Intelligent control method for supercritical CO2 injection based on AI model

Abstract

The invention discloses a supercritical CO2 injection intelligent regulation and control method based on an AI model, and relates to the technical field of fluid injection regulation and control, and the method comprises the steps: collecting real-time multi-source monitoring data, constructing a high-precision digital twinborn model, inputting the real-time multi-source monitoring data into the high-precision digital twinborn model for dynamic simulation, and outputting an underground pressure field and a CO2 concentration field; the control instruction set is issued to an electric control valve and a high-pressure pump of the injection well and an extraction pump of the extraction well, the injection pressure, the injection flow rate and the extraction pump power of CO2 are regulated and controlled, and a regulation and control event record is generated; and on the basis of regulation and control event records, acquiring updated real-time multi-source monitoring data, and feeding back the updated real-time multi-source monitoring data to the high-precision digital twinborn model to form closed-loop regulation and control. By integrating the variation assimilation framework of the reinforcement learning agent, online correction of the model parameters is realized, and the real-time response capability and the regulation and control accuracy of the dynamic change of the underground state are improved.

Description

Technical Field

[0001] This invention relates to the field of fluid injection control technology, and in particular to an intelligent control method for supercritical CO2 injection based on an AI model. Background Technology

[0002] In the field of supercritical CO2 injection enhanced coalbed methane extraction (CO2-ECBM), the development of digital twin technology, real-time multi-source monitoring technology, and multi-objective optimization algorithms has provided new technical approaches for the simulation, monitoring, and control of subsurface multiphase transport processes. High-fidelity numerical simulation and data assimilation methods can construct dynamic models of subsurface conditions, while optimization theories such as model predictive control provide a framework for real-time control of the injection process, aiming to synergistically improve CO2 sequestration efficiency and methane recovery rate.

[0003] However, existing technical solutions have shortcomings in achieving real-time, adaptive closed-loop control. The process between monitoring data, simulation prediction and control decision-making is usually a unidirectional, open-loop serial process. It lacks a closed-loop mechanism that can dynamically correct core model parameters online based on the actual formation response of the control action and optimize future control strategies in real time. This makes it difficult to quickly adapt to the spatiotemporal changes of complex underground heterogeneous geological conditions and multi-field coupled dynamic processes, resulting in insufficient real-time performance, accuracy and robustness of the control. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an AI-based intelligent control method for supercritical CO2 injection to solve the problems of control lag and insufficient accuracy caused by the lack of closed-loop feedback and online model correction mechanism in existing CO2 injection control methods.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] This invention provides an intelligent control method for supercritical CO2 injection based on an AI model. The method includes: acquiring real-time multi-source monitoring data and constructing a high-precision digital twin model; inputting the real-time multi-source monitoring data into the high-precision digital twin model for dynamic simulation, outputting a subsurface pressure field and a CO2 concentration field; assimilating the real-time multi-source monitoring data with the subsurface pressure field and CO2 concentration field, dynamically correcting the parameters of the high-precision digital twin model, and generating a high-confidence subsurface state field; based on the high-confidence subsurface state field, calculating a dynamic safety boundary characterized by an upper limit of injection pressure and a microseismic energy threshold within a future set time period, and outputting... The system generates a safety boundary prediction result; this result is used as a hard constraint and input into a lightweight digital shadow model. A Pareto optimal solution set satisfying the hard constraint is obtained through rolling optimization. A control strategy is selected from the Pareto optimal solution set based on preset preference weights, and this strategy is converted into a control command set. The control command set is then sent to the electric regulating valve and high-pressure pump in the injection well, as well as the extraction pump in the extraction well, to regulate the CO2 injection pressure, injection flow rate, and extraction pump power, generating a control event record. Based on the control event record, updated real-time multi-source monitoring data is collected and fed back to the high-precision digital twin model, forming a closed-loop control system.

[0008] As a preferred embodiment of the AI-based intelligent control method for supercritical CO2 injection described in this invention, the real-time multi-source monitoring data includes the injection well bottom pressure, the extraction well bottom pressure, the extraction wellhead methane concentration, and microseismic events.

[0009] As a preferred embodiment of the AI-based intelligent control method for supercritical CO2 injection described in this invention, the steps for constructing a high-precision digital twin model are as follows:

[0010] Based on the principles of porous media seepage mechanics and geomechanics, a physical information neural network framework is constructed.

[0011] By using real-time monitoring data from multiple sources as training samples and minimizing errors, a physical information neural network framework is pre-trained to complete the construction of a high-precision digital twin model.

[0012] As a preferred embodiment of the AI-based intelligent control method for supercritical CO2 injection described in this invention, the steps for outputting the underground pressure field and CO2 concentration field are as follows:

[0013] Spatiotemporal alignment and feature extraction are performed on real-time multi-source monitoring data to obtain standardized and feature-enhanced monitoring data sequences;

[0014] The standardized and feature-enhanced monitoring data sequence is input into a high-precision digital twin model, and the initial underground pressure field and CO2 concentration field distribution are obtained through forward inference calculation using a physical information neural network framework.

[0015] Physical consistency verification and spatiotemporal smoothing optimization are performed on the initial underground pressure field and CO2 concentration field distribution to output the underground pressure field and CO2 concentration field.

[0016] As a preferred embodiment of the AI-based intelligent control method for supercritical CO2 injection described in this invention, the steps for generating a high-confidence underground state field are as follows:

[0017] The underground pressure field and CO2 concentration field are used as background fields and compared with real-time multi-source monitoring data to generate a residual sequence between the background field and the real-time multi-source monitoring data.

[0018] The residual sequence is input into a variational assimilation framework that integrates reinforcement learning agents to calculate the mismatch index between the background field and real-time multi-source monitoring data.

[0019] With minimizing the mismatch index as the optimization objective, a distributed optimization algorithm is used to reverse-correct the parameters of the high-precision digital twin model, thereby obtaining the parameters of the optimized high-precision digital twin model.

[0020] The high-precision digital twin model is run using the optimized parameters to perform forward simulations, generating a high-confidence underground state field.

[0021] As a preferred embodiment of the AI-based intelligent control method for supercritical CO2 injection described in this invention, the steps for outputting the safety boundary prediction result are as follows:

[0022] The maximum principal stress gradient, CO2 saturation distribution uniformity, and pressure diffusion rate are extracted from the high-confidence underground state field, and then subjected to dimensional unification, dynamic weight allocation, and dynamic fusion to generate a quantitative safety feature vector.

[0023] The quantized safety feature vector is input into the spatiotemporal graph convolutional network, and the initial safety boundaries of the upper limit of injection pressure and the microseismic energy threshold are obtained through forward propagation calculation.

[0024] The initial safety boundary is dynamically calibrated using microseismic events, and the predicted safety boundary results are output.

[0025] As a preferred embodiment of the AI-based intelligent control method for supercritical CO2 injection described in this invention, the steps for obtaining the Pareto optimal solution set satisfying hard constraints through rolling optimization are as follows:

[0026] The safety boundary prediction results are used as inviolable hard constraints and input into the lightweight digital shadow model to set the boundary limits of the hard constraints.

[0027] Under hard constraint boundary conditions, a multi-objective optimization function is constructed to simultaneously maximize CO2 storage efficiency and methane recovery rate;

[0028] The rolling optimization algorithm is used to solve the multi-objective optimization function to obtain the Pareto optimal solution set that satisfies the hard constraints.

[0029] As a preferred embodiment of the AI-based intelligent control method for supercritical CO2 injection described in this invention, the step of converting the control strategy into a control instruction set includes the following steps.

[0030] Based on real-time multi-source monitoring data and preset preference weights, a multi-attribute decision-making method is used to select the control strategy with the highest comprehensive score from the Pareto optimal solution set.

[0031] The control strategy with the highest overall score is converted into a specific set of control instructions through predefined instruction mapping rules.

[0032] As a preferred embodiment of the AI-based intelligent control method for supercritical CO2 injection described in this invention, the steps for generating control event records are as follows:

[0033] A multivariable collaborative control algorithm is used to process the control instruction set and generate control signals for each actuator.

[0034] Control signals are sent to the corresponding electric regulating valves, high-pressure pumps, and extraction pumps to precisely regulate the CO2 injection pressure, injection flow rate, and extraction pump power, and generate control event records.

[0035] As a preferred embodiment of the AI-based intelligent control method for supercritical CO2 injection described in this invention, the step of feeding back to a high-precision digital twin model to form a closed-loop control is as follows:

[0036] Based on the execution timestamp and device response status in the control event log, the optimal sampling delay time and sampling frequency are determined, and a sampling parameter set is generated;

[0037] Based on the sampling parameter set, the updated real-time multi-source monitoring data is collected and time-synchronized to obtain a time-synchronized multimodal monitoring dataset.

[0038] The time-synchronized multimodal monitoring dataset is fed back to the high-precision digital twin model, and spatiotemporally attention-weighted data assimilation calculations are performed with the underground pressure field and CO2 concentration field to correct the parameters of the high-precision digital twin model and form a closed-loop control.

[0039] The beneficial effects of this invention are as follows: By integrating a variational assimilation framework with reinforcement learning agents, online correction of model parameters is achieved, improving the real-time response capability and control accuracy to dynamic changes in underground states; based on the residual sequence of measured data and model prediction fields, the mismatch index between the background field and monitoring data is calculated, and with the goal of minimizing the mismatch index, a distributed optimization algorithm is driven to reverse-correct key geomechanical and seepage parameters in the digital twin model, enabling the high-precision digital twin model to continuously approximate the actual evolution of underground physical states, improving the accuracy of the "high-confidence underground state field" and the adaptive correction capability of model parameters, fundamentally enhancing the perception and description accuracy of complex underground dynamic processes, and providing a reliable state estimation basis for safety boundary calculation and optimized control. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 A flowchart of an AI-based intelligent control method for supercritical CO2 injection.

[0042] Figure 2 This is a flowchart for outputting the underground pressure field and CO2 concentration field.

[0043] Figure 3 A flowchart for generating a high-confidence underground state field.

[0044] Figure 4 A flowchart for solving the Pareto optimal solution set using rolling optimization. Detailed Implementation

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0046] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0047] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0048] Reference Figures 1-4 This is one embodiment of the present invention, which provides an intelligent control method for supercritical CO2 injection based on an AI model, comprising the following steps:

[0049] S1. Collect real-time multi-source monitoring data and construct a high-precision digital twin model. Input the real-time multi-source monitoring data into the high-precision digital twin model for dynamic simulation and output the underground pressure field and CO2 concentration field.

[0050] Specifically, output the underground pressure field and CO2 concentration field, refer to Figure 2 As shown.

[0051] S1.1: Real-time multi-source monitoring data includes injection well bottom pressure, extraction well bottom pressure, extraction wellhead methane concentration, and microseismic events.

[0052] It should be noted that the injection well bottom pressure refers to the fluid pressure at the bottom of the CO2 injection well, which is obtained in real time through a pressure sensor installed at the bottom of the injection well.

[0053] The bottom pressure of the extraction well refers to the pressure at the bottom of the well used for extracting coalbed methane, which is collected in real time by the downhole pressure sensor of the extraction well.

[0054] The methane concentration at the wellhead refers to the volume or molar concentration of methane (CH4) in the gas discharged from the wellhead, which is obtained by online measurement using a gas composition analyzer (such as an infrared or laser methane detector) installed at the wellhead.

[0055] Microseismic events refer to seismic wave signals generated by minor ruptures or slippages in underground rock strata induced by CO2 injection. These signals are captured and located by microseismic monitoring equipment (including geophone arrays) deployed on the surface or underground.

[0056] S1.2: Based on the principles of porous media seepage mechanics and geomechanics, a physical information neural network framework is constructed;

[0057] Specifically, based on the seepage mechanics of porous media, a governing equation describing the flow of supercritical CO2 in porous media is established. At the same time, based on geomechanics, a governing equation describing the coupling relationship between formation stress-strain response and pore pressure is established. These two types of governing equations are embedded as physical constraints into the structure of a neural network, so that the node connection method and weight update rules of the neural network satisfy the physical laws specified by the seepage mechanics of porous media and geomechanics. This forms a physical information neural network framework that can simultaneously characterize the coupling behavior of fluid transport and rock mass deformation.

[0058] It should be noted that the physical information neural network framework includes at least an input layer, one or more hidden layers, and an output layer. The loss function is composed of the monitoring data fitting term and the physical equation residual term. After training using real-time multi-source monitoring data, the physical information neural network framework forms a high-precision digital twin model, with spatiotemporal location and well location information as inputs and underground pressure field and CO2 concentration field as outputs.

[0059] S1.3: Using real-time monitoring data from multiple sources as training samples, and aiming to minimize errors, the physical information neural network framework is pre-trained to complete the construction of a high-precision digital twin model.

[0060] Specifically, real-time multi-source monitoring data and the physical laws determined by the principles of porous media seepage mechanics and geomechanics in the physical information neural network framework are used together to adjust the internal parameters of the physical information neural network framework. During the adjustment process, the degree of difference between the real-time multi-source monitoring data and the prediction results of the physical information neural network framework is used as the optimization basis, and the internal parameters of the physical information neural network framework are continuously corrected until the degree of difference reaches the minimum, thereby completing the construction of a high-precision digital twin model.

[0061] S1.4: Perform spatiotemporal alignment and feature extraction processing on real-time multi-source monitoring data to obtain standardized and feature-enhanced monitoring data sequences;

[0062] Specifically, monitoring data from different sources are time-aligned using a unified time base, and linear interpolation is used to fill in missing data points. Spatial registration is performed based on the geographical coordinates of each monitoring point to ensure all monitoring data are in the same spatial reference system. Normalization methods are used to eliminate dimensional differences in variables such as pressure and concentration, making their numerical ranges comparable. Key attributes reflecting changes in subsurface conditions, including pressure change trends, concentration fluctuation characteristics, and the frequency of microseismic events, are extracted from the spatiotemporally aligned real-time multi-source monitoring data. The numerical ranges of the extracted key attributes are normalized to ensure that key attributes of different dimensions are within a comparable range and to enhance the expressive power of key attributes in the monitoring data sequence, thereby obtaining a standardized and feature-enhanced monitoring data sequence.

[0063] S1.5: Input the standardized and feature-enhanced monitoring data sequence into the high-precision digital twin model, and obtain the initial underground pressure field and CO2 concentration field distribution through forward inference calculation using the physical information neural network framework;

[0064] Specifically, when the standardized and feature-enhanced monitoring data sequences are combined with the high-precision digital twin model, the structural characteristics of the physical information neural network framework are utilized to enable the standardized and feature-enhanced monitoring data sequences to propagate in the high-precision digital twin model according to the physical relationships defined by the principles of porous media seepage mechanics and geomechanics. During the propagation process, the high-precision digital twin model generates an initial underground pressure field and CO2 concentration field distribution corresponding to the actual underground state based on the pressure, concentration, and microseismic information contained in the standardized and feature-enhanced monitoring data sequences.

[0065] S1.6: Perform physical consistency verification and spatiotemporal smoothing optimization on the initial underground pressure field and CO2 concentration field distribution, and output the underground pressure field and CO2 concentration field.

[0066] Specifically, based on the physical laws stipulated by the principles of porous media seepage mechanics and geomechanics, the initial underground pressure field and CO2 concentration field distribution are checked to see if they satisfy the mass conservation law, Darcy flow relation, and pressure-saturation coupling constraint. For areas that do not satisfy the physical laws, local adjustments are made to make them conform to the basic conservation law and constitutive relation. In the time dimension, a sliding window is used to coordinate the data of continuous time steps, and in the spatial dimension, the values ​​of adjacent grids are smoothed in the neighborhood to eliminate non-physical abrupt changes or oscillations, forming an underground pressure field and CO2 concentration field that conform to physical laws and are continuously changing in time and space.

[0067] It should be noted that the underground pressure field refers to the pressure distribution state that varies with spatial location and time in underground porous media formed by fluids (including supercritical CO2 and coalbed methane, etc.);

[0068] The CO2 concentration field refers to the phase saturation distribution of supercritical CO2, that is, the spatiotemporal distribution of the volume ratio of CO2 in the pore space; the phase saturation can be obtained by combining CO2 concentration with porosity, fluid density and phase relationship.

[0069] S2. Assimilate real-time multi-source monitoring data with the underground pressure field and CO2 concentration field, dynamically correct the parameters of the high-precision digital twin model, and generate a high-confidence underground state field.

[0070] Specifically, generating a high-confidence underground state field, referring to Figure 3 As shown.

[0071] S2.1: As Figure 3As shown, the underground pressure field and CO2 concentration field are used as background fields and compared with real-time multi-source monitoring data to generate a residual sequence between the background field and the real-time multi-source monitoring data.

[0072] Specifically, the underground pressure field and CO2 concentration field are used as background fields. Pressure values ​​corresponding to the bottom pressure of injection wells and extraction wells at the same location and time are extracted from the background fields. Based on the two-phase flow relationship between CO2 and methane in porous media and the adsorption-desorption characteristics of coal matrix on gas, the CO2 displacement state associated with the methane concentration at the extraction wellhead is derived from the CO2 concentration field, and combined with the spatial distribution information of microseismic events. The extracted and derived pressure values ​​and CO2 displacement states are compared item by item with the bottom pressure of injection wells, bottom pressure of extraction wells, methane concentration at the extraction wellhead, and microseismic events in real-time multi-source monitoring data. The numerical differences of each item at the same time and space location are recorded. The numerical differences of all items are integrated to form a residual sequence between the background field composed of the differences of each monitoring variable and the real-time multi-source monitoring data.

[0073] S2.2: Input the residual sequence into the variational assimilation framework that integrates reinforcement learning agents to calculate the mismatch index between the background field and real-time multi-source monitoring data;

[0074] ;

[0075] ;

[0076] In the formula, This index represents the degree of mismatch between the background field and real-time multi-source monitoring data. This represents the total number of spatiotemporal sampling points participating in the assimilation calculation. This represents the reciprocal of the total number of spatiotemporal sampling points participating in the assimilation calculation. Indicates the spatiotemporal sampling point index. Indicates the index of the monitored variable type. This represents the set of monitored variable types (bottom pressure of injection wells, bottom pressure of extraction wells, methane concentration at the extraction wellhead, and microseismic events). Indicates the type of monitoring variable The weighting coefficients, Indicates the first Types of monitored variables at each spatiotemporal sampling point The difference between the observed values ​​and the simulated values ​​of the background field. Indicates the type of monitoring variable The corresponding statistical standard deviation, Indicates the first Types of monitored variables at each spatiotemporal sampling point The observed values, Indicates the first The types of monitoring variables derived from the background field at each spatiotemporal sampling point The simulated value.

[0077] It should be noted that for any type of monitored variable Observed values Compared with simulated values Since they have the same dimensions (e.g., pressure is Pa, concentration is dimensionless fraction, and microseismic energy is J), the difference is... and The physical dimensions are consistent; The monitoring variables are normalized according to their importance and are dimensionless. Only the relative contribution of each variable to the total mismatch is changed without introducing a new dimension. The total number of spatiotemporal sampling points N is a count. Therefore, after weighted summation and averaging of all spatiotemporal sampling points and monitoring variables, the resulting mismatch index J is a dimensionless scalar, thus ensuring that the entire expression remains dimensionally consistent when different physical quantities participate simultaneously.

[0078] It should be noted that the types of monitored variables The weighting coefficients are set based on the importance of each monitoring variable in characterizing the subsurface state and its observation reliability. The weighting coefficients of all monitoring variables are normalized so that their sum is a unit value to ensure the relative balance of the mismatch index. The specific setting steps are as follows: assess the role of four types of monitoring variables—bottom pressure of injection wells, bottom pressure of extraction wells, methane concentration at the extraction wellhead, and microseismic events—in reflecting CO2 migration, formation stress response, and storage safety. Combining the historical observation noise level and sensor accuracy of each variable, higher weights are assigned to variables with high reliability and sensitivity to safety control, and lower weights are assigned to variables that are easily disturbed or indirectly reflect the subsurface state. The initial values ​​of each weighting coefficient are determined through backtracking calibration of historical assimilation effects. The exemplary value ranges are as follows: bottom-hole pressure of injection wells: 0.3–0.4; bottom-hole pressure of extraction wells: 0.2–0.3; methane concentration at the wellhead of extraction wells: 0.2–0.25; and microseismic events: 0.15–0.25. The values ​​are based on the directness of each variable to the physical process and the stability of the data. Bottom-hole pressure of injection wells directly controls the injection driving force, so it has the highest weight. Although microseismic events are critical, they are sparse and sudden, so their weight is relatively low. If the weight coefficient of a monitoring variable is lower than the lower limit, it may lead to insensitivity to the risk of rock mass fracture and weaken the early warning capability of the safety boundary. If the weight coefficient of a monitoring variable is higher than the upper limit, it may over-amplify the impact of gas composition fluctuations on model correction, causing distortion of the underground pressure field simulation and thus misleading the injection strategy.

[0079] S2.3: Taking the minimization of the mismatch index as the optimization objective, a distributed optimization algorithm is used to reverse correct the parameters of the high-precision digital twin model to obtain the parameters of the optimized high-precision digital twin model;

[0080] Specifically, the deviation direction between the current parameters of the high-precision digital twin model and the actual underground state is determined based on the mismatch index between the background field and real-time multi-source monitoring data. Under the coordination of a distributed optimization algorithm, the parameters of the high-precision digital twin model are divided into multiple subsets, and multiple parallel computation processes adjust their respective parameter subsets. Each adjustment updates the parameters based on local gradient information, and the update direction of each subset is coordinated through a global synchronization mechanism, so that the mismatch index continues to decrease. When the decrease in the mismatch index is lower than the preset mismatch index threshold in several consecutive iterations, it is determined that the parameters of the high-precision digital twin model have converged, and optimization stops. After multiple rounds of coordinated adjustment, the parameters of the high-precision digital twin model as a whole tend to the state that minimizes the mismatch index, thereby obtaining the parameters of the optimized high-precision digital twin model.

[0081] It should be noted that the mismatch index threshold is set based on a comprehensive evaluation of historical assimilation effects and the accuracy of underground state prediction. The specific setting steps are as follows: During the trial operation phase, multiple sets of mismatch indices are calculated using historical monitoring data and simulation results from a high-precision digital twin model; the correlation between different mismatch levels and underground state inaccuracies is analyzed in conjunction with actual microseismic events, pressure anomalies, and other safety events; a threshold range that can sensitively reflect model deviations without over-responding to measurement noise is determined as the mismatch index threshold; the exemplary value range is 0.8 to 1.5. The basis for this value is that when the mismatch is below 0.8, it indicates that the model is highly consistent with the measured data and is in a stable and controlled state; when it is above 1.5, it indicates that the model has significantly deviated from the actual underground state and parameter correction or safety warnings need to be triggered.

[0082] S2.4: Run the high-precision digital twin model using the optimized parameters to perform forward simulation and generate a high-confidence underground state field.

[0083] Specifically, the parameters of the optimized high-precision digital twin model are loaded into the high-precision digital twin model. Under the constraints of the principles of porous media seepage mechanics and geomechanics, the high-precision digital twin model combines the standardized and feature-enhanced monitoring data sequence of the current moment and advances the simulation process according to the propagation mode specified by the physical information neural network framework. The simulation process covers a future prediction period and advances step by step at a fixed time step until the prediction endpoint is reached. It reflects the evolution trend of the underground pressure field and CO2 concentration field, forming a high-confidence underground state field that is highly consistent with the actual underground state.

[0084] S3. Based on the high-confidence underground state field, calculate the dynamic safety boundary characterized by the upper limit of injection pressure and the microseismic energy threshold within a future set time period, and output the safety boundary prediction results.

[0085] S3.1: Extract the maximum principal stress gradient, CO2 saturation distribution uniformity, and pressure diffusion rate from the high-confidence underground state field, and perform dimensional unification, dynamic weight allocation, and dynamic fusion to generate a quantitative safety feature vector;

[0086] Specifically, the maximum principal stress gradient, CO2 saturation uniformity, and pressure diffusion rate are extracted from the high-confidence underground state field. Based on the underground pressure field and CO2 concentration field included in the high-confidence underground state field, the spatial variation characteristics of the maximum principal stress gradient are determined according to geomechanical relationships, the CO2 saturation uniformity is determined according to the CO2 concentration field, and the pressure diffusion rate is determined according to the temporal evolution characteristics of the underground pressure field. The maximum principal stress gradient, CO2 saturation uniformity, and pressure diffusion rate are normalized to map their values ​​to a unified comparable interval. Dynamic weights are allocated in real time according to the sensitivity of the maximum principal stress gradient, CO2 saturation uniformity, and pressure diffusion rate to safety risks under the current working conditions. The fusion process adopts a weighted linear combination method to synthesize a single safety feature vector from the normalized maximum principal stress gradient, CO2 saturation uniformity, and pressure diffusion rate according to the dynamic weights.

[0087] S3.2: Input the quantized safety feature vector into the spatiotemporal graph convolutional network, and obtain the initial safety boundaries of the upper limit of injection pressure and the microseismic energy threshold through forward propagation calculation;

[0088] Specifically, the quantified safety feature vectors are constructed into a graph structure according to the spatial location relationships of underground monitoring points. Each monitoring point serves as a node in the graph, and the connection relationships between nodes are determined based on geological connectivity or spatial proximity, with adjacency strength decreasing with increasing distance. Simultaneously, the quantified safety feature vectors from multiple consecutive time steps are stacked in chronological order to form a graph sequence with spatiotemporal dimensions. In the spatiotemporal graph convolutional network, neighborhood information is aggregated in the spatial dimension for the graph structure at each time step in the graph sequence, enabling each node to incorporate the safety feature information of its neighboring nodes. Based on this, the features of the same node at different time steps are correlated along the temporal dimension to capture the evolution trend of the safety state. Through the joint operation of spatial aggregation and temporal correlation, the spatiotemporal graph convolutional network generates upper limits for injection pressure and microseismic energy thresholds that match the current and recent underground mechanical and fluid conditions, forming the initial safety boundaries for the upper limits for injection pressure and microseismic energy thresholds.

[0089] It should be noted that the graph structure is based on the spatial location and geological connectivity of underground monitoring points. Each monitoring point is set as a graph node. Whether nodes are connected and the weight of edges are determined by spatial proximity and stratigraphic connectivity. The closer the distance or the stronger the geological connectivity, the higher the edge weight, thus forming an adjacency structure that reflects the actual physical relationship. The spatiotemporal graph convolutional network adopts a multi-layer spatiotemporal graph convolutional architecture, which includes several spatial graph convolutional layers and temporal convolutional layers. The number of layers is set according to the task complexity to balance expressive power and computational efficiency. The temporal convolutional layers act on the neighborhood of nodes to aggregate spatial information. The temporal convolution captures dynamic evolution features along the time step. The loss function is the mean square error between the predicted security boundary and the corresponding labels of historical real security events, which is used to train the network parameters end-to-end.

[0090] It should be noted that the exemplary range of the upper limit of injection pressure is 8 MPa to 15 MPa; the value is based on the safety margin of formation fracturing pressure and coal and rock mechanical strength, to ensure that the injection pressure is not too high, so as to avoid inducing fracture propagation or wellbore instability; if the injection pressure is lower than 8 MPa, the CO2 displacement power is insufficient, resulting in a decrease in methane desorption efficiency and a significant reduction in ECBM efficiency; if the injection pressure is higher than 15 MPa, it may exceed the minimum principal stress of the formation, triggering hydraulic fracturing or activating faults, causing CO2 leakage, increased microseismic activity, or even inducing perceptible earthquakes.

[0091] An exemplary range for the microseismic energy threshold is 10 times the energy released by a single microseismic event. 3 J to 10 5 J (i.e., the energy released by a microseismic event does not exceed 100,000 joules); the value is determined based on regional geological stability assessment and induced earthquake risk control standards, referencing historical microseismic activity levels and regulatory safety limits; if it is lower than 10... 3 J typically indicates weak formation response and controllable injection disturbance, but a persistently low value may reflect insufficient injection efficiency; a value higher than 10... 5 J indicates that the rock mass has undergone large-scale fracturing or fault slippage, posing a risk of inducing destructive earthquakes, which may endanger the safety of well site facilities and the integrity of underground storage, requiring immediate pressure reduction or suspension of injection.

[0092] S3.3: Dynamically calibrate the initial safety boundary using microseismic events and output the safety boundary prediction results.

[0093] Specifically, the system acquires microseismic events from real-time multi-source monitoring data, including the location, time, and energy magnitude of microseismic events. It then compares these microseismic events with the initial safety boundaries of the injection pressure upper limit and microseismic energy threshold in both time and space to determine if the events occur outside the allowable range of the initial safety boundaries. If the energy or spatial concentration of the microseismic events exceeds expectations, the injection pressure upper limit and microseismic energy threshold are appropriately lowered according to a preset proportional coefficient. If microseismic activity remains below the background level and its distribution is stable, the safety boundaries are allowed to be slightly raised, provided they do not exceed the historical maximum safety value. All adjustments are integrated with the safety boundaries of the previous moment through exponential smoothing to avoid abrupt changes and ensure a smooth transition of control commands. The adjustment process generates a safety boundary prediction result consistent with the current actual underground response.

[0094] It should be noted that the proportionality coefficient is set based on the correlation analysis between historical microseismic events and changes in subsurface conditions. The specific setting steps are as follows: collect historical data during operation, including the energy and frequency of microseismic events and their corresponding changes in subsurface pressure and CO2 concentration; determine the proportional relationship that needs to be adjusted between the upper limit of injection pressure and the microseismic energy threshold under different levels of microseismic activity based on historical data; and select the optimal value as the proportionality coefficient that can effectively respond to formation anomalies while avoiding over-regulation by repeatedly verifying the impact of different proportionality coefficients on the control effect. An exemplary value range is 0.1 to 0.5. When the proportionality coefficient is lower than 0.1, the adjustment range is too small and may not be able to respond to changes in formation pressure or microseismic risk in a timely manner. If it is higher than 0.5, it may lead to frequent and excessive control actions, increasing instability and affecting long-term storage efficiency.

[0095] The background level is set based on the average frequency and energy of microseismic events in long-term monitoring data. The specific steps are as follows: statistically analyze the energy distribution and frequency of all microseismic events over a relatively long historical period, and calculate the average microseismic energy and frequency over that period as the background level; periodically update the background level to reflect the latest geological conditions and operational status; an exemplary value range is for single microseismic events with energy levels in the range of 10... 2 J to 10 4 Between J, the frequency does not exceed once per hour; when the energy of a microseismic event is less than 10... 2 When the frequency is low, it indicates that the formation is in a relatively stable state and no immediate control measures are needed; if the energy of the microseismic event continues to exceed 10... 4 A significant increase in J or frequency indicates a potential risk of formation instability, requiring timely adjustments.

[0096] S4. The safety boundary prediction results are used as hard constraints and input into the lightweight digital shadow model. The Pareto optimal solution set that satisfies the hard constraints is solved by rolling optimization. The control strategy in the Pareto optimal solution set is selected according to the preset preference weights and the control strategy is converted into a control instruction set.

[0097] Specifically, rolling optimization is used to find the Pareto optimal solution set, refer to... Figure 4 As shown.

[0098] S4.1: As Figure 4 As shown, the safety boundary prediction results are used as inviolable hard constraints and input into the lightweight digital shadow model to set the boundary limits of the hard constraints.

[0099] Specifically, the upper limit of injection pressure and the specific limit of microseismic energy threshold are extracted from the safety boundary prediction results; the allowable range of variation of the control variables is set in the lightweight digital shadow model so that the CO2 injection pressure does not exceed the upper limit of injection pressure and the microseismic energy of microseismic events does not exceed the microseismic energy threshold, ensuring that all possible control strategies of the lightweight digital shadow model are strictly within the physical safety range specified by the safety boundary prediction results during operation, forming a hard constraint boundary limit.

[0100] It should be noted that the lightweight digital shadow model is constructed based on the high-precision digital twin model, extracting the input dimensions related to injection pressure, injection velocity, and pump power, as well as the output response relationships related to CO2 storage efficiency and methane recovery rate. On this basis, the network structure is simplified by removing the high-dimensional spatiotemporal coupling layer from the physical information neural network, retaining a low-complexity mapping architecture. The lightweight digital shadow model adopts a multilayer perceptron structure, with the current injection pressure, injection velocity, and pump power as inputs, and the predicted CO2 storage efficiency and methane recovery rate as outputs. The training loss is the mean square error between the output prediction value and the simulation result corresponding to the high-precision digital twin model. Hard constraints are achieved by pruning the feasible domain of the control variables during the optimization stage; that is, any candidate strategy that exceeds the upper limit of injection pressure or causes microseismic energy to exceed the limit is excluded during generation. Simultaneously, the upper limit of injection pressure and the microseismic energy threshold from the safety boundary prediction results are embedded into the model and set as insurmountable hard constraint boundaries, ensuring that all feasible solutions are within the safe operating range, forming a lightweight digital shadow model for rolling optimization.

[0101] The pre-training of the lightweight digital shadow model utilizes simulation data generated by running the high-precision digital twin model over historical time periods to obtain simulation results of CO2 storage efficiency and methane recovery rate under different combinations of injection pressure, injection flow rate, and pump power. Different combinations serve as input samples, and the corresponding simulation results serve as output labels. The parameters of the lightweight digital shadow model are adjusted so that it can approximately reproduce the response characteristics of the high-precision digital twin model regarding CO2 storage efficiency and methane recovery rate while maintaining low computational overhead. After pre-training, the lightweight digital shadow model has the ability to support multi-objective rolling optimization.

[0102] S4.2: Under hard constraint boundary conditions, construct a multi-objective optimization function that simultaneously maximizes CO2 storage efficiency and methane recovery rate;

[0103] Specifically, based on the upper limit of injection pressure and the microseismic energy threshold set in the lightweight digital shadow model, the feasible variation range of CO2 injection pressure, injection flow rate, and extraction pump power is determined. Simultaneously, physical limitations of on-site equipment are considered, including the maximum / minimum displacement of the injection pump, the upper limit of the rate of change of the regulating valve opening, the upper and lower limits of the extraction pump power, and the limit on the number of control command changes per unit time, collectively forming a complete feasible control domain. Within the feasible variation range, CO2 storage efficiency and methane recovery rate are treated as two independent optimization objectives, corresponding to the characterization of CO2 retention capacity and methane production capacity during supercritical CO2 injection-enhanced coalbed methane extraction, respectively. By placing the two objectives side-by-side in the same optimization framework, a multi-objective optimization function is formed, aiming to simultaneously improve CO2 storage efficiency and methane recovery rate.

[0104] S4.3: The rolling optimization algorithm is used to solve the multi-objective optimization function to obtain the Pareto optimal solution set that satisfies the hard constraints.

[0105] Specifically, an optimization window with a finite future timeframe is determined at the current moment. The window length covers several future control cycles, with each cycle corresponding to a fixed control step size. Within the optimization window, CO2 injection pressure, injection flow rate, and extraction pump power are used as control variables, with CO2 storage efficiency and methane recovery rate as the target directions. A series of control strategies are searched within the feasible variation range specified by the hard constraints. The rolling optimization algorithm, based on gradient information and non-dominated sorting mechanism, searches for non-dominated solutions on the Pareto front within the feasible region, seeking a balance point that improves CO2 storage efficiency without reducing methane recovery rate, or improves methane recovery rate without compromising CO2 storage efficiency. Each round of optimization executes only the control strategy at the current moment, advancing the optimization window by one time step and repeating the search process. After multiple rounds of rolling advancement, all non-dominated solutions are aggregated to form a Pareto optimal solution set that satisfies the hard constraints.

[0106] S5. Send the control command set to the electric regulating valve and high-pressure pump of the injection well and the extraction pump of the extraction well to regulate the CO2 injection pressure, injection flow rate and extraction pump power, and generate regulation event records.

[0107] S5.1: Based on real-time multi-source monitoring data and preset preference weights, a multi-attribute decision-making method is used to select the control strategy with the highest comprehensive score from the Pareto optimal solution set;

[0108] Specifically, the current injection well bottom pressure, extraction well bottom pressure, extraction wellhead methane concentration, and underground operating status reflected by microseismic events are obtained from real-time multi-source monitoring data. Combining the relative importance of CO2 storage efficiency and methane recovery rate in the preset preference weights, corresponding importance coefficients are assigned to the performance of each control strategy in the Pareto optimal solution set across the two target dimensions. The original values ​​of each control strategy on the two targets are normalized to eliminate dimensional differences. The product and sum of the normalized target values ​​and their corresponding importance coefficients are then used to obtain a comprehensive score for each control strategy. The comprehensive scores of all control strategies are compared, and the control strategy with the highest comprehensive score is selected.

[0109] It should be noted that the preset preference weights are based on the goal orientation of the supercritical CO2 injection enhanced coalbed methane extraction project. At the initial stage of the project, the priority relationship between CO2 storage efficiency and methane recovery rate is determined according to the core tasks of the current stage. Specifically, the setting steps are as follows: clarify whether the current operational stage prioritizes increasing methane production or enhancing CO2 storage capacity; allocate weight values ​​for the two objectives within the range of 0 to 1, ensuring that the sum of the two is 1. The weights of CO2 storage efficiency and methane recovery rate are complementary. An example value is a CO2 storage efficiency weight of 0.6 and a methane recovery rate weight of 0.4. The weighting is based on engineering management strategies. For example, during the carbon sequestration assessment period, CO2 sequestration efficiency is emphasized and given a higher weight; during the economic benefit priority period, methane recovery rate is emphasized and its weight is increased accordingly. If the weight of CO2 sequestration efficiency is lower than 0.6, the injected CO2 may not be effectively retained, weakening the geological sequestration effect and affecting the achievement of carbon emission reduction targets. If the weight of CO2 sequestration efficiency is higher than 0.6, the extraction intensity may be excessively restricted, resulting in methane not being extracted in time after desorption, reducing coalbed methane revenue, and even inducing micro-seismic risks due to pressure accumulation. When significant changes are detected in economic indicators (such as methane price, carbon trading revenue) or safety indicators (such as micro-seismic frequency, pressure approaching the upper limit), the weights can be manually adjusted by the operator or automatically switched to the weight combination of the corresponding stage by the pre-configured strategy.

[0110] S5.2: Convert the control strategy with the highest comprehensive score into a specific set of control instructions through predefined instruction mapping rules.

[0111] Specifically, based on the target values ​​of CO2 injection pressure, injection flow rate, and extraction pump power included in the control strategy with the highest comprehensive score, the corresponding electric regulating valve opening command, high-pressure pump speed command, and extraction pump power adjustment command are found in the predefined command mapping rules. The predefined command mapping rules are based on the correspondence between the physical characteristics of the equipment and the control variables, ensuring that each set of control strategy parameters has a unique combination of execution commands. Through the correspondence, a set of control commands that can be directly issued to the execution equipment is generated.

[0112] It should be noted that the command mapping rule refers to a one-to-one correspondence mapping table established between the target values ​​of CO2 injection pressure, injection flow rate, and extraction pump power in the control strategy, and the opening degree of the electric regulating valve of the injection well, the operating parameters of the high-pressure pump, and the power control signal of the extraction pump of the extraction well. After the correspondence mapping table is established during the equipment commissioning phase, it will be periodically calibrated or compensated online according to the actual wellhead temperature, pressure, and other operating parameters to correct execution deviations caused by changes in fluid properties or equipment aging, ensuring the accuracy of the command mapping. The specific setting steps are as follows: During the equipment installation and commissioning phase, the electric regulating valve of the injection well is fully... The stroke opening test records the steady-state values ​​of CO2 injection pressure and injection flow rate at different opening degrees. Simultaneously, multi-stage speed or displacement tests are conducted on the high-pressure pump to obtain its response relationship with injection flow rate and pressure. A power-pumping flow rate-bottom-well pressure linkage test is performed on the pumping pump of the extraction well to establish the correspondence between pumping pump power and extraction effect. The test results are compiled into a correspondence mapping table, forming command mapping rules to ensure that any given combination of CO2 injection pressure, injection flow rate, and pumping pump power can find a matching electric regulating valve opening command, high-pressure pump control command, and pumping pump power command in this correspondence mapping table.

[0113] S5.3: A multivariable collaborative control algorithm is used to process the control instruction set and generate control signals for each actuator;

[0114] Specifically, the system acquires the target values ​​of CO2 injection pressure, injection flow rate, and extraction pump power contained in the control command set. Based on the coupling relationship between the electric regulating valve and high-pressure pump of the injection well and the extraction pump of the extraction well, a multivariable collaborative control algorithm is used to coordinate the timing and adjustment range of each actuator. This avoids pressure fluctuations or flow imbalances caused by the independent action of a single actuator. The multivariable collaborative control algorithm is based on a multi-input multi-output feedback regulation mechanism. By monitoring the deviation between the actual output of each actuator and the target value in real time, it dynamically adjusts the output intensity of each control signal and introduces decoupling logic to reduce mutual interference between devices, thereby achieving overall stability of the injection and extraction process. During the coordination process, the multivariable collaborative control algorithm aims to maintain the stability of the overall injection-extraction process and generates control signals corresponding to the electric regulating valve, high-pressure pump, and extraction pump, respectively.

[0115] S5.4: Send control signals to the corresponding electric regulating valves, high-pressure pumps, and extraction pumps to precisely regulate the CO2 injection pressure, injection flow rate, and extraction pump power, and generate regulation event records.

[0116] Specifically, control signals are transmitted via industrial communication links to the electric regulating valve of the injection well, the high-pressure pump of the injection well, and the extraction pump of the extraction well. The electric regulating valve adjusts its opening according to the received control signals, the high-pressure pump adjusts its operating speed or flow rate according to the control signals, and the extraction pump changes its output power according to the control signals. The combined actions of these actions ensure that the CO2 injection pressure, injection flow rate, and extraction pump power reach the target values ​​required by the control strategy. After execution, the content of the control signals involved in this control, the name of the executing equipment, the execution timestamp, and the equipment response status are recorded to form a control event record.

[0117] S6. Based on the control event records, collect updated real-time multi-source monitoring data and feed it back to the high-precision digital twin model to form a closed-loop control.

[0118] S6.1: Based on the execution timestamp and device response status in the control event log, determine the optimal delay time and sampling frequency for sampling, and generate a sampling parameter set;

[0119] Specifically, the execution timestamps of the electric regulating valves, high-pressure pumps, and extraction pumps, as well as the time it takes for each device to reach a stable response state after execution, are extracted from the control event records. Based on the time constant required for the device's response state to stabilize from the execution of the action, an integer multiple of the time constant is used as the optimal sampling delay time to ensure that the underground fluid and stress field have fully responded to the control action. At the same time, based on the fluctuation decay period of the monitoring data during the device response process, the sampling frequency is set to a rate that can collect multiple data points within at least one fluctuation period to fully capture the dynamic process while avoiding data redundancy caused by excessively high frequencies. The determined optimal delay time and sampling frequency are combined to form a sampling parameter set.

[0120] S6.2: Based on the sampling parameter set, collect updated real-time multi-source monitoring data and perform time synchronization processing to obtain a time-synchronized multimodal monitoring dataset;

[0121] Specifically, according to the sampling frequency specified in the sampling parameter set, synchronous acquisition of injection well bottom pressure, extraction well bottom pressure, extraction wellhead methane concentration, and microseismic events is initiated after the optimal delay time is reached. During the acquisition process, a time label under a unified time reference is attached to each monitoring data to ensure that data from different sources have comparable time references. The injection well bottom pressure, extraction well bottom pressure, extraction wellhead methane concentration, and microseismic event data with time labels are aligned according to the time labels, so that the multi-source data at the same time are merged into a complete observation sample, forming a time-synchronized multimodal monitoring dataset.

[0122] S6.3: Feed the time-synchronized multimodal monitoring dataset back to the high-precision digital twin model, and perform spatiotemporal attention-weighted data assimilation calculations with the underground pressure field and CO2 concentration field to correct the parameters of the high-precision digital twin model and form a closed-loop control.

[0123] Specifically, the time-synchronized multimodal monitoring dataset is registered with the underground pressure field and CO2 concentration field currently simulated by the high-precision digital twin model at the same time and spatial location. The spatiotemporal attention weighting mechanism automatically assesses the importance of each monitoring point to the model state correction at the current moment by analyzing the spatiotemporal distribution characteristics of historical assimilation errors and current monitoring data. Data points near injection wells or microseismically active areas are given higher weights, while data points far from disturbance sources or in steady-state areas have lower weights. In the parameter correction stage, the high-precision digital twin model, based on the backpropagation path under physical constraints, transmits the weighted observation residuals layer by layer to the network weights related to key physical parameters such as seepage coefficient and pore elastic modulus, thereby achieving directional parameter correction. This brings the state of the high-precision digital twin model closer to the real underground state reflected by the time-synchronized multimodal monitoring dataset, allowing the updated state of the high-precision digital twin model to be used for the next round of control decisions, forming a closed-loop control.

[0124] In summary, this invention achieves online correction of model parameters through a variational assimilation framework integrating reinforcement learning agents, thereby improving the real-time response capability and control accuracy to dynamic changes in underground states. Based on the residual sequence of measured data and model prediction fields, it calculates the mismatch index between the background field and monitoring data, and drives a distributed optimization algorithm to reverse-correct key geomechanical and seepage parameters in the digital twin model with the goal of minimizing the mismatch index. This enables the high-precision digital twin model to continuously approximate the actual evolution of underground physical states, improves the accuracy of the "high-confidence underground state field," and enhances the adaptive correction capability of model parameters. This fundamentally strengthens the perception and description accuracy of complex underground dynamic processes, providing a reliable state estimation basis for safety boundary calculation and optimized control.

[0125] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent control of supercritical CO2 injection based on an AI model, characterized in that: include, Collect real-time multi-source monitoring data and construct a high-precision digital twin model. Input the real-time multi-source monitoring data into the high-precision digital twin model for dynamic simulation and output the underground pressure field and CO2 concentration field. Real-time multi-source monitoring data is assimilated with underground pressure field and CO2 concentration field for calculation, and the parameters of high-precision digital twin model are dynamically corrected to generate a high-confidence underground state field. Based on the high-confidence underground state field, the dynamic safety boundary characterized by the upper limit of injection pressure and the microseismic energy threshold is calculated within a future set time period, and the safety boundary prediction results are output. The safety boundary prediction results are used as hard constraints and input into a lightweight digital shadow model. The Pareto optimal solution set that satisfies the hard constraints is solved by rolling optimization. The control strategy in the Pareto optimal solution set is selected according to the preset preference weights and the control strategy is converted into a control instruction set. The control command set is sent to the electric regulating valve and high-pressure pump of the injection well and the extraction pump of the extraction well to regulate the CO2 injection pressure, injection flow rate and extraction pump power, and generate regulation event records. Based on the records of regulatory events, updated real-time multi-source monitoring data is collected and fed back to a high-precision digital twin model to form a closed-loop regulation.

2. The intelligent control method for supercritical CO2 injection based on an AI model as described in claim 1, characterized in that: The real-time multi-source monitoring data includes injection well bottom pressure, extraction well bottom pressure, extraction wellhead methane concentration, and microseismic events.

3. The intelligent control method for supercritical CO2 injection based on an AI model as described in claim 2, characterized in that: The steps for constructing a high-precision digital twin model are as follows: Based on the principles of porous media seepage mechanics and geomechanics, a physical information neural network framework is constructed. By using real-time monitoring data from multiple sources as training samples and minimizing errors, a physical information neural network framework is pre-trained to complete the construction of a high-precision digital twin model.

4. The AI-based intelligent control method for supercritical CO2 injection as described in claim 3, characterized in that: The steps for outputting the underground pressure field and CO2 concentration field are as follows. Spatiotemporal alignment and feature extraction are performed on real-time multi-source monitoring data to obtain standardized and feature-enhanced monitoring data sequences; The standardized and feature-enhanced monitoring data sequence is input into a high-precision digital twin model, and the initial underground pressure field and CO2 concentration field distribution are obtained through forward inference calculation using a physical information neural network framework. Physical consistency verification and spatiotemporal smoothing optimization are performed on the initial underground pressure field and CO2 concentration field distribution to output the underground pressure field and CO2 concentration field.

5. The AI-based intelligent control method for supercritical CO2 injection as described in claim 4, characterized in that: The steps for generating a high-confidence underground state field are as follows: The underground pressure field and CO2 concentration field are used as background fields and compared with real-time multi-source monitoring data to generate a residual sequence between the background field and the real-time multi-source monitoring data. The residual sequence is input into a variational assimilation framework that integrates reinforcement learning agents to calculate the mismatch index between the background field and real-time multi-source monitoring data. With minimizing the mismatch index as the optimization objective, a distributed optimization algorithm is used to reverse-correct the parameters of the high-precision digital twin model, thereby obtaining the parameters of the optimized high-precision digital twin model. The high-precision digital twin model is run using the optimized parameters to perform forward simulations, generating a high-confidence underground state field.

6. The AI-based intelligent control method for supercritical CO2 injection as described in claim 5, characterized in that: The steps for outputting the safety boundary prediction result are as follows: The maximum principal stress gradient, CO2 saturation distribution uniformity, and pressure diffusion rate are extracted from the high-confidence underground state field, and then subjected to dimensional unification, dynamic weight allocation, and dynamic fusion to generate a quantitative safety feature vector. The quantized safety feature vector is input into the spatiotemporal graph convolutional network, and the initial safety boundaries of the upper limit of injection pressure and the microseismic energy threshold are obtained through forward propagation calculation. The initial safety boundary is dynamically calibrated using microseismic events, and the predicted safety boundary results are output.

7. The AI-based intelligent control method for supercritical CO2 injection as described in claim 6, characterized in that: The steps for finding the Pareto optimal solution set that satisfies the hard constraints through rolling optimization are as follows: The safety boundary prediction results are used as inviolable hard constraints and input into the lightweight digital shadow model to set the boundary limits of the hard constraints. Under hard constraint boundary conditions, a multi-objective optimization function is constructed to simultaneously maximize CO2 storage efficiency and methane recovery rate; The rolling optimization algorithm is used to solve the multi-objective optimization function to obtain the Pareto optimal solution set that satisfies the hard constraints.

8. The AI-based intelligent control method for supercritical CO2 injection as described in claim 7, characterized in that: The steps for converting the control strategy into a control instruction set are as follows: Based on real-time multi-source monitoring data and preset preference weights, a multi-attribute decision-making method is used to select the control strategy with the highest comprehensive score from the Pareto optimal solution set. The control strategy with the highest overall score is converted into a specific set of control instructions through predefined instruction mapping rules.

9. The intelligent control method for supercritical CO2 injection based on an AI model as described in claim 8, characterized in that: The steps for generating the control event record are as follows: A multivariable collaborative control algorithm is used to process the control instruction set and generate control signals for each actuator. Control signals are sent to the corresponding electric regulating valves, high-pressure pumps, and extraction pumps to precisely regulate the CO2 injection pressure, injection flow rate, and extraction pump power, and generate control event records.

10. The AI-based intelligent control method for supercritical CO2 injection as described in claim 9, characterized in that: The feedback is sent to a high-precision digital twin model to form a closed-loop control, and the steps are as follows. Based on the execution timestamp and device response status in the control event log, the optimal sampling delay time and sampling frequency are determined, and a sampling parameter set is generated; Based on the sampling parameter set, the updated real-time multi-source monitoring data is collected and time-synchronized to obtain a time-synchronized multimodal monitoring dataset. The time-synchronized multimodal monitoring dataset is fed back to the high-precision digital twin model, and spatiotemporally attention-weighted data assimilation calculations are performed with the underground pressure field and CO2 concentration field to correct the parameters of the high-precision digital twin model and form a closed-loop control.

Images