Method, system, computer device and readable storage medium for evaluating potential and mechanism of carbon dioxide storage in deep saline aquifer
By combining numerical simulation and machine learning methods, a carbon dioxide sequestration potential assessment system for deep saline aquifers was constructed, which solved the problems of inaccurate assessment and high computational cost in traditional methods, and achieved efficient and accurate sequestration potential assessment and mechanism contribution analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2026-03-18
- Publication Date
- 2026-06-19
Smart Images
Figure CN122242015A_ABST
Abstract
Description
Technical Field
[0001] This invention specifically relates to a method, system, computer equipment, and readable storage medium for evaluating the carbon dioxide sequestration potential and mechanism of deep saline aquifers. Background Technology
[0002] Carbon dioxide is the primary source of greenhouse gases, accounting for approximately 76% of global greenhouse gas emissions. To mitigate the resulting climate change risks, efforts can be made to reduce fossil fuel use and employ carbon capture and storage (CCS) technologies to capture and store emissions long-term. Currently, CO2 can be stored in various geological media, including depleted oil and gas reservoirs, oceans, and brackish water layers. Among these, deep brackish water layers are considered to have the most significant storage potential due to their wide distribution, large storage capacity, and low competition from oil and gas resource development.
[0003] When storing CO2 in deep saline aquifers, supercritical CO2 injection is usually preferred because scCO2 has a higher density than gas, thus improving the storage efficiency per unit volume. In the initial injection phase, CO2 exists primarily as a mobile, free phase, potentially diffusing laterally within the reservoir and migrating upwards to the caprock interface under buoyancy, forming tectonic traps. During formation water erosion, some CO2 is trapped in pore spaces, forming residual gas traps, while others dissolve in the aqueous phase, forming dissolution traps, or react with formation minerals, forming mineral traps.
[0004] While related studies have demonstrated the significant potential for CO2 sequestration in saline aquifers, methods for assessing this potential remain a hot research topic. For example, Li et al. used numerical simulations to evaluate the CO2 sequestration potential in saline aquifers and analyzed hydrodynamic models to assess the impact of factors such as pressure and temperature on the CO2 sequestration mechanism. Li Xiaochun et al., based on the simplified CSLF formula, added data such as the proportion of usable saline aquifer to the total basin area and the proportion of saline aquifer thickness to sedimentary thickness, calculating the CO2 sequestration potential of saline aquifers in 24 major sedimentary basins in China to be approximately 1.435 × 10⁻⁶. 8 Tons. Q.Wang et al. established a multiphase flow numerical model based on MRST, quantitatively analyzed the influence of reservoir properties on CO2 migration and storage capacity, and pointed out that porosity and permeability are key control parameters. However, traditional storage potential assessment methods often have the following problems: (1) Traditional mathematical methods, such as fluid dynamics models, usually assume that the system is homogeneous and ignore the heterogeneity and multi-mechanism coupling process in the formation, which may lead to deviations in the assessment results; (2) Although numerical simulation methods can solve these problems to a certain extent, due to the involvement of complex physicochemical processes, the uncertainty of parameters and high computational cost, large-scale application is limited.
[0005] To address these shortcomings, machine learning (ML) has been proposed in recent years as a novel approach to improve the efficiency and accuracy of CO2 sequestration potential assessment. (Safaei) Farouji et al. systematically compared the performance of RF, ET, and RBF models in predicting CO2 capture efficiency, demonstrating that machine learning methods can effectively characterize complex nonlinear relationships and improve prediction accuracy. H. VoThanh et al. used RF, XGBoost, and SVR models to predict CO2 storage efficiency, showing that XGBoost performed best in prediction accuracy. To reduce the computational cost of numerical simulations, EMKanakaki et al. proposed a machine learning-driven grid block classification framework, which can effectively accelerate the simulation process while maintaining prediction accuracy. Tillero et al. built a CO2 storage prediction model based on ANN to quickly assess the CO2 capture and storage capacity of reservoirs. W. Lu et al. combined Geographic Information System (GIS) with five machine learning models to evaluate the suitability of saline aquifer storage, with model prediction accuracies generally exceeding 95%.
[0006] Building upon this foundation, research focus has gradually expanded from static capacity to the dynamic evolution of storage mechanisms. For example, A. Alqahtani et al. proposed a data-driven workflow integrating machine learning, Bayesian optimization, and global sensitivity analysis to quantitatively characterize the contributions of residual gas storage and tectonic storage, and further evaluated the impact of uncertainties in input geological parameters on storage results. S. Davoodi et al. used multiple algorithms such as LSSVM, ELM, and CNN to establish STI and RTI prediction models based on 6811 simulation records from global saline aquifer storage projects. The results showed that LSSVM performed best in terms of prediction accuracy and stability. MI Khan constructed a surrogate model for deep saline aquifer CO2 storage, achieving rapid prediction of different storage mechanisms under multiple scenarios, and verified the key controlling role of geological parameters such as permeability and porosity in the evolution of storage mechanisms by combining feature importance analysis.
[0007] Based on a comprehensive analysis of the aforementioned existing technologies, the applicant believes that traditional mathematical evaluation models in existing research technologies rely on idealized assumptions and are difficult to accurately characterize reservoir heterogeneity and multi-mechanism coupling processes. While numerical simulation methods are more refined, they usually face high time costs in the solution and parameter adjustment process. In addition, there is a lack of evaluation methods for short-term and long-term storage mechanisms, and there is no clear understanding of the evaluation methods for the storage potential of the combined effects of multiple storage mechanisms at different time scales. Summary of the Invention
[0008] In view of this, the present invention provides a method, system, computer equipment and readable storage medium for assessing the carbon dioxide sequestration potential and mechanism of deep saline aquifers, in order to solve the problems of the shortcomings of existing assessment methods, and at the same time fill the gap in the lack of engineering assessment methods that combine the effects of time scale and sequestration mechanism.
[0009] The technical solution is as follows: A method for assessing the carbon dioxide sequestration potential and mechanism of deep saline aquifers, the key of which includes the following steps: S1, obtain the geological and fluid parameters of the target saline aquifer, as well as the gas injection regime parameters; S2. Based on the parameters obtained in step S1, construct a multi-field coupled numerical simulation model for carbon dioxide sequestration in deep saline water layers. This numerical simulation model sets the parameter space for the injection regime. S3, performs multiple sets of numerical simulations within the injection system parameter space, outputs the decomposition results of the sequestration quantity and sequestration mechanism at different time scales, and constructs a sequestration quantity database; S4. Based on the storage volume database mentioned in step S3, a machine learning proxy model is constructed with the gas injection regime parameters as input features and the final storage volume as the output target. The optimal proxy model is then selected from multiple candidate models based on the model prediction performance index. S5. Based on the optimal surrogate model, interpretability analysis is introduced, and the contribution of each input regime parameter is calculated using the SHAP analysis method to obtain the feature importance ranking and influence direction of each gas injection regime parameter on the prediction result of the storage volume. S6. Based on the optimal surrogate model and Monte Carlo simulation method, the injection system parameters are randomly sampled, and the contribution of each sealing mechanism at different time scales is predicted by probability distribution, so as to obtain the probability distribution of the contribution of the sealing mechanism at at least two time scales. S7. Based on the results of steps S5 and S6, output the assessment results of the carbon dioxide storage potential of the target block, the evolution law of the contribution of the storage mechanism under multiple time scales, and the control law of the gas injection regime parameters.
[0010] The above scheme employs a hybrid approach combining numerical simulation and machine learning. Numerical simulation is used to establish multi-field coupled numerical models and construct a database. Machine learning is used to build surrogate models, significantly reducing computational costs and enabling rapid prediction of CO2 sequestration potential. Furthermore, Monte Carlo random sampling is introduced to propagate the uncertainty of injection mechanism parameters, transforming single prediction results into probability distributions and confidence intervals. This quantitatively characterizes the fluctuation range and dominant patterns of each sequestration mechanism's contribution at different time scales, providing a more engineering-usable and accurate basis for short-term and long-term prediction of CO2 sequestration potential in deep saline aquifers and optimization of injection strategies.
[0011] Another aspect of this application is a system for assessing the carbon dioxide sequestration potential and mechanisms in deep saline aquifers, characterized by comprising: The data acquisition module is used to acquire the geological and fluid parameters of the target saline aquifer, as well as the gas injection regime parameters; The numerical simulation module is used to establish a multi-field coupled numerical simulation model for carbon dioxide sequestration in the target saline aquifer and generate a sequestration volume database. The agent modeling module is used to build machine learning agent models and select the optimal agent model; The interpretability analysis module is used to calculate the contribution of each input feature to the prediction result of the sealed quantity using the SHAP analysis method; The uncertainty analysis module is used to predict the contribution distribution of various sealing mechanisms at different time scales using Monte Carlo simulation methods. The results output module is used to output at least the results of the storage potential assessment, the evolution law of the storage mechanism, and the control law of the gas injection system parameters.
[0012] This application also provides a computer device, the key feature of which is that the electronic device includes: Memory, which stores executable instructions; A processor that executes the executable instructions in the memory to implement the above-described method for assessing the potential and mechanism of carbon dioxide sequestration in deep saline aquifers.
[0013] Another aspect of this application provides a computer-readable storage medium, the key feature of which is that the computer-readable storage medium stores a computer program, which, when executed by a processor, implements the above-mentioned method for evaluating the potential and mechanism of carbon dioxide sequestration in deep saline aquifers.
[0014] Compared with the prior art, the beneficial effects of the present invention are: (1) By establishing a proxy model framework that combines numerical simulation and machine learning, the efficiency of sealing potential assessment is significantly improved while ensuring prediction accuracy. The single prediction time can reach the millisecond level, which is several orders of magnitude faster than the original numerical simulation.
[0015] (2) By introducing the SHAP analysis method, the influence of parameters such as injection pressure, injection time, injection rate and injection location on the amount of sealed can be quantitatively identified, thereby improving the interpretability of the model results.
[0016] (3) By introducing the Monte Carlo simulation method, a single prediction result can be extended into a probability distribution result, which can quantitatively assess the contribution fluctuation range of dissolution storage, residual gas storage, tectonic storage and mineralization storage at different time scales.
[0017] (4) It can reveal the evolution characteristics of the storage mechanism and the main control role of the gas injection system parameters at different time scales, providing a reliable basis for gas injection strategy optimization and storage safety improvement. Attached Figure Description
[0018] Figure 1 This is a schematic diagram of the evaluation method steps of the present invention; Figure 2 for Figure 1 The flowchart of the workflow of the method shown; Figure 3 for Figure 2 Framework diagram of the gradient boosting model in the middle range; Figure 4 for Figure 2 Framework diagram of the random forest model; Figure 5 for Figure 2 Multilayer perceptron model framework diagram; Figure 6 for Figure 2 Support Vector Regression Model Framework Diagram; Figure 7 for Figure 2 The enlarged diagram of the Monte Carlo process shown below; Figure 8 This is a geological columnar section showing the location of the target block and the target layer in the embodiments of this application; Figure 9 This is a schematic diagram of the porosity distribution in the target saline aquifer model in this embodiment; Figure 10 This is a schematic diagram of the permeability distribution in the target saline aquifer model in this embodiment; Figure 11 This is a schematic diagram of the gas-water phase permeability curve of the target saline aquifer in this embodiment; Figure 12 A schematic diagram of an experimental core sample from the target saline aquifer; Figure 13 X-ray diffraction pattern of experimental core samples from the target saline aquifer; Figure 14 A comparative diagram of mineral composition of experimental core samples from the target saline aquifer; Figure 15 This is a schematic diagram showing the changes in the amount of data stored at different time scales for the four storage mechanisms in this embodiment; Figure 16 This is a schematic diagram of carbon dioxide transport changes in this embodiment; Figure 17 This is a schematic diagram of the fitting and scatter plot of the random forest model for predicting the amount of stockpiled goods. Figure 18 A schematic diagram of the fitting and scatter plot for predicting the amount of stockpiled using the extreme gradient boosting model; Figure 19 A schematic diagram of the fitting and scatter plot for the prediction of the amount of stockpiled using a multilayer perceptron model; Figure 20 This is a schematic diagram showing the fitting and scatter plot of the support vector regression model for predicting the amount of stockpiled goods. Figure 21 A diagram showing the comparison of prediction errors on the test set for four machine learning models; Figure 22 This is a diagram illustrating the feature importance ranking results of SHAP feature importance analysis. Figure 23 A summary diagram of SHAP value results from SHAP feature importance analysis; Figure 24 Histograms of prediction frequency distributions at different time scales under four sealing mechanisms based on the optimal machine model (XGBoost) and Monte Carlo method; Figure 25 A schematic diagram illustrating the changing contribution of the sealing mechanism over time; Figure 26 The impact of injection time on the contribution of the sealing mechanism at different time scales; Figure 27 The impact of injection rate on the sealing mechanism at different time scales; Figure 28 The impact of injection location on the contribution of the sealing mechanism at different time scales; Figure 29 The impact of injection pressure on the contribution of the sequestration mechanism at different time scales. Detailed Implementation
[0019] The present invention will now be described in further detail with reference to the accompanying drawings.
[0020] refer to Figures 1 to 29 The method for assessing the carbon dioxide sequestration potential and mechanisms of deep saline aquifers, as shown, mainly includes the following steps: S1, obtain the geological and fluid parameters of the target saline aquifer, as well as the gas injection regime parameters; S2. Based on the parameters obtained in step S1, a multi-field coupled numerical simulation model for carbon dioxide sequestration in deep saline aquifers is constructed. This numerical simulation model sets the parameter space for the injection regime.
[0021] S3 performs multiple sets of numerical simulations within the injection system parameter space, outputs the decomposition results of the sequestration quantity and sequestration mechanism at different time scales, and constructs a sequestration quantity database.
[0022] S4. Based on the storage volume database described in step S3, a machine learning surrogate model is constructed with the gas injection regime parameters as input features and the final storage volume as the output target. The optimal surrogate model is then selected from multiple candidate models based on the model's prediction performance index.
[0023] S5. Based on the optimal surrogate model, interpretability analysis is introduced, and the contribution of each input regime parameter is calculated using the SHAP analysis method to obtain the ranking of the feature importance and influence direction of each gas injection regime parameter on the prediction result of the storage volume.
[0024] S6. Based on the optimal surrogate model and Monte Carlo simulation, the injection system parameters are randomly sampled, and the contribution of each sealing mechanism at different time scales is predicted by probability distribution, so as to obtain the contribution probability distribution of the sealing mechanism at at least two time scales.
[0025] S7. Based on the results of steps S5 and S6, output the assessment results of the carbon dioxide storage potential of the target block, the evolution law of the contribution of the storage mechanism under multiple time scales, and the control law of the gas injection regime parameters.
[0026] In specific implementation, the numerical simulation model mentioned in step S2 is a three-dimensional geological model coupled with thermal-fluid-chemical processes; the storage mechanism includes four mechanisms: dissolution storage, residual gas storage, tectonic storage, and mineralization storage.
[0027] In addition, step S2 also includes using the DOE volumetric method to quickly screen and estimate the storage potential at the block scale, and using the interval estimation results as numerical simulation parameter verification or prior constraints. The volumetric method uses the storage efficiency coefficient to parameterize various trapping mechanisms and transport processes to obtain the magnitude and interval of the storage potential.
[0028] In step S2, the numerical simulation model at least considers the multiphase flow process and the evolution of the storage mechanism of carbon dioxide injected in a supercritical state in the deep saline aquifer. Based on this, the numerical simulation model given in this application is a three-dimensional geological model with thermo-fluid-chemical coupling, that is, the temperature field, the chemical reaction field in the seepage field are coupled. In specific implementation, it is mainly based on the data of the refined three-dimensional geological model, reservoir rock relative permeability experiment, fluid PVT experiment, etc., and the CMG-GEM component simulator is used for numerical simulation.
[0029] The sequestration mechanisms include four types: dissolution sequestration, residual gas sequestration, tectonic sequestration, and mineralization sequestration. The decomposition results in step S3 include the sequestration amounts of the four mechanisms: dissolution sequestration refers to the amount of CO2 dissolved in the aqueous phase; residual gas sequestration refers to the amount of CO2 trapped in the pore space under capillary action to form immobile gas; tectonic sequestration refers to the amount of free phase CO2 that migrates under buoyancy and is trapped under the caprock; and mineralization sequestration refers to the amount of CO2 that undergoes geochemical reactions with minerals to form solid carbonates or equivalent solid products.
[0030] The gas injection regime parameters in step S1 include at least injection pressure, injection rate, injection time, injection location, and injection method. The injection method includes at least one or more of continuous injection, intermittent injection, or segmented injection. In step S2, multiple gas injection conditions are constructed by changing the injection rate, injection time, injection pressure, and injection location. The storage capacity database is formed based on the simulation results of each gas injection condition. The various gas injection conditions are decomposed into time scales, and the contribution rate of the storage mechanism at different time scales is output to form a contribution rate database.
[0031] In step S4, the candidate models include at least two of the following: Random Forest (RF), Extreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), and Support Vector Regression (SVR). Specifically, a training and test set partitioning method is used, and cross-validation is employed to evaluate the model's hyperparameters and generalization performance. The model prediction performance metrics in this application include the coefficient of determination. R 2 Root mean square error RMSE Mean absolute error MAE and mean absolute percentage error MAPE One or more of them.
[0032] refer to Figure 4 Random Forest (RF) is a Bagging-based ensemble learning method that reduces variance and improves generalization ability by constructing multiple independent regression decision trees and performing ensemble averaging. It can handle the nonlinear relationship between input parameters and the amount of data sealed, making it suitable for regression problems with small to medium sample sizes. In this invention, the Random Forest model is used as one of the candidate surrogate models, taking injection rate, injection time, injection pressure, and injection location as inputs and the predicted amount of data sealed as output.
[0033] refer to Figure 3Extreme Gradient Boosting (XGBoost) is an extreme gradient boosting tree model with regularization terms and second-order optimization information. Its basic idea is to progressively add (trees) to maximize the descent of the objective function. It primarily employs an iterative, staged training strategy, integrating multiple weak learners and progressively fitting the residuals, resulting in high regression accuracy and good generalization ability. In this invention, the XGBoost model is used to learn the complex nonlinear mapping relationship between gas injection regime parameters and sequestration volume, and serves as a candidate surrogate model in performance comparisons.
[0034] refer to Figure 5 The Multilayer Perceptron (MLP) model consists of an input layer, hidden layers, and an output layer, and learns complex mapping relationships through a nonlinear activation function. In this invention, the MLP model can fit the nonlinear relationship between the injected parameters and the sealed amount, and the results are then compared with other models.
[0035] refer to Figure 6 Support Vector Regression (SVR) is based on the principle of minimizing structural risk. It constructs a regression hyperplane in the feature space and introduces... Insensitive loss enables robust fitting of continuous variables; that is, the input is mapped to a high-dimensional feature space through a kernel function to achieve nonlinear regression prediction. In this invention, the SVR model serves as a candidate surrogate model for performance comparison and optimization with other models.
[0036] In specific implementation, R 2 Using the maximum and minimum error index as the selection criterion for the optimal surrogate model, this application preferentially selects the extreme gradient boosting model as the optimal surrogate model.
[0037] In step S5, the importance of a feature is determined by calculating the average absolute SHAP value of each input feature, and the contribution of the corresponding input feature to the prediction result of the sealed quantity is determined according to the sign of the SHAP value. In other words, the marginal contribution of injection pressure, injection time, injection rate, injection location and injection method to the sealed quantity is calculated by using the SHAP algorithm, and the average absolute SHAP value is output to quantify the importance of the feature.
[0038] In step S6, when randomly sampling the gas injection regime parameters, multiple independent samplings are performed based on a preset parameter range to form a Monte Carlo input sample set. The contribution of each sealing mechanism is characterized by at least one of frequency distribution, quantile statistics, or confidence interval, wherein the quantile statistics include at least P50.
[0039] Step S7 also includes providing suggestions for optimizing the injection regime based on the contribution evolution law and the control law of the gas injection regime parameters. The optimization suggestions include at least adjusting the injection duration and / or adjusting the injection position to increase the total amount of sequestration or increase the contribution of the target mechanism at the target time scale.
[0040] Of course, the main control role of different gas injection regime parameters can also be identified based on the changes in the contribution of the sealing mechanism at different time scales, and gas injection optimization suggestions can be generated accordingly. In specific implementation, the objective function can be the maximum total sealing volume and / or the maximum residual gas sealing ratio at the target time scale. Under the constraint of the upper limit of injection pressure, candidate injection time and injection location combinations are screened, and the top N parameter combinations are output. That is, a recommended scheme is generated based on the objective function and constraints.
[0041] Another aspect of this application provides a system for evaluating the carbon dioxide sequestration potential and mechanism of deep saline aquifers, which mainly includes the following modules: The data acquisition module is used to acquire the geological and fluid parameters of the target saline aquifer, as well as the gas injection regime parameters; The numerical simulation module is used to establish a multi-field coupled numerical simulation model for carbon dioxide sequestration in the target saline aquifer and generate a sequestration volume database. The agent modeling module is used to build machine learning agent models and select the optimal agent model; The interpretability analysis module is used to calculate the contribution of each input feature to the prediction result of the sealed quantity using the SHAP analysis method; The uncertainty analysis module is used to predict the contribution distribution of various sealing mechanisms at different time scales using Monte Carlo simulation methods. The results output module is used to output at least the results of the storage potential assessment, the evolution law of the storage mechanism, and the control law of the gas injection system parameters.
[0042] The prediction process of this application mainly utilizes computer calculations. Therefore, this application also provides a computer device, which includes a memory and a processor. The memory stores executable instructions, and the processor can run the executable instructions in the memory to realize the above-mentioned deep saline aquifer carbon dioxide sequestration potential and mechanism assessment method. Specifically, it mainly implements the calculation process of the above model and outputs the results. For easy and intuitive observation, the output results in this embodiment are in the form of charts.
[0043] In another aspect, this application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for evaluating the carbon dioxide sequestration potential and mechanism of deep saline aquifers. Specifically, the computer program mainly implements the calculation process of the aforementioned calculation model and outputs the results.
[0044] refer to Figures 1 to 29 The assessment process for the CO2 geological sequestration potential of the saline aquifer in the PL block using this application is as follows: Block PL (target block) is located in XX city in the central-eastern part of the Sichuan Basin, belonging to the central part of the gentle slope tectonic belt of the central Sichuan ancient uplift (e.g., Figure 8 (As shown). The lithology is mainly feldspathic lithic sandstone, followed by lithic sandstone and lithic feldspathic sandstone. The reservoir belongs to the delta front underwater distributary channel-estuary bar microfacies deposit, with a stratigraphic thickness of 80-120m and a reservoir thickness of 11.8-47.2m. The reservoir lithology is mainly medium- to fine-grained feldspathic lithic sandstone, and the reservoir type is fracture- to porous. The reservoir properties are characterized by low porosity and low permeability (porosity 8.1%-9.3%, permeability 0.11-1.76mD), and the reservoir is highly heterogeneous. The formation temperature is 77.08℃, and the formation pressure coefficient is 1.5-1.6. The burial depth is 2373m, and the original formation pressure is 35.86MPa.
[0045] In this embodiment, the target saline aquifer depth is between 2402 and 2703 m, the pressure is between 34.83 and 36.41 MPa, the pressure coefficient is between 1.5 and 1.6, and the average geothermal gradient is 0.91℃ / 100m. Overall, it is suitable for safe and long-term geological sequestration of CO2 in deep saline aquifers.
[0046] First, the CO2 sequestration capacity in deep saline aquifers is assessed using the DOE (Depth-of-Earth) volumetric method. G CO2 Make an estimate: (1) In the formula: A The effective distribution area of the saline aquifer is in km². 2 ; h The effective thickness of the saline aquifer is in meters (m). The average effective porosity of the saline aquifer is % G CO2 The effective CO2 sequestration capacity in deep saline aquifers, 10 4 t; ρ CO2 CO2 density under formation conditions, kg / m³ 3 The value is taken as 520 kg / m at depths of 800–3500 m. 3 ; E saline The storage efficiency (efficiency coefficient) is %, which is a dimensionless parameter. For sandstone, the typical value range is approximately 0.51–5.4.
[0047] In this embodiment, the above parameters are shown in Table 1: Table 1 ; The DOE volumetric method parameterizes the combined effects of multiple trapping mechanisms and transport processes using efficiency coefficients. It typically does not explicitly distinguish the individual contributions of mechanisms such as tectonics, residuals, dissolution, and mineralization. Therefore, it is suitable for rapid screening and capacity range estimation at the block scale. Calculations show that the saline aquifer storage potential of the PL block is between 259,000 and 2,749,000 tons.
[0048] Secondly, using the CMG-GEM component simulator, a numerical simulation model of the target layer was constructed based on data from a refined three-dimensional geological model, reservoir rock relative permeability experiments, and fluid PVT experiments. In this embodiment, the numerical simulation model consists of 87,840 grid cells, with 96, 61, and 15 grid cells in the I, J, and K directions, respectively. The grid parameters are shown in Table 2. Table 2 ; The average permeability and porosity of the saline aquifer are 1.5 mD and 8.5%, respectively. The current simulation is based on existing well patterns and includes 7 wells. The porosity and permeability distributions of the saline aquifer are shown below. Figure 10 and Figure 11 As shown in Table 3, the initial parameters of the model are as follows.
[0049] Table 3 ; In the model, the accumulation and migration process of free-phase CO2 is simulated to characterize the trapping mechanism, and its multiphase flow is characterized by Darcy's law: (2) In the formula: u α Darcy speed, m / s; p α Pressure, MPa; α Viscosity, mPa s; K Let m be the permeability tensor. 2 ; g The acceleration due to gravity is m / s². 2 .
[0050] Furthermore, the Land model is introduced to characterize the relative permeability lag effect, in order to more reasonably represent the residual gas capture process. The relationship between the residual gas saturation and the historical maximum gas saturation is as follows: (3) In the formula: S gr Residual gas saturation; S gi Initial gas saturation; CThis is the Land closure coefficient.
[0051] Meanwhile, a gas-liquid two-phase thermodynamic equilibrium relationship is established based on Henry's law to describe the dissolution behavior of CO2 in salt water: (4) In the formula: c Equilibrium concentration of CO2 in the aqueous phase, mol / m 3 ; H cp Henry's constant, mol / (m 3 Pa); p This is the partial pressure of the component in the gas phase, in Pa.
[0052] This was used to calculate the dissolution and distribution of CO2 in brine. Mineralization sequestration was achieved through geochemical simulation, based on mineral and clay analysis results from core samples (reference). Figures 12 to 14 The reacting minerals were identified, and their dissolution / precipitation kinetic parameters were defined using the embedded Wolery database. The reaction rate was calculated according to the Bethke equation, and the reaction area was dynamically updated with the number of moles of minerals. (5) In the formula: r m The reaction rate of the mineral is expressed in mol / s. A m The specific surface area of a mineral in m 2 ; k m (T) For temperature T Mineral reaction rate constant, mol / (m 2 s); a i The activity of the reactant component; n i This corresponds to the reaction order; Q m It is the ion activity product; K m This is the equilibrium constant.
[0053] The injected scCO2 fluid dissolves in groundwater in the formation and reacts with various minerals in the reservoir. These mineralization reactions are very slow, and the precipitation of some minerals can even affect the injection capacity. The amount of CO2 sequestered is almost negligible in the short term. As the time scale increases, the amount of mineral sequestered also increases. The main minerals involved in the reaction are quartz, potassium feldspar, plagioclase, and calcite. The main reactions set in the model are shown in Table 4. The following reactions can be achieved by setting the main components of the mineralization reaction.
[0054] Table 4 ; This model simulates a timescale of 200 years. Numerical simulation results based on the existing well network and gas injection regime show that after selecting the keywords GHGAQU, GHGGAS, GHGLIA, GHGMNR, GHGSCRIT, GHGSOL, and GHGTHY in the model, the numerical simulation will output the amounts of dissolved, tectonic, residual gas, and mineralized sequestration after the calculation is completed. The sum of the values at the end of the simulation time is the total sequestration amount after 200 years. The calculated sequestration potential of the study area is approximately 1.35 million tons. By comparing the results with the DOE volumetric method, the model's calculation results are within the range of the DOE volumetric method's calculation results, which can preliminarily determine the accuracy of the model's calculation results and the necessity of further evaluation mechanisms.
[0055] The amount of stockpiled under the four mechanisms continues to change (e.g.) Figure 15 As shown), structural sealing is dominated in the early stages (1-50 years): during the initial injection, free-phase CO2 plumes are formed and gradually migrate towards higher structural regions (e.g., Figure 16 (As shown). As the plume continues to migrate, free-phase CO2 is gradually trapped by capillary action and dissolved upon contact with brine, causing CO2, which originally existed in the form of structural aggregates, to gradually transform into other more stable sequestration forms. Residual gas sequestration continues to increase over time and becomes the main contributor in the middle and late stages, reflecting the process by which the hysteresis effect causes CO2 to be divided into unconnected gas clusters by capillary forces and gradually form residual gas. Dissolution sequestration is relatively stable overall, entering a stable stage after a rapid increase in the early stage. Since the study area is dominated by sandstone minerals such as quartz and feldspar, the mineral reaction rate is slow, and the content of carbonate minerals is low, resulting in the smallest contribution of mineralization sequestration, with no significant changes over a century-long timescale. This is consistent with the understanding that mineralization usually accumulates significantly over a longer timescale.
[0056] Although reservoir permeability, porosity, formation pressure, and temperature all affect CO2 sequestration and migration in saline aquifers, the directly controllable parameters in engineering are often limited to injection regime parameters. Therefore, this application focuses on the effect of injection regime parameters on sequestration volume. As mentioned earlier, a surrogate modeling module is constructed based on numerical simulation research. This module is mainly used to construct machine learning surrogate models and select the optimal surrogate model. The machine learning surrogate models include Random Forest (RF), Extreme Gradient Boosting (XGBoost), Multilayer Perceptron (MLP), and Support Vector Regression (SVR). The core reason for this is its fast computation speed and high efficiency. The model's single prediction time can reach the millisecond level, and the computation speed is improved by five orders of magnitude compared to the original numerical simulation method (i.e., a speed improvement of 5 × 10⁻⁶). 5(times). Given that single-degree simulations typically take tens of hours, surrogate models can significantly reduce time costs and enable efficient iteration.
[0057] In this embodiment, the main injection regime parameters are injection rate, injection time, bottom hole injection pressure, and injection mode. The injection mode reflects the perforation method at different formations and employs one-hot encoding, using a 3-bit state register to represent low-position injection, high-position injection, and all-position injection, respectively. The output variable of this dataset is the CO2 sequestration amount at different time scales.
[0058] For each set of original data, the values of each injected institutional parameter are selected from the adjustment range (see Table 5). A total of 100 datasets are obtained through Latin hypercube sampling, and these datasets are combined for numerical simulation. Latin hypercube sampling (LHS) is an efficient sampling method for multidimensional parameter spaces. By stratifying and rationally combining samples for each parameter, it makes the samples more evenly distributed across all dimensions. Compared to simple random sampling, LHS can more fully cover the parameter space, reduce clustering and redundancy, and improve the training efficiency and generalization ability of alternative models with the same sample size.
[0059] Table 5 ; The dataset (sample set) was constructed by varying the injection rate, injection time, injection pressure, and injection location in a numerical simulation model to create multiple gas injection conditions. Based on the simulation results of each gas injection condition, a database of the sealed volume was formed, as shown in Table 6 below: Table 6 ; This application utilizes four machine learning algorithms to extract the complex relationship between model input features and stored data. The dataset is divided into training and test sets. The training set, comprising 80% of the total data points, is used for model development and training; while the test set, containing 20% of the data, is used to evaluate the model's prediction accuracy and generalization reliability. To avoid the randomness of a single partition and improve model generalization performance, k-fold cross-validation is employed during model training to robustly evaluate the key hyperparameters and training process of different models, reducing the randomness of a single data partition and improving the model's generalization performance on unknown samples.
[0060] refer to Figures 17 to 20 To verify the generalization ability of the surrogate model, the prediction results of the four models on the training and test sets were analyzed. The prediction results fit the regression curve and scatter distribution (blue dots in the scatter plot represent training set data values, and red dots represent test set data values).
[0061] Overall, the XGBoost model demonstrated the best generalization ability, fitting well to both the training and test sets, with the test set data points being more concentrated in a specific distribution. y=x Near the test point, the best-fit line deviates less from the ideal line, indicating that the model has a more balanced ability to characterize different levels of storage. The MLP and RF models perform slightly worse, both of which can capture the main fluctuation trends, but a certain degree of prediction deviation can still be observed near local abrupt changes and extreme values. The dispersion of the test set points is slightly larger than that of XGBoost, indicating that its error control in some intervals is relatively insufficient. In contrast, the SVR model has a more obvious deviation in the test segment, indicating that its ability to characterize the complex relationship between the model input features and the storage amount is relatively limited. The scatter distribution is discrete and the deviation of the fitted line from y=x is more obvious, reflecting that it has a certain degree of overfitting or sensitivity to changes in data distribution.
[0062] To visually demonstrate the accuracy of the model's predictions, refer to Figure 21 The relative error distributions of the four models in prediction were compared. Overall, the XGBoost model had the narrowest box and the median closest to 0, with a more concentrated error distribution and fewer outliers, indicating that its overall prediction error was smaller and its stability was the best. The other models had more positive outliers and the median showed a certain degree of negative bias, indicating that they were more likely to overestimate or fluctuate more on some samples. The SVR model had the largest error dispersion and relatively weak generalization ability.
[0063] Based on Table 7, considering the differences in prediction performance metrics among different models, the XGBoost model outperforms other models in... MAE , RMSE and MAPE The model achieved good predictive results in terms of performance metrics, which further highlights its excellent generalization ability.
[0064] Table 7 ; Secondly, to further improve the determinism and interpretability of the method, this embodiment uses the preferred surrogate model (XGBoost model) and SHAP analysis method to further reveal the contribution of each institutional parameter to the model output. This is mainly based on the trained XGBoost model, using the SHAP algorithm to calculate the marginal contribution of each input feature to the prediction result, and using the average absolute SHAP value sorting and summary chart visualization to quantitatively characterize the importance of different parameters and their influence on the model output.
[0065] Specifically, the program reads data from the database (Table 6) file, uses all columns except the target variable (final sealed quantity) as input features, and uses the target column as the prediction target. The samples are divided into training and test sets in an 8:2 ratio. An XGBoost regression model is constructed based on the input features and the target variable, where the relevant parameters are: n_estimators=100, indicating that the number of weak learners (i.e., decision trees) is 100; max_depth=5, indicating that the maximum depth of a single tree is 5; and learning_rate=0.1, indicating the learning step size for each iteration.
[0066] Based on this, a SHAP interpreter suitable for tree models is constructed using `shap.TreeExplainer(model)`, and then the SHAP value matrix corresponding to all samples and all features is calculated, such as... Figure 23 As shown, for any sample, each feature has a corresponding SHAP value. Observation reveals that the original feature values transition from red to blue, indicating that the feature's influence on the prediction results gradually increases as the value increases. In SHAP analysis, features with high SHAP values contribute significantly to the model's predictions, and their positive or negative values indicate the direction of the feature's influence on the prediction results. A positive SHAP value means that an increase in the feature value leads to an increase in the prediction result, while a negative SHAP value does the opposite. For example, the feature value range for injection pressure is relatively wide, and the SHAP value shows a linear growth trend as the value increases, indicating that higher injection pressure significantly improves the model output. Although injection time is less significant than injection pressure, its SHAP value fluctuates less, showing that its influence is relatively stable and limited. The SHAP values for injection rate and injection location show a more dispersed distribution, further verifying their weak influence on the model's prediction results. Especially for injection location, its SHAP value is almost constant, indicating that its contribution to the model in this study is negligible.
[0067] Secondly, based on the constructed uncertainty analysis module, the Monte Carlo simulation method is used to predict the contribution distribution of each sealing mechanism at different time scales, as follows: First, based on a database of contribution rates from different mechanisms output in numerical simulations, the optimal model is trained to further improve its accuracy. The optimal surrogate model (XGBoost) is also a multi-output surrogate model. This model takes four injection regime parameters as input: injection rate, injection time, injection pressure, and injection location, and outputs the predicted contribution values of the four sequestration mechanisms (dissolution, residual gas, tectonics, and mineralization) at four time scales (10 years, 50 years, 100 years, and 200 years). Each variable is independently and uniformly sampled to generate 5000 sets of input samples (as shown in Table 9, where D represents dissolution sequestration, R represents residual gas sequestration, S represents tectonics sequestration, and M represents mineralization sequestration).
[0068] In practice, data is first read from the database (Table 8), selecting the first four columns as input variables and the fifth and subsequent columns as output variables. The data is then divided into training and test sets in an 8:2 ratio, with XGBoost selected as the prediction model during the Monte Carlo simulation. Regarding input parameter sampling, the code uses the minimum and maximum values of each input variable in the training set as sampling boundaries, assuming a uniform distribution within this range. The sampling size is set to n_mc = 5000, representing 5000 sets of random input samples. The sampling method involves independent and uniform sampling within the training sample range for each input variable. The generated random input matrix X_mc has dimensions (5000, 4), where 4 represents the number of input variables. These samples are then input into the trained XGBoost surrogate model to obtain the prediction result Y_mc = best_model.predict(X_mc). The output matrix has dimensions (5000, n_mc). targets ), corresponding to the predicted values of each output variable under 5000 sets of input conditions (Table 9, where X1-X4 are injection rate, injection time, injection pressure and injection location respectively, Dis_S corresponds to 10D in Table 8, and the rest correspond one-to-one).
[0069] Table 8 ; Table 9 ; Finally, statistical analysis was performed on the prediction results of different sealing mechanisms at various time scales, i.e., the analysis was conducted on Table 9. The prediction contribution of each sealing mechanism at each time scale was displayed using frequency distribution histograms (e.g., Figure 24 As shown in the figure, the dynamic evolution characteristics of each mechanism during short-term and long-term storage are revealed.
[0070] Then, np.percentile(vals, [5, 50, 95]) is used to calculate P5, P50, and P95 respectively. Here, P50 represents the median, which reflects the typical level of the output; P5 and P95 represent the boundaries of the low-value and high-value regions respectively, which can be used to characterize the main fluctuation range and uncertainty interval of the prediction results. The quantile calculation results of the four sealing mechanisms are shown in Table 10.
[0071] In the initial stage (10 years), the contribution of dissolution-based sequestration (SSS) was relatively limited, with a median contribution (P50) of 16%, indicating that the contribution of dissolution to CO2 sequestration was small in the short term. Residual gas sequestration showed a wider distribution, with a P50 of 25%, while tectonic sequestration had a P50 of 58%, becoming the dominant sequestration mechanism. Mineralization sequestration contributed very little, with a P50 of 1%, which was almost negligible in the short term, indicating a weak impact of mineralization on CO2 sequestration. As time progressed, in the 50-year simulation, the contribution of dissolution-based SSS increased slightly, with a P50 of 18%, but remained at a low level. The contribution of residual gas sequestration continued to increase, reaching a P50 of 48%, while tectonic sequestration had a P50 of 31%, and its contribution began to decline rapidly. Mineralization sequestration remained weak, with its contribution remaining essentially unchanged. In the 100-year and 200-year simulations, the contribution of residual gas sequestration continued to increase. In 200 years, the P50 of residual gas sequestration was 60%, indicating that the residual gas sequestration mechanism gradually played a greater role over time. The contribution of tectonic sequestration continued to decline, eventually reaching a P50 of 14%.
[0072] Table 10 ; Combination Figure 25 This study demonstrates the evolution of the contributions of four sequestration mechanisms over a 200-year timescale. From 10 to 200 years, the contribution of residual gas sequestration shows a gradual increasing trend, with its fluctuation range expanding over time. Considering the reservoir's low porosity and low permeability, as the injection duration increases, some injected gas gradually migrates from the near-wellbore area to the far-field; when it enters smaller pores, capillary forces restrict its further migration, leading to an increase in the amount of trapped gas. Structural sequestration (SGS) has a relatively high contribution at 10 years, but its contribution decreases over time and exhibits significant fluctuations. This indicates that as the injection pressure increases, the amount of gas entering the reservoir increases and migrates towards the top of the reservoir under buoyancy, eventually being trapped by the caprock. In contrast, the contributions of dissolution sequestration and mineral sequestration remain relatively stable overall. Multiple long-term numerical simulations and field observations suggest that on millennia-long or longer timescales, dissolution sequestration may become one of the main sequestration mechanisms after residual gas sequestration, with its contribution significantly increasing. For mineral sequestration, due to its control by extremely slow geochemical reactions, significant growth typically takes a very long time to occur, even under long-term storage conditions. The results of this analysis provide important theoretical basis for optimizing CO2 sequestration engineering; in particular, the time-varying evolution of sequestration mechanisms at different time scales can provide guidance for formulating practical storage strategies in engineering applications.
[0073] refer to Figures 26 to 29In summary, it can be seen that the influence of different injection scheme parameters on the contribution ratio of CO2 sequestration mechanism exhibits a significant time-scale dependence. Overall, mineral sequestration accounts for only a very small proportion and hardly changes with the injection operation.
[0074] Among the various injection parameters, injection duration and injection location have a stronger regulatory effect, while injection rate and injection pressure have relatively limited influence. For injection rates (10,000–200,000 m³ / s),... 3 / d), the contribution proportions of each sequestration mechanism changed only slightly across different time scales. Taking 10 years as an example, the proportion of tectonic sequestration increased from about 54% to 58%, dissolution sequestration decreased from about 19% to 16%, and residual gas sequestration remained almost unchanged. Furthermore, residual gas sequestration continued to be the dominant mechanism during the 100–200 year period.
[0075] In contrast, the duration of injection has the most significant impact on the contribution ratio in the short term (1–10 years). On a 10-year timescale, the proportion of tectonic sequestration increases from approximately 37% to 64%, while residual gas sequestration and dissolution sequestration decrease from approximately 33% to 20% and from approximately 27% to 15%, respectively. This indicates that the duration of injection causes changes in the contributions of tectonic sequestration, residual gas sequestration, and dissolution sequestration by 12%–27%. The same trend persists as the timescale extends to 50–200 years, but the magnitude of change gradually decreases. For example, in 200 years, the proportion of tectonic sequestration increases only from approximately 10% to 19%.
[0076] The overall impact of injection pressure (40–52 MPa) is moderate, exhibiting a generally monotonic increase with a relatively small range of variation. On a 10-year timescale, pressure-induced structural storage fluctuations are approximately 9%; on a 50-year timescale, this fluctuation increases to approximately 10%. Over the 100–200 year period, pressure-driven structural storage variations range from approximately 7%–8%, indicating a limited ability to alter the long-term dominant storage mechanism.
[0077] It is noteworthy that the injection location has the most significant impact on the medium- to long-term contribution ratio, exhibiting a non-monotonic transition characteristic. When CO2 is injected at higher layers, it is more likely to accumulate in structurally high regions due to buoyancy, resulting in a higher proportion of structurally stored gas. However, when the injection location changes from higher layers to lower layers, the proportion of residual gas stored gas increases significantly, while the proportion of structurally stored gas decreases simultaneously: at 100 years, residual gas stored gas increases from approximately 50% to 61%, while structurally stored gas decreases from approximately 32% to 16%; at 200 years, residual gas stored gas further increases from approximately 56% to 67%, while structurally stored gas decreases from approximately 25% to 8%. When the injection location further changes from lower layers to the entire layer, the proportion of residual gas stored gas decreases while the proportion of structurally stored gas increases. This indicates that the entire layer injection promotes a more dispersed distribution of CO2 between layers, thereby weakening the advantage of lower layers in enhancing residual gas stored gas. The fundamental reason for this shift is that changes in the injection layer affect the transport channels and spatial distribution patterns of CO2, thereby causing differences in the contribution ratios of different storage mechanisms.
[0078] The results indicate that the injection scheme parameters not only determine the distribution of each sealing contribution in the short term, but also affect the evolution path of the sealing mechanism on medium- and long-term scales. It should be noted that the importance ranking of the injection parameters differs from the SHAP results. This suggests that the sealing mechanism influenced by different injection parameters exhibits more complex evolutionary behavior across multiple time scales, and the proposed process can more comprehensively characterize and explain this complexity.
[0079] Finally, the output module can output the above evaluation results, the evolution law of the sealing mechanism, and the control law of institutional parameters, while providing suggestions for optimizing the injection system, as follows: This application takes the PL block saline aquifer as an example, establishing a numerical simulation model for CO2 sequestration to simulate four sequestration mechanisms, and constructing a database based on the simulation results. Subsequently, surrogate models of the numerical simulation model were constructed using four machine learning models, and the impact of gas injection regime parameters on sequestration potential was evaluated using the SHAP algorithm. Furthermore, Monte Carlo simulation was used to quantitatively characterize the short-term and long-term CO2 sequestration patterns in the saline aquifer. The main conclusions are as follows: (1) Through numerical simulation studies, the CO2 sequestration potential of the saline aquifer in the PL block is initially assessed to be 1.35 million tons, covering four sequestration mechanisms: dissolution sequestration, residual gas sequestration, tectonic sequestration and mineralization sequestration.
[0080] (2) A surrogate model for the numerical simulation model was established using four machine learning models: RF, XGBoost, MLP, and SVR. The results showed that the XGBoost model outperformed the MLP, RF, and SVR models, and was the best in predicting storage potential, with a coefficient of determination R. 2The value reached 0.938. SHAP interpretability analysis showed that injection pressure, injection duration, injection rate, and injection location had a significant impact on CO2 storage potential.
[0081] (3) The evolution of the sequestration mechanism at different time scales was studied using Monte Carlo simulations. In the short term (1–10 years), tectonic sequestration was dominant; as time progressed, the contribution of residual gas sequestration gradually increased, while the contribution of tectonic sequestration gradually decreased. The contributions of dissolution sequestration and mineral sequestration were relatively stable, at approximately 16%–20% and 1%–2%, respectively.
[0082] (4) The contribution of injection scheme parameters to the storage mechanism has a significant time scale dependence. On a 10-year timescale, the injection duration is the dominant factor controlling the storage ratio, which can cause a change of 12%–27% in the contribution of tectonic storage, residual gas storage and dissolution storage. In other words, on a shorter timescale, the injection duration has a more significant impact on the contribution of tectonic storage, residual gas storage and dissolution storage. On a 100–200-year timescale, the injection location becomes the key control factor, causing a change of 10%–17% in the contribution of tectonic storage and residual gas storage. That is, on a longer timescale, the injection location has a more significant control effect on the conversion between tectonic storage and residual gas storage.
[0083] Recommendation: In practical engineering applications, in the short term, shortening the injection duration can promote the transformation from structural sequestration to residual gas sequestration and dissolution sequestration; in the long term, injecting CO2 into lower structural locations can promote the transformation from structural sequestration to residual gas sequestration, thereby achieving a more stable and safer sequestration effect.
[0084] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention. Those skilled in the art, under the guidance of the present invention, can make various similar representations without departing from the spirit and claims of the present invention, and such modifications all fall within the protection scope of the present invention.
Claims
1. A method for assessing the carbon dioxide sequestration potential and mechanism of deep saline aquifers, characterized in that, Includes the following steps: S1, obtain the geological and fluid parameters of the target saline aquifer, as well as the gas injection regime parameters; S2. Based on the parameters obtained in step S1, construct a multi-field coupled numerical simulation model for carbon dioxide sequestration in deep saline water layers. This numerical simulation model sets the parameter space for the injection regime. S3, performs multiple sets of numerical simulations within the injection system parameter space, outputs the decomposition results of the sequestration quantity and sequestration mechanism at different time scales, and constructs a sequestration quantity database; S4. Based on the storage volume database mentioned in step S3, a machine learning proxy model is constructed with the gas injection regime parameters as input features and the final storage volume as the output target. The optimal proxy model is then selected from multiple candidate models based on the model prediction performance index. S5. Based on the optimal surrogate model, interpretability analysis is introduced, and the contribution of each input regime parameter is calculated using the SHAP analysis method to obtain the feature importance ranking and influence direction of each gas injection regime parameter on the prediction result of the storage volume. S6. Based on the optimal surrogate model and Monte Carlo simulation method, the injection system parameters are randomly sampled, and the contribution of each sealing mechanism at different time scales is predicted by probability distribution, so as to obtain the probability distribution of the contribution of the sealing mechanism at at least two time scales. S7. Based on the results of steps S5 and S6, output the assessment results of the carbon dioxide storage potential of the target block, the evolution law of the contribution of the storage mechanism under multiple time scales, and the control law of the gas injection regime parameters.
2. The method according to claim 1, characterized in that: The numerical simulation model mentioned in step S2 is a three-dimensional geological model coupled with thermal-fluidization. The storage mechanisms include four types: dissolution storage, residual gas storage, tectonic storage, and mineralization storage.
3. The method according to claim 1 or 2, characterized in that: The gas injection regime parameters mentioned in step S1 include at least the injection pressure, injection rate, injection time, injection location, and injection method; In step S2, multiple gas injection conditions are constructed by changing the injection rate, injection time, injection pressure, and injection location, and the sealed volume database is formed based on the simulation results of each gas injection condition.
4. The method according to claim 1 or 2, characterized in that: The candidate models mentioned in step S4 include at least two of the following: random forest model, extreme gradient boosting model, multilayer perceptron model, and support vector regression model; The model's predictive performance metrics include the coefficient of determination. R 2 Root mean square error RMSE Mean absolute error MAE and mean absolute percentage error MAPE One or more of them.
5. The method according to claim 1, characterized in that: In step S5, the importance of each input feature is determined by calculating the average absolute SHAP value of each input feature, and the contribution of the corresponding input feature to the prediction result of the sealed quantity is determined according to the sign of the SHAP value.
6. The method according to claim 1, characterized in that: In step S6, the contribution of each sealing mechanism is characterized by at least one of frequency distribution, quantile statistics, or confidence interval.
7. The method according to claim 1, characterized in that: Step S7 further includes providing suggestions for optimizing the injection regime based on the contribution evolution law and the control law of the gas injection regime parameters. The optimization suggestions include at least adjusting the injection duration and / or adjusting the injection position to increase the total amount of sequestration or increase the contribution of the target mechanism at the target time scale.
8. A system for evaluating the potential and mechanism of carbon dioxide sequestration in deep saline aquifers, characterized in that, include: The data acquisition module is used to acquire the geological and fluid parameters of the target saline aquifer, as well as the gas injection regime parameters; The numerical simulation module is used to establish a multi-field coupled numerical simulation model for carbon dioxide sequestration in the target saline aquifer and generate a sequestration volume database. The agent modeling module is used to build machine learning agent models and select the optimal agent model; The interpretability analysis module is used to calculate the contribution of each input feature to the prediction result of the sealed quantity using the SHAP analysis method; The uncertainty analysis module is used to predict the contribution distribution of various sealing mechanisms at different time scales using Monte Carlo simulation methods. The results output module is used to output at least the results of the storage potential assessment, the evolution law of the storage mechanism, and the control law of the gas injection system parameters.
9. A computer device, characterized in that, include: Memory, which stores executable instructions; A processor that executes the executable instructions in the memory to implement the deep saline aquifer carbon dioxide sequestration potential and mechanism assessment method according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for assessing the carbon dioxide sequestration potential and mechanism of deep saline aquifers as described in any one of claims 1 to 7.