A system for producing liquid food grade carbon dioxide

Abstract

The application relates to the technical field of carbon dioxide production, in particular to a production system for preparing liquid food-grade carbon dioxide, which comprises a raw material processing module, which performs washing and separation operations on carbon dioxide crude gas and collects and outputs physical sensing signals in the washing and separation operation process; and a component prediction module, which receives the physical sensing signals, converts the physical sensing signals into a state matrix representing a heat and mass transfer relationship, performs feature extraction on the state matrix through a graph neural network with a thermodynamic equation constraint, and outputs a gas component prediction instruction and the state matrix. In the application, the thermodynamic mechanism model and the graph neural network are combined to constrainively predict impurity components in the carbon dioxide, and the application can still maintain a relatively stable prediction result under the condition that a sensor fluctuates or data is incomplete, so that the reliability and continuity of overall control of the system are improved.

Description

Technical Field

[0001] This invention relates to the field of carbon dioxide production technology, and in particular to a system for producing liquid food-grade carbon dioxide. Background Technology

[0002] In the production of food-grade carbon dioxide, it is usually necessary to monitor and control impurities such as oxygen and volatile organic compounds in the raw gas in real time to ensure the stable operation of subsequent purification and liquefaction processes. Traditional methods often rely on online analytical instruments or simple data models based on experience for component estimation. However, due to the measurement lag, noise interference, and asynchronous sampling of sensors in industrial settings, direct measurement or single data-driven methods are difficult to reflect changes in gas composition in a timely and accurate manner. Therefore, soft measurement methods that combine mechanistic models and data-driven algorithms are gradually being introduced to improve the predictive capabilities under complex operating conditions.

[0003] However, existing technologies mostly use pure data-driven models for component prediction, which do not adequately consider thermodynamic constraints and process mechanisms. When sensor data fluctuates or becomes abnormal, prediction deviations are likely to occur, which in turn affects the stability of system control. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a system for producing liquid food-grade carbon dioxide, aiming to improve the problem that existing technologies often use purely data-driven models for component prediction, which do not adequately consider thermodynamic constraints and process mechanisms.

[0005] This invention provides the following technical solution: a system for producing liquid food-grade carbon dioxide, comprising:

[0006] The raw material processing module performs washing and separation operations on the crude carbon dioxide gas, and collects and outputs physical sensor signals during the washing and separation process.

[0007] The composition prediction module receives the physical sensing signal, converts the physical sensing signal into a state matrix representing the heat and mass transfer relationship, extracts features from the state matrix through a graph neural network constrained by thermodynamic equations, and outputs a gas composition prediction command and the state matrix.

[0008] The purification control module receives the gas composition prediction command and the state matrix, adjusts the reactant injection ratio of the catalytic oxidation reactor according to the gas composition prediction command, solves the dynamic concentration front parameter corresponding to the state matrix through the adsorption kinetic equation, controls the molecular sieve regeneration according to the dynamic concentration front parameter, and outputs the purified gas flow and purification state parameters.

[0009] The energy efficiency optimization module receives the purification state parameters, obtains electricity price information and meteorological information, inputs the purification state parameters, the electricity price information and the meteorological information into a reinforcement learning model with carbon dioxide gas-liquid phase boundary as action constraint for processing, and outputs control strategy instructions;

[0010] The compression liquefaction module receives the purified airflow and adjusts the interstage pressure parameters of the multi-stage compressor and the condensing temperature parameters of the refrigeration unit according to the control strategy command, thereby converting the purified airflow into liquid carbon dioxide.

[0011] Preferably, in the raw material processing module, the step of collecting and outputting the physical sensor signals during the washing and separation operation specifically includes the following steps:

[0012] Temperature, pressure, and flow signals were collected at the inlet and outlet of the scrubbing tower and at the exhaust end of the gas-liquid separator, respectively.

[0013] The temperature sensing signal, the pressure sensing signal, and the flow sensing signal are timestamped according to a preset sampling frequency, packaged and encapsulated into multidimensional time series data, and the multidimensional time series data is output as a physical sensing signal.

[0014] Preferably, in the composition prediction module, the step of extracting features from the state matrix using a graph neural network constrained by thermodynamic equations and outputting a gas composition prediction command and the state matrix specifically includes the following steps:

[0015] Construct an initial graph network structure that includes the process topology relationships of the production system;

[0016] The state matrix is ​​mapped to the node attributes of the initial graph network structure. During the propagation between network layers, the gas-liquid equilibrium thermodynamic equation is superimposed as a loss function penalty term for feature extraction, resulting in the converged node hidden layer feature vector.

[0017] The node hidden layer feature vector is mapped to the impurity concentration prediction value through a fully connected layer. A gas composition prediction instruction is generated based on the impurity concentration prediction value, and the extracted state matrix and the gas composition prediction instruction are output.

[0018] Preferably, the feature extraction using the superimposed gas-liquid equilibrium thermodynamic equation as a loss function penalty term specifically includes the following steps:

[0019] Construct the Penn-Robinson state equation that includes the eccentricity factor and critical property parameters;

[0020] Substitute the intermediate output tensor during the network forward propagation process into the Penn-Robinson equation of state to calculate the fugacity difference between the theoretical thermodynamic phase state and the actual measured phase state;

[0021] The fugacity difference is converted into a mean squared error term and added to the basic loss function of the graph neural network. The inter-layer weight parameters of the graph neural network are then updated using the backpropagation algorithm.

[0022] Preferably, in the purification control module, the step of solving the dynamic concentration front parameter corresponding to the state matrix through the adsorption kinetic equation and controlling the molecular sieve regeneration based on the dynamic concentration front parameter specifically includes the following steps:

[0023] Extract the intake air moisture characteristics from the state matrix and substitute the intake air moisture characteristics into the adsorption mass transfer equation in partial differential form;

[0024] Calculate the spatial distribution gradient of moisture concentration inside the molecular sieve, and calculate the physical spatial coordinates of the spatial distribution gradient reaching the set value as dynamic concentration front parameters;

[0025] Determine whether the dynamic concentration front parameter has moved to the set boundary position of the molecular sieve bed. If it reaches the set boundary position, trigger the bed switching command and start the heating and regeneration program of the corresponding molecular sieve.

[0026] Preferably, substituting the intake air moisture characteristics into the partial differential form of the adsorption mass transfer equation specifically includes the following steps:

[0027] Establish a linear driving force mass transfer model equation that includes gas phase mass conservation parameters and solid phase adsorption rate parameters;

[0028] The physical axial length of the molecular sieve bed is discretized into a set number of grid nodes, and the inlet moisture characteristics are used as initial boundary conditions input into the linear driving force mass transfer model equation.

[0029] The discretized linear driving force mass transfer model equation is solved using the upwind difference algorithm, and the adsorbate distribution data of each grid node is output and the spatial distribution gradient is calculated.

[0030] Preferably, in the energy efficiency optimization module, the step of inputting the purification state parameters, the electricity price information, and the meteorological information into a reinforcement learning model with carbon dioxide gas-liquid phase boundary as action constraint for processing, and outputting control strategy instructions specifically includes the following steps:

[0031] The purification state parameters, the electricity price information, and the meteorological information are concatenated to form the state space vector of the reinforcement learning model;

[0032] Based on the physical property parameters of carbon dioxide, a gas-liquid phase envelope curve equation between the triple point and the critical point is constructed, and the gas-liquid phase envelope curve equation is loaded into the action space of the reinforcement learning model as a penalty boundary mask.

[0033] Within the action domain where the penalty boundary mask is not triggered, the strategy iteration is performed using the system's per-ton electricity cost as the reward function, generating and outputting control strategy instructions.

[0034] Preferably, loading the gas-liquid phase envelope curve equation as a penalty boundary mask into the action space of the reinforcement learning model specifically includes the following steps:

[0035] Extract the initial continuous action vector generated by the output layer of the reinforcement learning model;

[0036] Calculate the Euclidean geometric distance from the thermodynamic coordinates pointed to by the initial continuous motion vector to the boundary of the gas-liquid phase envelope curve equation;

[0037] A logarithmic barrier function based on the Euclidean geometric distance is constructed. The logarithmic barrier function is used to truncate and remap the initial continuous action vector in the preset threshold region to generate the actual action vector.

[0038] Preferably, in the compression liquefaction module, adjusting the interstage pressure parameters of the multi-stage compressor and the condensing temperature parameters of the refrigeration unit according to the control strategy command specifically includes the following steps:

[0039] The control strategy instructions are analyzed to separate the pressure distribution matrix for the multi-stage compressor and the temperature setpoint for the refrigeration unit.

[0040] Based on the pressure distribution matrix, the opening degree of the inlet guide vane of the multi-stage compressor and the opening degree of the interstage cooler water valve are adjusted to output the interstage pressure parameters of the multi-stage compressor.

[0041] Based on the temperature setpoint, the refrigerant circulation flow rate and expansion valve opening in the refrigeration unit are adjusted to cool the purified airflow to below the gas-liquid phase change temperature.

[0042] The present invention has the following beneficial effects:

[0043] 1. In this invention, by combining a thermodynamic mechanism model with a graph neural network, the impurity components in carbon dioxide are predicted under constraints. Even when there are fluctuations in the sensor or incomplete data, the prediction results can still be kept relatively stable, thereby improving the reliability and continuity of the overall system control.

[0044] 2. This invention uses an adsorption mass transfer kinetics model to numerically solve the internal water migration process of molecular sieves and combines dynamic concentration front parameters to determine the timing of regeneration, making the regeneration process more in line with actual working conditions, reducing energy consumption and ineffective regeneration to a certain extent.

[0045] 3. This invention introduces a reinforcement learning method with safety boundary constraints to optimize the energy efficiency of the compression and liquefaction process. Under the premise of avoiding the operating state from approaching the risk region of phase change, the load of the equipment is dynamically adjusted, which helps to achieve a relatively balanced energy consumption control and operational stability. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the architecture of a production system for preparing liquid food-grade carbon dioxide proposed in this invention. Detailed Implementation

[0047] The technical solutions in 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.

[0048] In a first embodiment of the present invention, the present invention provides a system for producing liquid food-grade carbon dioxide, such as... Figure 1 As shown, it includes:

[0049] The raw material processing module performs washing and separation operations on the crude carbon dioxide gas, and collects and outputs physical sensor signals during the washing and separation process.

[0050] Furthermore, in the raw material processing module, the acquisition and output of physical sensor signals during the washing and separation processes specifically includes the following steps:

[0051] Temperature, pressure, and flow signals were collected at the inlet and outlet of the scrubbing tower and at the exhaust end of the gas-liquid separator, respectively.

[0052] Temperature, pressure, and flow sensing signals are timestamped according to a preset sampling frequency, packaged and encapsulated into multidimensional time series data, and then output as physical sensing signals.

[0053] Specifically, the raw material processing module is mainly used to remove water-soluble impurities and free liquid water from the carbon dioxide crude gas and to collect basic process parameters. The carbon dioxide crude gas is first introduced into the bottom of the scrubbing tower, where it comes into counter-current contact with the scrubbing water sprayed from the top of the tower for scrubbing and cooling. The gas discharged from the top of the scrubbing tower is then sent to a gas-liquid separator for physical gas-liquid separation. In this process, the system uses transmitters deployed on the pipelines to perform data acquisition. Specifically, temperature transmitters, pressure transmitters, and mass flow meters are used at the crude gas inlet of the scrubbing tower, the scrubbing gas outlet of the scrubbing tower, and the exhaust end of the gas-liquid separator to continuously collect temperature, pressure, and flow sensor signals at each process node. Due to differences in communication protocols between different field instruments, the original acquired sensor signals are presented as time-asynchronous discrete data streams.

[0054] To meet the timing consistency requirements of input features in downstream neural network models, after receiving the raw asynchronous sensor data, the system executes a clock synchronization and resampling interpolation algorithm based on the microprocessor to align the temperature, pressure, and flow sensing signals according to a preset sampling frequency. The system first obtains the alignment start timestamp during system operation. And set the target resampling period as Construct the target timestamp sequence In the formula Integer step size index greater than or equal to zero. For the sensed physical quantities collected at each process node, at the target timestamp... Alignment data below The calculation is performed using a linear interpolation algorithm, and the specific algorithm formula is as follows:

[0055] ;

[0056] In the formula, Represents the target timestamp The aligned values ​​calculated by down-interpolation correspond to the units of Celsius (°C), megapascals (MPa), and standard cubic meters per hour, depending on the specific data acquisition object. . The target timestamp represents the discrete time step, in seconds (s). and These represent the timestamps immediately adjacent to the target timestamp in the original asynchronous data stream. The two actual instrument sampling timestamps before and after, and satisfying The timing truncation constraint condition, and Corresponding to the original instruments Time and Measurements of temperature, pressure, or flow rate transmitted in real time.

[0057] After the alignment process described above, the individual physical quantity data of each process node are forcibly mapped to the same physical time section. The system further uses the same target timestamp. The inlet and outlet temperature alignment data of the scrubbing tower, the exhaust pressure alignment data of the separator, and the flow alignment data of the corresponding nodes are concatenated into feature vectors and packaged into a multidimensional state vector. As the system control loop sequence continues to advance, the multidimensional state vectors generated at multiple time steps are combined column-wise to form a multidimensional time series data matrix. The system finally outputs the completed multidimensional time series data matrix as a physical sensing signal.

[0058] By unifying the clock reference and physical dimensions of multi-source sensors through linear interpolation algorithm, a low-level feature matrix with time dimension alignment and clear values ​​is provided for downstream networks, preventing prediction deviations induced by physical signal timing misalignment or dimensional ambiguity.

[0059] The composition prediction module receives physical sensing signals, converts the physical sensing signals into a state matrix representing the heat and mass transfer relationship, extracts features from the state matrix through a graph neural network constrained by thermodynamic equations, and outputs a gas composition prediction command and a state matrix.

[0060] Furthermore, in the composition prediction module, the feature extraction of the state matrix using a graph neural network constrained by thermodynamic equations, and the output of the gas composition prediction command and state matrix specifically include the following steps:

[0061] Construct an initial graph network structure that includes the process topology relationships of the production system;

[0062] The state matrix is ​​mapped to the node attributes of the initial graph network structure. During the propagation between network layers, the gas-liquid equilibrium thermodynamic equation is superimposed as a loss function penalty term for feature extraction, resulting in the converged node hidden layer feature vector.

[0063] The hidden layer feature vectors of nodes are mapped to impurity concentration prediction values ​​through a fully connected layer. Based on the impurity concentration prediction values, a gas composition prediction instruction is generated, and the extracted feature state matrix and the gas composition prediction instruction are output.

[0064] Furthermore, the feature extraction using the superimposed gas-liquid equilibrium thermodynamic equation as a loss function penalty term specifically includes the following steps:

[0065] Construct the Penn-Robinson state equation that includes the eccentricity factor and critical property parameters;

[0066] Substitute the intermediate output tensor during the network forward propagation process into the Penn-Robinson equation of state to calculate the fugacity difference between the theoretical thermodynamic phase state and the actual measured phase state;

[0067] The fugacity difference is converted into a mean squared error term and added to the basic loss function of the graph neural network. The inter-layer weight parameters of the graph neural network are then updated through the backpropagation algorithm.

[0068] Specifically, the component prediction module receives the physical sensing signals output from the previous module and converts the multidimensional time-series data matrix into a state matrix representing the fluid heat and mass transfer relationship. In the specific application scenario of producing food-grade carbon dioxide, gas flow rate, temperature, and pressure directly determine the mass transfer efficiency of the gas-liquid separation interface and the conversion rate of the catalytic oxidation reaction, constituting an intrinsic physical correlation between the input process characteristics and the final output impurity concentration. The system constructs an initial graph network structure based on the actual physical pipeline connections of the scrubbing tower, multi-stage compressor, and catalytic oxidation reactor in the real production line. The system maps the scrubbing tower, compressor, and reactor as independent graph network nodes, maps the gas pipelines connecting each device as directed edges, and constructs an adjacency matrix representing the physical topology. The system decomposes the multidimensional time-series data matrix according to the spatial location of the equipment, and projects the temperature-aligned data, pressure-aligned data, and flow-aligned data of the corresponding physical nodes into a high-dimensional feature space through a multilayer perceptron network, mapping them into dynamic feature vectors of the corresponding graph nodes, thereby completing the construction of the state matrix.

[0069] The system inputs the state matrix into a graph neural network constrained by thermodynamic equations to initiate forward propagation calculations. During the inter-layer propagation phase, the graph neural network aggregates feature information of adjacent process nodes along the pipeline flow path to the edge using the adjacency matrix. The specific algorithm formula for node feature information transfer and update is as follows:

[0070] ;

[0071] In the formula, This represents the layer index number of the graph neural network. Represents the index number of a node in the target graph network. This represents the index number of the neighboring node that is adjacent to the target graph network node. This represents the target graph network node at the next network level. The hidden layer feature vector generated by the layer. Represents the neighboring nodes at the current network level The hidden layer feature vectors of the layer. Represents the current network level The trainable weight matrix of the layer. Representing nodes in the adjacency matrix With nodes The numerical values ​​of physical topology elements in the flow connection relationship. This represents the activation function used to introduce the nonlinear mapping capability of the network. The topological scope is the set of all local neighbor nodes associated with a node in the target graph network.

[0072] To constrain the solution space of the AI ​​model in complex chemical engineering scenarios, the system superimposes the gas-liquid equilibrium thermodynamic equation as a loss function penalty term in the network hidden layer. The system extracts intermediate output tensors during the network forward propagation process and uses a linear decoding layer to resolve these tensors into derived temperature and pressure vectors for each process diagram node at future time steps. The system constructs a Penn-Robinson cubic equation of state with built-in eccentricity factors and critical temperature and pressure parameters for multi-component mixtures of carbon dioxide, oxygen, and volatile organic compounds. To accurately calculate the thermodynamic state of the mixture, the system introduces van der Waals mixing rules to construct a mixture attraction constant and a mixture repulsion constant containing binary interaction parameters. The specific joint calculation formula is as follows:

[0073] ;

[0074] ;

[0075] In the formula, It represents the total number of substances contained in the gas mixture. and These represent the summation index numbers for different material components. The gravitational constant of a multi-component mixture is calculated according to the van der Waals mixing rule, and is expressed in Pascals to the power of six meters per mole square. This represents the repulsion constant of a multi-component mixture, expressed in cubic meters per mole. and Representing the components of the substance With material components The mole fraction in a gas mixture, dimensionless. and Representing the components of the substance With material components As the gravitational constant in the pure matter state, its unit is Pascals to the power of six meters per mole square. Representative material components As the repulsive constant in the pure matter state, its unit is cubic meters per mole. Representative characterizing material components With material components The binary interaction constant, dimensionless, represents the deviation of intermolecular forces between molecules.

[0076] In specific food-grade carbon dioxide purification scenarios, for a mixed gas system consisting of carbon dioxide, oxygen, and volatile organic compounds (VOCs), the binary interaction constant... These are empirical property parameters that are highly dependent on temperature and substance pairing, and can be directly obtained by accessing the NIST fluid thermodynamics database or using the built-in property regression tools in chemical engineering simulation software such as AspenPlus. To reduce implementation difficulty and ensure smooth model initialization, the system pre-sets reference values ​​for typical mixed components: among which, the binary interaction parameter between carbon dioxide and oxygen... A typical value is 0.012; the binary interaction parameter between carbon dioxide and typical volatile organic compounds (taking methane and ethanol as examples). The preferred numerical range is 0.050 to 0.125; the intermolecular forces between oxygen and volatile organic compounds have a small deviation, and their interaction parameters are relatively small. The default setting is 0.

[0077] The system combines the derived temperature and pressure vectors with the mixture's attractive and repulsive constants, substituting them into the Penn-Robinson cubic equation of state to solve for the thermodynamic compressibility factor at the corresponding process node state. This allows for the calculation of the predicted fugacity value for the theoretical thermodynamic phase. Based on actual sensor measurements, the system calculates the baseline fugacity value under real-world conditions and then calculates the thermodynamic fugacity penalty term between the predicted and baseline values. The specific algorithm formula is as follows:

[0078] ;

[0079] In the formula, The mean square error penalty term representing the calculated thermodynamic fugacity. This represents the total number of graph network nodes participating in phase state evaluation in the initial graph network structure. The summation index number represents the node of the graph network. The first value represents the temperature and pressure derived from the analysis of the network intermediate tensor, combined with the Penn-Robinson equation. The predicted fugacity values ​​for each node, in megapascals (MPA). The first value is calculated based on the actual physical measurement value of the sensor. The baseline fugacity value of each node, in megapascals (MPA).

[0080] The system extracts the true concentration labels of oxygen and volatile organic compounds from historical real-world laboratory data of the factory, and constructs a basic prediction loss function for impurity concentration. The system then linearly weights and fuses the thermodynamic fugacity mean square error penalty term with the basic prediction loss function for impurity concentration to generate a global loss function to guide model weight updates. The specific algorithm formula is as follows:

[0081] ;

[0082] In the formula, Represents the global loss function for joint training. This represents the total number of samples in the training sample batch. The summation index number represents the data sample. The first one representing the output of the network forward propagation The predicted impurity concentration for each sample, including the concentration values ​​of oxygen and volatile organic compounds, in parts per million. Representing the The actual impurity concentration label for each sample, in parts per million. The dimensionless hyperparameter penalty weight represents the strength of the thermodynamic physical constraint. The mean square error penalty term for the thermodynamic fugacity calculated above is set to a preferred value range of 0.01 to 0.1 in order to balance the prediction accuracy driven by network data and the consistency of physical mechanism during the iterative optimization process.

[0083] The system calculates the gradient of the global loss function using the backpropagation algorithm. To prevent gradient explosion caused by the highly nonlinear nature of the Penn-Robinson cubic equation when differentiating across phase boundaries, a gradient pruning mechanism is superimposed during the backpropagation calculation. This mechanism forcibly truncates the highest norm of the calculated global loss function gradient tensor to within a set norm threshold, with the preferred value range being 1.0 to 5.0. The system uses a gradient descent optimizer to update the inter-layer weight parameters of the graph neural network along the negative gradient direction until the model converges, outputting the converged node hidden layer feature vector. The system inputs the converged node hidden layer feature vector into a fully connected layer network for feature dimensionality reduction and nonlinear mapping, outputting the final impurity concentration prediction value before entering the subsequent deep purification process. Based on the final impurity concentration prediction value, the system generates a gas composition prediction command to trigger the feedforward adjustment action of the downstream catalytic oxidation reactor. Finally, the system outputs the state matrix processed by network feature extraction, along with the gas composition prediction command, to the purification control module.

[0084] By constructing a graph network message passing mechanism and introducing binary interaction parameters of mixed gases, the multi-component phase equilibrium thermodynamic equation is transformed into a loss function penalty term and implanted into the graph neural network. At the same time, gradient pruning technology is combined to ensure the training stability of the physical information neural network, constrain the algorithm prediction boundary, and avoid the concentration prediction distortion caused by pure data-driven models that violate the law of conservation of matter when sensor signals fluctuate.

[0085] The purification control module receives gas composition prediction commands and state matrix, adjusts the reactant injection ratio of the catalytic oxidation reactor according to the gas composition prediction commands, solves the dynamic concentration front parameters corresponding to the state matrix through the adsorption kinetic equation, controls the regeneration of molecular sieves based on the dynamic concentration front parameters, and outputs the purified gas flow and purification state parameters.

[0086] Furthermore, in the purification control module, the dynamic concentration front parameters corresponding to the state matrix are solved by the adsorption kinetic equation, and the molecular sieve regeneration is controlled based on the dynamic concentration front parameters, specifically including the following steps:

[0087] Extract the intake air moisture characteristics from the state matrix and substitute them into the adsorption mass transfer equation in partial differential form.

[0088] Calculate the spatial distribution gradient of moisture concentration inside the molecular sieve, and calculate the physical spatial coordinates of the spatial distribution gradient when it reaches the set value as dynamic concentration front parameters.

[0089] Determine whether the dynamic concentration front parameter has moved to the set boundary position of the molecular sieve bed. If it reaches the set boundary position, trigger the bed switching command and start the heating and regeneration program of the corresponding molecular sieve.

[0090] Furthermore, substituting the intake air moisture characteristics into the partial differential form of the adsorption mass transfer equation specifically includes the following steps:

[0091] Establish a linear driving force mass transfer model equation that includes gas phase mass conservation parameters and solid phase adsorption rate parameters;

[0092] The physical axial length of the molecular sieve bed is discretized into a set number of grid nodes, and the inlet moisture characteristics are used as the initial boundary conditions to input the linear driving force mass transfer model equation.

[0093] The discretized linear driving force mass transfer model equations are solved using the upwind differential algorithm, and the adsorbate distribution data of each grid node are output and the spatial distribution gradient is calculated.

[0094] Specifically, the purification control module receives the gas composition prediction command and state matrix output from the previous module. The system parses the gas composition prediction command to obtain the predicted oxygen concentration and volatile organic compound (VOC) concentration values, and simultaneously extracts the corresponding carbon dioxide primary gas main pipeline flow data from the state matrix. Based on the predicted VOC concentration and the ideal conversion equivalence ratio of the chemical oxidation reaction, the system constructs a feedforward compensation algebraic model for the catalytic oxidation reactor to calculate the reactant injection compensation value. The specific calculation formula is as follows:

[0095] ;

[0096] In the formula, This represents the calculated reactant injection compensation value, expressed in standard cubic meters per hour. This represents the flow rate data of the main carbon dioxide pipeline extracted from the state matrix, in standard cubic meters per hour. The ideal conversion equivalence constant represents the ideal conversion equivalence constant for the complete combustion of the corresponding volatile organic compounds. To ensure that the organic compounds are completely oxidized and to control the residual oxygen content, the preferred value range of the ideal conversion equivalence constant is 1.05 to 1.15. This represents the predicted concentration of volatile organic compounds in the gas composition prediction command, expressed in volume fraction. The oxygen prediction concentration value in the gas composition prediction command is expressed as a volume fraction. The system generates a low-level analog-to-digital conversion signal based on the reactant injection compensation value, adjusting the valve opening of the micro-oxygen mass flow meter to control the reactant injection ratio, thus ensuring that organic impurities in the carbon dioxide crude gas are fully converted into carbon dioxide and water.

[0097] After catalytic oxidation, the system proceeds to the molecular sieve deep dehydration stage. The system extracts the inlet gas moisture characteristics from the state matrix, which include the temperature, pressure, and initial humidity data of the inlet gas flow. To accurately track the microscopic migration trajectory of moisture within the porous medium, a linear driving force mass transfer model equation is established, incorporating gas-phase mass conservation parameters and solid-phase adsorption rate parameters.

[0098] The system discretizes the physical axial length of the molecular sieve bed into a predetermined number of grid nodes and executes the initial mapping settings for the numerical algorithm. The system forcibly maps the initial humidity data in the inlet gas moisture characteristics to a constant first-type boundary condition at the inlet of the first-layer grid of the algorithm, and sets the spatial derivative of the gas phase concentration at the outlet of the end grid of the molecular sieve bed to zero, constructing a Neumann second-type boundary condition that satisfies the continuity of fluid mass transfer. The system sets the initial values ​​of the gas phase moisture concentration and the initial values ​​of the solid phase moisture adsorption within all grid nodes as the background residual reference values ​​after deep regeneration of the molecular sieve. These background residual reference values ​​are obtained through calibration and conversion using measured dew point data from the regeneration tail gas.

[0099] After algorithm initialization, the system initiates a double-nested loop computational logic to solve the linear driving force mass transfer model equations. The system constructs an outer discrete time-step loop and an inner physical grid space loop. To ensure numerical convergence and absolute stability of the explicit difference algorithm during time-domain iteration, the system strictly constrains the ratio of time step to spatial step according to the Courant-Friedrich-Lévy condition, ensuring that the product of the fluid gap velocity and the time step of the discrete solution iteration is less than or equal to the physical spatial distance between adjacent grid nodes. Under the premise of satisfying the Courant-Friedrich-Lévy stability constraint, the system utilizes the upwind difference algorithm to access the concentration gradient information of upstream grid nodes to guarantee the physical stability of the numerical iteration. At each discrete time step, the system traverses the grid nodes sequentially from the inlet to the outlet, synchronously updating the gas phase moisture concentration and the solid phase moisture adsorption amount. The specific joint formula for the discrete solution algorithm is as follows:

[0100] ;

[0101] ;

[0102] In the formula, The index number representing the discrete time step. Represents the index number of a spatial grid node. This represents the next time step of the grid node. Updated value of vapor phase moisture concentration at the location, in moles per cubic meter. and These represent the current time step. Lower grid node and upstream adjacent grid nodes The current calculated value of the vapor phase moisture concentration at the specified location, in moles per cubic meter. The physical time step represents the discrete solution iteration, in seconds, with a preferred value range of 0.1 seconds to 1.0 seconds. The fluid gap velocity is calculated based on the characteristics of the intake air moisture and the cross-sectional area of ​​the pipe, and is expressed in meters per second. This represents the physical spatial distance between adjacent grid nodes, in meters, with a preferred value range of 0.01 meters to 0.05 meters. The dimensionless parameter representing the porosity of the molecular sieve bed has a preferred value range of 0.35 to 0.45. It represents the mass transfer rate constant within the solid phase, expressed in negative first power of seconds, with a preferred value range of 0.001 to 0.01. This represents the theoretical equilibrium solid-phase adsorption capacity derived from the mapping of local pressure and temperature characteristics at grid nodes, expressed in moles per cubic meter. To achieve accurate numerical calculations of the theoretical equilibrium solid-phase adsorption capacity, the system employs the Langmuir isothermal adsorption model to construct the mapping relationship. The specific algebraic calculation formula is as follows:

[0103] ;

[0104] In the formula, This represents the local vapor phase moisture partial pressure at grid node i at the current time step, in megapascals (MPa). It can be expressed as the vapor phase moisture concentration at the grid node. It was derived by converting the ideal gas law. This represents the saturated adsorption constant of the molecular sieve for water, expressed in moles per cubic meter. For a typical water-molecular sieve system in a deep carbon dioxide dehydration process, the typical saturated adsorption value is 15.2 mol / m³. The Langmuir adsorption affinity constant, representing the relationship between the system temperature and the temperature, is expressed in negative first-order megapascals (MPa⁻¹), with a typical measured value of approximately 0.45 MPa⁻¹ under calibration conditions at 25°C. By introducing the aforementioned explicit isotherm model parameters, closed-loop calculations in the time-step iterations of the numerical simulation are ensured. and These represent the next time step and the current time step, respectively. Lower grid node The actual amount of solid-phase moisture adsorbed at the location, expressed in moles per cubic meter.

[0105] The system outputs the updated actual solid-phase moisture adsorption capacity of each grid node as adsorbate distribution data through a two-level iterative calculation. The system then substitutes the adsorbate distribution data of adjacent grid nodes into the spatial difference calculation model to calculate the spatial distribution gradient of moisture concentration within the molecular sieve. The specific gradient calculation formula is as follows:

[0106] ;

[0107] In the formula, The index number representing the discrete time step. Represents the index number of a spatial grid node. Represents the current time step Lower grid node Spatial gradient at a given location, expressed in moles per cubic meter per meter. and Representing grid nodes respectively and upstream adjacent grid nodes The actual amount of solid-phase moisture adsorbed at the location, expressed in moles per cubic meter. Represents the physical spatial distance between adjacent grid nodes, in meters.

[0108] The system traverses the grid nodes along the spatial axis from the inlet to the outlet, calculating and recording the physical spatial coordinates of the first point where the spatial distribution gradient reaches a set gradient threshold as the dynamic concentration front parameter. Considering the differences in pore size and packing density among different types of molecular sieves, the set gradient threshold is set as a relative dynamic threshold. The system extracts the maximum spatial distribution gradient within the entire bed at the current time step in real time and calculates the set gradient threshold by multiplying the maximum spatial distribution gradient by a set percentage. The preferred value range for the set percentage is 10% to 30%. The dynamic concentration front parameter accurately characterizes the mass transfer boundary location in the macroscopic physical world where a large amount of water is undergoing gas-solid phase transfer.

[0109] The system determines whether the dynamic concentration front parameter has moved to the set boundary position of the molecular sieve bed. If the dynamic concentration front parameter reaches the set boundary position, the system converts the value exceeding the limit event into a level toggling signal of the underlying programmable logic controller, triggering a physical bed switching command and starting the heating and regeneration program of the corresponding molecular sieve. After completing the impurity removal and dehydration process, the system outputs purified gas flow that meets the subsequent liquefaction standards. The system packages the molecular sieve bed operating temperature, the mass flow meter valve opening / closing status, and the dynamic concentration front parameter into purification status parameters, and synchronously outputs the purified gas flow and purification status parameters to the energy efficiency optimization module.

[0110] Feedforward control and dehydration actions, through a numerical mass transfer deduction algorithm with clear boundary conditions and two-layer loop logic, directly map the abstract spatial gradient calculation into the physical valve action of the programmable logic controller, realizing accurate prediction of the molecular sieve regeneration timing and preventing water penetration and excessive energy consumption.

[0111] The energy efficiency optimization module receives purification state parameters, obtains electricity price information and meteorological information, inputs the purification state parameters, electricity price information and meteorological information into a reinforcement learning model with carbon dioxide gas-liquid phase boundary as action constraint for processing, and outputs control strategy instructions.

[0112] Furthermore, in the energy efficiency optimization module, the purification state parameters, electricity price information, and meteorological information are input into a reinforcement learning model with carbon dioxide gas-liquid phase boundary as the action constraint for processing. The output control strategy command specifically includes the following steps:

[0113] The purified state parameters, electricity price information, and meteorological information are concatenated into the state space vector of the reinforcement learning model.

[0114] Based on the physical properties of carbon dioxide, a gas-liquid phase envelope curve equation between the triple point and the critical point is constructed, and the gas-liquid phase envelope curve equation is loaded into the action space of the reinforcement learning model as a penalty boundary mask.

[0115] Within the action domain where the penalty boundary mask is not triggered, the strategy iteration is performed using the system's per-ton electricity cost as the reward function, generating and outputting control strategy instructions.

[0116] Furthermore, loading the gas-liquid phase envelope curve equation as a penalty boundary mask into the action space of the reinforcement learning model specifically includes the following steps:

[0117] Extract the initial continuous action vector generated by the output layer of the reinforcement learning model;

[0118] Calculate the Euclidean geometric distance from the thermodynamic coordinates pointed to by the initial continuous motion vector to the boundary of the gas-liquid phase envelope curve equation;

[0119] A logarithmic barrier function based on Euclidean geometric distance is constructed. The logarithmic barrier function is used to truncate and remap the initial continuous action vectors in the preset threshold region to generate the actual action vectors.

[0120] Specifically, the energy efficiency optimization module receives the purified airflow and purification state parameters synchronously output by the previous module, and simultaneously obtains time-of-use electricity price information from the external power grid and meteorological information of the plant environment through the industrial internet interface. The meteorological information specifically includes the ambient dry-bulb temperature and relative humidity, which affect the cooling tower's heat exchange efficiency. The system performs numerical normalization preprocessing on the purification state parameters, electricity price information, and meteorological information, and then sequentially concatenates them into a state space vector for the reinforcement learning model at discrete time steps. This state space vector comprehensively and quantitatively represents the system's current physical operating conditions, external environmental boundaries, and energy cost constraints, serving as the input feature matrix for the reinforcement learning model's agent to perceive the global environment.

[0121] To achieve high-dimensional nonlinear optimization of system energy consumption, a deep deterministic policy gradient algorithm architecture is constructed, comprising an action network and an evaluation network. Both the action and evaluation networks employ a multilayer perceptron structure with three hidden layers. The system sets the number of neurons in the three hidden layers to 256, 128, and 64 respectively, and uses a linear rectified activation function in the hidden layers. The action network uses a hyperbolic tangent activation function in the output layer to generate initial continuous action vectors.

[0122] To prevent the algorithm from outputting actions that would cause carbon dioxide to overflow and form dry ice that clogs the pipeline when exploring high-efficiency strategies, the system constructs a gas-liquid phase envelope curve equation between the triple point and the critical point based on the physical properties of carbon dioxide. To enable the computer to perform distance calculations at the continuous phase boundary, the system introduces the Antoine equation as the mathematical basis for the gas-liquid phase envelope curve. The specific boundary calculation equation is as follows:

[0123] ;

[0124] In the formula, This represents the boundary temperature variable on the gas-liquid phase envelope curve, and its unit is Kelvin. This represents the saturated vapor pressure value corresponding to the boundary temperature variable, in megapascals (MPa). , as well as These represent dimensionless fitting constants for the physical properties of carbon dioxide within a specific temperature range. The system extracts the initial continuous action vector generated by the action network output layer at the current discrete time step. This initial continuous action vector contains the target interstage pressure for the downstream multi-stage compressor and the target condensing temperature for the refrigeration unit. The system maps the target condensing temperature and target interstage pressure pointed to by the initial continuous action vector into two-dimensional thermodynamic coordinates. The Antoine equation is used to calculate the Euclidean geometric distance from the thermodynamic coordinates to the boundary of the gas-liquid phase envelope curve equation. The specific distance calculation formula is as follows:

[0125] ;

[0126] In the formula, This represents the index number of the discrete time step in the reinforcement learning algorithm. Represents the discrete time step The Euclidean geometric distance from the calculated thermodynamic coordinates to the boundary of the gas-liquid phase envelope curve is dimensionless. Represents discrete time step The initial continuous action vector refers to the target condensation temperature, in Kelvin. Represents discrete time step The initial continuous motion vector refers to the target interstage pressure, in megapascals (MPA). and These represent the triple point temperature and critical point temperature of carbon dioxide, respectively, and are measured in Kelvin. A dimensionless constant representing the weighting ratio of different physical magnitudes of equilibrium temperature and pressure.

[0127] The system adds a differentiable safety layer after the output layer of the action network, and hardcodes the gas-liquid phase envelope curve equation and the logarithmic barrier function into the differentiable safety layer. The system uses the logarithmic barrier function to truncate and remap the gradients of initial continuous action vectors within a preset threshold region. The specific action mapping and boundary penalty formulas are as follows:

[0128] ;

[0129] In the formula, The index number represents the discrete time step. Represents the discrete time step The actual action vector is generated after being truncated by the differentiable safety layer and the gradient reset mapping. Represents the discrete time step The initial continuous action vector generated by the lower action network. Represents the discrete time step The Euclidean geometric distance calculated below is dimensionless. The parameter representing the boundary distance of the preset threshold area that triggers the security protection mechanism is dimensionless, and the preferred value range is 0.05 to 0.15. It represents the absolute limit of safe physical distance that must not be crossed. It is dimensionless, and the preferred numerical range is 0.01 to 0.02. This represents the dimensionless update step size constant for reverse gradient reset. The gradient operator representing the partial derivative of the logarithmic barrier function with respect to the initial continuous action vector approaches infinity as the distance nears the limit of safe physical distance. The gradient vector calculated based on the partial derivative forces the algorithm's actions to mathematically move away from the dangerous phase transition region. During backpropagation, the differentiable safety layer directly propagates the penalty gradient back to the action network according to the chain rule, driving the action network to automatically avoid the dangerous state space.

[0130] Within the action domain where the penalty boundary mask is not triggered, the system executes policy iterations using the system's per-ton electricity cost as the reward function. To ensure the code writability of the reinforcement learning algorithm in the industrial controller, the system constructs a quantified per-ton electricity cost reward function feedback model. The specific reward calculation formula is as follows:

[0131] ;

[0132] In the formula, This represents the index number of the discrete time step in the reinforcement learning algorithm. Represents the discrete time step The calculated system's per-ton electricity cost reward function feedback value is expressed in yuan per ton. Since lower production costs indicate better energy efficiency, a negative sign is introduced to reverse the reward direction. This represents the total power consumption of the compressor and refrigeration unit, calculated based on the actual execution action vector. The unit is kilowatt. This represents the real-time electricity price extracted from the electricity price information, in yuan per kilowatt-hour. This represents the predicted mass flow rate converted into liquid carbon dioxide per unit time, expressed in tons per hour.

[0133] To construct a closed-loop environmental feedback model during the reinforcement learning training phase, enabling the agent to obtain reward signals without trial and error on real devices, the system incorporates a steady-state deduction equation based on physical mechanisms to calculate the aforementioned parameters in real time. and Based on the system's mass conservation law, the predicted carbon dioxide production value The mass flow rate of the purified gas stream, directly equivalent to that measured in the previous purification process. Total power consumption. The power is then derived from the multi-stage compressor. Power projection with refrigeration unit It is composed of superposition.

[0134] The power of the multi-stage compressor is estimated based on the polytropic compression thermodynamics formula, specifically:

[0135] ;

[0136] In the formula, The target interstage pressure extracted from the actual execution action vector is used as an exhaust end variable, with the unit being megapascals. and These are the real-time inlet pressure (in megapascals) and absolute temperature (in Kelvin) of the compressor, obtained from the state space vector, respectively. This is the gas constant of carbon dioxide, measured in joules per kilogram (Kelvin). ); is the polytropic index of carbon dioxide, a dimensionless constant; It is the combined mechanical and isentropic efficiency constant of the compressor, a dimensionless constant; the constant 3600 is used in conjunction with the gas constant unit to complete the dimensional conversion of mass flow rate from tons per hour to the basic SI unit, and the final power to kilowatts (kW).

[0137] The estimated power of the refrigeration unit is based on the inverse Carnot cycle performance coefficient, which incorporates a reduction in work efficiency. The specific formula is as follows:

[0138] ;

[0139] In the formula, The target condensation temperature extracted from the actual action vector, in Kelvin; The external environment dry-bulb temperature is extracted from meteorological information and is expressed in Kelvin. To correspond to the latent heat of gas-liquid phase change under the operating conditions, the unified dimension is inherited from the aforementioned liquefaction module, with the unit being joules per kilogram (J / kg). ); The thermodynamic perfection coefficient of the refrigeration unit is a dimensionless constant. Used to convert the final power value to the unit of kilowatt (kW).

[0140] During model training and iteration, the system initializes a fixed-capacity experience replay pool to store historical state transition data. To fully explore the state space, Ornstein-Uhlenbeck random noise is superimposed on the initial continuous action vectors. The system employs a uniform sampling strategy to extract batch data from the experience replay pool, calculates the temporal difference error of the evaluation network, and updates the evaluation network weights. The system updates the action network weights along the direction of increasing reward feedback values ​​from the evaluation network output, and sets the target network's soft update coefficient to 0.005 for synchronous updates, prompting the model to prioritize increasing the cooling load during periods of low electricity prices and low ambient temperatures. The system converts the converged actual execution action vectors into control policy instructions recognizable by the underlying hardware and outputs them to the downstream compression and liquefaction stages.

[0141] The compression liquefaction module receives the purified airflow and adjusts the interstage pressure parameters of the multi-stage compressor and the condensing temperature parameters of the refrigeration unit according to the control strategy instructions, converting the purified airflow into liquid carbon dioxide.

[0142] Furthermore, in the compression-liquefaction module, adjusting the interstage pressure parameters of the multi-stage compressor and the condensing temperature parameters of the refrigeration unit according to the control strategy instructions specifically includes the following steps:

[0143] Analyze the control strategy instructions to separate the pressure distribution matrix for the multi-stage compressor and the temperature setpoint for the refrigeration unit;

[0144] Based on the pressure distribution matrix, adjust the inlet guide vane opening of the multi-stage compressor and the water valve opening of the interstage cooler to output the interstage pressure parameters of the multi-stage compressor.

[0145] Based on the temperature setpoint, adjust the refrigerant circulation flow rate and expansion valve opening in the refrigeration unit to cool the purified airflow below the gas-liquid phase change temperature.

[0146] Specifically, the compression-liquefaction module receives the purified airflow synchronously output from the previous module, along with control strategy instructions containing actual execution action vectors. The system parses the control strategy instructions and uses a low-level decoding algorithm to decompose the continuous execution action vector space, separating the target interstage pressure value for the multi-stage compressor and the target condensing temperature value for the refrigeration unit.

[0147] The system extracts the target interstage pressure value and collects the actual measured pressure values ​​in the pipelines of each stage of the multi-stage compressor. The system constructs an incremental proportional-integral control algorithm model, and based on the pressure deviation between the target interstage pressure value and the actual measured pressure value, independently calculates the control command values ​​for the inlet guide vanes of the multi-stage compressor and the water valves of the interstage cooler. The specific combined control algorithm formula is as follows:

[0148] ;

[0149] ;

[0150] In the formula, The index number represents the discrete time step. and These represent the current time step. Compared with the previous time step The calculated control command values ​​for the corresponding actuators are expressed as percentages. This represents the proportional gain coefficient constant. This represents the integral gain coefficient constant. and These represent the current time step. Compared with the previous time step The pressure deviation is expressed in megapascals (MPA). This represents the current time step, which is strictly inherited from the output instruction of the previous module. The target interstage pressure value, in megapascals (MPa). This represents the current time step acquired by the pressure transmitter. The actual measured pressure value is expressed in megapascals (MPa).

[0151] The system employs the Ziegler-Nichols step response method to tune the proportional gain constant and integral gain constant. To prevent system oscillations caused by integral saturation when the actuator opening reaches its physical limit, an integral anti-saturation logic mechanism is added to the incremental proportional-integral control algorithm model. When the calculated control command value exceeds the maximum fully open limit and minimum fully closed limit range of the physical valve, the system forcibly truncates the actual output control command value to the corresponding limit boundary and freezes the accumulation of the integral term corresponding to the integral gain constant within this discrete time step. Based on the control command value after anti-saturation processing, the system adjusts the inlet guide vane opening of the multi-stage compressor and the interstage cooler water valve opening, outputting and maintaining the multi-stage compressor interstage pressure parameters to achieve the set target.

[0152] Based on the target condensation temperature obtained from the separation, the system constructs a phase change heat conservation calculation model to calculate the refrigerant circulation flow rate required to cool the purified gas flow below the gas-liquid phase change temperature. To ensure the absolute accuracy of the thermodynamic property parameter calls, the system pre-installs carbon dioxide and refrigerant property data tables based on the National Institute of Standards and Technology's fluid thermodynamic property database in its underlying memory. The system extracts the mass flow rate and actual measured temperature of the purified gas flow and substitutes them into the phase change heat conservation calculation model. The specific refrigerant flow rate calculation formula is as follows:

[0153] ;

[0154] In the formula, The index number represents the discrete time step. Represents the discrete time step The target value of refrigerant circulation mass flow rate is calculated below, in kilograms per second. Represents the purified airflow at discrete time steps The mass flow rate is expressed in kilograms per second. This represents the isobaric specific heat capacity of the purified airflow, obtained by dynamically looking up a table based on the pressure and temperature conditions at the current discrete time step. The unit is joules per kilogram Kelvin. This represents the actual measured temperature of the purified airflow before cooling, expressed in Kelvin. This represents the current time step, which is strictly inherited from the output instruction of the previous module. The target condensation temperature value, in Kelvin. This represents the latent heat of phase change of carbon dioxide at the current discrete time step under the set operating conditions, expressed in joules per kilogram. This represents the saturated vapor enthalpy of the refrigerant at the evaporator outlet, expressed in joules per kilogram. It represents the saturated liquid enthalpy of the refrigerant at the evaporator inlet, expressed in joules per kilogram.

[0155] When acquiring the aforementioned dynamic physical property parameters, the system utilizes a two-dimensional bicubic spline interpolation algorithm. Based on the real-time collected pressure and temperature coordinates, it performs nonlinear smooth interpolation extraction in preset carbon dioxide and refrigerant property data tables to generate the isobaric specific heat capacity, latent heat of gas-liquid phase change, and enthalpy values ​​corresponding to the discrete time steps. Based on the target refrigerant circulation mass flow rate, the system generates variable frequency drive and electromagnetic drive signals to synchronously adjust the compressor's operating frequency in the refrigeration unit to change the refrigerant circulation flow rate, and adjusts the electronic expansion valve opening to precisely match the heat exchange superheat. After purification, the airflow exchanges heat with the refrigerant in the condenser, its temperature decreasing below the gas-liquid phase change temperature and transforming into liquid carbon dioxide fluid.

[0156] The compression and liquefaction steps seamlessly transform the top-level reinforcement learning optimization strategy into the underlying physical driving action through an incremental error compensation algorithm with an integral anti-saturation mechanism and a dynamic heat conservation equation with high-precision physical property spline interpolation, thus realizing a continuous and stable phase transition from high-purity gas flow to liquid carbon dioxide.

[0157] Finally, it should be noted that the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A system for producing liquid food-grade carbon dioxide, characterized in that, include: The raw material processing module performs washing and separation operations on the crude carbon dioxide gas, and collects and outputs physical sensor signals during the washing and separation process. The composition prediction module receives the physical sensing signal, converts the physical sensing signal into a state matrix representing the heat and mass transfer relationship, extracts features from the state matrix through a graph neural network constrained by thermodynamic equations, and outputs a gas composition prediction command and the state matrix. The purification control module receives the gas composition prediction command and the state matrix, adjusts the reactant injection ratio of the catalytic oxidation reactor according to the gas composition prediction command, solves the dynamic concentration front parameter corresponding to the state matrix through the adsorption kinetic equation, controls the molecular sieve regeneration according to the dynamic concentration front parameter, and outputs the purified gas flow and purification state parameters. The energy efficiency optimization module receives the purification state parameters, obtains electricity price information and meteorological information, inputs the purification state parameters, the electricity price information and the meteorological information into a reinforcement learning model with carbon dioxide gas-liquid phase boundary as action constraint for processing, and outputs control strategy instructions; The compression liquefaction module receives the purified airflow and adjusts the interstage pressure parameters of the multi-stage compressor and the condensing temperature parameters of the refrigeration unit according to the control strategy command, thereby converting the purified airflow into liquid carbon dioxide.

2. The system for producing liquid food-grade carbon dioxide according to claim 1, characterized in that, In the raw material processing module, the process of collecting and outputting physical sensor signals during the washing and separation operation specifically includes the following steps: Temperature, pressure, and flow signals were collected at the inlet and outlet of the scrubbing tower and at the exhaust end of the gas-liquid separator, respectively. The temperature sensing signal, the pressure sensing signal, and the flow sensing signal are timestamped according to a preset sampling frequency, packaged and encapsulated into multidimensional time series data, and the multidimensional time series data is output as a physical sensing signal.

3. The system for producing liquid food-grade carbon dioxide according to claim 1, characterized in that, In the composition prediction module, the step of extracting features from the state matrix using a graph neural network constrained by thermodynamic equations and outputting a gas composition prediction command and the state matrix specifically includes the following steps: Construct an initial graph network structure that includes the process topology relationships of the production system; The state matrix is ​​mapped to the node attributes of the initial graph network structure. During the propagation between network layers, the gas-liquid equilibrium thermodynamic equation is superimposed as a loss function penalty term for feature extraction, resulting in the converged node hidden layer feature vector. The node hidden layer feature vector is mapped to the impurity concentration prediction value through a fully connected layer. A gas composition prediction instruction is generated based on the impurity concentration prediction value, and the extracted state matrix and the gas composition prediction instruction are output.

4. The system for producing liquid food-grade carbon dioxide according to claim 3, characterized in that, The feature extraction process, which uses the superimposed gas-liquid equilibrium thermodynamic equations as a loss function penalty term, specifically includes the following steps: Construct the Penn-Robinson state equation that includes the eccentricity factor and critical property parameters; Substitute the intermediate output tensor during the network forward propagation process into the Penn-Robinson equation of state to calculate the fugacity difference between the theoretical thermodynamic phase state and the actual measured phase state; The fugacity difference is converted into a mean squared error term and added to the basic loss function of the graph neural network. The inter-layer weight parameters of the graph neural network are then updated using the backpropagation algorithm.

5. The system for producing liquid food-grade carbon dioxide according to claim 1, characterized in that, In the purification control module, the step of solving the dynamic concentration front parameter corresponding to the state matrix through the adsorption kinetic equation and controlling the molecular sieve regeneration based on the dynamic concentration front parameter specifically includes the following steps: Extract the intake air moisture characteristics from the state matrix and substitute the intake air moisture characteristics into the adsorption mass transfer equation in partial differential form; Calculate the spatial distribution gradient of moisture concentration inside the molecular sieve, and calculate the physical spatial coordinates of the spatial distribution gradient reaching the set value as dynamic concentration front parameters; Determine whether the dynamic concentration front parameter has moved to the set boundary position of the molecular sieve bed. If it reaches the set boundary position, trigger the bed switching command and start the heating and regeneration program of the corresponding molecular sieve.

6. A system for producing liquid food-grade carbon dioxide according to claim 5, characterized in that, Substituting the inlet air moisture characteristics into the partial differential form of the adsorption mass transfer equation specifically includes the following steps: Establish a linear driving force mass transfer model equation that includes gas phase mass conservation parameters and solid phase adsorption rate parameters; The physical axial length of the molecular sieve bed is discretized into a set number of grid nodes, and the inlet moisture characteristics are used as initial boundary conditions input into the linear driving force mass transfer model equation. The discretized linear driving force mass transfer model equation is solved using the upwind difference algorithm, and the adsorbate distribution data of each grid node is output and the spatial distribution gradient is calculated.

7. The system for producing liquid food-grade carbon dioxide according to claim 1, characterized in that, In the energy efficiency optimization module, the process of inputting the purification state parameters, the electricity price information, and the meteorological information into a reinforcement learning model with carbon dioxide gas-liquid phase boundary as the action constraint for processing, and outputting control strategy instructions specifically includes the following steps: The purification state parameters, the electricity price information, and the meteorological information are concatenated to form the state space vector of the reinforcement learning model; Based on the physical property parameters of carbon dioxide, a gas-liquid phase envelope curve equation between the triple point and the critical point is constructed, and the gas-liquid phase envelope curve equation is loaded into the action space of the reinforcement learning model as a penalty boundary mask. Within the action domain where the penalty boundary mask is not triggered, the strategy iteration is performed using the system's per-ton electricity cost as the reward function, generating and outputting control strategy instructions.

8. A system for producing liquid food-grade carbon dioxide according to claim 7, characterized in that, The step of loading the gas-liquid phase envelope curve equation as a penalty boundary mask into the action space of the reinforcement learning model specifically includes the following steps: Extract the initial continuous action vector generated by the output layer of the reinforcement learning model; Calculate the Euclidean geometric distance from the thermodynamic coordinates pointed to by the initial continuous motion vector to the boundary of the gas-liquid phase envelope curve equation; A logarithmic barrier function based on the Euclidean geometric distance is constructed. The logarithmic barrier function is used to truncate and remap the initial continuous action vector in the preset threshold region to generate the actual action vector.

9. A system for producing liquid food-grade carbon dioxide according to claim 1, characterized in that, In the compression liquefaction module, adjusting the interstage pressure parameters of the multi-stage compressor and the condensing temperature parameters of the refrigeration unit according to the control strategy command specifically includes the following steps: The control strategy instructions are analyzed to separate the pressure distribution matrix for the multi-stage compressor and the temperature setpoint for the refrigeration unit. Based on the pressure distribution matrix, the opening degree of the inlet guide vane of the multi-stage compressor and the opening degree of the interstage cooler water valve are adjusted to output the interstage pressure parameters of the multi-stage compressor. Based on the temperature setpoint, the refrigerant circulation flow rate and expansion valve opening in the refrigeration unit are adjusted to cool the purified airflow to below the gas-liquid phase change temperature.

Images