytic methanol synthesis unit 140 has its feed gas inlet connected to the CO2 outlet of the rich liquid regeneration unit 130 and an external hydrogen-rich gas supply pipeline. The external hydrogen-rich gas can be hydrogen produced as a byproduct of coal chemical industry, coke oven gas, or hydrogen produced by water electrolysis. Preferably, before entering the reactor, the external hydrogen-rich gas needs to undergo purification treatment such as desulfurization and dechlorination in a pretreatment unit to ensure that its impurity content meets the requirements of the methanol synthesis catalyst.

[0077] The heat pump waste heat recovery unit 150 is used to recover the heat of methanol synthesis reaction and then supply it to the rich liquor regeneration unit 130 after upgrading.

[0078] The AI ​​intelligent control unit 160 is used to collect the process parameters of the aforementioned units and to perform coordinated optimization control of the actuators of each unit based on the improved model predictive control algorithm.

[0079] More specifically, the specific structure and connection relationships of each unit are as follows:

[0080] (I) CO2 Absorption Unit 110

[0081] A packed absorber tower is adopted, with its bottom flue gas inlet connected to the clean flue gas pipeline after desulfurization and denitrification in the thermal power plant, its top absorbent inlet connected to the lean liquid phase outlet of the liquid-liquid phase separation unit, and its bottom rich absorbent outlet connected to the inlet of the liquid-liquid phase separation unit. The absorber tower is equipped with a flue gas flow sensor, a flue gas CO2 concentration sensor, an absorbent temperature sensor, a bottom liquid level sensor, and an absorbent feed regulating valve to collect unit operating parameters and control the absorbent feed rate.

[0082] CO2 absorption unit 110 is used to absorb CO2 in flue gas from thermal power plants using a non-aqueous phase change absorbent to form a rich absorbent liquid.

[0083] As a preferred embodiment of the present invention, the present invention proposes a low-energy liquid-liquid phase-separated non-aqueous phase change absorbent for CO2 capture. The non-aqueous phase change absorbent in the system is composed of a lipophilic tertiary amine, a sterically hindered amine activator, and a hydrophobic organic solvent in a certain mass ratio. The components and their proportions are as follows:

[0084] (1) Lipophilic tertiary amine: N-methylcyclohexylamine (MCS), accounting for 40%~60% by mass, is the main component for CO2 absorption;

[0085] (2) Steric hindered amine activator: 2-amino-2-methyl-1-propanol (AMP), with a mass ratio of 10%~20%, is used to improve the CO2 absorption reaction rate;

[0086] (3) Hydrophobic organic solvent: n-octanol, with a mass ratio of 20%~40%, is a phase separation control component.

[0087] The core characteristics of this non-aqueous phase change absorbent are: it is a homogeneous and transparent solution at room temperature (20-40℃); after absorbing CO2, the polarity of the carbamate generated in the system is significantly enhanced, and it spontaneously undergoes liquid-liquid phase separation with the hydrophobic solvent system, forming an upper lean liquid phase and a lower rich liquid phase; the volume of the rich liquid phase accounts for only 30%-40% of the total volume of the absorbent, and it enriches more than 95% of the CO2 loading in the system; under typical operating conditions of flue gas from thermal power plants with an absorption temperature of 40℃ and a CO2 partial pressure of 15kPa, the CO2 molar loading is ≥1.2mol / mol amine, which is far higher than the industry standard of 0.5mol / mol amine for traditional monoethanolamine (MEA) absorbents.

[0088] (II) Liquid-liquid phase separation unit 120

[0089] A horizontal gravity phase separator is adopted, with its rich liquid phase outlet connected to the inlet of the regeneration tower of the rich liquid regeneration unit, and the lean liquid phase outlet flowing back to the top inlet of the CO2 absorption unit after passing through a cooler. The phase separator is equipped with a two-phase interface sensor, a rich liquid CO2 loading rate online analyzer, and a lean liquid amine concentration sensor to collect phase separation process parameters and monitor absorbent performance.

[0090] (III) Rich liquid regeneration unit 130

[0091] A packed regeneration tower is adopted, with the heat source side of the reboiler at the bottom of the tower connected to the condenser end of the heat pump waste heat recovery unit. The desorbed gas outlet at the top of the tower is cooled and separated into gas and liquid, and then connected to the feed gas inlet of the CO2 catalytic synthesis methanol unit. The regeneration tower is equipped with temperature sensors, pressure sensors, reboiler heat flow sensors, steam feed regulating valves, and frequency converters for the rich liquid feed pump, which are used to collect regeneration process parameters and control the rich liquid feed rate and regeneration heat source.

[0092] (iv) CO2 catalytic synthesis of methanol unit 140

[0093] A tubular fixed-bed reactor is adopted, with its feed gas inlet connected to an external hydrogen-rich gas source (hydrogen from coal chemical byproducts / coke oven gas) pipeline. The reactor shell side is connected to the evaporation end of the heat pump waste heat recovery unit, and the reactor outlet is equipped with a methanol separation and purification device. Inside the reactor, along the feed gas flow direction, an impurity adsorption protective layer and a methanol synthesis catalyst layer are arranged sequentially. The impurity adsorption protective layer is filled with modified activated carbon adsorbent to remove trace amounts of sulfur dioxide (SOx), nitrogen oxides (NOx), and amine impurities entrained in CO2, thereby extending the service life of the main catalyst. The synthesis unit is equipped with reactor temperature sensors, pressure sensors, synthesis gas flow sensors, an online methanol yield analyzer, a reaction heat flow sensor, and a synthesis gas feed regulating valve to collect synthesis process parameters and control the synthesis gas feed rate.

[0094] (V) Heat pump waste heat recovery unit 150

[0095] 1. A high-temperature water source heat pump system is adopted. Its evaporation end recovers the exothermic reaction heat of the methanol synthesis reactor, and the condensation end transfers the upgraded high-grade heat energy to the reboiler of the regeneration tower to provide heat for the CO2 desorption process.

[0096] 2. The heat pump system is equipped with evaporator / condenser temperature sensors, compressor frequency sensors, heating capacity sensors, and compressor frequency converters. It is also equipped with an auxiliary steam heater connected in parallel with the reboiler heat source side to supplement heat when the operating conditions fluctuate greatly.

[0097] (vi) AI Intelligent Control Unit 160

[0098] The AI ​​intelligent control unit 160 has its signal input terminal electrically connected to the signal output terminals of all the aforementioned sensors and online analyzers, and its signal output terminal electrically connected to the control terminals of all the aforementioned regulating valves and frequency converters. The AI ​​intelligent control unit 160 is equipped with the improved model predictive control algorithm proposed in this invention, as well as an LSTM deep learning absorbent ratio optimization module (LSTM is a long short-term memory network), which is used to realize closed-loop optimization control of the entire system and adapt to the process characteristics of large inertia and large lag in chemical processes.

[0099] To address this, the present invention proposes an AI intelligent control algorithm based on improved model predictive control (hereinafter referred to as: improved model predictive control algorithm). This improved model predictive control algorithm is mounted on the AI ​​intelligent control unit 160. It draws on the core technologies of improved model predictive control, such as multi-step predictive rolling optimization, weightless factor cost function design, piecewise control quantity combination optimization, and robust full-order observer. It is adapted and improved in combination with the characteristics of chemical process control to construct a control architecture of "hybrid modeling + state observation + rolling optimization". The specific implementation steps of the improved model predictive control algorithm include: state observation and delay compensation, multi-objective reference value conversion and normalization, dual-mode control quantity optimization and action time calculation, rolling optimization and feedback correction.

[0100] Furthermore, the AI ​​intelligent control unit 160 includes a state observation and delay compensation module 161, a multi-objective reference value conversion and normalization module 162, a dual-mode control quantity optimization and action time calculation module 163, and a rolling optimization and feedback correction module 164, all configured to implement the improved model predictive control algorithm. The state observation and delay compensation module 161 is used to establish a system state-space prediction model and compensate for system delays. The multi-objective reference value conversion and normalization module 162 is used to convert multiple control objectives into a single comprehensive state reference vector and normalize all state variables. The dual-mode control quantity optimization and action time calculation module 163 is used to determine the first optimal control quantity and the second optimal control quantity within a control cycle and calculate the optimal action time of the first optimal control quantity. The rolling optimization and feedback correction module 164 is used to repeatedly execute the optimization process in each control cycle based on the real-time state estimate. Figure 2 The diagram shown is a schematic representation of the internal module structure of the AI ​​intelligent control unit 160 in an embodiment of the present invention. Figure 3 The diagram shown is a flowchart of the improved model predictive control algorithm in an embodiment of the present invention.

[0101] 1. State observation and delay compensation module 161

[0102] This invention employs a hybrid modeling method combining system identification and mechanism correction to establish the state-space equations of the continuous domain, while simultaneously performing discretization and delay compensation for the delay characteristics of chemical processes.

[0103] Furthermore, the state observation and delay compensation module 161 is specifically used for: establishing a continuous domain state-space equation using a hybrid modeling method of system identification and mechanism correction, serving as the basis for the system state-space prediction model; discretizing the continuous domain state-space equation using the Hennew prediction-correction method and implementing one-step delay compensation; wherein, the continuous domain state-space equation includes state equations and output equations, the state equations are used to describe the relationship between the state vector and time, and the output equations are used to establish the mapping relationship between the state vector and the output vector; the state vector includes the CO2 concentration of the flue gas at the absorber outlet, the CO2 loading rate of the rich liquid, and the regeneration tower... The system parameters include: reactor temperature, methanol synthesis reactor bed temperature, heat pump system output, effective amine concentration in the absorbent, system heat loss coefficient, and catalyst relative activity. Control input vectors include the absorbent feed regulating valve opening, rich liquid feed pump frequency, reboiler steam regulating valve opening, heat pump compressor frequency, and syngas feed regulating valve opening. Measurable disturbance vectors include inlet flue gas flow rate, inlet flue gas CO2 concentration, and hydrogen concentration from the hydrogen-rich gas source. Output vectors include CO2 capture rate, system unit energy consumption, and methanol single-pass yield. The coefficient matrices in the state equation and output equation are identified through on-site operating data from the thermal power plant and corrected using chemical process mechanisms. The specific process for establishing the system state-space prediction model and completing discretization and delay compensation is as follows:

[0104] (1) Establish the continuous domain state-space equation as the basis of the prediction model. The equation is:

[0105] ;

[0106] In the formula, each symbol is defined as follows:

[0107] : State vector The parameters are, in order: CO2 concentration in flue gas at the absorber outlet, CO2 loading rate in rich liquid, temperature of the regeneration tower bottom, bed temperature of the methanol synthesis reactor, heating capacity of the heat pump system, effective amine concentration in the absorbent, system heat loss coefficient, and relative catalyst activity.

[0108] : Control input vector, The parameters are, in order: absorbent feed regulating valve opening, rich liquid feed pump frequency, reboiler steam regulating valve opening, heat pump compressor frequency, and synthesis gas feed regulating valve opening.

[0109] : Measurable perturbation vector The values ​​are, in order, inlet flue gas flow rate, inlet flue gas CO2 concentration, and hydrogen concentration from the hydrogen-rich source.

[0110] : Control the output vector, The following are, in order: CO2 capture rate, system unit energy consumption, and methanol single-pass yield;

[0111] System state matrix : Control input matrix, Output matrix, : Disturbance matrix; The above matrix is ​​obtained by identifying the initial matrix through on-site operation data of thermal power plants, and then corrected by combining chemical process mechanisms (mass transfer reaction kinetics, heat balance equations) to adapt to the operating conditions of equipment aging and parameter drift.

[0112] (2) The Henu prediction-correction method is used to discretize the continuous state equation, and one-step delay compensation is achieved simultaneously. The discretized state prediction equation is as follows:

[0113] ;

[0114] In the formula, The system control cycle is set to 5–30 seconds based on the thermal inertia characteristics of the chemical process, adapting to the conventional interruption cycle of industrial distributed control systems (DCS) / programmable logic controllers (PLCs). : Current control moment, : Next control moment; : The initial value of the predicted state vector at the next moment; : Precise predicted value after prediction and correction; The current control input vector. : The perturbation vector at the current moment.

[0115] 2. Multi-objective reference value conversion and normalization module 162

[0116] To address the problem of traditional MPC multi-objective weight factors relying on engineering experience for tuning, and to eliminate the dimensional differences of chemical state variables, the multi-objective reference value conversion and normalization module 162 of this invention is equipped with a normalization preprocessing step and completes the conversion of multiple objectives into a single reference vector.

[0117] Furthermore, the multi-objective reference value conversion and normalization module 162 is specifically used for: normalizing the state variables by scaling all state variables to a predetermined numerical range through linear transformation to eliminate dimensional differences; determining the system's process constraint boundaries, which include the lower limit of CO2 capture rate, the upper limit of system unit energy consumption, and the lower limit of methanol single-pass yield; and constructing a normalized comprehensive state reference vector composed of the normalized optimal reference values ​​of each state variable, based on the optimal parameters of the system under rated operating conditions and in conjunction with the process constraint boundaries, as a unified reference benchmark for multi-objective control. The specific steps are as follows:

[0118] 2.1 Normalize the state variables by scaling them to the [0,1] interval using a linear transformation. This eliminates differences in dimensions and orders of magnitude, ensuring the physical validity of subsequent vector operations. The normalization formula is as follows:

[0119] ;

[0120] In the formula, : No. Normalized values ​​of each state variable; : No. The actual values ​​of each state variable; : No. The lower limit of industrial operation for each state variable; : No. The upper limit of industrial operation for each state variable; , These limits are predetermined based on process design specifications, equipment design parameters, and safe operating procedures. For example, the temperature of the regeneration tower bottom. The lower and upper operating limits can be set to 100℃ and 140℃ respectively. Below the lower limit, desorption may be incomplete, while above the upper limit may lead to thermal degradation of the absorbent. The normalized state vector is denoted as... The observed values ​​and reference values ​​are simultaneously normalized.

[0121] 2.2 Complete the conversion from multi-objective reference values ​​to a single integrated state reference vector, specifically as follows:

[0122] (1) Determine the system process constraint boundaries: capture rate System unit energy consumption ton Methanol single-pass yield ;

[0123] (2) Based on the optimal parameters of the system under rated operating conditions, and in conjunction with the above-mentioned constraint boundaries, a normalized comprehensive state reference vector is constructed. The expression is:

[0124] ;

[0125] In the formula, The normalized integrated state reference vector consists of the normalized optimal reference values ​​of eight state variables and serves as a unified reference benchmark for multi-objective control. : No. The optimal reference values ​​for each state variable correspond sequentially to the carbon dioxide in the flue gas at the absorber outlet. Concentration, rich solution The optimal setpoints for loading rate, regeneration tower bottom temperature, methanol synthesis reactor bed temperature, heat pump system heating capacity, effective amine concentration in absorbent, system heat loss coefficient, and catalyst relative activity are determined by system mechanism analysis and optimal on-site operating data. Vector transpose operation converts a row vector into a column vector, adapting to the computational requirements of state-space models; : No. The normalized calculation formula for the optimal reference value of each state variable scales the optimal reference value to... Intervals eliminate dimensional differences.

[0126] 3. Dual-mode control quantity optimization and action time calculation module 163

[0127] This invention optimizes the traditional "dual vector" motor control into a dual-mode control quantity combination optimization adapted to chemical processes. Within one control cycle, it uses two sets of optimal control quantity combinations and optimizes their action time, taking into account both dynamic response speed and steady-state control accuracy. At the same time, it optimizes the search strategy to reduce industrial computing load.

[0128] Furthermore, the dual-mode control quantity optimization and action time calculation module 163 is specifically used for: constructing a cost function to measure the deviation between the normalized comprehensive state reference vector at a specified future time and the normalized comprehensive state reference vector; selecting the control quantity combination that minimizes the cost function value as the first optimal control quantity by traversing candidate control quantities and calculating the cost function value; determining the candidate set of the second optimal control quantity based on the single-step adjustment principle, whereby only a single control quantity is allowed to switch gears once relative to the first optimal control quantity, while the remaining control quantities remain unchanged; assuming that the action time of the first optimal control quantity within a single control cycle is the optimal action time to be solved, and the action time of the second candidate control quantity is the remaining time of the control cycle, combined with... The normalized state prediction model derives the normalized state prediction value at the specified future time. Substituting this value into the cost function, the optimal action time is analytically calculated by ensuring the partial derivative of the cost function with respect to the optimal action time is zero. The optimal action time is then limited to a value within a single control cycle. All second candidate control variables are iterated over, and their corresponding optimal action times and cost function values ​​are calculated. The second candidate control variable that minimizes the cost function value is selected as the second optimal control variable. The first and second optimal control variables are then used as the final dual-mode control variable combination output. Within the current control cycle, the first optimal control variable is applied first and maintained for the optimal action time, followed by the application of the second optimal control variable for the remaining time of the control cycle.

[0129] The specific steps are as follows:

[0130] 3.1. Initial selection of the first optimal control variable, specifically:

[0131] (1) with Using the time-normalized integrated state reference vector as the objective, construct a cost function without weighting factors:

[0132] ;

[0133] In the formula, Cost function without weighting factors, used to measure The degree of deviation between the normalized predicted value of the system state at any given time and the normalized comprehensive state reference vector is the evaluation criterion for the optimal selection of control variables. The 2-norm square operation is used to calculate the square of the Euclidean distance between two vectors, avoiding the square root operation to simplify the calculation, while ensuring the non-negativity of the bias evaluation; : The normalized predicted values ​​of the system state at time t, corresponding to the eight state variables in sequence. The normalized prediction result at time step is obtained recursively by combining the discretized prediction equation with the normalization process; : Current control time, which is the time node identifier for discretized control; The second control moment is the target moment of the prediction model, which is used to predict the system state in advance to optimize the control strategy.

[0134] To reduce computational load, critical control variables are pre-screened through sensitivity analysis, and only three candidate levels (decrease, maintain, increase) are set for critical control variables, while non-critical control variables are kept at their current values; alternatively, quadratic programming (QP) can be used instead of full traversal optimization.

[0135] (3) Traverse the candidate control variables and calculate the cost function value, and select the appropriate control variable. The combination of control variables with the smallest value is taken as the first optimal control variable. Its normalized value is denoted as .

[0136] in, The first optimal control variable is the one that, after traversing all candidate control variables, makes the cost function... The combination of control quantities with the smallest value corresponds to the optimal combination of absorbent feed regulating valve opening, rich liquid feed pump frequency, reboiler steam regulating valve opening, heat pump compressor frequency, and syngas feed regulating valve opening, in turn. The normalized value of the first optimal control quantity, Each component is scaled to the [0,1] interval to adapt to the normalized state prediction model operation.

[0137] 3.2. Determine the candidate set of the second optimal control quantity, specifically as follows:

[0138] Following the "single-step adjustment" principle of chemical processes, the second candidate control quantity is only adjusted from the value of the control variable. The selection is made from adjacent combinations, meaning that only a single control variable is allowed to switch gears once while the remaining control variables remain unchanged. This also includes control variables that maintain their full value, ultimately forming a candidate set of no more than 6 groups, significantly reducing the computational load for optimization. The normalized value of the second candidate control variable is denoted as... .

[0139] 3.3 Analytical solution for optimal action time, specifically:

[0140] (1) Assuming a single control cycle Inside, The duration of action is Second candidate control quantity The duration of action is ;

[0141] (2) Combining the normalized state prediction model, we obtain Time-normalized state prediction:

[0142]

[0143] In the formula, : The normalized predicted value of the system state at any given time consists of the normalized predicted results of 8 core state variables and is the prediction target for control quantity optimization. : The normalized value of the system state at time t is The basic values ​​for time-state prediction; The system control cycle is set to 5~30s based on the thermal inertia characteristics of the chemical process, adapting to the normal interruption cycle of industrial distributed control systems (DCS) / programmable logic controllers (PLCs). The normalized system state matrix is ​​obtained by normalizing the original system state matrix and is adapted for operations on normalized state variables. The normalized perturbation matrix is ​​obtained by normalizing the original perturbation matrix. Correlation normalization can measure perturbations and state changes. : The time-normalized measurable perturbation vectors are, in order, the inlet flue gas flow rate and the inlet flue gas carbon dioxide (CO2). The normalized values ​​of hydrogen concentration and hydrogen concentration from hydrogen-rich sources; The optimal action time of the first optimal control variable is within a single control cycle. Execution time; The normalized control input matrix is ​​obtained by normalizing the original control input matrix, and it correlates the normalized control quantity with the state change. The normalized value of the first optimal control quantity is the value that makes the cost function... The minimum combination of control variables is normalized. The normalized value of the second candidate control quantity is the normalized result of the combination of candidate control quantities under the single-step adjustment principle.

[0144] (3) Substitute the above equation into the cost function and solve for the cost function. The analytical solution for the optimal action time is obtained:

[0145]

[0146] In the formula, Cost function For optimal action time The partial derivatives of can be obtained by solving for solutions that equal zero. The optimal value; Normalized state tracking error, which is the normalized integrated state reference vector and... The difference between the normalized state values ​​at any given time; The vector dot product operation is used to calculate the dot product of two vectors, and the result is a scalar. The 2-norm square operation is used to calculate the square of the Euclidean distance of a vector, simplifying the calculation and ensuring non-negativity.

[0147] (4) Perform amplitude limiting to meet the requirements. This ensures the rationality of industrial implementation.

[0148] : The limiting constraint ensures that the optimal action time is within a single control cycle, which conforms to industrial execution logic.

[0149] 3.4 Determine the optimal combination of dual-mode control variables, specifically as follows:

[0150] Iterate through all candidate second control variables, calculate the optimal action time and cost function value for each combination, and select... The group with the smallest value ( As the final control output; within the current control cycle, the output is first... effect Time, then output The remaining time is used to achieve smooth, segmented control.

[0151] The second optimal control variable is the one that, after iterating through the second candidate control variables, makes the cost function... The combination of control quantities with the smallest value; The optimal dual-mode control quantity combination is the segmented control quantity combination that is ultimately output by this invention. The cost function without weighting factors is used as a criterion for evaluating the quality of control quantity combinations. A smaller value indicates better control performance; Remaining time: deducted within a single control cycle. The duration afterwards, i.e. ,for Execution time.

[0152] 4. Rolling optimization and feedback correction module 164

[0153] To eliminate prediction errors caused by model mismatch and external disturbances, and to ensure the long-term stable operation of the control system, this invention sets up a rolling optimization and feedback correction stage.

[0154] Furthermore, the rolling optimization and feedback correction module 164 is specifically used for: obtaining the real-time state estimate of the system through the full-order closed-loop state observer in each control cycle, and performing feedback correction on the state estimate in combination with the field measured output value to update the initial value of the state prediction model; based on the corrected initial state value, triggering the multi-objective reference value conversion and normalization module 162 and the dual-mode control quantity optimization and action time calculation module 163 to re-execute the corresponding operations to output the updated control quantity; cyclically executing the feedback correction and rolling optimization process to realize the real-time closed-loop optimization control of the system. The specific steps are as follows:

[0155] 4.1 Design a full-order closed-loop state observer

[0156] To address the issues of some system state variables being unmeasurable online and parameters drifting with operating conditions, this invention designs a closed-loop full-order state observer to achieve accurate estimation of all state variables while suppressing parameter disturbances and measurement noise.

[0157] Furthermore, the mathematical model of the full-order closed-loop state observer includes the observer state equation and the observer output equation. The observer state equation describes the relationship between the state observation estimate and time, while the observer output equation establishes the mapping relationship between the state observation estimate and the output observation estimate. The observer state equation contains the observer gain matrix, which is obtained by solving the pole placement method or the Riccati equation and is used to adjust the convergence speed and stability of the observer. Specifically:

[0158] (1) Establish the mathematical model of the full-order closed-loop state observer, with the following equations:

[0159] ;

[0160] In the formula, : Observed estimates of the state vector; The time derivative of the state vector observation estimate reflects the rate of change of the observation estimate over time. : The observed estimate of the output vector; The control output vectors are, in order, CO2 capture rate, system unit energy consumption, and methanol single-pass yield; : Observer gain matrix.

[0161] (2) Observer gain matrix The observer poles are determined by solving the pole placement method or the Riccati equation, and are individually tuned according to the convergence rate requirements of each state variable (temperature, concentration, load rate, etc.). The observer poles are placed on the negative real axis to the left of the poles of the system state matrix to ensure asymptotic stability and fast convergence of the observer, which is suitable for the control requirements of chemical multivariable systems.

[0162] 4.2 Scrolling Optimization and Feedback Correction

[0163] (1) Feedback correction: In each control cycle, the real-time state estimate of the system is obtained through the full-order observer, and the state estimate is corrected by feedback in combination with the field measured output value to update the initial value of the state prediction model;

[0164] (2) Rolling optimization: Based on the corrected initial state value, the multi-objective reference value conversion and normalization process and the dual-mode control quantity optimization and action time calculation are re-executed to complete the multi-objective reference value conversion, dual-mode control quantity optimization and action time calculation, and output the new control quantity;

[0165] (3) Repeatedly execute the above feedback correction and rolling optimization steps to realize real-time closed-loop optimization control of the system and adapt to the wide operating condition fluctuation characteristics of the chemical process.

[0166] As a preferred embodiment of the present invention, the AI ​​intelligent control unit 160 further includes an LSTM deep learning absorbent ratio optimization module 165 that works in conjunction with the improved model predictive control algorithm.

[0167] like Figure 4 The diagram shown is a schematic of the network structure of the LSTM deep learning absorbent ratio optimization module in this embodiment of the invention. This LSTM deep learning absorbent ratio optimization module 165 employs a three-layer LSTM network structure. Its input parameters include inlet flue gas CO2 concentration, flue gas flow rate, flue gas temperature, absorbent circulation rate, and rich liquid CO2 loading rate. The output parameter is the optimal replenishment ratio of N-methylcyclohexylamine (MCS), 2-amino-2-methyl-1-propanol (AMP), and n-octanol. The training dataset for this module comes from 12 consecutive months of operational data from a thermal power plant and laboratory absorbent performance test data (valid data volume ≥ 100,000 sets). Model training and iterative optimization are completed through supervised learning.

[0168] In actual operation, the module's real-time output of the optimal replenishment ratio command serves two purposes. First, it acts as a feedforward signal, working in conjunction with the output of the improved model predictive control algorithm on the actuators of the absorbent replenishment system. Second, the effects of absorbent replenishment (such as changes in the CO2 loading rate of the rich liquid) serve as feedback, updating the system state through the AI ​​intelligent control unit 160 for optimization of model predictive control in the next cycle. This feedforward-feedback approach achieves true synergy between dynamic optimization of absorbent performance and global system operation optimization, ensuring the long-term stability of absorbent phase separation performance and CO2 absorption efficiency. Specific implementation examples:

[0170] Example 1: Formulation and Performance Verification of Low-Energy Liquid-Liquid Separation Non-Aqueous Phase Change Absorbent

[0171] S1. Preparation of raw materials

[0172] Industrial-grade raw materials with a purity of ≥99% were selected, specifically: N-methylcyclohexylamine (MCS), 2-amino-2-methyl-1-propanol (AMP), and n-octanol.

[0173] S2, Specific preparation process

[0174] This embodiment 1 provides two specific formulas and uniform preparation conditions, wherein formula 1 is the preferred formula and formula 2 is the comparative formula.

[0175] S21, Standardized Preparation Conditions

[0176] Preheat the reactor to 30°C, add N-methylcyclohexylamine (MCS) and 2-amino-2-methyl-1-propanol (AMP) in sequence according to the formula, turn on the mechanical stirrer at a speed of 300 r / min, and stir for 30 min until 2-amino-2-methyl-1-propanol (AMP) is completely dissolved;

[0177] While maintaining a constant stirring speed, slowly add n-octanol dropwise over 15 minutes. After the addition is complete, continue stirring for 60 minutes.

[0178] After standing for 20 minutes, a homogeneous and transparent non-aqueous phase change absorbent is obtained. It should be sealed and stored away from light.

[0179] S22, Formula 1 (Preferred Formula)

[0180] Mass ratio: N-methylcyclohexylamine (MCS): 2-amino-2-methyl-1-propanol (AMP): n-octanol = 50:15:35; Prepare the finished absorbent according to the above preparation conditions.

[0181] S23, Formula 2 (Comparative Formula)

[0182] Mass ratio: N-methylcyclohexylamine (MCS): 2-amino-2-methyl-1-propanol (AMP): n-octanol = 45:20:35; Prepare the finished absorbent according to the above preparation conditions.

[0183] S3, Performance Verification Test

[0184] Under typical operating conditions of flue gas in thermal power plants (absorption temperature 40℃, CO2 partial pressure 15kPa, flue gas flow rate 1000m³ / h), the performance of two absorbents and the traditional monoethanolamine (MEA) absorbent was compared and tested. The test results are shown in Table 1.

[0185] Table 1 Results of Absorbent Performance Verification

[0186]

[0187] S4. Verification Conclusion

[0188] The non-aqueous phase change absorbent of this invention has a CO2 loading capacity that is much higher than that of traditional monoethanolamine (MEA) absorbents, reduces regeneration energy consumption by more than 50%, has excellent phase separation performance, and has a CO2 enrichment rate of ≥95% in the rich liquid phase, which fully meets the process requirements for CO2 capture in flue gas of thermal power plants.

[0189] Example 2: Acquisition and Verification of System State Space Matrix

[0190] S1. Construct a simplified state-space model

[0191] To address the core control objectives of the combined production system, a simplified state-space model is constructed using three core state variables, two core control inputs, and one measurable disturbance, thereby reducing modeling complexity. The specific variable definitions are as follows:

[0192] State vector : (Temperature of the regeneration tower, °C) (CO2 capture rate, %) (Methanol single-pass yield, %)

[0193] Control input vector : (Frequency of rich liquid feed pump, Hz) (Heat pump compressor frequency, Hz);

[0194] Measurable perturbation vector : (Inlet flue gas CO2 concentration, );

[0195] Output vector All state variables are measurable / observable.

[0196] S2, Matrix Acquisition Method

[0197] The system matrix is ​​obtained by using the recursive least squares (RLS) method combined with mechanism correction. The specific steps are as follows:

[0198] S21. Data Acquisition: Collect 1,000 sets of operational data from the thermal power plant site, with a sampling period of 30 seconds, covering the full operating range of flue gas CO2 concentration from 8% to 15% and system load from 50% to 100%.

[0199] S22. Initial Matrix Construction: Based on heat balance and mass transfer reaction kinetics, determine the system state matrix. Control input matrix Perturbation matrix Output matrix The initial structure;

[0200] S23. Parameter identification: The recursive least squares (RLS) method is used to identify the parameters of the initial matrix. The number of iterations is 1000 and the convergence threshold is 1e-6.

[0201] S24. Mechanism Correction: The identification results were corrected based on laboratory test data to obtain the final matrix under typical operating conditions (flue gas CO2 concentration 12%, system load 80%).

[0202] ; ;

[0203] ; .

[0204] S3, Model Validation

[0205] Substitute the above matrix into the discretized state equation proposed in this invention, input the field-measured control variables and disturbance data, obtain the predicted values ​​of the state variables, compare the predicted values ​​with the field-measured values, calculate the root mean square error (RMSE), and the verification results are shown in Table 2.

[0206] Table 2 Validation results of the simplified state-space model

[0207]

[0208] S4. Verification Conclusion

[0209] The state-space model obtained by the hybrid modeling method of system identification and mechanism correction adopted in this invention has a root mean square error (RMSE) of ≤1 between the predicted and measured values, and the prediction accuracy meets the actual requirements of industrial control.

[0210] Example 3: Pole placement design of a full-order closed-loop state observer

[0211] This embodiment uses the pole placement method as an example to describe in detail the design of a full-order closed-loop state observer. Those skilled in the art should understand that other known methods in the art, such as solving the Riccati equation, can also be used to obtain the observer gain matrix. Taking the simplified state-space model of Embodiment 2 as an example, this invention uses the pole placement method to design the observer gain matrix. The specific implementation steps are as follows:

[0212] S1. Determine the desired poles of the observer.

[0213] Based on the convergence speed requirements of the chemical process, the time for the state observation value to converge to the true value is set to ≤ 5 control cycles (control cycle in this embodiment). =30s, i.e., convergence time ≤150s); place the desired pole of the observer on the negative real axis to the left of the system pole, and select the desired pole. for:

[0214] ;

[0215] The actual convergence time corresponding to the expected pole is approximately 120 seconds, which meets the preset convergence speed requirement.

[0216] S2. Deriving the characteristic equation

[0217] S21. The characteristic equation of the observer is:

[0218] ;

[0219] In the formula, each symbol is defined as follows: Determinant operation: used to solve for the determinant value of a matrix, and is the core operation for constructing characteristic equations; Complex variables are the core variables in complex frequency domain analysis, used to describe the dynamic characteristics of a system. : Identity matrix, and system state matrix Identity matrices of the same dimension ensure dimension matching for matrix operations; The system state matrix is ​​the core state matrix of the flue gas CO2 capture-methanol cogeneration system in a thermal power plant, reflecting the changing relationships of the internal state of the system. The observer gain matrix, obtained by the pole placement method or the Riccati equation, is used to adjust the convergence characteristics of the observer. Output matrix: Establishes a linear mapping relationship between the system state vector and the output vector; The closed-loop state matrix of the observer determines the dynamic performance and convergence of the observer; : Multiplication operation, for Continuous multiplication of terms from 1 to 3 is used to adapt to the operation of a simplified 3D state-space model. The observer's first The desired poles are preset observer poles, located on the negative real axis to the left of the system poles, to ensure asymptotic stability of the observer; The first-order factor of the observer characteristic polynomial is obtained by subtracting the complex variable from the corresponding expected pole. The characteristic polynomial of the observer, whose roots are the actual poles of the observer, determines the convergence speed and stability of the observer.

[0220] S22. Substituting the desired pole into the right side and expanding, we obtain the characteristic polynomial:

[0221] ;

[0222] S3. Solve for the observer gain matrix.

[0223] S31. Define the observer gain matrix. Example 2 , Substituting the matrix into the left side of the observer's characteristic equation and expanding it yields the characteristic polynomial containing the gain matrix elements;

[0224] By comparing the coefficients of the characteristic polynomials on both sides, and solving the system of cubic equations, the observer gain matrix can be obtained.

[0225] ;

[0226] S4. Observer Performance Verification

[0227] The above gain matrix Substituting the full-order closed-loop state observer model designed in this invention, and inputting output data containing measurement noise (noise amplitude ±0.5%), the convergence and robustness of the observer are tested. The test results show that the state estimate of the observer converges to the true value within 120s, and the steady-state error is ≤0.3%, which can effectively suppress measurement noise and meet the robustness requirements of industrial control.

[0228] Example 4: Training and Validation of the LSTM Deep Learning Absorber Ratio Optimization Module

[0229] S1, Network Structure Details

[0230] The LSTM absorbent ratio optimization module of this invention adopts a 3-layer LSTM network structure, and the specific parameters are as follows:

[0231] (1) Input layer: 5 feature dimensions, namely, inlet flue gas CO2 concentration, flue gas flow rate, flue gas temperature, absorbent circulation rate, and rich liquid CO2 loading rate;

[0232] (2) Hidden layers: 64 LSTM neurons in the first layer, 32 LSTM neurons in the second layer, and 16 LSTM neurons in the third layer;

[0233] (3) Output layer: Fully connected layer with output dimension 3, which are the mass percentages of N-methylcyclohexylamine (MCS), 2-amino-2-methyl-1-propanol (AMP), and n-octanol respectively.

[0234] S2, Model Training Hyperparameters

[0235] The LSTM model of this invention is trained using supervised learning, and the specific training hyperparameters are as follows:

[0236] S21. Dataset: 12 months of on-site operation data from thermal power plants + laboratory absorbent performance test data, totaling 100,000 sets; divided into training set, validation set, and test set in a 7:2:1 ratio.

[0237] S22. Data preprocessing: All feature data are normalized to mean, missing values ​​are filled by linear interpolation, and outliers are removed by the 3σ criterion.

[0238] S23, Optimizer: Adaptive Moment Estimator (Adam) Optimizer;

[0239] S24, Learning Rate: Initial learning rate 0.001, decaying by 10% every 10 training epochs;

[0240] S25. Batch size: 64;

[0241] S26, Number of training epochs: 50;

[0242] S27. Loss Function: Mean Squared Error (MSE) Loss Function;

[0243] S28, Regularization: L2 regularization, weight decay coefficient 0.0001;

[0244] S29. Early stopping strategy: If the validation set loss does not decrease for 5 consecutive training epochs, stop training to avoid model overfitting.

[0245] S3, Model Training Results

[0246] After the model is trained, the loss values ​​for the training set, validation set, and test set are calculated, and the results are shown in Table 3.

[0247] Table 3 LSTM Model Training Loss Values

[0248]

[0249] S4. On-site verification

[0250] The trained LSTM model was deployed to the AI ​​intelligent control unit 160 at the thermal power plant site, and the optimal replenishment ratio of absorbent was output in real time. The field operation verification results showed that the CO2 loading of absorbent fluctuated by ≤0.05mol / mol amine, the volume ratio of rich liquid phase after phase separation fluctuated by ≤2%, and the overall performance of absorbent remained stable, verifying the effectiveness and practicality of the LSTM deep learning absorbent ratio optimization module of this invention.

[0251] In summary, the above embodiments supplement the specific operable details of the technical solution of the present invention from four core dimensions: the preparation process and performance verification of the non-aqueous phase change absorbent, the acquisition and verification of the system state space matrix, the pole configuration design of the full-order closed-loop state observer, and the training and verification of the LSTM deep learning absorbent ratio optimization module 165. They also verify the reproducibility, industrial applicability, and engineering feasibility of the technical solution of the present invention.

[0252] In practical applications, those skilled in the art can make adaptive adjustments to the proportioning parameters, model matrix, control cycle, etc. of the present invention according to the specific operating conditions of the thermal power plant (such as flue gas composition, equipment parameters, load range), and all adjusted technical solutions fall within the protection scope of the present invention.

[0253] It should also be noted that, in the embodiments of this application, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0254] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of this application may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown in this application, but is to be accorded the widest scope consistent with the principles and novel features disclosed in the embodiments of this application.

Claims

1. A low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants, characterized in that, It includes a CO2 absorption unit, a liquid-liquid phase separation unit, a rich liquid regeneration unit, a CO2 catalytic methanol synthesis unit connected in sequence, a heat pump waste heat recovery unit connected to the CO2 catalytic methanol synthesis unit and the rich liquid regeneration unit respectively, and an AI intelligent control unit electrically connected to each of the aforementioned units respectively. The CO2 absorption unit is used to absorb CO2 in the flue gas of thermal power plants using a non-aqueous phase change absorbent to form a rich absorbent liquid. The liquid-liquid phase separation unit is used to spontaneously separate the rich absorbent into a lean liquid phase and a rich liquid phase, and return the lean liquid phase to the CO2 absorption unit. The rich liquid regeneration unit is used to heat and regenerate the rich liquid phase to release CO2 gas. The CO2 catalytic methanol synthesis unit is used to synthesize methanol from CO2 gas and external hydrogen-rich gas. The heat pump waste heat recovery unit is used to recover the heat of methanol synthesis reaction and then supply it to the rich liquid regeneration unit after upgrading. The AI ​​intelligent control unit is used to collect the process parameters of the aforementioned units and to perform coordinated optimization control of the actuators of each unit based on the improved model predictive control algorithm.

2. The low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants according to claim 1, characterized in that, The non-aqueous phase change absorbent is composed of a lipophilic tertiary amine, a hindered amine activator, and a hydrophobic organic solvent. The lipophilic tertiary amine is N-methylcyclohexylamine, accounting for 40% to 60% of the absorbent by mass; the hindered amine activator is 2-amino-2-methyl-1-propanol, accounting for 10% to 20% of the absorbent by mass. The hydrophobic organic solvent is n-octanol, which accounts for 20% to 40% of the mass of the absorbent.

3. The low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants according to claim 2, characterized in that, The non-aqueous phase change absorbent is a homogeneous solution at 20~40℃. After absorbing CO2, liquid-liquid phase separation occurs to form a rich liquid phase and a poor liquid phase. The volume of the rich liquid phase accounts for 30%~40% of the total volume of the absorbent, and it enriches more than 95% of the CO2 loading in the system.

4. The low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants according to claim 1, characterized in that, The AI ​​intelligent control unit includes a state observation and delay compensation module, a multi-objective reference value conversion and normalization module, a dual-mode control quantity optimization and action time calculation module, and a rolling optimization and feedback correction module, all designed to implement the improved model predictive control algorithm. The state observation and delay compensation module is used to establish a system state space prediction model and compensate for system delay. The system state space prediction model includes a state vector, a control input vector, a measurable disturbance vector, and an output vector. The multi-objective reference value conversion and normalization module is used to convert multiple control objectives into a single comprehensive state reference vector and to normalize all state variables. The dual-mode control quantity optimization and action time calculation module is used to determine the first optimal control quantity and the second optimal control quantity within a control cycle, and to calculate the optimal action time of the first optimal control quantity. The rolling optimization and feedback correction module is used to repeatedly execute the optimization process in each control cycle based on the real-time state estimate.

5. The low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants according to claim 4, characterized in that, The state observation and delay compensation module is specifically used for: A hybrid modeling approach combining system identification and mechanism correction is used to establish continuous domain state-space equations, which serve as the basis for the system state-space prediction model. The Heinrich prediction-correction method is used to discretize the continuous domain state space equations and achieve one-step delay compensation. The continuous domain state space equation includes a state equation and an output equation. The state equation is used to describe the relationship between the state vector and time, and the output equation is used to establish the mapping relationship between the state vector and the output vector. The state vector includes the CO2 concentration in the flue gas at the absorber outlet, the CO2 loading rate in the rich liquid, the temperature of the regeneration tower bottom, the bed temperature of the methanol synthesis reactor, the heating capacity of the heat pump system, the effective amine concentration in the absorbent, the system heat loss coefficient, and the relative activity of the catalyst. The control input vector includes the opening degree of the absorbent feed regulating valve, the frequency of the rich liquid feed pump, the opening degree of the reboiler steam regulating valve, the frequency of the heat pump compressor, and the opening degree of the syngas feed regulating valve. The measurable disturbance vector includes the inlet flue gas flow rate, the inlet flue gas CO2 concentration, and the hydrogen concentration of the hydrogen-rich source. The output vector includes CO2 capture rate, system unit energy consumption, and methanol single-pass yield. The coefficient matrices in the state equation and output equation are obtained by identifying data from the on-site operation of thermal power plants and correcting them in conjunction with the chemical process mechanism.

6. The low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants according to claim 4, characterized in that, The multi-objective reference value conversion and normalization module is specifically used for: The state variables are normalized by scaling all state variables to a predetermined numerical range through a linear transformation to eliminate dimensional differences. The process constraint boundaries of the system are determined, including the lower limit of CO2 capture rate, the upper limit of system unit energy consumption, and the lower limit of methanol single-pass yield. Based on the optimal parameters of the system under rated operating conditions, and combined with the process constraint boundaries, a normalized comprehensive state reference vector composed of the normalized optimal reference values ​​of each state variable is constructed as a unified reference benchmark for multi-objective control.

7. The low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants according to claim 4, characterized in that, The dual-mode control quantity optimization and action time calculation module is specifically used for: Using the normalized composite state reference vector at a specified future time as the target, a cost function is constructed to measure the degree of deviation between the normalized predicted value of the system state and the normalized composite state reference vector. By traversing the candidate control variables and calculating the cost function value, the combination of control variables that minimizes the cost function value is selected as the first optimal control variable. The candidate set of the second optimal control quantity is determined based on the single-step adjustment principle, which means that only a single control quantity is allowed to switch gears once relative to the first optimal control quantity, while the other control quantities remain unchanged. Assuming that the action time of the first optimal control variable within a single control cycle is the optimal action time to be solved, and the action time of the second candidate control variable is the remaining time of the control cycle, the normalized state prediction value at the specified future time is derived by combining the normalized state prediction model. Substitute the normalized state prediction value at the specified future time into the cost function, solve for the partial derivative of the cost function with respect to the optimal action time to obtain the optimal action time analytically, and then perform amplitude limiting processing on the optimal action time to make its value within a single control cycle. Iterate through all the second candidate control variables, calculate the corresponding optimal action time and cost function value, select the second candidate control variable that minimizes the cost function value as the second optimal control variable, and use the first optimal control variable and the second optimal control variable as the final dual-mode control variable combination output. Within the current control cycle, the first optimal control quantity is applied and maintained for the optimal duration, and then the second optimal control quantity is applied and maintained for the remaining time of the control cycle.

8. The low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants according to claim 4, characterized in that, The rolling optimization and feedback correction module is specifically used for: Within each control cycle, the real-time state estimate of the system is obtained through the full-order closed-loop state observer, and the state estimate is fed back and corrected in combination with the field measured output value to update the initial value of the state prediction model. Based on the corrected initial state value, the multi-target reference value conversion and normalization module and the dual-mode control quantity optimization and action time calculation module are triggered to re-execute the corresponding operations to output the updated control quantity; The feedback correction and rolling optimization process is executed cyclically to achieve real-time closed-loop optimization control of the system.

9. The low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants according to claim 8, characterized in that, The mathematical model of the full-order closed-loop state observer includes the observer state equation and the observer output equation. The observer state equation is used to describe the relationship between the state observation estimate and time, and the observer output equation is used to establish the mapping relationship between the state observation estimate and the output observation estimate. The observer state equation includes an observer gain matrix, which is obtained by solving the pole placement method or the Riccati equation and is used to adjust the convergence speed and stability of the observer.

10. The low-energy-consumption phase change CO2 capture-methanol co-production process system for flue gas from thermal power plants according to claim 4, characterized in that, The AI ​​intelligent control unit also includes an LSTM deep learning absorbent ratio optimization module that works in conjunction with the improved model predictive control algorithm. The LSTM deep learning absorbent ratio optimization module adopts a three-layer LSTM network structure. Its input parameters include inlet flue gas CO2 concentration, flue gas flow rate, flue gas temperature, absorbent circulation volume and rich liquid CO2 loading rate. The output parameter is the optimal replenishment ratio of N-methylcyclohexylamine, 2-amino-2-methyl-1-propanol and n-octanol. The training dataset for the LSTM deep learning absorbent ratio optimization module comes from on-site operation data of thermal power plants and laboratory absorbent performance test data. The model training and iterative optimization are completed through supervised learning. The LSTM deep learning absorbent ratio optimization module outputs the optimal replenishment ratio command in real time to control the operation of the absorbent replenishment system.

Images