A low power consumption high purity oxygen separation method and system

Abstract

The present application relates to the technical field of gas separation process simulation optimization, and particularly relates to a low-power-consumption high-purity oxygen separation method and system. The method comprises the following steps: obtaining physical layer data and adsorption bed data; constructing a virtual simulation model according to the adsorption bed data to obtain an adsorption bed model; performing simulation calculation according to the physical layer data and the adsorption bed model to obtain adsorption bed layer data; performing graph constraint calculation according to the physical layer data and the adsorption bed layer data to obtain virtual-real difference data; performing outlet oxygen concentration prediction according to the adsorption bed layer data to obtain outlet oxygen concentration prediction data; performing differential feature extraction according to the outlet oxygen concentration prediction data and the physical layer data to obtain prediction deviation data; obtaining regeneration step data according to the virtual-real difference data and the prediction deviation data; and obtaining energy constraint control data according to the regeneration step data and the adsorption bed layer data. The present application realizes high-purity output and low-power-consumption control of the oxygen separation process.

Description

Technical Field

[0001] This invention relates to the field of gas separation process simulation and optimization technology, and in particular to a low-power, high-purity oxygen separation method and system. Background Technology

[0002] Existing oxygen separation systems are mostly based on pressure swing adsorption (PSA) or temperature swing adsorption mechanisms. They separate oxygen from other gases (such as nitrogen and argon) by periodically adjusting the pressure or temperature of the adsorbent material. Although these systems are widely used in industrial and medical applications, their overall performance is still limited by the responsiveness of the basic model and the robustness of the control strategy. The main problems include: simulation models rely on static parameters, making dynamic adaptation difficult. Current adsorption bed modeling typically uses fixed mass transfer diffusion coefficients, constant bed heat capacity, and simplified boundary conditions, neglecting state changes caused by factors such as inlet temperature, adsorption rate, or equipment aging during actual operation. This results in insufficient model accuracy and difficulty in accurately predicting mass migration and energy exchange characteristics during adsorption. This static modeling approach struggles to correct simulation deviations in a timely manner when faced with multiple operating condition switching, external disturbances, or operational state drift, easily leading to fluctuations in the system's output oxygen concentration, affecting separation purity and energy efficiency. Summary of the Invention

[0003] To address the aforementioned technical problems, this invention proposes a low-power, high-purity oxygen separation method and system, thereby resolving at least one of the aforementioned technical issues.

[0004] This application provides a low-power, high-purity oxygen separation method, the method comprising:

[0005] S1. Obtain physical layer data and adsorption bed data; construct a virtual simulation model based on the adsorption bed data to obtain the adsorption bed model; perform simulation calculations based on the physical layer data and the adsorption bed model to obtain the adsorption bed data.

[0006] S2. Perform graph constraint calculations based on physical layer data and adsorption bed data to obtain virtual-real difference data; predict outlet oxygen concentration based on adsorption bed data to obtain outlet oxygen concentration prediction data; extract differential features based on outlet oxygen concentration prediction data and physical layer data to obtain prediction deviation data.

[0007] S3. Perform constrained convex optimization based on the virtual-real difference data and prediction deviation data to obtain the regeneration step size data;

[0008] S4. Based on the regeneration step size data, perform energy consumption simulation on the adsorption bed data to obtain energy constraint control data.

[0009] This invention improves the accuracy and controllability of the adsorption process by fusing physical layer sensing data with an adsorption bed simulation model. By constructing a graph-constrained correction path and an outlet oxygen concentration prediction model, it effectively identifies nonlinear differences between the adsorption state and actual operation. Combined with multi-scale differential feature extraction and prediction deviation analysis, it achieves dynamic sensing and error localization of oxygen concentration fluctuation trends. Based on residual-driven constrained convex optimization, it can adaptively adjust the regeneration step size to reduce system energy consumption while ensuring oxygen purity. Energy-constrained control is achieved using energy consumption simulation feedback, improving the energy efficiency ratio and operational stability of the oxygen separation process.

[0010] Optionally, the virtual simulation model construction includes:

[0011] Adsorption isotherms and capacity calibrations were performed based on the adsorption bed data to obtain calibration data.

[0012] Mass transfer flow parameters are inverted from the calibration data to obtain mass transfer flow data;

[0013] An energy coupling model is constructed based on mass transfer flow data to obtain a thermal coupling model;

[0014] The thermal coupling model is discretized to obtain a discretized model;

[0015] The adsorption bed model is obtained by solving the grid based on the discretized model.

[0016] This invention achieves accurate extraction of fundamental adsorption characteristic parameters by performing isothermal and capacity calibration on adsorption bed data, providing data support for subsequent modeling. Through the mass transfer and flow parameter inversion process, the diffusion and flow behavior of gas in porous media under different operating conditions can be effectively characterized, improving the model's fit to actual operating conditions. Combined with the construction of a thermally coupled model, the system can simultaneously consider the coupling relationship between mass transfer and thermal effects, enhancing the ability to express the impact of thermodynamic changes on adsorption performance. By discretizing and meshing the thermally coupled model, the numerical stability and solution accuracy of the simulation model are ensured. The resulting adsorption bed model can more realistically reflect the dynamic evolution characteristics of the adsorption process, improving system response efficiency and the feasibility of control strategies.

[0017] Optionally, the simulation calculation includes:

[0018] Boundary-driven reconstruction is performed based on physical layer data and adsorption bed model to obtain boundary reconstruction model;

[0019] Based on the boundary reconstruction model, collaborative coupling calculations are performed to obtain collaborative coupling data.

[0020] Based on the collaborative coupling data, simulation deviation correction data is obtained by performing deviation correction.

[0021] The deviation correction data are mapped to adsorption bed level variables to obtain adsorption bed data.

[0022] In this invention, boundary-driven reconstruction enables dynamic boundary condition updates to the adsorption bed model based on real-time sensor data collected at the physical layer, thereby enhancing the model's adaptability to changes in actual operating conditions. Cooperative coupling computation unifies the solution of multiple physical fields (such as mass transfer, heat conduction, and fluid dynamics), effectively improving the simulation accuracy of multi-factor interactions. The simulation deviation correction loop performed by the system can compare the residuals between measured data and simulation prediction results for in-model correction, thereby reducing simulation errors. Through adsorption bed hierarchical variable mapping operations, the correction results are finely mapped to each adsorption unit of the model, improving the simulation results in terms of spatial resolution and temporal dynamic tracking.

[0023] Optionally, the graph constraint calculation includes:

[0024] Based on the physical layer data and the adsorption bed data, a graph is constructed to obtain the structural physical graph data;

[0025] Graph coupling error mapping is performed on the structural physical graph data to obtain graph-level residual data;

[0026] Graph convolution propagation is performed on the graph-level residual data to obtain graph convolution data;

[0027] The amplitude of graph domain residual propagation is calculated based on graph convolution data to obtain virtual-real difference data.

[0028] This invention constructs structural physical graph data, transforming the spatial structure and variable relationships in physical layer data and adsorption bed data into a graph structure representation, thus preserving and modeling multivariate interaction relationships. Through graph-coupled error mapping, local residual features between model simulation and actual operation can be identified on the spatial topology, effectively revealing the structural distribution of model errors. Graph convolution propagation enables the layer-by-layer diffusion and feature enhancement of error information in the graph structure, which is beneficial for uncovering potential error correlation paths. Graph domain residual propagation amplitude calculation can accurately measure the error response intensity between different nodes and regions, thereby achieving high-resolution identification of virtual and real differences and improving the basic accuracy of model correction and control strategy optimization.

[0029] Optionally, the amplitude of the graph domain residual propagation includes:

[0030] Based on the graph convolution data, node residual vectors and region residual vectors are generated to obtain node residual data and region residual data respectively;

[0031] The residual data of nodes is obtained by averaging the residuals of adjacent nodes based on the regional residual data;

[0032] The residual transmission amplitude is calculated based on the weighted residual data to obtain the residual transmission amplitude data.

[0033] Based on the residual transmission amplitude data, the difference path is extracted from the node residual data and the regional residual data to obtain the node difference path data and the regional difference path data.

[0034] Based on the difference path data of nodes and the difference path data of regions, difference similarity processing is performed to obtain virtual and real difference data.

[0035] This invention achieves a detailed characterization of local disturbances and regional structural errors through dual-scale representation of node residuals and regional residuals. By averaging the residuals of adjacent nodes, it can integrate the error information of adjacent nodes in the local topology, avoiding the interference of isolated errors on the overall judgment and improving the stability and representativeness of the residual representation. The graph domain residual propagation amplitude calculation performed by the system not only preserves the spatial intensity characteristics of the original residuals but also realizes the directional propagation of error signals on the structure graph, effectively revealing the path and structural diffusion law of errors. Through difference path extraction and similarity processing, it can identify error clusters with consistent response characteristics, improving the pertinence and accuracy of difference compensation and model correction.

[0036] Optionally, the outlet oxygen concentration prediction includes:

[0037] Feature coding is performed based on the adsorption bed data to obtain feature-coded data;

[0038] Physical embedded network prediction is performed on the feature-encoded data to obtain physical network prediction data;

[0039] Confidence output is generated based on physical network prediction data, resulting in predicted outlet oxygen concentration data.

[0040] This invention utilizes feature encoding operations on adsorption bed data to structurally represent the spatiotemporal distribution characteristics and physical state variables during the adsorption process, providing high-dimensional and identifiable inputs for prediction. Modeling with a physically embedded network not only integrates the advantages of data-driven and physical constraints, overcoming the problems of traditional black-box models' sensitivity to boundary conditions and weak generalization ability, but also ensures that the prediction results possess physical interpretability and engineering consistency. The system performs confidence output processing, which can quantitatively estimate the prediction credibility of the model in different input spaces, effectively enhancing the model's stability and fault tolerance in complex dynamic environments.

[0041] Optionally, the differential feature extraction includes:

[0042] Multi-scale differential data is obtained by constructing a multi-scale differential model based on the predicted oxygen concentration at the outlet and the physical layer data.

[0043] The continuity of change direction and the difference drift ratio between adjacent intervals are calculated for multi-scale difference data to obtain continuity data and difference drift ratio data, respectively.

[0044] The coupling feature significance data is obtained by screening the coupling feature significance based on the coherence data and the difference drift ratio data;

[0045] Local error segments are extracted from the data with significant coupling features to obtain error segment data;

[0046] Multi-scale perturbation features are extracted from the error segment data to obtain the prediction deviation data.

[0047] This invention constructs multi-scale differential data to capture fine-grained deviation characteristics between predicted and physical observations of outlet oxygen concentration at different time scales. By using the coherence of the change direction and the difference drift ratio between adjacent intervals, it systematically identifies the offset trend and instability between prediction and measurement, thereby constructing a highly sensitive feature set that can distinguish perturbation trends. The coupled feature saliency screening step filters out irrelevant or redundant indicators in the multi-dimensional feature space, strengthening the expression of error-driven structural response patterns. Local error segment extraction and perturbation feature extraction focus on the dynamic evolution of significant error intervals, enhancing the system's ability to perceive deviation mechanisms under factors such as boundary changes, mass transfer perturbations, or abnormal loading.

[0048] Optionally, S3 includes:

[0049] The feasible region is constructed based on the difference between virtual and real data and the prediction deviation data to obtain the feasible region data;

[0050] Physical constraints are added to the feasible region data to obtain physical constraint data; perturbation constraints are added to the physical constraint data to obtain perturbation constraint data.

[0051] The perturbation constraint data is rewritten using a Lagrangian structure to obtain the constraint projection data.

[0052] Robust solution is performed based on the constraint projection data to obtain the regeneration step size data.

[0053] This invention constructs an optimized feasible region using virtual-real difference data and prediction deviation data, which can reasonably define the effective boundary of adjustable parameters based on the understanding of the adsorption bed state, avoiding the introduction of invalid or overfitting solutions. By gradually embedding physical constraints and perturbation constraints, it not only preserves the physical consistency of system operation and the rationality of boundary conditions, but also enhances the adaptability to dynamic disturbance factors (such as gas flow rate fluctuations, temperature disturbances, etc.). The Lagrange structure rewriting constructs a projection mapping of the constraint space based on the explicit expression of constraint conditions, making the convergence direction of the solution space physically interpretable and perturbation responsive. Based on the robust solution strategy after projection, the regeneration step size can be dynamically adjusted, so that the regeneration strategy still has stability, energy saving and response accuracy under operating conditions.

[0054] Optionally, S4 includes:

[0055] The regeneration step size vector is deconstructed based on the regeneration step size data to obtain the vector deconstruction data.

[0056] The stage performance consumption data is obtained by calculating the stage performance consumption data of the adsorption bed data based on the vector deconstruction data.

[0057] Based on the stage performance consumption data, cumulative simulation is performed to obtain energy constraint control data.

[0058] In this invention, by performing vector deconstruction on the regeneration step size data, the overall regeneration control strategy can be refined into specific time series and operational amplitude information, thereby accurately reproducing the actual execution behavior of each stage. By performing stage performance consumption calculation on the adsorption bed data, the energy consumption corresponding to different regeneration step sizes can be carefully evaluated, revealing the energy efficiency bottlenecks and optimization potential of each stage in the system operation. By performing cumulative simulation through stage performance consumption data, not only can the cumulative energy consumption trend be dynamically tracked, but also the system constraints can be combined to implement regulation, thereby forming an energy constraint control mechanism with feedback capability.

[0059] Optionally, this application also provides a low-power, high-purity oxygen separation system for performing the low-power, high-purity oxygen separation method described above, the low-power, high-purity oxygen separation system comprising:

[0060] The digital twin adsorption bed modeling module is used to acquire physical layer data and adsorption bed data; construct a virtual simulation model based on the adsorption bed data to obtain the adsorption bed model; and perform simulation calculations based on the physical layer data and the adsorption bed model to obtain the adsorption bed layer data.

[0061] The difference deviation extraction module is used to perform graph constraint correction based on physical layer data and adsorption bed data to obtain virtual-real difference data; to predict the outlet oxygen concentration based on adsorption bed data to obtain outlet oxygen concentration prediction data; and to extract differential features based on the outlet oxygen concentration prediction data and physical layer data to obtain prediction deviation data.

[0062] The constrained convex optimization module is used to perform constrained convex optimization based on the virtual-real difference data and the prediction deviation data to obtain the regeneration step size data.

[0063] The energy consumption simulation module is used to perform energy consumption simulation on the adsorption bed data based on the regeneration step size data, and obtain energy constraint control data.

[0064] The purpose of this invention is to construct a virtual simulation model integrating physical layer data and adsorption bed parameters in step S1, and to achieve high-precision simulation reconstruction of adsorption mass transfer behavior through boundary-driven and coupled solution, providing a reliable foundation for optimization. Next, in step S2, the structural differences between the actual operation and the simulation model are expressed as graphical residuals using graph constraints, and local prediction deviations and perturbation characteristics are accurately identified by combining oxygen concentration prediction and multi-scale difference extraction. Step S3 integrates the virtual-real difference and deviation data to construct a feasible region under both physical and perturbation constraints, and enhances the robustness of the solution through Lagrange structure rewriting, thereby obtaining the optimal regeneration step size that meets energy efficiency constraints. In step S4, based on the vector deconstruction of the regeneration step size and the staged energy consumption accumulation calculation, the energy consumption trend under different strategies is accurately simulated, forming an energy consumption prediction and control mechanism with closed-loop control capabilities. Attached Figure Description

[0065] Other features, objects, and advantages of this application will become more apparent from the following detailed description of the non-limiting embodiments, taken with reference to the accompanying drawings:

[0066] Figure 1 A flowchart illustrating the steps of a low-power, high-purity oxygen separation method according to one embodiment is shown.

[0067] Figure 2 A flowchart illustrating the steps of a virtual simulation model construction method according to one embodiment is shown.

[0068] Figure 3 A flowchart illustrating the steps of a graph constraint correction method according to an embodiment is shown.

[0069] Figure 4 A flowchart illustrating the steps of a constrained convex optimization method according to one embodiment is shown.

[0070] Figure 5 A flowchart illustrating the steps of an energy consumption simulation method according to one embodiment is shown.

[0071] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0072] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0073] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. Functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0074] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0075] Please see Figures 1 to 5 This application provides a low-power, high-purity oxygen separation method, the method comprising:

[0076] S1. Obtain physical layer data and adsorption bed data; construct a virtual simulation model based on the adsorption bed data to obtain the adsorption bed model; perform simulation calculations based on the physical layer data and the adsorption bed model to obtain the adsorption bed data.

[0077] Specifically, the system collects physical layer data including: inlet gas velocity, gas component molar concentration, inlet temperature, inlet pressure, effective bed length, and adsorbent particle size distribution; simultaneously, it collects adsorption bed data, including adsorbent particle specific surface area, porosity, adsorption material type, and corresponding isothermal adsorption model parameters (such as the saturation value and adsorption constant of the Langmuir model, or the exponential correction parameters of the Toth model). In the simulation modeling phase, the adsorption process is modeled. This modeling couples the mass transfer equation, momentum conservation equation, and energy conservation equation, where the mass transfer equation is expressed as: ,in Let be the gas phase concentration of the i-th gas. For time parameters, For the effective diffusion coefficient, The axial position of the bed. Let be the axial linear velocity of the gas. The porosity of the bed, For the first The amount of gas adsorbed per unit volume on the adsorbent is calculated. A multiphysics coupled model of the adsorption bed is constructed by simultaneously solving the heat conservation equation and the pressure drop equation using the above equations. The system divides the above continuous equations into spatial grids with a fixed step size (e.g., every 1 cm) along the spatial axis and discretizes them in time with a fixed step size (e.g., every 0.1 seconds). In the numerical solution process, the Crank-Nicolson method or the explicit finite difference method is used for calculation. The boundary conditions are set as Dirichlet boundaries based on the inlet concentration and Neumann flux-free boundaries at the bed outlet. The system outputs the gas concentration distribution, temperature field distribution, and pressure distribution within the adsorption bed at each time step within the simulation period, and generates adsorption bed data.

[0078] S2. Perform graph constraint calculations based on physical layer data and adsorption bed data to obtain virtual-real difference data; predict outlet oxygen concentration based on adsorption bed data to obtain outlet oxygen concentration prediction data; extract differential features based on outlet oxygen concentration prediction data and physical layer data to obtain prediction deviation data.

[0079] Specifically, a physical topology map of the adsorption bed is constructed based on the adsorption bed data, where nodes represent unit points obtained by spatially discrete partitioning of the bed (e.g., location points obtained by partitioning with a step size of 1 cm), and edges represent physical adjacency and mass transfer continuity relationships; for each node, its physical observation value is set to... The simulation output value is The difference between the two is denoted as the residual. The residuals of all nodes constitute a residual vector; a graph convolutional network is used to smooth the residual vector in the graph domain to obtain a smoothed residual vector. ,in Let be the adjacency matrix of the graph. For the node residual vector, Given a trainable convolutional weight matrix, ReLU is the activation function; extract the smoothed residual magnitude corresponding to each node. , serving as an indicator of the difference between virtual and real at that node;

[0080] The system extracts feature variables, such as the inlet oxygen concentration sequence, the temperature distribution sequence in the middle of the bed, and the pressure fluctuation trend within a recent period (e.g., the previous 10 seconds), and concatenates them into an input vector. This input vector is then fed into an integrated physical embedding module, which includes embedding strategies such as normalization for gas constants, heat ratios, and molar mass, and weighted aggregation of physical quantities. The embedded features are then fed into a multilayer perceptron (MLP) model for prediction, outputting a predicted oxygen outlet concentration. ,in This is the predicted value for the export oxygen concentration. It is a multilayer perceptron model. For physical embedding functions, The input vector is the predicted sequence of outlet oxygen concentration. Time difference is performed, and the difference order k can take multiple scales (e.g., 1, 2, 4 seconds). The corresponding k-th order difference is defined as: ,in Let be the k-th order difference value, representing the predicted change in concentration over the next k seconds. The predicted concentration value at time point t+k. At the current time point, It is the difference order. The predicted concentration value at time point t;

[0081] The system calculates the drift ratio index for each order of difference based on predicted outlet oxygen concentration data and physical layer data. ,in The k-th order drift ratio index For the k-th order difference value, This represents the (k-1)th order difference value. To calculate the coherence of the direction of change, in a predicted sequence of total time length T, if the difference values ​​at a certain time point t and the next time point t+1 have the same sign (i.e., the direction of change is consistent), the coherence increases. Specifically: ,in As a consistency indicator, To predict the total duration of the sequence, Index for the current time point, This is an indicator function that returns either 0 or 1. For symbolic functions, For the current k-th order difference, This is the k-th difference value at the next time step. This serves as the index for the current time point. Based on the drift ratio and coherence index, the largest and second-largest clusters are selected using threshold judgment or clustering calculations to identify time segments in the predicted sequence that exhibit significant fluctuations and inconsistent directions, thus serving as key error segments.

[0082] S3. Perform constrained convex optimization based on the virtual-real difference data and prediction deviation data to obtain the regeneration step size data;

[0083] Specifically, let the regeneration control variable be a step size vector, representing the unit adsorption time or pressure adjustment interval at different time steps; let the residual between virtual simulation and physical observation be... The difference index is composed of graph-constrained propagation; let the error between the predicted outlet oxygen concentration and the actual measured concentration be... The system calculates based on the following weighted residual objective function: ,in This is an operation to minimize the step size vector. The regeneration step size variable to be optimized. The residual weighting coefficient is the difference between virtual and real data. The prediction bias weighting coefficient is used. The system will adjust the step size within the range of physical feasibility and system stability, and set the following constraints: (1) Boundary constraints: Let the upper and lower limits of the step size be . and Then each item (No. (a step-size vector) should satisfy: This constraint prevents non-physical jumps in the adsorption process, such as overheating due to too short a time or adsorption saturation due to too long a time; (2) Mass transfer stability constraint: To avoid excessive fluctuations between adjacent time steps, the maximum increment of the step size change is set to Then the difference between any two adjacent step sizes must satisfy: This constraint prevents drastic fluctuations in the system response and maintains the asymptotic nature of the adsorption bed operation. Based on the above objective function and constraints, the system is calculated using a constraint optimization function in the form of Lagrange multipliers. ,in Let Lagrange be the objective function. Let Lagrange be the objective function. As a Lagrange multiplier of the upper realm, Let be the lower bound Lagrange multipliers. The function is optimized iteratively within the feasible region using the constrained projection method or the Lagrange-KKT conditions to obtain the optimal regeneration step size vector that satisfies the constraints. ,in To optimize the output of the obtained regeneration strategy, To minimize the solution within the feasible region, The optimal regeneration step size vector. The feasible solution domain is defined for all constraints.

[0084] S4. Based on the regeneration step size data, perform energy consumption simulation on the adsorption bed data to obtain energy constraint control data.

[0085] Specifically, the optimal regeneration step size vector obtained is deconstructed in stages and mapped to a stage control time series according to the time series. Each of them Indicates the adsorption bed system in the first... The regeneration duration of each stage (such as the operation time of desorption, purging, or backflushing stages) serves as the basis for energy consumption calculations. For each regeneration stage... Based on the physical parameters of the adsorption bed and the control strategy, the unit operating energy consumption for this stage is calculated. According to the operating power of the air compressor used in the current stage and the stage operating time, the energy consumed by the compressor is estimated as follows: ,in The energy consumption of the compressor in stage i is... In the stage Instantaneous compression power at any given moment The energy consumption of the compressor in stage i is... Let be the duration of stage i. If the regeneration process includes a hot stripping or heating step, then let be... Specific heat capacity of the gas; This refers to the regeneration temperature (regeneration gas or bed temperature). Where is the ambient temperature; m is the regeneration gas mass flow rate per unit time; then the energy consumption required for heating in this stage is: The total energy consumption for each stage is: The system accumulates energy consumption values ​​across all stages to obtain the stage energy consumption sequence and total energy consumption. Combining the current adsorption bed operating status, sensor feedback information, and historical energy consumption models, the system performs cumulative energy consumption simulation to determine whether the current regeneration step size control strategy meets the constraints of the energy consumption budget threshold and the target oxygen concentration achievement rate. If the cumulative energy consumption exceeds the set threshold, or the oxygen concentration fails to reach the target, the system dynamically generates energy constraint control data based on the aforementioned energy consumption feedback. This data includes the following instruction fields: the target oxygen concentration for the current stage (calculated by the system's oxygen concentration prediction module); the estimated cumulative energy consumption for the current stage (obtained from the simulation); the recommended regeneration step size adjustment value (used to reduce energy consumption while ensuring oxygen concentration); and control status flags (such as "maintain existing control" or "suggest shortening the stripping time"). This energy constraint control instruction can be pushed to the control execution module in real time.

[0086] Optionally, the virtual simulation model construction includes:

[0087] S11. Perform adsorption isotherm and capacity calibration based on adsorption bed data to obtain calibration data;

[0088] Specifically, the adsorption bed was experimentally operated under isothermal conditions, with several typical temperature points (e.g., 25°C, 35°C, 45°C) preferably set to correspond to different actual operating environments. At each temperature point, the partial pressure of the oxygen-nitrogen mixture was controlled, and the adsorption capacity of the adsorbent for the two gases was recorded under different oxygen and nitrogen partial pressures, constructing an adsorption isotherm dataset. Based on the above experimental data, the system underwent curve fitting. For any component i of oxygen or nitrogen (where i is... The adsorption capacity per unit mass of adsorbent With partial pressure The relationship between them is: ,in To achieve the specified voltage division Below, the adsorption capacity of a unit mass of adsorbent for component i is... Let be the maximum saturated adsorption capacity of the adsorbent for component i, and represent the theoretical maximum loading. Let be the adsorption affinity coefficient of component i, representing the affinity strength of that component for the adsorbent surface. Let be the partial pressure of component i in the gas mixture. Based on the fitting results of the above isothermal model, the system outputs a set of adsorption characteristic parameters for oxygen and nitrogen under different temperature conditions.

[0089] S12. Perform mass transfer flow parameter inversion on the calibration data to obtain mass transfer flow data;

[0090] Specifically, the system preferably establishes mass transfer kinetic equations based on a linear driving model or a void diffusion model to represent the diffusion mass transfer process of gas molecules on the surface of adsorbed particles. ,in The adsorption rate, This represents the actual adsorption capacity of the adsorbent for component i, expressed in units of 1 / s. This represents the equilibrium adsorption capacity of component i at the current partial pressure, expressed in mol / kg. This represents the actual adsorption capacity of the adsorbent for component i, expressed in mol / kg. Regarding flow behavior modeling, the system models the pressure drop within the adsorption bed, which is expressed as: ,in For unit pressure drop, The dynamic viscosity of the gas. The porosity of the bed, The diameter of the adsorbent particles is [missing information]. For gas density, Let be the axial linear velocity of the gas. After acquiring macroscopic experimental observation data or preliminary simulation output data, the system uses the following observations for parameter inversion: the pressure difference between the inlet and outlet of the adsorption bed; the adsorption breakthrough curve (i.e., the trend of component concentration changing at the outlet over time); and the mass transfer equilibrium time for the adsorption process to reach steady state. The system employs a least squares optimization algorithm (preferably the Levenberg-Marquardt nonlinear fitting method) to jointly invert and solve the above mass transfer and flow parameters, obtaining the parameter set that best matches the following model, including the effective mass transfer coefficient of each component; the average axial velocity of the gas within the bed; and the bed friction coefficient or pressure drop coefficient.

[0091] S13. Construct an energy coupling model based on mass transfer flow data to obtain a thermal coupling model;

[0092] Specifically, mass transfer is coupled with the energy equation to construct a complete non-isothermal adsorption kinetic model. The mass conservation equation (for gas phase components) is as follows: ,in Let be the concentration of the i-th component in the gas phase. For time parameters, Let be the effective diffusion coefficient of the i-th component. For spatial coordinates, The average gas velocity, The porosity of the bed, Let be the adsorption amount per unit mass of the i-th component. Let be the density of the solid per unit volume. The energy conservation equation is: ,in For gas density, The specific heat capacity of a gas at constant pressure. For temperature, For time parameters, Thermal conductivity, For spatial coordinates, The average gas velocity, For component ordinal terms, The adsorption heat of the i-th component is... Let be the adsorption amount per unit mass of the i-th component. The two conservation equations above are modeled by sharing the adsorption amount. and its time derivative Achieve tight coupling. The system presets the gas inlet boundary, such as setting a constant pressure and temperature input; and sets the wall boundary with adiabatic or heat exchange conditions.

[0093] S14. Discretize the thermal coupling model to obtain a discretized model;

[0094] Specifically, the system performs spatial discretization. The axial length of the adsorption bed is divided into equally spaced grid nodes, with one computational node set every 1 centimeter, for a total of N nodes. For the second-order spatial derivative terms in the thermally coupled model (e.g., the second-order derivative of temperature along the axial direction), a central difference formula is used for numerical approximation. The spatial derivatives of state variables such as gas concentrations and adsorption amounts of each component are also uniformly discretized using a central difference approximation. The system discretizes the time dimension. The system sets the time step to Δt; a time weighting factor is set between each time step and the next. ;set up =0.5, for each discrete node's state variable (such as temperature or gas concentration), its time variation term is numerically calculated according to the following format: ,in For the first The spatial discrete node at the th The state variable values ​​at each time step include any one of the following: bed temperature, gas phase component concentration, or solid phase adsorption amount. For the first The spatial discrete node at the th The corresponding state variable values ​​at each time step. The time step size is the interval between two adjacent discrete moments. As a time-weighted factor, For the first The source term function at each spatially discrete node is an algebraic function obtained by spatial discretization of the thermally coupled model. The source term function includes at least one or more of the following: heat conduction term, convection heat transfer term, adsorption heat release term, and gas-phase mass transfer term. In the first The vector set consisting of the state variables at all discrete nodes in space at each time step. For the first Each discrete moment in time. In the first The vector set consisting of the state variables at all discrete nodes in space at each time step. For the first At each discrete time point, the system establishes numerical iterative relationships for the following variables at each grid node: bed temperature at each node; gas phase component concentration at each node; solid phase adsorption at each node; and numerical source terms corresponding to each physical process (such as mass transfer, convection, and heat release).

[0095] S15. Solve the grid according to the discretized model to obtain the adsorption bed model.

[0096] Specifically, the system pre-sets the initial temperature distribution, initial gas concentration, and initial adsorption capacity of the adsorption bed based on data collected by physical sensors. The system uses iterative solvers such as Gauss-Seidel or BiCGStab to handle the coupled partial differential equations. Throughout the simulation period, the system continuously outputs the physical state parameters of each node in the adsorption bed, including gas component concentration vectors (e.g., oxygen and nitrogen concentrations), bed temperature values, and adsorption capacity at each node. With a total simulation duration of, for example, 60 seconds and a time step of 0.1 seconds, the number of iterations is 600.

[0097] Optionally, the simulation calculation includes:

[0098] Boundary-driven reconstruction is performed based on physical layer data and adsorption bed model to obtain boundary reconstruction model;

[0099] Specifically, the system dynamically adjusts the inlet and outlet boundary conditions of the adsorption bed model by combining real-time physical monitoring data to match the actual operating state. The system collects key physical measurement parameters during adsorption bed operation, including but not limited to inlet and outlet pressure, inlet gas temperature, inlet oxygen and nitrogen concentrations, and bed structure parameters. Based on the operating mode (pressurization, vacuuming, desorption), the system determines the boundary type as a constant pressure boundary, constant flow boundary, or adiabatic boundary. The system performs boundary reconstruction, such as directly setting real-time gas concentration data as the initial concentration value of the inlet node. For example, for the i-th component gas, its inlet concentration value is set to the input data at time t and continuously updated over time; the inlet gas temperature acquired by the sensor is set as the initial temperature value of the boundary node, while the outlet boundary is set as an adiabatic boundary, i.e., the temperature spatial derivative is zero.

[0100] Based on the boundary reconstruction model, collaborative coupling calculations are performed to obtain collaborative coupling data.

[0101] Specifically, after setting the boundary driving conditions, the gas-solid mass transfer, energy exchange, and concentration distribution during the adsorption process are dynamically solved over the entire domain. The system performs joint modeling and simultaneous solution based on the mass transfer coupling dimension and the energy coupling dimension. The system establishes a set of coupled control equations to simultaneously describe the evolution of gas concentration and bed temperature, namely:

[0102] ;

[0103] in Let be the gas phase concentration (i-th component). For time parameters, The axial flow velocity is... These are spatial coordinates, representing the spatial position along the length of the bed. is the axial diffusion coefficient (for the i-th component). The porosity of the bed, denoted as the solid phase adsorption amount (i-th component). For temperature, Where is the thermal diffusivity, For component index variables, The heat of adsorption (for the i-th component). For gas density, The specific heat at constant pressure of the gas is used. The Crank-Nicolson semi-implicit method with a time step of 0.05 s is used for solution; spatial discretization is performed using an equally spaced grid of 1 cm, and collaborative iterative updates are used to obtain collaboratively coupled data.

[0104] Based on the collaborative coupling data, simulation deviation correction data is obtained by performing deviation correction.

[0105] Specifically, residual analysis is performed between the model output (such as the outlet oxygen concentration) and the physical layer measurements: ,in For oxygen concentration residual, To predict oxygen concentration for the model, To predict oxygen concentration for the model. To evaluate the mass transfer coefficient in the cooperative coupling model. Add a residual compensation term to the initial concentration distribution or boundary temperature: ,in This is the corrected mass transfer coefficient. The original mass transfer coefficient, To control the learning rate, the adjustment range is controlled. This represents the residual oxygen concentration. Alternatively, error minimization optimization can be performed through a parameter calibration module, using Kalman filtering or a genetic algorithm to adjust the model parameters.

[0106] The deviation correction data are mapped to adsorption bed level variables to obtain adsorption bed data.

[0107] Specifically, based on deviation correction data, the system spatially stratifies the adsorption bed along the axial direction and extracts physical quantity indicators in each segment. The system divides the adsorption bed into multiple continuous spatial segments at equal intervals along its main axis (i.e., the gas flow direction). Within each segment, the system performs integral averaging calculations of physical variables based on the corrected temperature distribution, gas concentration distribution, and adsorption load distribution data.

[0108] Optionally, the graph constraint calculation includes:

[0109] S21. Construct a diagram based on the physical layer data and the adsorption bed data to obtain the structural physical diagram data;

[0110] Specifically, the physical states of each segment of the adsorption bed in space are mapped as nodes, and the physical dependencies between them (such as heat transfer and mass transfer) are mapped as edges, thus constructing a structural physics graph. The system refines the adsorption bed along the axial direction (i.e., the main airflow direction) as the node constituent units of the graph, that is, the adsorption bed is divided into N equally spaced spatial units, for example, each unit is 5 cm. Each spatial unit is a graph node, and the physical state information corresponding to each node includes the following parameters, which constitute the node feature vector: oxygen concentration (unit: mole fraction or volume fraction); nitrogen concentration; temperature (unit: degrees Celsius or Kelvin); oxygen adsorption load in the solid phase (unit: mol / kg); nitrogen adsorption load in the solid phase; pressure value (unit: Pascal). For each pair of adjacent spatial units, a bidirectional connection edge is established in the graph; for specific cases, such as cross-segment channels between inlet and outlet, return paths, etc., edges are established between non-adjacent nodes; the edges are accompanied by the following physical characteristic indicators, such as spatial spacing, effective diffusion coefficient, thermal conductivity, and local velocity.

[0111] S22. Perform graph coupling error mapping on the structural physical graph data to obtain graph-level residual data;

[0112] Specifically, the system performs node-level residual calculations on the differences between real-time physical layer data and simulation data, and propagates this calculation to the graph structure to obtain the structural deviation distribution. The system performs difference processing on the real-time monitoring data and simulation output of each node to obtain a residual vector. The system then performs modulo operations on the residual vectors of each node to obtain the overall residual strength index. Finally, the system sequentially combines the residual strength values ​​of all nodes to construct a graph-level residual matrix.

[0113] S23. Perform graph convolution propagation on the graph-level residual data to obtain graph convolution data;

[0114] Specifically, residual information is propagated through a graph convolutional neural network to obtain graph convolutional data. Graph convolution propagation: ,in For the first The node feature representation matrix of the layer, For activation function, For degree matrix, It is an adjacency matrix. The node feature representation matrix of the l-th layer, i.e., the graph-level residual data. Let F be the weight matrix of the l-th layer. The system adopts a 2-layer graph convolutional propagation structure to cover the local and second-adjacent propagation ranges; the output dimension is set to F=8 or F=16, representing the convolutional representation dimension of each node.

[0115] S24. Calculate the amplitude of graph domain residual propagation based on graph convolution data to obtain virtual-real difference data.

[0116] Specifically, the residual propagation strength after convolution is extracted to construct a systematic metric for the difference between virtual and real data. The norm of the convolution output vector at each node represents the residual information propagation strength. ,in Let be the residual propagation amplitude of the i-th node. Let be the graph convolution output vector of the i-th node. The system performs statistical analysis on the conduction amplitude data of all nodes and sets dynamic thresholds to identify abnormal regions. Specifically, this includes calculating the overall mean and standard deviation of the conduction amplitude of all nodes; setting the mean plus one standard deviation as the discrimination threshold; and identifying nodes with conduction amplitudes greater than this threshold as regions with significant virtual-real differences, i.e., risk nodes where the simulation model deviates from the true state. The system combines the residual conduction amplitudes of all nodes into a one-dimensional vector, the length of which is equal to the total number of nodes in the graph structure.

[0117] Optionally, the amplitude of the graph domain residual propagation includes:

[0118] Based on the graph convolution data, node residual vectors and region residual vectors are generated to obtain node residual data and region residual data respectively;

[0119] Specifically, based on the output of the graph convolutional neural network, the system generates a corresponding residual feature vector for each graph node. Then, it aggregates multiple nodes according to a preset region division rule to generate region-level residual feature vectors. The system acquires the output data of the graph convolutional neural network, denoted as a matrix of graph convolution output results. In this matrix, each row represents the convolution output of a node, i.e., the residual feature vector of the corresponding node. Specifically, for the i-th node, its residual feature vector can be represented as an F-dimensional vector; the values ​​of each dimension of the vector represent the different characteristic responses of the node during the propagation process of the graph neural network (such as thermal anomalies, gas concentration change rates, mass transfer dissimilatory degrees, etc.); the node residual vector is used to characterize the local deviation pattern and propagation intensity at that location. The system organizes the residual vectors of all nodes according to the node number order to form a complete node residual dataset. The system divides the nodes in the graph into several regions and aggregates the node residual vectors within each region to generate region residual feature vectors. The system divides the adsorption bed into several segments along its axis, with each 10 consecutive nodes forming a region, resulting in multiple regional units. For each region, the system fuses the residual vectors of all nodes within it using a dimensional mean approach, i.e., for each dimension, the average value of the residual values ​​of all nodes in that dimension within the region is taken, resulting in a region-level residual vector that is also F-dimensional.

[0120] The residual data of nodes is obtained by averaging the residuals of adjacent nodes based on the regional residual data;

[0121] Specifically, for any node in the graph structure, the system constructs a neighbor set consisting of all its adjacent nodes. The system then performs a dimension-wise averaging of the residual vectors of all nodes in this neighbor set to obtain the node's average neighbor residual vector. The arithmetic mean of the residual vectors of all nodes in the neighbor set is then calculated; that is, for each dimension, the mean residual value of all adjacent nodes in that dimension is calculated. The resulting vector has the same dimension as the node residual vector. For each node, a weighted fusion value of its own residual and the average neighbor residual is calculated. ,in Let be the weighted residual vector of the i-th node. These are the residual fusion weighting coefficients. Let be the original residual vector of the i-th node. Let be the average residual vector of the adjacency set of the i-th node. The system performs the above weighting operation on all nodes to obtain a set of residual weighted datasets.

[0122] The residual transmission amplitude is calculated based on the weighted residual data to obtain the residual transmission amplitude data.

[0123] Specifically, for each node's weighted residual vector, its norm is calculated as the propagation amplitude: ,in Let be the residual propagation amplitude at node i. Let be the weighted residual vector of the i-th node. Let f be the residual value of the i-th node in the f-th feature dimension. Indexed by feature dimensions, This represents the total number of residual feature dimensions. The system sequentially performs the above residual norm calculation on all nodes, resulting in a one-dimensional vector composed of the residual transmission amplitudes of all nodes, i.e., the residual transmission amplitude data.

[0124] Based on the residual transmission amplitude data, the difference path is extracted from the node residual data and the regional residual data to obtain the node difference path data and the regional difference path data.

[0125] Specifically, setting dynamic thresholds ,in , The mean and standard deviation of the propagation amplitude of all nodes; if the residual propagation amplitude of a certain node is... If a node is identified as a high residual propagation node, then the system uses all high residual propagation nodes as starting nodes. The system performs a depth-first search (DFS) or breadth-first search (BFS) algorithm in the graph structure to find links between high residual nodes connected in the graph topology, forming high residual propagation paths between nodes. The system only retains links with a path length greater than or equal to 2 to ensure the continuity and propagation significance of the extracted paths. The set of extracted paths constitutes node difference path data. For each of the aforementioned regions, the norm of its corresponding regional residual vector is calculated, and a dynamic threshold at the region level is set, for example, by adding the standard deviation to the mean of the residual magnitudes of all regions. Regions that meet this threshold are identified as high residual propagation regions. For multiple high residual propagation regions, the system determines whether there is a spatial proximity relationship or a graph topological adjacency relationship between them (e.g., located in adjacent layers, adjacent modules, or connected by nodes in the structure graph). If an adjacency relationship exists, the connection between these regions constitutes a regional difference path, indicating that the error has a cross-regional diffusion trend.

[0126] Based on the difference path data of nodes and the difference path data of regions, difference similarity processing is performed to obtain virtual and real difference data.

[0127] Specifically, for each unit contained in the node difference path and the regional difference path, its structural location, parameter response type (such as temperature-type residual or velocity-type residual), and direction information (upstream → downstream or countercurrent) are extracted. Clustering algorithms (such as KMeans or DBSCAN) are used to cluster these path feature vectors to find clusters of difference types. Each cluster represents a typical virtual-real inconsistency pattern. The path labels or feature indexes of each type of cluster are merged to generate a virtual-real difference data structure, including difference type labels (Type A: structural hysteresis type, Type B: mass transfer error type, etc.); node or region numbers of the affected range; corresponding physical quantity terms and difference intensity values.

[0128] Optionally, the outlet oxygen concentration prediction includes:

[0129] Feature coding is performed based on the adsorption bed data to obtain feature-coded data;

[0130] Specifically, the input data includes the inlet / outlet pressure of each bed, adsorption time, bed temperature distribution, adsorbent type and particle size parameters, and historical oxygen concentration from the previous cycle. For continuous variables (such as temperature and pressure), normalization or standardization is used; for periodic behaviors (such as cycle number and flow stage), one-hot encoding or cycle position encoding is used.

[0131] Physical embedded network prediction is performed on the feature-encoded data to obtain physical network prediction data;

[0132] Specifically, the input vector is the feature encoding data generated in the previous step; it is input into a hidden layer structure containing several fully connected hidden layers, each using the ReLU (Modified Linear Unit) activation function, and processed by a Dropout random deactivation module. During neural network training, the system sets a set of physical consistency loss terms: ,in For physical consistency loss, This is a predicted value for oxygen concentration. For time variables, The system gas flow rate (can be set to constant or input variable). For the spatial location of the bed, Here, represents the diffusion coefficient (an adsorbent-related constant). The system's training objective function is a weighted sum of the following two parts: ,in For the total loss function, This represents the mean square error between the predicted and observed oxygen concentration values. For predicted values, For the true value, The weighting coefficient for physical consistency loss. This refers to the physical consistency loss defined above. After training and optimization, the system outputs the oxygen concentration prediction results.

[0133] Confidence output is generated based on physical network prediction data, resulting in predicted outlet oxygen concentration data.

[0134] Specifically, while keeping the Dropout layer enabled in the neural network, n independent forward propagation processes (e.g., n=30) are performed on the same input feature encoded data to obtain n predicted sequences of outlet oxygen concentration. The average of these n predictions is then used as the predicted outlet oxygen concentration. Based on the variance of the predictions, the system calculates the error bound at a 95% confidence level, expressed as the confidence interval amplitude. ,in The confidence interval amplitude, To predict the number of times, For the i-th prediction index, Let i be the predicted outlet oxygen concentration value. The predicted mean of the export oxygen concentration is given. The predicted mean and confidence interval amplitudes are combined to form a structured prediction result.

[0135] Optionally, the differential feature extraction includes:

[0136] Multi-scale differential data is obtained by constructing a multi-scale differential model based on the predicted oxygen concentration at the outlet and the physical layer data.

[0137] Specifically, the system first acquires a continuous predicted sequence of outlet oxygen concentration. Multiple time-granularity windows are set (e.g., 1 second, 10 seconds, 60 seconds), and the sliding difference is calculated at each time scale. The system performs differential operations at the same time granularity for key physical parameters of the adsorption bed, including temperature, pressure, and gas flow rate. The system combines features such as oxygen concentration differences, temperature differences, pressure differences, and flow rate differences at different time scales to construct a multi-scale difference tensor.

[0138] The continuity of change direction and the difference drift ratio between adjacent intervals are calculated for multi-scale difference data to obtain continuity data and difference drift ratio data, respectively.

[0139] Specifically, the calculation of the continuity of the direction of change: ,in For the continuity of the direction of change, The number of differential samples, For differential indexes, For indicator functions, The sign of the i-th difference term (positive value is +1, negative value is -1, and zero is 0). For the i-th difference term, For the first The sign of each difference term (positive value is +1, negative value is -1, zero is 0). Difference drift ratio between adjacent intervals: ,in This is the differential drift ratio. This is the current difference value. This is the time difference score from the previous time step. It is a function with maximum value. To prevent the minimum value of the constant when divided by zero (such as...) ).

[0140] The coupling feature significance data is obtained by screening the coupling feature significance based on the coherence data and the difference drift ratio data;

[0141] Specifically, the feature coupling significance index is calculated. ,in The feature coupling significance index, This is the consistency weighting coefficient. For consistency data, The differential drift ratio weighting coefficient is used. It is a natural constant. This is the differential drift ratio data. Set the coupling threshold. For example, 0.7-0.9, if If a time point is identified as a significant change point, then the system will output the selected significant feature point information as the following two types of data: a significant change time index set, which records the time step numbers of all points identified as significant; and a significant coupled feature tensor, which contains the feature data corresponding to each significant time point, including difference values, physical state values, and prediction residuals.

[0142] Local error segments are extracted from the data with significant coupling features to obtain error segment data;

[0143] Specifically, the system scans the obtained significant change time index set in chronological order and merges it into several time segments according to the following rules: If the time interval between any two adjacent significant time points is less than a preset threshold (e.g., 5 seconds), they are determined to be in the same segment; multiple local time segments are obtained through continuous merging, and the time points within each segment constitute a set of candidate error evaluation intervals. For each merged time segment, the mean square error between the predicted oxygen concentration value and the actual physical oxygen concentration value within that segment is calculated. The system calculates the average prediction error globally and uses it as a benchmark value. Only when the mean square error of a segment is greater than the average prediction error is the segment determined to be a valid error segment and retained. The system outputs all error segments that meet the conditions in list form, forming an error segment dataset.

[0144] Multi-scale perturbation features are extracted from the error segment data to obtain the prediction deviation data.

[0145] Specifically, within each error segment, the following perturbation features are extracted, such as local range calculation. ,in For local range, The maximum predicted value, The maximum predicted value, This is a predicted value for oxygen concentration. For time indexing, The system defines the k-th error segment. It performs a Fast Fourier Transform on the predicted value sequence within the segment and extracts the main peak value from the spectral amplitude to obtain the local spectral power peak. The system presets multiple time scale windows and calculates the sliding window variance of the oxygen concentration prediction values ​​at different scales. For all error segments, the system splices and integrates the perturbation features extracted from each segment to form a high-dimensional difference feature matrix.

[0146] Optionally, S3 includes:

[0147] S31. Construct the feasible region based on the virtual-real difference data and the prediction deviation data to obtain the feasible region data;

[0148] Specifically, the virtual-to-real difference data and prediction bias data are concatenated / merged into a joint migration space. Based on all observation points in the joint migration space, the system defines the feasible region as follows: ,in For the set of feasible regions, The vector of variables after perturbation. For the real number space, To optimize variable dimensions, For simulation of nominal values, The feasible region dynamic shrinkage factor can be adjusted based on the error variance, such as... , For adjustment coefficients, For variance, The variance of the joint offset space. The generated feasible region data includes a set of dynamic boundaries in a multidimensional variable space; each boundary set corresponds to the constraint range of a class of key model inputs or intermediate state variables.

[0149] S32. Add physical constraints to the feasible region data to obtain physical constraint data; add perturbation constraints to the physical constraint data to obtain perturbation constraint data.

[0150] Specifically, based on the operational boundaries and engineering constraints of the target system, the system defines the following set of physical constraint expressions to define the physical feasibility range of the constraint variables, such as pressure limit constraints (setting the system operating pressure between the upper and lower limits of safe pressure); temperature stability constraints (the temperature change between adjacent time steps must be less than or equal to the maximum allowable temperature fluctuation threshold); and regeneration time constraints (the duration of each system regeneration cycle must be greater than or equal to its minimum physical operating limit). These constraints are formalized into a set of inequalities and fused with the feasible region reconstruction expression. This yields feasible region data containing specific physical upper and lower limits and operational boundaries.

[0151] The system constructs perturbation predictions, representing the variable's tolerance to perturbations at future times. Its construction is based on the standard deviation estimate within the historical window and the current trend drift calculation, as follows: ,in For the disturbance tolerance bandwidth, Let N be the standard deviation of the variable over the past N time points. For the historical sliding window variable sequence, This is the tolerance adjustment amplification factor (e.g., 1.5, 2.0). This represents the rate of change of the variable's trend (e.g., the moving average difference or the absolute value of the first difference). Based on the above perturbation predictions, for each optimization variable... Add the following bandwidth range: , and Let be the minimum and maximum values ​​of the quantities within the physically feasible region. The system output is the optimization variable constraint space with interference tolerance bandwidth.

[0152] S33. Rewrite the perturbation constraint data using a Lagrangian structure to obtain the constraint projection data;

[0153] Specifically, the system performs the Lagrange objective function calculation: ,in For Lagrange functions, The objective function is, for example, minimum energy consumption or maximum oxygen concentration. To optimize the variable vector, For constraint numbering index, For the Lagrange multiplier corresponding to the i-th constraint, These include perturbation and physical constraints, such as pressure range limits, temperature fluctuation limits, regeneration cycle lower limits, and perturbation tolerance bandwidth. The system solves the problem based on the KKT conditions in nonlinear constraint optimization, among which necessary conditions include... ,in Let be the gradient vector of the Lagrangian function with respect to the optimization variable x. Based on the above, a numerical optimization method is used to project the variable space: in the high-dimensional space, the search path of the variable is projected onto the constraint space to satisfy all perturbations and physical constraints; the output is the constraint projection solution structure, which is the set of all optimization variable solutions that satisfy the perturbation constraints and physical constraints.

[0154] S34. Perform robust calculations based on the constraint projection data to obtain the regeneration step size data.

[0155] Specifically, under constrained projection, robust optimization solvers (such as SOCP and interior point methods) are used to solve the following problems: ,in The minimize operator indicates that the goal of this optimization problem is to find a set of values ​​for the variable x that minimizes the objective function f(x). To optimize the variable vector, let represent the set of control parameters to be solved. Let be the objective function. To satisfy the following constraints, The feasible region of the perturbation constraint projection refers to the set of all feasible variables that satisfy the Lagrangian constraints, and the objective function is... This is the energy consumption or oxygen purity deviation function, i.e., the mean square error between the target oxygen concentration and the simulation prediction. In the solution results, the system extracts the solution quantity with respect to the time dimension from the optimization variable vector, denoted as the regeneration step size vector, which represents the optimal allocation of regeneration time for each stage.

[0156] Optionally, S4 includes:

[0157] S41. Deconstruct the regeneration step size vector based on the regeneration step size data to obtain the vector deconstruction data;

[0158] Specifically, each element in the regeneration step length vector represents the optimal control duration for the corresponding time period during the regeneration process; the vector length corresponds to the total number of time segments within a complete adsorption-regeneration cycle. Based on the adsorption-desorption process structure, the system decomposes the overall vector into several stage sub-vectors with engineering semantics, including but not limited to: adsorption stage step length sub-vectors (e.g., high-pressure adsorption time period); desorption stage step length sub-vectors (e.g., low-pressure evacuation and gas release); cleaning stage step length sub-vectors (e.g., purging with inert gas); and pressurization stage step length sub-vectors (e.g., compressed gas backfilling or pre-filling). The above stage splicing operation can be represented as the regeneration step length vector being composed of the spliced ​​stages. Within each stage sub-vector, the system further refines the time segments according to the operation type. For example, the adsorption stage can be subdivided into multiple process units such as high-pressure isothermal adsorption and instantaneous ventilation replenishment; the desorption stage can be subdivided into specific steps such as evacuation and condensation recovery; and the cleaning stage can include various operation strategies such as forward washing and backwashing. Each sub-vector is further decomposed into several step length units, forming the smallest control time period granularity. The output is a set of time control vector data with completed stage labeling and granularity division.

[0159] S42. Calculate the stage performance consumption of the adsorption bed data based on the vector deconstruction data to obtain the stage performance consumption data.

[0160] Specifically, for each stage, the system acquires the adsorption bed operation response data for the corresponding time period, including heating power per unit time; vacuum pump power; compressor power; and synchronously acquired state variables such as bed pressure drop curves and temperature change curves. For each stage, the energy consumption value is calculated using the following physical integral expression: ,in Let be the energy consumption value for stage i, representing the total energy consumed by the adsorption bed during the operation of this stage. Let be the starting time point of the i-th stage, representing the starting point of the energy consumption integral for that stage. Let be the duration (step size) of the i-th stage. The power consumed for heating per unit time. This refers to the power consumption of the vacuum pump during the vacuuming process. This refers to the power consumed by the compressor during the gas compression process. is the integration variable, representing any moment on the continuous time axis. The system output includes energy consumption assessment calculation results for multiple stages, yielding stage performance consumption data.

[0161] S43. Perform cumulative simulation based on the stage performance consumption data to obtain energy constraint control data.

[0162] Specifically, the system calculates the total energy consumption over the entire lifecycle as the sum of the results from each stage: ,in Total energy consumption over the entire cycle represents the cumulative energy consumption of the oxygen separation system during a complete operating cycle. This serves as a stage index, used to identify the i-th operational stage (e.g., adsorption, desorption, evacuation, pressurization, etc.) in the oxygen separation process. The total number of stages represents the number of stages divided into in a complete operating cycle. This represents the energy consumption value for stage i. The system sets a maximum energy consumption upper limit, requiring that the total energy consumption within the simulation cycle must not exceed this upper limit. The system sets a minimum target purity threshold, requiring that the predicted oxygen outlet purity for the current cycle is not lower than this threshold. The system employs a rolling time window simulation mechanism combined with a step size adjustment strategy. The entire oxygen separation process is divided into several time windows, and stage simulations are performed within each window, while simultaneously predicting the trends of subsequent stages. If the constraints of the upper limit of energy consumption or the lower limit of purity are not met within any simulation cycle, the system will fine-tune the control parameters (especially the time step) of the relevant stage. For stages that do not meet the constraints, the system executes the following callback strategy: the current stage step size parameter is increased or decreased by a small percentage (e.g., ±5%); the simulation of that stage is re-performed, and the impact on the total energy consumption curve and outlet gas purity is observed; if the multi-objective constraints are met, the step size configuration is updated; otherwise, the iterative callback continues. The system obtains a set of control parameter data optimized under constraints, including the optimal step size configuration, predicted purity index and corresponding energy consumption assessment results for each stage; this output is defined as energy-constrained control data.

[0163] Optionally, this application also provides a low-power, high-purity oxygen separation system for performing the low-power, high-purity oxygen separation method described above, the low-power, high-purity oxygen separation system comprising:

[0164] The digital twin adsorption bed modeling module is used to acquire physical layer data and adsorption bed data; construct a virtual simulation model based on the adsorption bed data to obtain the adsorption bed model; and perform simulation calculations based on the physical layer data and the adsorption bed model to obtain the adsorption bed layer data.

[0165] The difference deviation extraction module is used to perform graph constraint correction based on physical layer data and adsorption bed data to obtain virtual-real difference data; to predict the outlet oxygen concentration based on adsorption bed data to obtain outlet oxygen concentration prediction data; and to extract differential features based on the outlet oxygen concentration prediction data and physical layer data to obtain prediction deviation data.

[0166] The constrained convex optimization module is used to perform constrained convex optimization based on the virtual-real difference data and the prediction deviation data to obtain the regeneration step size data.

[0167] The energy consumption simulation module is used to perform energy consumption simulation on the adsorption bed data based on the regeneration step size data, and obtain energy constraint control data.

[0168] Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended application documents rather than the foregoing description. Thus, it is intended that all variations falling within the meaning and scope of the equivalents of the application documents be incorporated into the invention.

[0169] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A low-power, high-purity oxygen separation method, characterized in that, The method includes: S1. Obtain physical layer data and adsorption bed data; construct a virtual simulation model based on the adsorption bed data to obtain the adsorption bed model; perform simulation calculations based on the physical layer data and the adsorption bed model to obtain the adsorption bed data. S2. Perform graph constraint calculations based on physical layer data and adsorption bed data to obtain virtual-real difference data; predict outlet oxygen concentration based on adsorption bed data to obtain outlet oxygen concentration prediction data; extract differential features based on outlet oxygen concentration prediction data and physical layer data to obtain prediction deviation data. S3. Perform constrained convex optimization based on the virtual-real difference data and prediction deviation data to obtain the regeneration step size data; S4. Based on the regeneration step size data, perform energy consumption simulation on the adsorption bed data to obtain energy constraint control data, and push it to the control execution module in real time. The graph constraint calculation includes: Based on the physical layer data and the adsorption bed data, a graph is constructed to obtain the structural physical graph data. The graph construction includes mapping each segment of the physical state of the adsorption bed in space to nodes, mapping the physical dependencies between them to edges, and constructing the structural physical graph. The system performs graph coupling error mapping on the structural physical graph data to obtain graph-level residual data. The graph coupling error mapping includes calculating the residual at the node level between the real-time data and simulation data of the physical layer and propagating it to the graph structure to obtain the structural deviation distribution. The system performs difference processing on the real-time monitoring data and simulation output of each node to obtain the residual vector. The system performs modulus operation on the residual vector of each node to obtain the overall residual strength index. The system combines the residual strength values ​​of all nodes in order to construct the graph-level residual matrix. Graph convolution propagation is performed on the graph-level residual data to obtain graph convolution data; The graph domain residual propagation amplitude is calculated based on the graph convolution data to obtain the virtual-real difference data; S3 includes: The feasible region is constructed based on the difference between virtual and real data and the prediction deviation data to obtain the feasible region data; Physical constraints are added to the feasible region data to obtain physical constraint data; perturbation constraints are added to the physical constraint data to obtain perturbation constraint data. The perturbation constraint data is rewritten using a Lagrangian structure to obtain the constraint projection data. Robust solution is performed based on the constraint projection data to obtain the regeneration step size data.

2. The method according to claim 1, characterized in that, The virtual simulation model construction includes: Adsorption isotherms and capacity calibrations were performed based on the adsorption bed data to obtain calibration data. Mass transfer flow parameters are inverted from the calibration data to obtain mass transfer flow data; An energy coupling model is constructed based on mass transfer flow data to obtain a thermal coupling model; The thermal coupling model is discretized to obtain a discretized model; The adsorption bed model is obtained by solving the grid based on the discretized model.

3. The method according to claim 1, characterized in that, The simulation calculations include: Boundary-driven reconstruction is performed based on physical layer data and adsorption bed model to obtain boundary reconstruction model; Based on the boundary reconstruction model, collaborative coupling calculations are performed to obtain collaborative coupling data. Based on the cooperative coupling data, simulation deviation correction is performed to obtain deviation correction data; The deviation correction data are mapped to adsorption bed level variables to obtain adsorption bed data.

4. The method according to claim 1, characterized in that, The amplitude of the residual propagation in the graphic domain includes: Based on the graph convolution data, node residual vectors and region residual vectors are generated to obtain node residual data and region residual data respectively; The residual data of nodes is obtained by averaging the residuals of adjacent nodes based on the regional residual data; The residual transmission amplitude is calculated based on the weighted residual data to obtain the residual transmission amplitude data. Based on the residual transmission amplitude data, the difference path is extracted from the node residual data and the regional residual data to obtain the node difference path data and the regional difference path data. Based on the difference path data of nodes and the difference path data of regions, difference similarity processing is performed to obtain virtual and real difference data.

5. The method according to claim 1, characterized in that, The predicted outlet oxygen concentration includes: Feature coding is performed based on the adsorption bed data to obtain feature-coded data; Physical embedded network prediction is performed on the feature-encoded data to obtain physical network prediction data; Based on the physical network prediction data, confidence output is performed to obtain the predicted oxygen concentration at the outlet.

6. The method according to claim 1, characterized in that, The differential feature extraction includes: Multi-scale differential data is obtained by constructing a multi-scale differential model based on the predicted oxygen concentration at the outlet and the physical layer data. The continuity of change direction and the difference drift ratio between adjacent intervals are calculated for multi-scale difference data to obtain continuity data and difference drift ratio data, respectively. The coupling feature significance data is obtained by screening the coupling feature significance based on the coherence data and the difference drift ratio data; Local error segments are extracted from the data with significant coupling features to obtain error segment data; Multi-scale perturbation features are extracted from the error segment data to obtain the prediction deviation data.

7. The method according to claim 1, characterized in that, S4 include: The regeneration step size vector is deconstructed based on the regeneration step size data to obtain the vector deconstruction data. The stage performance consumption data is obtained by calculating the stage performance consumption data of the adsorption bed data based on the vector deconstruction data. Based on the stage performance consumption data, cumulative simulation is performed to obtain energy constraint control data.

8. A low-power, high-purity oxygen separation system, characterized in that, For performing the low-power, high-purity oxygen separation method as described in claim 1, the low-power, high-purity oxygen separation system comprises: The digital twin adsorption bed modeling module is used to acquire physical layer data and adsorption bed data; construct a virtual simulation model based on the adsorption bed data to obtain the adsorption bed model; and perform simulation calculations based on the physical layer data and the adsorption bed model to obtain the adsorption bed layer data. The difference deviation extraction module is used to perform graph constraint correction based on physical layer data and adsorption bed data to obtain virtual-real difference data; to predict the outlet oxygen concentration based on adsorption bed data to obtain outlet oxygen concentration prediction data; and to extract differential features based on the outlet oxygen concentration prediction data and physical layer data to obtain prediction deviation data. The constrained convex optimization module is used to perform constrained convex optimization based on the virtual-real difference data and the prediction deviation data to obtain the regeneration step size data. The energy consumption simulation module is used to perform energy consumption simulation on the adsorption bed data based on the regeneration step size data, and obtain energy constraint control data.

Images