Intelligent management and control method and system for carbon dioxide sequestration

Abstract

The invention provides an intelligent management and control method and system for carbon dioxide sequestration, and the method comprises the steps: obtaining multi-source data, carrying out the space-time alignment processing, feature extraction processing and multi-scale hierarchical data fusion processing of the multi-source data, obtaining dynamic parameter field data, carrying out the time sequence feature extraction and spatial feature coding based on the dynamic parameter field data, and carrying out the time sequence feature extraction and spatial feature coding. Generating a spatial-temporal characteristic matrix, inputting the spatial-temporal characteristic matrix into a pre-trained short-term leakage prediction model, and outputting a high-risk area coordinate of which the leakage probability is greater than a preset value in a future first duration, obtaining reservoir pressure, cover layer stress, fault activation index and geomechanical simulation data, inputting the data into a long-term stability evaluation model obtained through pre-training, outputting gas injection adjustment amplitude and recommended dosage of a repairing agent in a second time period in the future, and determining the reservoir pressure gradient and the earth surface deformation rate based on the gas injection adjustment amplitude, the recommended dosage of the repairing agent, the reservoir pressure gradient and the earth surface deformation rate. An executable repair strategy is generated through a multi-objective optimization algorithm, and a repair instruction set is issued according to the repair strategy to execute targeted repair, so that the leakage positioning and risk area identification precision is improved, and a'prediction-decision-repair 'full-process automatic closed loop is formed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of carbon capture and storage, and particularly relates to a carbon dioxide storage intelligent management and control method and system. BACKGROUND

[0002] With the acceleration of global carbon neutralization process, CO2 geological storage technology has become a key means to cope with climate change. However, this technology faces many challenges in practical application: first, factors such as reservoir pressure change and micro-fracture development during the storage process may lead to CO2 leakage, which not only affects the storage effect, but also may pose a serious threat to the surrounding ecological environment; second, the existing monitoring system has problems such as incomplete data collection and insufficient analysis accuracy, making it difficult to discover potential risks in a timely and accurate manner; finally, traditional management and control methods rely on manual experience for judgment, and have slow response speed and insufficient decision-making scientificity.

[0003] The current technology mainly has three limitations: 1) difficulty in integrating multi-source heterogeneous data, the data obtained by different monitoring methods (such as seismic exploration, optical fiber sensing, satellite remote sensing, etc.) differ significantly in time and space scales and accuracy, making it difficult to form a unified and accurate reservoir state model; 2) insufficient risk early warning capability, the existing system mostly adopts a static threshold alarm mechanism, which cannot accurately predict dynamic risks under complex geological conditions; 3) low efficiency of emergency response, it often takes several hours or even several days from discovering problems to implementing repair, and there is a lack of effective effect evaluation and system optimization mechanism. SUMMARY

[0004] In order to solve the problems of the prior art, the present application provides a carbon dioxide storage intelligent management and control method and system.

[0005] The application provides a carbon dioxide storage intelligent management and control method, comprising the following steps: acquiring multi-source data, performing space-time alignment processing, feature extraction processing and multi-scale hierarchical data fusion processing on the multi-source data to obtain dynamic parameter field data; wherein the multi-source data comprises four-dimensional seismic data, distributed optical fiber acoustic wave signals, ground millimeter-level deformation data, ground CO2 concentration anomaly data and temperature gradient change data, and the dynamic parameter field data comprises CO2 saturation field, reservoir pressure field, micro-fracture density and ground leakage probability; time sequence feature extraction and spatial feature coding are performed based on the dynamic parameter field data to generate a space-time feature matrix; wherein the space-time feature matrix comprises pressure statistics, CO2 saturation gradient field and deformation-pressure coupling coefficient; the space-time feature matrix is input into a pre-trained short-term leakage prediction model to output high-risk region coordinates with a leakage probability greater than a preset value within a first time length in the future; wherein the short-term leakage prediction model is obtained by inputting the space-time feature matrix of historical data into an LSTM network and taking historical leakage labels as output labels for training; reservoir pressure, cap rock stress, fault activation index and geomechanics simulation data are acquired and input into a pre-trained long-term stability evaluation model to output gas injection adjustment amplitude and repair agent recommended dosage within a second time length in the future; wherein the long-term stability evaluation model is obtained by reinforcement learning based on defined state space and action space using a proximal policy optimization framework; an executable repair strategy is generated by a multi-objective optimization algorithm based on the gas injection adjustment amplitude, the repair agent recommended dosage, reservoir pressure gradient and ground deformation rate, and repair instruction sets are issued according to the repair strategy to perform targeted repair.

[0006] According to the carbon dioxide storage intelligent management and control method provided by the application, the multi-scale hierarchical data fusion processing comprises: obtaining and outputting a pore pressure field and a CO2 saturation field based on four-dimensional seismic data through reservoir parameter inversion; obtaining an enhanced reservoir model comprising a CO2 saturation field, a pore pressure field and a micro-fracture density distribution based on the pore pressure field, the CO2 saturation field and DAS acoustic signal features through wavelet domain fusion and micro-fracture correction; and obtaining a corrected four-dimensional dynamic parameter field by deformation-pressure correlation modeling and leakage probability spatial fusion based on the enhanced reservoir model, an InSAR deformation field, a CO2 leakage probability thermal map and a temperature gradient field, wherein the corrected four-dimensional dynamic parameter field integrates reservoir internal pressure distribution, CO2 migration path and ground leakage risk thermal map.

[0007] According to the carbon dioxide storage intelligent management and control method provided by the application, the LSTM network comprises two layers of LSTM, an attention mechanism module is inserted between the layers, a Sigmoid activation function is used in the output layer, and a random inactivation layer with a dropout rate of 0.2 is added.

[0008] The long-term stability evaluation model has a state space including reservoir pressure, caprock stress, fault activation index and repair agent dosage, and an action space including gas injection adjustment information and repair agent injection information, and a reward function is set by balancing stability, economy and safety during training.

[0009] The method further comprises verifying the short-term leakage prediction model according to an AUC index and a false alarm rate, verifying the long-term stability evaluation model according to a cumulative leakage amount index, and periodically updating the short-term leakage prediction model and the long-term stability evaluation model by using a federated learning framework.

[0010] The multi-objective optimization algorithm has a target function expressed as min (repair cost + λ * environmental risk), wherein λ represents an environmental risk penalty weight coefficient, a constraint condition is that a pressure recovery rate is greater than a first preset value and a leakage probability is less than a second preset value, and the repair instruction set includes downhole nanorobot injection coordinates, sealing material type and dosage, and ground laser mineralization device working parameters.

[0011] The method further comprises selecting a sealing material by using a permeability matching algorithm, which is expressed as: wherein: represents a material coefficient, represents a leakage zone permeability; the ground laser mineralization device working parameters include laser power, and the method further comprises determining the laser power by dynamic adaptation with a CO2 concentration, which is expressed as: wherein, represents laser power, represents an energy conversion coefficient, represents a carbon dioxide concentration, represents a leakage area.

[0012] In the execution process of the targeted repair, the method further comprises, for downhole repair, injecting adaptive gel according to a permeability matching result and controlling a gel solidification time required by a time control equation; for ground repair, adjusting laser power in real time to ensure that a mineralization rate meets a set requirement; and after the execution of the targeted repair is completed, the method further comprises extracting a pressure recovery rate and a leakage concentration decay rate and determining a repair effect according to the pressure recovery rate and the leakage concentration decay rate.

[0013] According to the carbon dioxide storage intelligent management and control method provided in the application, after the targeted repair is completed, the method further comprises: obtaining a repair efficiency report and real-time monitoring data, feeding back the repair efficiency report and the real-time monitoring data to the short-term leakage prediction model and the long-term stability evaluation model, so that the short-term leakage prediction model is subjected to enhanced training, and the long-term stability evaluation model is subjected to reward function weight adjustment; based on a federal learning framework, after training at each local node, parameters are uploaded to a central server for aggregation, and the short-term leakage prediction model and the long-term stability evaluation model that are optimized are output.

[0014] The application further provides a carbon dioxide storage intelligent management and control system, comprising the following modules: a data fusion module, configured to: obtain multi-source data, perform space-time alignment processing, feature extraction processing and multi-scale hierarchical data fusion processing on the multi-source data, and obtain dynamic parameter field data; wherein the multi-source data comprises four-dimensional seismic data, distributed optical fiber acoustic wave signals, ground millimeter-level deformation data, ground CO2 concentration anomaly data and temperature gradient change data, and the dynamic parameter field data comprises a CO2 saturation field, a reservoir pressure field, microfracture density and a ground leakage probability; a feature processing module, configured to: perform time series feature extraction and spatial feature coding based on the dynamic parameter field data, and generate a space-time feature matrix; wherein the space-time feature matrix comprises pressure statistics, a CO2 saturation gradient field and a deformation-pressure coupling coefficient; a short-term risk prediction module, configured to: input the space-time feature matrix into a pre-trained short-term leakage prediction model, and output high-risk region coordinates in which a leakage probability is greater than a preset value within a first time length in the future; wherein the short-term leakage prediction model is trained by inputting a space-time feature matrix of historical data into an LSTM network and taking a historical leakage label as an output label; a long-term risk prediction module, configured to: obtain reservoir pressure, caprock stress, fault activation index and geomechanics simulation data, and input the data into a pre-trained long-term stability evaluation model, and output an injection adjustment amplitude and a recommended amount of repair agent within a second time length in the future; wherein the long-term stability evaluation model is trained by reinforcement learning based on a defined state space and action space using a proximal policy optimization framework; and a repair processing module, configured to: based on the injection adjustment amplitude, the recommended amount of repair agent, a reservoir pressure gradient and a ground deformation rate, generate an executable repair strategy through a multi-objective optimization algorithm, and issue repair instruction sets according to the repair strategy to perform targeted repair.

[0015] The application further provides an electronic device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the carbon dioxide storage intelligent management and control method according to any one of the above-mentioned methods when executing the computer program.

[0016] The application further provides a non-transitory computer-readable storage medium, which stores a computer program, and the computer program is executed by a processor to implement the carbon dioxide storage intelligent management and control method.

[0017] The application further provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the carbon dioxide storage intelligent management and control method.

[0018] The carbon dioxide storage intelligent management and control method and system provided by the application, by acquiring multi-source data, performing space-time alignment processing, feature extraction processing and multi-scale hierarchical data fusion processing on the multi-source data, obtaining dynamic parameter field data, performing time sequence feature extraction and spatial feature coding based on the dynamic parameter field data, generating a space-time feature matrix, inputting the space-time feature matrix into a pre-trained short-term leakage prediction model, outputting high-risk region coordinates with a leakage probability greater than a preset value within a first time length in the future, acquiring reservoir pressure, caprock stress, fault activation index and geomechanical simulation data and inputting the data into a pre-trained long-term stability evaluation model, outputting gas injection adjustment amplitude and recommended amount of repair agent within a second time length in the future, based on the gas injection adjustment amplitude, the recommended amount of repair agent, the reservoir pressure gradient and the surface deformation rate, generating an executable repair strategy through a multi-objective optimization algorithm, issuing repair instruction sets according to the repair strategy to execute targeted repair, and realizing accurate three-dimensional modeling of underground reservoirs and surface deformation through hierarchical fusion technology, improve the leakage positioning and risk area identification accuracy, and form a full-process automatic closed loop of "prediction-decision-repair". BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the application or prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0020] Figure 1 is one of the flowcharts of the carbon dioxide storage intelligent management and control method provided by the application.

[0021] Figure 2 is another flowchart of the carbon dioxide storage intelligent management and control method provided by the application.

[0022] Figure 3 is a structural schematic diagram of the carbon dioxide storage intelligent management and control system provided by the application.

[0023] Figure 4 is a structural schematic diagram of the electronic device provided by the application. DETAILED DESCRIPTION

[0024] In order to make the objects, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below in combination with the drawings in the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the protection scope of the present application.

[0025] Key term definition: Carbon dioxide storage: Carbon dioxide storage refers to a process of storing captured carbon dioxide in a specific geological structure or marine environment for a long time by technical means, so as to isolate it from the atmosphere and reduce the concentration of greenhouse gases.

[0026] Carbon dioxide sweep range monitoring: The core goal of carbon dioxide sweep range monitoring is to dynamically track the migration path and concentration gradient of CO2 in a specific medium (such as underground rock layers, atmospheric layers, enclosed spaces), and to clarify the physical boundary of its diffusion. In carbon storage projects, the migration range of CO2 plume (i.e. the continuous distribution of CO2 in the ground) needs to be monitored to evaluate whether it exceeds the safe storage area.

[0027] Multi-modal perception: By integrating multiple heterogeneous sensors and monitoring technologies, multi-dimensional dynamic data (such as underground data, surface data, environmental data) of the carbon dioxide storage library are obtained in real time, heterogeneous data fusion algorithms are used to eliminate the spatial and temporal scale differences of multi-source data, and a unified dynamic feature matrix is constructed; edge computing nodes are deployed to realize data preprocessing and compressed transmission.

[0028] Autonomous decision-making: Based on the real-time analysis of multi-modal perception data by artificial intelligence models, a leakage risk assessment result and a repair strategy are dynamically generated.

[0029] Targeted repair: For the predicted leakage path or high-risk area, self-repairing materials are precisely placed to plug the seepage channel, and a multi-level repair strategy (first plugging the main channel, then reinforcing the microcracks) is designed; the repair effect is verified by a digital twin model and the scheme is dynamically adjusted.

[0030] Intelligent closed-loop system: A control system that realizes the automation of the whole process of "perception → analysis → decision-making → execution → feedback", the core logic of which is: multi-modal perception data are input into the AI decision engine in real time; repair instructions are generated and intelligent valves / unmanned aerial vehicles are driven to execute; after repair, data are fed back to the system to verify the effect and optimize subsequent decision-making.

[0031] In view of the bottleneck of the prior art, the application innovatively puts forward a closed-loop management and control system of "perception-decision-execution-optimization". Through multi-modal data fusion technology, the information island is broken, and a high-precision reservoir digital twin model is constructed; the deep learning algorithm is used to realize intelligent risk prediction; combined with automatic repair equipment for rapid response; and a continuous optimization mechanism based on actual operation data is established. This system not only greatly improves the safety and reliability of the sealing process, but also provides key technical support for realizing large-scale CO2 sealing.

[0032] The present application aims to solve the problems of passivity, inefficiency and high cost in the existing carbon dioxide storage leakage monitoring and repair technology. The traditional method relies on static geological models and artificial experience judgment, which is difficult to capture dynamic leakage risk in real time, and the repair response is lagging, resulting in insufficient sealing safety. In view of the above defects, the present application constructs a "perception-decision-repair-verification" whole-process closed-loop system through multi-modal data fusion (underground and surface data), AI-driven dynamic risk assessment model (combination of LSTM and reinforcement learning) and intelligent triggering mechanism of targeted self-repairing material, realizes active early warning, accurate positioning and rapid plugging of leakage risk, and improves the long-term stability and operation economy of the storage.

[0033] Figure 1 is one of the process schematic diagrams of the carbon dioxide storage intelligent management and control method provided by the present application, as Figure 1 shown, the method comprises: Step S1, acquiring multi-source data, performing spatio-temporal alignment processing, feature extraction processing and multi-scale hierarchical data fusion processing on the multi-source data to obtain dynamic parameter field data; wherein the multi-source data includes four-dimensional seismic data, distributed optical fiber acoustic signal, surface millimeter-level deformation data, surface CO2 concentration anomaly data and temperature gradient change data, and the dynamic parameter field data includes CO2 saturation field, reservoir pressure field, micro-fracture density and surface leakage probability.

[0034] Acquiring multi-source data, performing spatio-temporal alignment processing, feature extraction processing and multi-scale hierarchical data fusion processing on the multi-source data to obtain dynamic parameter field data.

[0035] Among them, the multi-source data includes: ① Four-dimensional seismic data Source: Three-dimensional seismic exploration (time-lapse seismic) implemented in stages before sealing, during injection and after sealing.

[0036] Data form: Three-dimensional reservoir grid data (horizontal resolution 100m x 100m, vertical 10m), time series (year / quarter).

[0037] Key parameters: Seismic wave amplitude, velocity, differential imaging index (NRMS value).

[0038] ②Distributed Acoustic Sensing (DAS) Source: Fiber-optic sensor network deployed along injection wells, monitoring wells, and reservoir pathways.

[0039] Data form: High-frequency time series (sampling rate ≥ 1 kHz), spatial coverage of key reservoir pathways.

[0040] Key parameters: Acoustic energy density, microseismic event spatiotemporal coordinates (50-500 Hz frequency band).

[0041] ③Ground deformation data at millimeter level (InSAR, Interferometric Radar) Source: Satellite Synthetic Aperture Radar Interferometry (e.g., Sentinel-1).

[0042] Data form: Two-dimensional ground deformation field (resolution 100 m x 100 m), time series (revisit period 6-12 days).

[0043] Key parameters: Deformation rate (mm / year), deformation gradient direction.

[0044] ④Ground CO2 concentration anomaly data Source: Multi-spectral sensor (e.g., short-wave infrared spectrometer) mounted on a drone.

[0045] Data form: Two-dimensional point cloud or grid data (resolution 10 m x 10 m).

[0046] Key parameters: CO2 concentration (ppm) inverted based on Beer-Lambert's law, leakage probability thermal map.

[0047] ⑤Temperature gradient change data Source: Infrared thermal imager on a drone + ground temperature sensors.

[0048] Data form: Two-dimensional temperature field (resolution 1 m x 1 m), time series (hourly level).

[0049] Key parameters: Temperature gradient (°C / m), abnormal temperature rise area coordinates.

[0050] Multi-source data spatiotemporal alignment and preprocessing: Input data: 4D seismic data: reservoir grid data (horizontal resolution 100 m x 100 m, vertical 10 m), time series (annual / quarterly timestamps, e.g. at the midpoint of the exploration period). DAS acoustic signals: high-frequency time series (sampling rate > 1 kHz, spatially distributed along the well path). InSAR deformation data: original interferogram sequence (revisit period 6-12 days), deformation rate in mm / year (converted to daily). UAV data: CO2 concentration point cloud (resolution 10 m x 10 m, scanned hourly) infrared temperature field (resolution 1 m x 1 m, scanned hourly); ground sensors: real-time temperature / pressure sequence (sampling rate 10 Hz).

[0051] Operation flow: Time reference synchronization (based on DAS millisecond-level timestamps): InSAR deformation data: Based on the original interferogram cumulative deformation sequence, generate continuous daily deformation (error < 0.1 mm / day) by cubic spline interpolation, and then linearly interpolate to millisecond-level timestamps.

[0052] Example: convert annual rate 18.25 mm / year to 0.05 mm / day, and align to the value at time t0.

[0053] 4D seismic data: Take the statistical characteristics (e.g. daily frequency of microseismic events) of DAS data within the exploration period (e.g. quarterly) as the representative value of that period, and associate it with the seismic timestamp.

[0054] UAV data: Perform time linear interpolation on CO2 concentration and temperature field (weighting the data of the previous and next two scans), generating millisecond-level instantaneous field.

[0055] Ground sensors: directly resample to match DAS timestamps.

[0056] Spatial grid unification (target grid: 100 m x 100 m seismic grid): DAS path data: mapped to the seismic grid by Kriging interpolation, preserving acoustic energy density characteristics (RMSE < 5%).

[0057] UAV data: CO2 concentration point cloud -> Gaussian kernel smoothing (σ = 20 m) to the seismic grid, generating a leakage probability heat map.

[0058] Temperature field -> first calculate the gradient field, then Gaussian smoothing and down-sampling.

[0059] InSAR deformation field: directly resample to 100 m grid, PCA separation to construct motion interference.

[0060] Key precision control: Time alignment error: ≤1 ms (clock synchronization protocol).

[0061] Spatial interpolation error: RMSE ≤5% (cross-validation optimization).

[0062] Multimodal data feature extraction and noise reduction: Feature extraction on time-lapse seismic difference imaging data (NRMS value), reservoir pressure field inversion through rock physics model: ( Vp is the P-wave velocity, Vs is the S-wave velocity, and ρ is the density of the rock) Wavelet packet decomposition (WPD) is performed on the original high-frequency acoustic time series (including mechanical noise) to extract microseismic event features (frequency band 50-500 Hz), and DBSCAN clustering algorithm is used to mark CO2 migration path (clustering radius ≤50 m). The DBSCAN clustering algorithm marks the spatiotemporal coordinates of microseismic events, and its essence is to indirectly invert the CO2 migration path by analyzing the distribution pattern of rock failure signals.

[0063] Principal component analysis (PCA) is performed on the InSAR deformation rate field (including tectonic movement interference) of surface deformation data to separate tectonic deformation (low frequency) and leakage deformation (high frequency), and extract deformation gradient direction ( ) and CO2 plume migration direction correlation. CO2 plume migration direction and CO2 migration path are different scales of the same dynamic process. CO2 plume migration direction is the instantaneous flow direction (vector field) of underground CO2 plume in the reservoir, and CO2 migration path is the cumulative trajectory (linear set) of CO2 plume advancing with time.

[0064] The original data of unmanned aerial vehicle multispectral and infrared thermal imaging are processed to invert CO2 concentration and temperature data, and CO2 concentration is calculated based on Beer-Lambert law: ( σ is the absorption cross section, L is the optical path, and I0 is the incident / transmitted light intensity). The temperature gradient field is calculated: , T represents temperature, and y represents the horizontal and vertical coordinates.

[0065] Multiscale hierarchical data fusion: First layer: build a reference framework (low resolution layer); second layer: embed high-frequency details (medium resolution layer); third layer: surface data correction (high resolution layer).

[0066] The final dynamic parameter field data includes CO2 saturation field, reservoir pressure field, micro-fracture density and surface leakage probability.

[0067] In step S2, time sequence feature extraction and spatial feature coding are performed based on the dynamic parameter field data to generate a space-time feature matrix; the space-time feature matrix includes pressure statistics, CO2 saturation gradient field and deformation-pressure coupling coefficient.

[0068] Input data preparation and feature engineering: Input data: ① Dynamic parameter field: fused four-dimensional CO2 saturation field (resolution 5m~100m); reservoir pressure field, micro-fracture density, and surface leakage probability (based on multi-modal fusion output). ② Auxiliary data: geological structure map (fault distribution, caprock thickness), historical operation record (gas injection rate, remediation agent dosage), and environmental sensitive area (water source, residential area coordinates).

[0069] Operation process: ① Time sequence feature extraction: Based on the reservoir pressure field in the dynamic parameter field, the pressure change rate (mean, variance, trend slope), and microseismic frequency (mean, variance, trend slope) are extracted with a 24-hour sliding window.

[0070] ② Spatial feature coding: calculate CO2 saturation gradient field: , wherein S represents CO2 volume saturation, x, y, z represent three-dimensional coordinates; calculate deformation-pressure correlation coefficient (Pearson correlation coefficient): , wherein D represents deformation, and ΔP represents pressure change.

[0071] Integrate the time sequence feature extraction results (pressure / microseismic statistics) and spatial feature coding results (gradient field, correlation coefficient) to construct a multi-dimensional feature matrix as the input of the AI model.

[0072] The complete feature vector of the i-th grid at time t is: wherein represents the mean of pressure, represents the variance of pressure, represents the trend slope of pressure; represents the mean of microseismic frequency, represents the variance of microseismic frequency, represents the trend slope of microseismic frequency.

[0073] Step S3, inputting the spatio-temporal feature matrix into a pre-trained short-term leakage prediction model to output high-risk region coordinates in which the leakage probability in a first time length in the future is greater than a preset value; wherein the short-term leakage prediction model is obtained by inputting the spatio-temporal feature matrix of historical data into an LSTM network and taking historical leakage labels as output labels.

[0074] The short-term leakage prediction model is used to output high-risk region coordinates in which the leakage probability in a first time length in the future is greater than a preset value, and the first time length in the future is, for example, 72 hours. The short-term leakage prediction model is obtained by inputting the spatio-temporal feature matrix of historical data into an LSTM network and taking historical leakage labels as output labels.

[0075] Short-term leakage prediction model construction (LSTM network): Input data: spatio-temporal feature matrix (time series statistics + spatial gradient); historical leakage labels (0 / 1 labeled data).

[0076] Operation process: ① Network structure design. ② Loss function and training: binary cross-entropy (Binary Cross-Entropy), formula: Wherein is the true label (0 or 1), is the predicted probability, is the number of samples.

[0077] Optimizer: Adam algorithm (efficient processing of sparse gradient, low memory occupation), early stopping method to monitor validation set loss.

[0078] Training rounds: usually set to 1000 iterations, and the early stopping method (Early Stopping) is used to monitor the validation set loss.

[0079] Finally, the leakage probability (0~1) in the next 72 hours is output, and when the probability is greater than or equal to 0.5, an early warning is triggered, and high-risk region coordinates are output.

[0080] Step S4, obtaining reservoir pressure, caprock stress, fault activation index and geomechanical simulation data and inputting them into a pre-trained long-term stability evaluation model to output gas injection adjustment amplitude and recommended amount of repair agent in a second time length in the future; wherein the long-term stability evaluation model is obtained by reinforcement learning according to a defined state space and action space using a proximal policy optimization framework.

[0081] The long-term stability evaluation model is used to output the gas injection adjustment range and the recommended amount of repair agent within a second future time length, such as 10 years in the future. The long-term stability evaluation model is obtained by reinforcement learning training according to the defined state space and action space using a proximal policy optimization framework.

[0082] Mechanism of training model with simulation data: 1. Construct a digital twin environment Based on actual geological parameters (reservoir pressure, caprock stress, fault distribution, etc.), a geomechanical simulation model is established to simulate the evolution of the reservoir for up to 10 years under different gas injection / repair strategies.

[0083] The simulation outputs dynamic state data (such as pressure changes, fault activation index, CO2 leakage amount, etc.) as the training environment for reinforcement learning.

[0084] 2. Reinforcement learning framework design Including state space (State) design, action space design, reward function design.

[0085] 3. PPO algorithm training process Actor-Critic network structure: Policy network (Actor): input current state, output action probability distribution (select gas injection or repair action).

[0086] Value network (Critic): evaluate the long-term return of the current state, guide the optimization direction of the policy.

[0087] Interactive training: Each training round simulates a 5-year storage process, and the state-action sequence is generated monthly.

[0088] The reward value is fed back through the simulation environment, and the policy gradient is updated to update the network weights.

[0089] Use proximal policy optimization (PPO) to constrain the update range to avoid unstable training.

[0090] Method for generating 10-year optimal strategy: 1. Strategy generation principle The trained Actor network learns the mapping rule from state to action: For example: when the caprock stress increase amplitude > threshold, automatically trigger the "reduce gas injection rate + increase repair agent injection" action.

[0091] Through simulation to deduce the state in the next 10 years, call the Actor network to generate the action sequence monthly / yearly.

[0092] 2. Multi-objective optimization output Under the condition of meeting the constraint condition (such as 10 years cumulative leakage ≤1000 tons): Dynamic balance between sealing safety (minimize leakage risk) and economy (repair agent consumption ≤500 m³ / year).

[0093] Obtain reservoir pressure, caprock stress, fault activation index and geomechanics simulation data and input into the long-term stability evaluation model pre-trained to output the gas injection adjustment amplitude and repair agent recommended consumption in the future second time length.

[0094] Step S5, based on the gas injection adjustment amplitude, the repair agent recommended consumption, the reservoir pressure gradient and the surface deformation rate, an executable repair strategy is generated through a multi-objective optimization algorithm, and repair instruction set is issued according to the repair strategy to execute targeted repair. Table 1 is an explanation of the parameters related to the repair strategy.

[0095] Table 1

[0096] The long-term stability evaluation output gas injection adjustment amplitude (such as-10%), repair agent recommended consumption (such as 50 m³ / month) and dynamic parameter field reservoir pressure gradient, surface deformation rate synergistically act through a multi-objective optimization algorithm to generate an executable repair strategy, and repair instruction set is issued according to the repair strategy to execute targeted repair.

[0097] The carbon dioxide storage intelligent management and control method provided by the application obtains multi-source data, performs spatio-temporal alignment processing, feature extraction processing and multi-scale hierarchical data fusion processing on the multi-source data to obtain dynamic parameter field data, performs time series feature extraction and spatial feature coding based on the dynamic parameter field data to generate a spatio-temporal feature matrix, inputs the spatio-temporal feature matrix into a short-term leakage prediction model pre-trained to output high-risk area coordinates with a leakage probability greater than a preset value in a future first time length, obtains reservoir pressure, caprock stress, fault activation index and geomechanics simulation data and inputs them into a long-term stability evaluation model pre-trained to output gas injection adjustment amplitude and repair agent recommended consumption in a future second time length, based on the gas injection adjustment amplitude, the repair agent recommended consumption, the reservoir pressure gradient and the surface deformation rate, an executable repair strategy is generated through a multi-objective optimization algorithm, and repair instruction set is issued according to the repair strategy to execute targeted repair, through hierarchical fusion technology, accurate three-dimensional modeling of underground reservoir and surface deformation is realized, leakage positioning and risk area identification accuracy are improved, and a full-process automatic closed loop of “prediction-decision-repair” is formed.

[0098] The application provides a carbon dioxide storage intelligent management and control method, and the multi-scale hierarchical data fusion processing comprises the following steps: obtaining and outputting a pore pressure field and a CO2 saturation field through reservoir parameter inversion based on four-dimensional seismic data; obtaining an enhanced reservoir model through wavelet domain fusion and micro-fracture correction based on the pore pressure field, the CO2 saturation field and DAS acoustic signal characteristics, wherein the enhanced reservoir model comprises a CO2 saturation field, a pore pressure field and a micro-fracture density distribution; and obtaining a corrected four-dimensional dynamic parameter field through deformation-pressure correlation modeling and leakage probability spatial fusion based on the enhanced reservoir model, an InSAR deformation field, a CO2 leakage probability thermal map and a temperature gradient field, wherein the corrected four-dimensional dynamic parameter field is integrated with a reservoir internal pressure distribution, a CO2 migration path and a surface leakage risk thermal map.

[0099] 1. First layer: low-resolution reference layer (constructing reservoir dynamic field) Input: four-dimensional seismic data body (containing three-dimensional seismic imaging results before storage, during gas injection and after storage, horizontal resolution is 100m x 100m, and vertical resolution is 10m); differential imaging index (NRMS value three-dimensional grid, obtained by calculating time-lapse seismic data).

[0100] Operation process: ① Reservoir parameter inversion: The seismic wave velocity (Vp), the shear wave velocity (Vs) and the density (rho) are converted into the pore pressure field by using a rock physics model: (empirical coefficient calibration) The CO2 saturation field is inverted based on the NRMS value: ② Macro modeling: The CO2 plume boundary is extracted: The reservoir pressure distribution map (resolution is 100m) is generated.

[0101] Output: reservoir CO2 saturation field , pore pressure field 2. Second layer: embedding medium-resolution acoustic high-frequency details (DAS data fusion) Input: reservoir CO2 saturation field , pore pressure field output by the first layer; DAS acoustic signal characteristics (microseismic event space-time coordinate set) Operation process: ​​​① Wavelet domain fusion: 3-level discrete wavelet transform (DWT) is applied to seismic data and DAS signals respectively, using Daubechies wavelet basis (db4).

[0102] Decomposition components: Low-frequency approximation component (LLL): seismic data (carrying reservoir macrostructure).

[0103] High-frequency detail components (HHL, HLH, LHH): DAS data (reflecting microfractures and pressure fluctuations).

[0104] Fusion formula: .

[0105] wherein, represents the LLL component of seismic data, represents the HHL component of DAS data, represents the HLH component of DAS data, represents the LHH component of DAS data. represents the inverse discrete wavelet transform operation.

[0106] ② Microfracture correction: project the microseismic event coordinates located by DAS to the fusion grid, and keep in the area where NRMS>10%, otherwise move to the nearest high NRMS point (suppress false fractures).

[0107] Output: enhanced reservoir model (resolution 10m level) containing: CO2 saturation field , pressure field , microfracture density distribution.

[0108] 3. Third layer: high-resolution surface correction (InSAR and UAV data fusion) Input: enhanced reservoir model output by the second layer, InSAR deformation field (leakage-related deformation rate ), UAV data: CO2 leakage probability thermal map , temperature gradient field Operation process: ① Deformation-pressure correlation modeling: establish the quantitative relationship between deformation rate and reservoir pressure change : wherein, represents the theoretical value of surface deformation rate predicted based on reservoir pressure change. Coefficient is fitted by least squares based on historical gas injection data.

[0109] When the measured , it is determined as a boundary anomaly and triggers correction ②Leakage probability spatial fusion: Weighted formula: wherein, represents the surface leakage risk comprehensive weight, represents the surface CO2 concentration, represents the reservoir cap rock top boundary depth.

[0110] For The area with >0.5 is subjected to grid local encryption (the resolution is improved to 5m) Output: The modified four-dimensional dynamic parameter field integrates: the reservoir internal pressure distribution, the CO2 migration path and the surface leakage risk thermal map.

[0111] The carbon dioxide storage intelligent management and control method provided by the application obtains and outputs a pore pressure field and a CO2 saturation field through reservoir parameter inversion based on four-dimensional seismic data, obtains an enhanced reservoir model through wavelet domain fusion and microfracture correction based on the pore pressure field, the CO2 saturation field and DAS acoustic signal characteristics, the enhanced reservoir model includes a CO2 saturation field, a pore pressure field and a microfracture density distribution, obtains a modified four-dimensional dynamic parameter field through deformation-pressure correlation modeling and leakage probability spatial fusion based on the enhanced reservoir model, an InSAR deformation field, a CO2 leakage probability thermal map and a temperature gradient field, and the modified four-dimensional dynamic parameter field integrates a reservoir internal pressure distribution, a CO2 migration path and a surface leakage risk thermal map, integrates four-dimensional seismic, optical fiber acoustic, satellite deformation and unmanned aerial vehicle data, realizes accurate three-dimensional modeling of underground reservoirs and surface deformation through hierarchical fusion technology, and improves leakage positioning and risk area identification accuracy.

[0112] According to the carbon dioxide storage intelligent management and control method provided by the application, the LSTM network includes two layers of LSTM, an attention mechanism module is inserted between layers, a Sigmoid activation function is used in an output layer, and a random inactivation layer with a dropout rate of 0.2 is added.

[0113] The training of the short-term leakage prediction model adopts an LSTM network. Network structure design: 2 layers of LSTM (the number of hidden units is greater than or equal to 128), an attention mechanism module is inserted between layers, a Sigmoid activation function is used in an output layer. Dropout layer: a random inactivation layer with a dropout rate of 0.2 is added to prevent overfitting. Through the attention mechanism and the Dropout layer, substantial improvement is realized, and the sensitivity and generalization ability of the model to leakage risk are improved.

[0114] The carbon dioxide storage intelligent management and control method provided by the application, by setting an LSTM network including two layers of LSTM, inserting an attention mechanism module between the layers, using a Sigmoid activation function for the output layer, and adding a random inactivation layer with a dropout rate of 0.2, the sensitivity of the model to the leakage risk and the generalization ability are improved.

[0115] According to the carbon dioxide storage intelligent management and control method provided by the application, the state space of the long-term stability evaluation model includes reservoir pressure, caprock stress, fault activation index and repair agent dosage, the action space includes gas injection adjustment information and repair agent injection information, and the reward function is set by balancing stability, economy and safety during training.

[0116] Long-term stability evaluation model construction (reinforcement learning): Input data: dynamic parameter field (reservoir pressure, caprock stress, fault activation index, geomechanical simulation data (simulate 10 years of reservoir evolution)).

[0117] Operation process: first, environment modeling is performed, the state space (reservoir pressure, caprock stress, fault activation index, repair agent dosage) and the action space (gas injection adjustment, repair agent injection) are defined, a multi-objective reward function is designed, and stability, economy and safety are balanced. Then, using the proximal policy optimization (PPO) framework, the Actor-Critic network generates a strategy, the model is trained based on geomechanical simulation data, and 10 years of reservoir evolution are simulated. Finally, the optimal gas injection and repair strategy for 10 years (such as annual repair agent dosage ≤500 m³) is generated.

[0118] State space (State): Including key parameters such as reservoir pressure, caprock stress, fault activation index, and historical repair agent dosage.

[0119] Action space (Action): Gas injection rate adjustment (-10%, 0, +10%); Repair agent injection amount (0-100 m³ / month).

[0120] Reward function (Reward): Positive reward: inhibit caprock stress change / fault activation (such as + reward when stress is stable).

[0121] Negative punishment: excessive use of repair agent (economic penalty), CO2 leakage (high penalty).

[0122] Different effects of repair agent dosage: state space: repair agent dosage as a historical cumulative indicator, reflecting the degree of intervention of the system. Action space: repair agent dosage as an adjustable decision variable, directly controlling the repair behavior.

[0123] a. Environment Definition: State: Reservoir pressure, Caprock stress, Fault activation index.

[0124] Action: Adjust injection rate, Initiate preventive remediation injection. Adjust injection rate (-10%, 0, +10%). Control preventive remediation agent injection volume (0-100 m³ / month).

[0125] Reward: Stability reward: Inhibit caprock stress change and fault activation. Economic penalty: Reduce remediation agent usage. Leakage penalty: Significant deduction of reward upon severe leakage.

[0126] b. Algorithm and Network Design: PPO (Proximal Policy Optimization) based on Actor-Critic framework, balancing exploration and exploitation.

[0127] Network Structure: Actor: Input state, output action probability distribution (injection adjustment and remediation agent usage).

[0128] Critic: Evaluate state long-term reward, guide policy optimization.

[0129] c. Training Process: Data Generation: Simulate reservoir evolution under different injection strategies through geomechanical simulation (10-year time span).

[0130] Interactive Sampling: Simulate 5-year storage process per training round, iterate state and action monthly.

[0131] Parameter Update: Batch process data, optimize network weights, target is to maximize cumulative reward.

[0132] A complete training cycle is called a "round", each round simulates a 5-year (60-month) carbon dioxide storage process.

[0133] Objective: Through simulating long-term storage processes (such as geological deformation, pressure accumulation, etc. slow time-varying effects), evaluate the impact of different strategies on long-term reservoir stability (such as 10-year cumulative leakage ≤1000 tons).

[0134] Decision Frequency: The model makes a decision every month, i.e. every time step of 1 month, it performs the following operations: ① State Observation Obtain the current reservoir state: pressure value, caprock stress, fault activation index, remediation agent historical usage, etc.

[0135] ② Action Execution Action according to the policy network (Actor) output: adjust the injection rate (-10% / 0 / +10%), inject repair agent (0-100 m³ / month).

[0136] ③ Environmental feedback Calculate multi-objective reward: Stability reward: suppress the stress mutation of the caprock (e.g. stress change ≤0.1 MPa) Economic penalty: repair agent usage exceeds the baseline value (e.g. >50 m³ / month) Leakage penalty: simulate the amount of CO2 leakage (e.g. 1 ton of leakage deducts 10% of the reward value) ④ State update According to the action result, update the reservoir state to the next month through the geomechanical simulation model.

[0137] The carbon dioxide storage intelligent management and control method provided by the application improves the prediction accuracy of the long-term stability evaluation model by reasonably setting the state space, action space, and reward function.

[0138] According to the carbon dioxide storage intelligent management and control method provided by the application, the method further comprises: verifying the short-term leakage prediction model according to the AUC index and the false positive rate, verifying the long-term stability evaluation model according to the cumulative leakage amount index, and periodically updating the short-term leakage prediction model and the long-term stability evaluation model using a federated learning framework.

[0139] Model verification and dynamic update: Input data: measured pressure data, microseismic positioning results, newly collected deformation and CO2 concentration data.

[0140] Operation process: This step mainly verifies and updates the short-term model and the long-term model. Verification index: short-term model: AUC ≥0.85, false positive rate ≤10%; long-term model: cumulative leakage amount ≤1000 tons (10 years). Using a federated learning framework, the model is updated quarterly, and the deformation-pressure coefficient (k, b) is adaptively optimized.

[0141] The verification target of "cumulative leakage amount ≤1000 tons (10 years)" is achieved through a dynamic proxy index and a prediction backtracking mechanism, rather than relying on 10 years of measured data. The essence of long-term model verification: predicted value verification rather than measured value.

[0142] Adaptive optimization of the deformation-pressure coefficient (k, b) is achieved through incremental learning dynamic parameter tuning, rather than full-amount model reconstruction.

[0143] Initial establishment: fitting a linear relationship using historical data (least squares method).

[0144] Real-time calibration: update parameters by Recursive Least Squares (RLS) every time a new data set is acquired where, represents a new pressure monitoring data set, represents a new deformation monitoring data set, is a dynamic gain matrix, is a background construction motion correction term.

[0145] The carbon dioxide storage intelligent management and control method provided by the application verifies the short-term leakage prediction model according to the AUC index and the false positive rate, verifies the long-term stability evaluation model according to the cumulative leakage amount index, and periodically updates the short-term leakage prediction model and the long-term stability evaluation model by using a federal learning framework, thereby realizing verification and updating of the model.

[0146] According to the carbon dioxide storage intelligent management and control method provided by the application, the objective function of the multi-objective optimization algorithm is expressed as min (repair cost + λ * environmental risk), λ represents an environmental risk penalty weight coefficient, and the constraint condition is that the pressure recovery rate is greater than a first preset value and the leakage probability is less than a second preset value; the repair instruction set includes downhole nanorobot injection coordinates, sealing material types and dosages, and ground laser mineralization device working parameters.

[0147] Repair strategy generation and instruction issuance: Input data: ① AI decision result: high-risk area coordinates of short-term leakage prediction probability (≥0.5). The output of long-term stability evaluation is the gas injection adjustment amplitude (such as -10%) and the recommended dosage of repair agent (such as 50 m³ / month). ② Dynamic parameter field: leakage point permeability distribution, reservoir pressure gradient, and surface deformation rate.

[0148] Operation process: first, multi-objective optimization, objective function: , constraint condition: pressure recovery rate ≥ 90%, leakage probability ≤ 5%. Then, repair means matching is performed: downhole repair parameters and ground repair parameters are determined. Finally, the repair instruction set is output: downhole nanorobot injection coordinates, sealing material types and dosages; ground laser mineralization device working parameters (power, irradiation range, duration).

[0149] Table 2 shows the meanings of different values of the environmental risk penalty weight coefficient λ.

[0150] Table 2

[0151] ​Multi-objective optimization: balance repair cost, environmental risk and effect at the macro level through objective function, and determine resource allocation principles (such as: prioritize processing high pressure gradient area, repair agent dosage floating range).

[0152] The carbon dioxide storage intelligent management and control method provided by the application improves the repair effect by reasonably setting the objective function and constraint condition of the multi-objective optimization algorithm and determining the repair instruction set.

[0153] According to the carbon dioxide storage intelligent management and control method provided by the application, the method further comprises: selecting a sealing material through a permeability matching algorithm, which is represented as: , wherein: represents the material coefficient, represents the permeability of the leakage area; the ground laser mineralization device working parameter comprises laser power, and the method further comprises: determining the laser power through dynamic adaptation with the CO2 concentration, which is represented as: , wherein, represents the laser power, represents the energy conversion coefficient, represents the carbon dioxide concentration, represents the leakage area.

[0154] The downhole repair parameters and the ground repair parameters are determined by performing repair means matching. Wherein: ① Downhole repair: the permeability matching algorithm selects a sealing material (gel, nanoparticle): is the material coefficient, is the permeability of the leakage area); ② Ground repair: dynamic adaptation of laser power and CO2 concentration: is the energy conversion coefficient, is the leakage area), and the CO2 mineralization rate is ≥1 t / h.

[0155] Repair means matching: at the micro level, according to the specific constraints (such as leakage point coordinates, repair agent dosage range) output by the optimization result, and combining physical parameters (permeability, CO2 concentration), an operable scheme (such as material type, laser power) is selected. Table 3 shows the use of different parameters in repair means matching.

[0156] Table 3

[0157] Gel and nanoparticle have significantly different material coefficient ranges, which directly determine their applicable scenarios in permeability matching.

[0158] Gel: mainly for low-to-medium permeability leakage, with the advantages of low cost and high environmental friendliness;​​ Nanoparticles: Dominant in high-osmotic leakage, advantages are fast response, high temperature resistance.

[0159] Table 4 is the performance of different sealing materials.

[0160] Table 4

[0161] The core mechanism to ensure that the mineralization rate is greater than or equal to 1 t / h is to control the dynamic adaptation of laser power: Table 5 is the meaning information of the related parameters.

[0162] Table 5

[0163] The determination of the amount of sealing material is the result of dynamic collaborative decision-making of long-term model output and short-term leakage risk: Table 6 is the meaning information of the related parameters.

[0164] Table 6

[0165] The calculation basis of the irradiation range is: : irradiation range; : leakage area (real-time mapping by unmanned aerial vehicle infrared imaging + InSAR deformation boundary); : safety expansion coefficient (default 0.2, i.e. expand 20% boundary, prevent plume diffusion).

[0166] The calculation basis of the duration is: : duration; : target mineralization amount (short-term predicted leakage amount x safety factor 1.5); Mineralization rate: dynamically guaranteed by laser power ≥1t / h; : buffer time (default 10% total time, to deal with device response delay).

[0167] The carbon dioxide sealing intelligent management and control method provided by the application selects sealing materials through a transmittance matching algorithm, determines laser power through dynamic adaptation with CO2 concentration, and gives a determination method of other repair parameters, further ensuring the repair effect.

[0168] According to the carbon dioxide storage intelligent management and control method provided by the application, in the execution process of the targeted repair, the method further comprises: for downhole repair, injecting adaptive gel according to the permeability matching result, and controlling the gel solidification time according to the time control equation; for ground repair, by adjusting the laser power in real time, the mineralization rate is ensured to meet the set requirement; after the execution of the targeted repair is completed, the method further comprises: extracting the pressure recovery rate and the leakage concentration decay rate, and determining the repair effect according to the pressure recovery rate and the leakage concentration decay rate.

[0169] Targeted repair execution: Input data: repair instruction set (coordinates, materials, equipment parameters).

[0170] Operation process: ① Downhole repair execution: nano robot cluster control: guide the robot to the leakage point through the magnetic navigation system (positioning accuracy ≤0.1 m). According to the permeability matching result, inject adaptive gel, and trigger the time control equation: : gel solidification time (unit: seconds).

[0171] : the volume of the injected plugging material (unit: cubic meters).

[0172] : the flow rate of the plugging material in the wellbore (unit: m / s).

[0173] : the radius of the injection well (unit: meters).

[0174] : the cross-sectional area of the injection well (i.e. the channel area of fluid flow).

[0175] ② Ground repair execution: laser-induced mineralization: 10.6 μm wavelength laser irradiation of the leakage area, and excitation of minerals (such as olivine) to react with CO2: Adjust the laser power in real time to ensure that the mineralization rate is ≥1 t / h.

[0176] Repair effect real-time monitoring: Input data: downhole pressure gauge, temperature sensor (sampling rate 10 Hz), ground multispectral sensor, InSAR deformation data.

[0177] Operation process: extract the key indicators of pressure recovery rate and leakage concentration decay rate: , wherein, is the pressure recovery rate, is the pressure before repair, Ppost Pbase wherein, LDR Cpre Cpost if , the repair is determined to be ineffective, triggering a secondary decision. Output the repair effectiveness report (success: pressure recovery rate ≥ 90%, concentration drop ≥ 80%; partial success: pressure recovery rate 70%~90%, start dynamic weight adjustment; failure: trigger emergency protocol (such as increasing the number of repair robots)).

[0178] In the input data, downhole data includes pressure data (from downhole pressure gauges) and temperature data (from downhole temperature sensors); surface data includes surface multispectral sensor data (for monitoring CO2 concentration) and InSAR deformation data (monitoring surface deformation).

[0179] Use of temperature data: 1. Dynamic control of repair process Downhole temperature sensors monitor the curing state of the sealing material (such as self-adaptive gel) in real time: Gel curing releases heat, and the temperature gradient changes ( ) reflect the progress of material reaction.

[0180] If the temperature does not reach the threshold value (such as temperature rise ≥ 5℃) within the expected time , it is determined that the material injection is abnormal, triggering adjustment instructions.

[0181] Ground-based infrared thermal imaging (carried by drones) monitors the laser mineralization area: The reaction of minerals (such as olivine) with CO2 is an exothermic process, and the temperature field changes verify the mineralization efficiency.

[0182] The mineralization rate is positively correlated with the temperature rise, and when the temperature does not reach the expected value (such as local temperature rise ≥ 10℃), the laser power is dynamically increased .

[0183] 2. Auxiliary verification of repair effect Calculate the temperature recovery rate: Tbackground .

[0184] Success criteria: abnormal temperature rise area disappears (RT ≥ 85%).

[0185] Use of deformation data (InSAR): 1. Geological stability assessment Monitoring the deformation rate of the surface after repair ): When the repair is successful, the deformation rate should return to the baseline level (e.g., ≤1 mm / year).

[0186] If the deformation continues (e.g., ≥3 mm / year), it indicates that the reservoir pressure has not been completely released, and there is a risk of secondary leakage.

[0187] Deformation-pressure correlation verification: The deformation rate after repair should be strongly correlated with the degree of reservoir pressure recovery (e.g., ≥3 mm / year): When the Pearson correlation coefficient r < 0.5, it is determined that the repair is ineffective.

[0188] 2. Accurate positioning of the repair boundary InSAR deformation field and UAV CO2 concentration thermal map superposition: The deformation center coincides with the high-value area of concentration → confirms that the main leakage point has been plugged; The deformation edge continues to spread → indicates that the micro-cracks are extended, and additional repair is needed.

[0189] The system verifies the cross-verification results by temperature + deformation + pressure + concentration four-dimensional data, as shown in Table 7.

[0190] Table 7

[0191] In real-time monitoring of repair effectiveness, the system divides the repair status according to multi-dimensional indicators, and the concentration decay rate (Rc) is a necessary condition for partial success determination. Table 8 shows the relationship between repair status and parameters.

[0192] Table 8

[0193] Implementation mechanism of dynamic weight adjustment: When it is determined to be partially successful, the system starts dynamic weight adjustment to optimize the parameter weights of the subsequent decision model. The mathematical essence is to adjust the balance between economy and safety in the reward function through reinforcement learning. The specific process is as follows: 1. Adjustment object: reward function weight parameters In the reinforcement learning model, the reward function is defined as: : stability reward (negatively correlated with caprock stress change);​​ : repair cost (unit: ten thousand yuan / m³); : leakage penalty (proportional to the scale of leakage); : weight coefficient (initial value ).

[0194] 2. Dynamic adjustment formula According to the dominant failure type of partial success, the weight is adjusted in proportion: : economic weight Dynamic adjustment amount; : learning rate, control weight adjustment range (default ); : pressure recovery rate (pressure recovery degree after repair); : economic weight before adjustment.

[0195] The carbon dioxide storage intelligent management and control method provided by the application injects self-adaptive gel according to the permeability matching result through downhole repair, and controls the time required for gel solidification according to the time control equation, for ground repair, through real-time adjustment of laser power, ensures that the mineralization rate meets the set requirements, improves the repair reliability; by extracting the pressure recovery rate and the leakage concentration decay rate, and determining the repair effect according to the pressure recovery rate and the leakage concentration decay rate, the real-time monitoring of the repair effect is realized.

[0196] According to the carbon dioxide storage intelligent management and control method provided by the application, after the targeted repair is completed, the method further comprises: obtaining a repair efficiency report and real-time monitoring data, feeding back the repair efficiency report and the real-time monitoring data to the short-term leakage prediction model and the long-term stability evaluation model, so that the short-term leakage prediction model is trained and the long-term stability evaluation model is adjusted; based on the federal learning framework, after training at each local node, the parameters are encrypted and uploaded to the central server for aggregation, and the optimized short-term leakage prediction model and long-term stability evaluation model are output.

[0197] Closed-loop feedback and model updating: Input data: repair efficiency report, real-time monitoring data (pressure, concentration, deformation).

[0198] Operation flow: First, feedback data to decision model: ① Short-term model update: add repaired pressure and microseismic data to LSTM training set, incrementally train model. ② Long-term model update: dynamic adjustment of reinforcement learning reward function weight (e.g. leakage penalty weight w4 increases by 20%). Next, federal learning framework, each node locally trains and uploads encrypted parameters to central server for aggregation, update cycle: every quarter or after repair event. Finally, output optimized decision model: leakage prediction accuracy improves (AUC ≥ 0.9), repair agent dosage reduces by 15% (under the same leakage size).

[0199] Table 8 is a related explanation of the training nodes.

[0200] Table 8

[0201] Purpose of parameter upload and aggregation: 1. Solve data island and privacy problems Original data does not leave the local: sensitive data (such as reservoir geological structure, leakage coordinates) is only stored in edge nodes and not uploaded to the cloud.

[0202] Only encrypted parameters are transmitted: each node uploads the local training model gradient increment (such as LSTM weight difference ΔW) after homomorphic encryption.

[0203] 2. Realize global model collaborative optimization Aggregation process: ① Parameter weighted average: cloud assigns weights to nodes based on data quality (e.g. higher AUC value, higher weight): , where represents the global model parameter change generated by cloud aggregation of all local updates, represents the AUC value of the i th node, represents the model parameter change generated by the th local node (such as edge device or client) during training.

[0204] ② Dynamic fusion: update global model .

[0205] 3. Improve system adaptive ability Cross-scenario generalization: integrate node data from different geological conditions (such as sandstone / salt cavern reservoirs) to enhance model robustness.

[0206] Continuous evolution: within the quarterly update cycle, leakage prediction AUC increases from 0.88 to 0.91.

[0207] In addition, system integration and visualization can also be carried out: system integration integrates multi-modal sensing data processing, AI decision model and targeted repair execution module into a unified platform through standardized interface protocol (such as RESTAPI), realizes the cooperative operation of edge real-time data preprocessing and cloud complex calculation, and visualizes the CO2 saturation field, leakage risk thermal map and repair equipment state based on three-dimensional geographic information system (GIS), combines Web cockpit to display key parameters such as pressure and temperature in real time, and provides intuitive monitoring and emergency intervention interface for operators through hierarchical alarm (two levels of yellow / red) and historical data playback function, to ensure efficient operation of the whole system under the security framework of encrypted transmission and blockchain storage.

[0208] The carbon dioxide storage intelligent management method provided by the application obtains a repair efficiency report and real-time monitoring data, feeds back the repair efficiency report and the real-time monitoring data to the short-term leakage prediction model and the long-term stability evaluation model, so that the short-term leakage prediction model is subjected to enhanced training, and the long-term stability evaluation model is subjected to reward function weight adjustment, based on a federal learning framework, after training at each local node, the parameters are encrypted and uploaded to a central server for aggregation, and the short-term leakage prediction model and the long-term stability evaluation model after optimization are output, based on repair effect monitoring data, AI model parameters are dynamically updated, and multi-node collaborative optimization is realized through federal learning, to ensure continuous improvement of system prediction accuracy, and form a "monitoring-repair-learning" self-iteration capability.

[0209] Figure 2 is a second flowchart of the carbon dioxide storage intelligent management method provided by the application. As shown in Figure 2 , the method comprises: Multi-modal sensing data processing and fusion, including: spatio-temporal alignment and preprocessing, feature extraction and noise reduction, multi-scale hierarchical data fusion; AI-driven autonomous decision model construction, including: input data preparation and feature engineering, short-term leakage prediction model construction, long-term stability evaluation model construction, model verification and dynamic update; Targeted repair execution and closed-loop feedback, including: repair strategy generation and instruction issuance, targeted repair execution, real-time monitoring of repair effect; wherein, if the repair fails, the model is updated; System integration and visualization, including: real-time monitoring and alarm, three-dimensional GIS dynamic display, blockchain storage and security framework.

[0210] The processing flow of the carbon dioxide storage intelligent management method provided by the application will be described below through a specific example.

[0211] Specific implementation takes a certain block of an oilfield in eastern China as an example, and the block widely carries out carbon dioxide flooding and storage. In the region, the method is used to carry out related work: Step 1: Spatiotemporal alignment and preprocessing of multi-source data Input data and parameters: ① Four-dimensional seismic data: horizontal resolution 100m x 100m, vertical 10m, time series (2018-2023, annual update). ② DAS data: 3 optical fiber paths laid along the gas injection well, sampling rate 1.2kHz, covering the key path of the reservoir (total length 15km). ③ InSAR data: Sentinel-1 satellite data (revisit period 12 days), deformation field resolution 100m x 100m (2020-2023). ④ UAV CO2 concentration data: shortwave infrared spectrometer acquisition, resolution 10m x 10m (2022 September leak event area data). ⑤ Temperature gradient data: ground sensor network + UAV thermal imaging, resolution 1m x 1m (updated every hour).

[0212] Operation: First, time alignment, cubic spline interpolation is used for InSAR data to generate daily deformation sequence (error ≤1ms), and time stamp is aligned with seismic and DAS. Then, spatial unification, DAS acoustic signals are mapped to seismic grid (100m x 100m) by Kriging interpolation, with root mean square error (RMSE) ≤5%.

[0213] UAV CO2 concentration data is smoothed to seismic grid using Gaussian kernel function (σ=20m), and the coverage area of the leakage probability heat map is 20km².

[0214] Step 2: Multi-modal data feature extraction and noise reduction From seismic data, underground pressure changes are inverted, CO2 accumulation areas are identified, microseismic events are extracted from acoustic signals, CO2 migration paths are marked, natural disturbances in satellite deformation data are separated, and CO2 concentration and temperature anomaly areas are calculated. Finally, 52 microseismic events (frequency band 80-300Hz) are extracted, 3 CO2 migration paths are marked, and the peak CO2 concentration in the leakage area is 420ppm.

[0215] Step 3: Data integration and precision improvement Combine low-resolution seismic data with high-precision acoustic and surface data to generate a more detailed underground model. Correct the reservoir boundary and locate high-risk leakage areas (precision improved to 5 meters). The high-risk area is 1.2 square kilometers, and the microfracture positioning error is ≤18 meters.

[0216] Step 4: AI model input preparation Extracting pressure change trends, CO2 diffusion rates, and other characteristics to generate data tables for AI models. Daily pressure change rate: 0.02 MPa / hour.

[0217] Step 5: Short-term Leakage Prediction Using AI models (LSTM) to predict leakage risks within the next 3 days, triggering alerts in high-risk areas. Alert area coordinates: East 123.5°, North 32.8°.

[0218] Step 6: Long-term Stability Assessment Simulating reservoir changes over the next 10 years to develop optimal gas injection rates and repair agent usage. Repair agent annual usage ≤450 cubic meters, 10-year cumulative leakage ≤800 tons.

[0219] Step 7: Model Verification and Update Regularly verify model accuracy with new data and automatically optimize parameters. Leakage prediction accuracy: 91% (originally 88%).

[0220] Step 8: Develop Repair Solutions Based on AI prediction results, generate specific instructions for downhole plugging and surface laser repair. Downhole plugging material usage: 120 cubic meters. Laser power: 1.2 kW, repair area 500 square meters.

[0221] Step 9: Execute Repair Operations Control nanorobots to accurately inject plugging materials, and surface lasers to irradiate CO2 leakage areas to accelerate mineralization. Downhole positioning accuracy: 0.08 meters, repair time 4.5 hours.

[0222] Step 10: Monitor Repair Effectiveness Monitor pressure recovery and CO2 concentration reduction in real-time to determine if the repair is successful. Pressure recovery rate: 92%, CO2 concentration reduction 85%.

[0223] Step 11: Feedback and Optimization Feedback repair results to the AI model to improve the accuracy of subsequent predictions and decisions. Repair agent usage reduced by 18%, and alert response time shortened to 15 minutes.

[0224] Step 12: System Integration and Display Integrate all functions into a visualization platform, displaying real-time reservoir status, risk heat maps, and repair progress. Achieve dynamic 3D map updates, record operation logs on the blockchain, and ensure data security.

[0225] The carbon dioxide storage intelligent management and control method provided by the application aims at the problems of space-time mismatch and low resolution of multi-source data in CO2 storage monitoring, adopts multi-scale hierarchical fusion technology (seismic data benchmark layer + DAS high-frequency embedding + surface data correction), combines Kriging interpolation and wavelet transform, realizes unified modeling of underground-surface data (the resolution is improved from 100 m to 5 m level), and the error of leakage path identification is ≤20 m; aiming at the problems of low efficiency of storage risk dynamic prediction and long-term management decision, the LSTM-attention mechanism model is used to predict short-term leakage (72 hours AUC≥0.9), combined with the reinforcement learning PPO framework to generate 10-year optimization strategy (repairing agent consumption is reduced by 15%), realizes the whole process automation of “prediction-decision”, and the false alarm rate is ≤10%; aiming at the problem that the effect after repair depends on manual intervention and the system has insufficient self-adaptive ability, the federal learning and real-time feedback closed-loop mechanism (repair efficiency data driven model updating) is adopted, the leakage penalty weight and repair parameters are dynamically adjusted, the self-evolution ability of the system (AUC quarterly improvement 4%, response time ≤15 minutes) is realized, and the whole cycle management and control is continuously optimized.

[0226] The carbon dioxide storage intelligent management and control system provided by the application is described below, and the carbon dioxide storage intelligent management and control system described below can be correspondingly referred to the carbon dioxide storage intelligent management and control method described above.

[0227] Figure 3 is a structural schematic diagram of the carbon dioxide storage intelligent management and control system provided by the application. As Figure 3As shown, the system comprises a data fusion module 10, a feature processing module 20, a short-term risk prediction module 30, a long-term risk prediction module 40, and a repair processing module 50, wherein: the data fusion module 10 is configured to: acquire multi-source data, perform spatio-temporal alignment processing, feature extraction processing, and multi-scale hierarchical data fusion processing on the multi-source data to obtain dynamic parameter field data; wherein the multi-source data comprises four-dimensional seismic data, distributed fiber acoustic wave signals, surface millimeter-level deformation data, surface CO2 concentration anomaly data, and temperature gradient change data, and the dynamic parameter field data comprises CO2 saturation field, reservoir pressure field, micro-fracture density, and surface leakage probability; the feature processing module 20 is configured to: perform time series feature extraction and spatial feature encoding based on the dynamic parameter field data to generate a spatio-temporal feature matrix; wherein the spatio-temporal feature matrix comprises pressure statistics, CO2 saturation gradient field, and deformation-pressure coupling coefficient; the short-term risk prediction module 30 is configured to: input the spatio-temporal feature matrix into a pre-trained short-term leakage prediction model to output high-risk region coordinates with a leakage probability greater than a preset value within a first time length in the future; wherein the short-term leakage prediction model is trained by inputting the spatio-temporal feature matrix of historical data into an LSTM network and taking historical leakage labels as output labels; the long-term risk prediction module 40 is configured to: acquire reservoir pressure, caprock stress, fault activation index, and geomechanics simulation data and input them into a pre-trained long-term stability evaluation model to output gas injection adjustment amplitude and repair agent recommended dosage within a second time length in the future; wherein the long-term stability evaluation model is trained by reinforcement learning according to a defined state space and action space using a proximal policy optimization framework; and the repair processing module 50 is configured to: based on the gas injection adjustment amplitude, the repair agent recommended dosage, reservoir pressure gradient, and surface deformation rate, generate an executable repair strategy through a multi-objective optimization algorithm, and issue repair instructions according to the repair strategy to perform targeted repair.

[0228] The carbon dioxide storage intelligent management and control system provided by the application obtains multi-source data, performs space-time alignment processing, feature extraction processing and multi-scale hierarchical data fusion processing on the multi-source data, obtains dynamic parameter field data, performs time sequence feature extraction and spatial feature coding based on the dynamic parameter field data, generates a space-time feature matrix, inputs the space-time feature matrix into a short-term leakage prediction model obtained by pre-training, outputs high-risk region coordinates in which a leakage probability in a first time length in the future is greater than a preset value, obtains reservoir pressure, caprock stress, fault activation index and geomechanical simulation data and inputs the geomechanical simulation data into a long-term stability evaluation model obtained by pre-training, outputs an injection adjustment amplitude and a recommended amount of repair agent in a second time length in the future, generates an executable repair strategy based on the injection adjustment amplitude, the recommended amount of repair agent, a reservoir pressure gradient and a surface deformation rate through a multi-objective optimization algorithm, issues a repair instruction set according to the repair strategy to perform targeted repair, and through hierarchical fusion technology, accurate three-dimensional modeling of the underground reservoir and surface deformation is realized, the leakage positioning and risk region identification accuracy are improved, and a full-process automatic closed loop of "prediction, decision and repair" is formed.

[0229] Figure 4 An example of a schematic diagram of the physical structure of an electronic device is shown in FIG. 1. Figure 4As shown, the electronic device can include a processor 410, a communications interface 420, a memory 430, and a communications bus 440, wherein the processor 410, the communications interface 420, and the memory 430 complete communication with each other through the communications bus 440. The processor 410 can invoke a logical instruction in the memory 430 to execute a carbon dioxide storage intelligent management and control method, which includes: acquiring multi-source data, performing spatio-temporal alignment processing, feature extraction processing, and multi-scale hierarchical data fusion processing on the multi-source data to obtain dynamic parameter field data; wherein the multi-source data includes four-dimensional seismic data, distributed optical fiber acoustic wave signals, surface millimeter-level deformation data, surface CO2 concentration anomaly data, and temperature gradient change data, and the dynamic parameter field data includes CO2 saturation field, reservoir pressure field, micro-fracture density, and surface leakage probability; performing time series feature extraction and spatial feature encoding based on the dynamic parameter field data to generate a spatio-temporal feature matrix; wherein the spatio-temporal feature matrix includes pressure statistics, CO2 saturation gradient field, and deformation-pressure coupling coefficient; inputting the spatio-temporal feature matrix into a pre-trained short-term leakage prediction model to output high-risk area coordinates with a leakage probability greater than a preset value within a first time length in the future; wherein the short-term leakage prediction model is trained by inputting the spatio-temporal feature matrix of historical data into an LSTM network and taking historical leakage labels as output labels; acquiring reservoir pressure, caprock stress, fault activation index, and geomechanical simulation data and inputting them into a pre-trained long-term stability evaluation model to output gas injection adjustment amplitude and repair agent recommended dosage within a second time length in the future; wherein the long-term stability evaluation model is trained by reinforcement learning according to a defined state space and action space using a proximal policy optimization framework; based on the gas injection adjustment amplitude, the repair agent recommended dosage, the reservoir pressure gradient, and the surface deformation rate, an executable repair strategy is generated through a multi-objective optimization algorithm, and repair instruction sets are issued according to the repair strategy to execute targeted repair.

[0230] In addition, the logic instructions in the memory 430 described above can be implemented in the form of a software function unit and sold or used as an independent product, which can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or the parts that contribute to the prior art or parts of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium, including a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the method described in various embodiments of the present application. The foregoing storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disk, and various media that can store program codes.

[0231] In another aspect, the present application also provides a computer program product, the computer program product comprising a computer program, the computer program being stored on a non-transitory computer readable storage medium, and the computer program being executable by a processor to cause a computer to perform the carbon dioxide storage intelligent management and control method provided by the above method, the method comprising: acquiring multi-source data, performing spatio-temporal alignment processing, feature extraction processing and multi-scale hierarchical data fusion processing on the multi-source data to obtain dynamic parameter field data; wherein the multi-source data comprises four-dimensional seismic data, distributed optical fiber acoustic signal, ground millimeter-level deformation data, ground CO2 concentration anomaly data and temperature gradient change data, and the dynamic parameter field data comprises CO2 saturation field, reservoir pressure field, micro-fracture density and ground leakage probability; performing time series feature extraction and spatial feature encoding based on the dynamic parameter field data to generate a spatio-temporal feature matrix; wherein the spatio-temporal feature matrix comprises pressure statistics, CO2 saturation gradient field and deformation-pressure coupling coefficient; inputting the spatio-temporal feature matrix into a pre-trained short-term leakage prediction model to output high-risk region coordinates with a leakage probability greater than a preset value within a first time length in the future; wherein the short-term leakage prediction model is trained by inputting the spatio-temporal feature matrix of historical data into an LSTM network and taking historical leakage labels as output labels; acquiring reservoir pressure, caprock stress, fault activation index and geomechanical simulation data and inputting them into a pre-trained long-term stability evaluation model to output gas injection adjustment amplitude and repair agent recommended dosage within a second time length in the future; wherein the long-term stability evaluation model is trained by reinforcement learning according to the defined state space and action space using a proximal policy optimization framework; generating an executable repair strategy based on the gas injection adjustment amplitude, the repair agent recommended dosage, the reservoir pressure gradient and the ground deformation rate through a multi-objective optimization algorithm, and issuing repair instruction sets according to the repair strategy to perform targeted repair.

[0232] In another aspect, the present application also provides a non-transitory computer readable storage medium having stored thereon a computer program, which, when executed by a processor, implements the carbon dioxide storage intelligent management and control method provided by the above method, which comprises: acquiring multi-source data, performing spatio-temporal alignment processing, feature extraction processing and multi-scale hierarchical data fusion processing on the multi-source data to obtain dynamic parameter field data; wherein the multi-source data comprises four-dimensional seismic data, distributed optical fiber acoustic wave signals, ground millimeter-level deformation data, ground CO2 concentration anomaly data and temperature gradient change data, and the dynamic parameter field data comprises CO2 saturation field, reservoir pressure field, micro-fracture density and ground leakage probability; performing time series feature extraction and spatial feature encoding based on the dynamic parameter field data to generate a spatio-temporal feature matrix; wherein the spatio-temporal feature matrix comprises pressure statistics, CO2 saturation gradient field and deformation-pressure coupling coefficient; inputting the spatio-temporal feature matrix into a pre-trained short-term leakage prediction model to output high-risk region coordinates with a leakage probability greater than a preset value within a first time length in the future; wherein the short-term leakage prediction model is trained by inputting the spatio-temporal feature matrix of historical data into an LSTM network and taking historical leakage labels as output labels; acquiring reservoir pressure, caprock stress, fault activation index and geomechanical simulation data and inputting them into a pre-trained long-term stability evaluation model to output gas injection adjustment amplitude and repair agent recommended dosage within a second time length in the future; wherein the long-term stability evaluation model is trained by reinforcement learning according to the defined state space and action space using a proximal policy optimization framework; based on the gas injection adjustment amplitude, the repair agent recommended dosage, the reservoir pressure gradient and the ground deformation rate, an executable repair strategy is generated through a multi-objective optimization algorithm, and repair instruction set is issued according to the repair strategy to execute targeted repair.

[0233] The device embodiments described above are only schematic, wherein the units shown as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the present embodiment scheme. Those skilled in the art can understand and implement without creative labor.

[0234] Those skilled in the art can clearly understand the technical solutions of the various embodiments from the above description of the embodiments, and the various embodiments can be implemented by means of software with the necessary general hardware platforms, and of course, can also be implemented by hardware. Based on such understanding, the above technical solutions, essentially or in other words, the part of the prior art that makes a contribution, can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0235] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, rather than limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for some technical features therein; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for intelligent control of carbon dioxide storage, characterized in that, include: Multi-source data is acquired, and spatiotemporal alignment, feature extraction, and multi-scale hierarchical data fusion are performed on the multi-source data to obtain dynamic parameter field data. The multi-source data includes four-dimensional seismic data, distributed fiber optic acoustic signals, surface millimeter-level deformation data, surface CO2 concentration anomaly data, and temperature gradient change data. The dynamic parameter field data includes CO2 saturation field, reservoir pressure field, microfracture density, and surface leakage probability. Based on the dynamic parameter field data, temporal features are extracted and spatial features are encoded to generate a spatiotemporal feature matrix; wherein, the spatiotemporal feature matrix includes pressure statistics, CO2 saturation gradient field and deformation-pressure coupling coefficient; The spatiotemporal feature matrix is ​​input into a pre-trained short-term leakage prediction model, which outputs the coordinates of high-risk areas with a leakage probability greater than a preset value within the first time period in the future. The short-term leakage prediction model is trained by inputting the spatiotemporal feature matrix of historical data into an LSTM network and using historical leakage labels as output labels. The reservoir pressure, caprock stress, fault activation index, and geomechanical simulation data are acquired and input into a pre-trained long-term stability assessment model. The model outputs the gas injection adjustment range and recommended dosage of remediation agent within the next second time period. The long-term stability assessment model is obtained by training through reinforcement learning based on a defined state space and action space using a proximal strategy optimization framework. Based on the gas injection adjustment range, the recommended dosage of the remediation agent, the reservoir pressure gradient, and the surface deformation rate, an executable remediation strategy is generated through a multi-objective optimization algorithm. The remediation instruction set is then issued according to the remediation strategy to execute targeted remediation.

2. The intelligent control method for carbon dioxide storage according to claim 1, characterized in that, The multi-scale hierarchical data fusion processing includes: Pore ​​pressure field and CO2 saturation field are obtained and output by reservoir parameter inversion based on four-dimensional seismic data; Based on the pore pressure field, the CO2 saturation field, and the characteristics of the DAS acoustic signal, an enhanced reservoir model is obtained through wavelet domain fusion and microfracture correction. The enhanced reservoir model includes the CO2 saturation field, the pore pressure field, and the microfracture density distribution. Based on the enhanced reservoir model, InSAR deformation field, CO2 leakage probability thermogram and temperature gradient field, a corrected four-dimensional dynamic parameter field is obtained through deformation-pressure correlation modeling and leakage probability space fusion. The corrected four-dimensional dynamic parameter field integrates the reservoir internal pressure distribution, CO2 migration path and surface leakage risk thermogram.

3. The intelligent control method for carbon dioxide storage according to claim 1, characterized in that, The LSTM network consists of two LSTM layers with an attention mechanism module inserted between the layers. The output layer uses the Sigmoid activation function and has a randomly deactivated layer with a dropout rate of 0.

2.

4. The intelligent control method for carbon dioxide storage according to claim 1, characterized in that, The state space of the long-term stability assessment model includes reservoir pressure, caprock stress, fault activation index, and remediation agent dosage. The action space includes gas injection adjustment information and remediation agent injection information. During training, a reward function is set by balancing stability, economy, and safety.

5. The intelligent control method for carbon dioxide storage according to claim 1, characterized in that, The method further includes: The short-term leakage prediction model was validated based on the AUC metric and the false alarm rate, and the long-term stability assessment model was validated based on the cumulative leakage rate. The federated learning framework was used to periodically update the short-term leakage prediction model and the long-term stability assessment model.

6. The intelligent control method for carbon dioxide storage according to claim 1, characterized in that, The objective function of the multi-objective optimization algorithm is expressed as min(repair cost + λ * environmental risk), where λ represents the environmental risk penalty weight coefficient, and the constraint is that the pressure recovery rate is greater than the first preset value and the leakage probability is less than the second preset value. The repair instruction set includes the downhole nanorobot injection coordinates, the type and amount of plugging material, and the operating parameters of the surface laser mineralization device.

7. The intelligent control method for carbon dioxide storage according to claim 6, characterized in that, The method further includes: The selection of sealing materials using a permeability matching algorithm is expressed as follows: ,in: Indicates the material coefficient. Indicates the permeability of the leaked area; The operating parameters of the ground-based laser mineralization device include laser power, and the method further includes: determining the laser power by dynamically adapting it to CO2 concentration, expressed as: ,in, Indicates laser power. Indicates the energy conversion coefficient. Indicates carbon dioxide concentration. Indicates the area of ​​leakage.

8. The intelligent control method for carbon dioxide storage according to claim 1, characterized in that, During the targeted remediation process, the method further includes: for downhole remediation, injecting an adaptive gel based on permeability matching results and controlling the gel solidification time according to a time control equation; for surface remediation, ensuring that the mineralization rate meets the set requirements by adjusting the laser power in real time. After the targeted repair is completed, the method further includes: extracting the pressure recovery rate and the leakage concentration decay rate, and determining the repair effect based on the pressure recovery rate and the leakage concentration decay rate.

9. The intelligent control method for carbon dioxide storage according to claim 1, characterized in that, After the targeted repair is completed, the method further includes: Obtain repair performance reports and real-time monitoring data, and feed the repair performance reports and real-time monitoring data back to the short-term leakage prediction model and the long-term stability assessment model for enhanced training of the short-term leakage prediction model and for adjusting the reward function weights of the long-term stability assessment model. Based on the federated learning framework, after training on each local node, the parameters are encrypted and uploaded to the central server for aggregation, and the optimized short-term leakage prediction model and the long-term stability evaluation model are output.

10. A smart control system for carbon dioxide storage, characterized in that, include: The data fusion module is used to: acquire multi-source data, perform spatiotemporal alignment processing, feature extraction processing, and multi-scale hierarchical data fusion processing on the multi-source data to obtain dynamic parameter field data; wherein, the multi-source data includes four-dimensional seismic data, distributed fiber optic acoustic signals, surface millimeter-level deformation data, surface CO2 concentration anomaly data, and temperature gradient change data, and the dynamic parameter field data includes CO2 saturation field, reservoir pressure field, microfracture density, and surface leakage probability; The feature processing module is used to: extract temporal features and encode spatial features based on the dynamic parameter field data to generate a spatiotemporal feature matrix; wherein, the spatiotemporal feature matrix includes pressure statistics, CO2 saturation gradient field and deformation-pressure coupling coefficient; The short-term risk prediction module is used to: input the spatiotemporal feature matrix into a pre-trained short-term leakage prediction model, and output the coordinates of high-risk areas with a leakage probability greater than a preset value within the first time period in the future; wherein, the short-term leakage prediction model is trained by inputting the spatiotemporal feature matrix of historical data into an LSTM network and using historical leakage labels as output labels. The long-term risk prediction module is used to: acquire reservoir pressure, caprock stress, fault activation index, and geomechanical simulation data and input them into a pre-trained long-term stability assessment model; and output the gas injection adjustment range and recommended dosage of remediation agent in the second future time period. The long-term stability assessment model is obtained by training through reinforcement learning based on a defined state space and action space using a proximal strategy optimization framework. The remediation module is used to: generate an executable remediation strategy based on the gas injection adjustment range, the recommended dosage of the remediation agent, the reservoir pressure gradient, and the surface deformation rate, using a multi-objective optimization algorithm, and issue a set of remediation instructions according to the remediation strategy to perform targeted remediation.

Images