Method for predicting CO2 leakage after fracture of CO2 geological storage cover layer and reservoir breakthrough

Abstract

The invention belongs to the technical field of CO2 geological sequestration, and particularly relates to a method for predicting CO2 leakage after fracture of a CO2 geological sequestration cover layer and breakthrough of a reservoir stratum.According to the method, by establishing a CO2 sequestration model, sequestration processes under different CO2 injection amounts are simulated; correlation is established between strain and pressure in the cover layer and gas saturation in the cover layer and between strain and pressure in the reservoir layer through machine learning, correlation between the CO2 injection amount and strain and pressure response is proved, it is indicated that strain and pressure monitoring can track the CO2 injection process, safe injection is guaranteed, and the CO2 injection efficiency is improved. And the CO2 leakage risk caused by the geomechanical effect is reduced. The method is convenient to operate, after machine learning is completed, the trained model is deployed in a production environment, CO2 leakage can be judged only by processing data monitored by the DAS through a computer, the problems of cap fracture and reservoir breakthrough are found in advance, and the corresponding position is determined.

Description

Technical Field

[0001] The invention relates to the technical field of CO2 geological storage, and is a method for predicting CO2 leakage after a CO2 geological storage cap rock rupture or a reservoir breakthrough. Background Art

[0002] CO2 is a greenhouse gas that exists naturally in the atmosphere and is also emitted by human activities. Although CO2 has a lower global warming potential than other gases such as methane, its emissions make it a major cause of global warming and ocean acidification. Therefore, CO2 emissions must be reduced, and one of the important measures is CO2 storage. CO2 storage is a method of reducing CO2 emissions in the atmosphere by injecting it into a reservoir deep underground or by other means, which is used to curb global warming.

[0003] Long-term storage of CO2 may cause the tensile strength of the caprock and reservoir to weaken, thereby inducing the leakage of the stored CO2. Once a large amount of CO2 leaks, it will cause serious adverse consequences. CO2 leakage may cause the formation at the storage point to sink, thereby triggering micro-earthquakes and destroying the stability of the entire soil ecosystem; CO2 leakage may cause groundwater acidification, which will cause soil acidification after circulating with soil water and damage the soil environment; CO2 leakage overflows the surface and cannot achieve the goal of reducing CO2 emissions. Therefore, it is of great significance to monitor the CO2 storage effect and give early warning of CO2 leakage. DAS is a sensing system that uses optical fiber as a sensing sensitive element and mission signal medium. It uses distributed acoustic sensing vertical seismic profile (DAS VSP) technology to monitor the amount of CO2 storage and has great potential in reservoir monitoring.

[0004] Chinese patent document CN118425413A (202410254374.X) discloses a real-time monitoring method and system for carbon dioxide storage, which uses monitoring sensors and detectors deployed on the surface, CO2 transmission pipelines, and reservoirs to obtain real-time data, and compare the data with the three set thresholds to see whether the corresponding alarm module is triggered. The monitoring method adopted by this patent faces the problems of high cost and difficulty in construction, and it is difficult to achieve centralized data management and sharing.

[0005] Chinese patent document CN115906409A (202211296793.7) discloses a method and system for predicting and evaluating the leakage risk of carbon dioxide storage, and proposes a CO2 storage layer model; according to the pressure value and coordinates in the model, it is judged whether there is a leakage risk in the carbon dioxide storage site. This patent only simulates the reservoir, and fails to fully consider the interaction and influence of the reservoir and the cap layer in the CO2 geological storage system, ignoring the good sealing ability of the cap layer on the CO2 in the reservoir. Summary of the invention

[0006] The main purpose of the present invention is to provide a method for predicting CO2 leakage after cap rock rupture and reservoir breakthrough in CO2 geological storage, and proposes a method for predicting CO2 leakage using machine learning, which can timely detect cap rock rupture, reservoir breakthrough phenomena and specific locations, and provide technical support for subsequent response measures.

[0007] The technical problem to be solved by the present invention is achieved by adopting the following technical solution: A method for predicting CO2 leakage after the rupture of the cap rock and reservoir breakthrough of CO2 geological storage, comprising the following steps:

[0008] S1. Establish a CO2 storage model. The CO2 storage model is a closed saline layer. The CO2 storage model includes a cover layer, a cap layer and a reservoir layer arranged in sequence from top to bottom, and the rock properties of the cover layer, the cap layer and the reservoir layer are set respectively;

[0009] Injection wells are provided in the CO2 storage model;

[0010] Setting a crack model in the caprock and setting the crack opening stress threshold;

[0011] S2. Simulate the storage process under different CO2 injection rates and different reservoir permeabilities;

[0012] S3, extracting the strain and pressure data corresponding to different positions in the cap layer at different injection times during the CO2 injection process, and simultaneously extracting the gas saturation data at the corresponding positions of the cap layer and the reservoir closest to the extraction point of the cap layer, and combining the strain and pressure data corresponding to different positions of the cap layer extracted at the same injection time and the gas saturation data at the corresponding nearest position of the cap layer to form a first feature data set, and combining the strain and pressure data corresponding to different positions of the cap layer extracted at the same injection time and the gas saturation data at the corresponding nearest position of the reservoir to form a second feature data set;

[0013] Divide the first feature data set and the second feature data set into a training set, a validation set and a test set respectively;

[0014] S4, using the extracted first feature data set and the second feature data set to train the machine learning model respectively, with the strain and pressure data as input data;

[0015] The gas saturation is used as the judgment standard. If the gas saturation is greater than 0, it means that carbon dioxide exists and the output is 1; if the gas saturation is equal to 0, it means that there is no carbon dioxide and the output is 0;

[0016] S5. When the accuracy of the predicted value of the test set reaches a set threshold, the threshold here is set inside the machine learning model. The specific method is the existing technology and will not be described here. A trained machine learning model is obtained. The first feature data set is trained to obtain a cap rock rupture prediction model, and the second feature data set is trained to obtain a wellbore breakthrough prediction model;

[0017] S6. Using the strain and pressure data in the cap layer collected by the DAS monitoring system, the cap layer rupture prediction model and the wellbore breakthrough prediction model are used to predict whether the cap layer will rupture, whether the reservoir will break through, and their corresponding positions.

[0018] Preferably, in step S6, specifically, a DAS monitoring system in a monitoring well or on the outer wall of an injection well is used to collect strain and pressure data at various locations in the cap layer, and input the data into a cap layer rupture prediction model. If the output is 1, carbon dioxide exists in the cap layer closest to the corresponding measurement position of the cap layer, and the cap layer has been ruptured; if the output is 0, carbon dioxide does not exist in the cap layer closest to the corresponding measurement position of the cap layer, and the cap layer has not been ruptured;

[0019] The DAS monitoring system in the monitoring well or on the outer wall of the injection well is used to collect strain and pressure data from various locations in the cap rock and input them into the wellbore breakthrough prediction model. If the output is 1, then there is carbon dioxide in the reservoir closest to the corresponding measurement position of the cap rock, and the reservoir has been breached; if the output is 0, then there is no carbon dioxide in the reservoir closest to the corresponding measurement position of the cap rock, and the reservoir has not been breached.

[0020] In the actual prediction process, the permeability data of the cap rock can be input into the cap rock rupture prediction model or the wellbore breakthrough prediction model, and combined with the actual injection amount and actual injection time of carbon dioxide to assist in improving the accuracy of the prediction.

[0021] Preferably, in step S4, the machine learning model includes an MLP model;

[0022] 1) Design the network structure: Design an MLP model consisting of a fully connected layer (Dense), which can extract features from the input data and make predictions;

[0023] The input layer of the model receives normalized strain and pressure data;

[0024] The model contains two hidden layers. The first hidden layer has 64 neurons and the second hidden layer has 32 neurons. Both use the ReLU activation function to introduce nonlinearity.

[0025] 2) Feature extraction: The first hidden layer uses 128 neurons to capture complex patterns in the input data. The complex patterns here are the processing methods of the MLP model and have no specific meaning;

[0026] The second hidden layer uses 64 neurons to further refine the features. This step is a deep learning process and is a state-of-the-art technology.

[0027] 3) Data output: For discrete responses (classification tasks), the output layer uses one neuron with a sigmoid activation function, which is suitable for binary classification problems and outputs 0 (not leaked) or 1 (leaked) to predict the event outcome.

[0028] Preferably, in step S4, the machine learning model includes a CNN model;

[0029] 1) Design network structure: Design a CNN model containing a one-dimensional convolutional layer (Conv1D) and a maximum pooling layer (MaxPooling1D) to extract spatiotemporal features from the input data;

[0030] The input layer receives the strain and pressure data after reshaping. The data shape is (batch_size, 1, 9), where 1 represents one dimension; 9 represents the number of features, including 1 layer of time, 1 layer of permeability, 1 layer of injection volume, 3 layers of strain and 3 layers of pressure data; "reshaping" is an existing technology, which can be implemented by deep learning code. The purpose of reshaping is to match the code with the data. The above 9 features are all input data;

[0031] 2) Feature extraction: The first convolutional layer uses 64 filters with a kernel size of 3. The ReLU activation function is used to introduce nonlinearity, and the output size is maintained by "same" padding;

[0032] This is followed by a max pooling layer, which is used to reduce the spatial size of features and increase the abstraction ability of the model;

[0033] After the flattening layer, the data is fed into two fully connected layers (Dense), with 128 and 1 neurons respectively, to continue extracting features and performing nonlinear transformations.

[0034] The last fully connected layer uses the sigmoid activation function to meet the needs of the binary classification problem;

[0035] 3) Model training: The discrete response model was trained for 100 epochs. In order to prevent overfitting, the early stopping method was set. The early stopping method is an existing technology and a convergence condition.

[0036] Preferably, in step S4, the machine learning model includes a dual CNN model;

[0037] 1) Design network structure: Design a dual CNN model consisting of a one-dimensional convolutional layer (Conv1D) and a maximum pooling layer (MaxPooling1D) to extract spatiotemporal features from the input data;

[0038] The input layer receives the reshaped strain and pressure data, and the data shape is (batch_size, 1, 9), where 1 represents one dimension; 9 represents the number of features, including 1 layer of time, 1 layer of permeability, 1 layer of injection volume, 3 layers of strain and 3 layers of pressure data;

[0039] 2) Feature extraction: The first convolutional layer of the model uses 64 filters with a kernel size of 3. The ReLU activation function is used to introduce nonlinearity, and the output size is maintained by "same" padding;

[0040] This is followed by a max pooling layer, which is used to reduce the spatial size of features and increase the abstraction ability of the model;

[0041] Then comes the second convolutional layer, which also uses 64 filters, a kernel size of 3, a ReLU activation function, and the same padding;

[0042] This is followed by a second max pooling layer;

[0043] After the flattening layer, the data is fed into two fully connected layers (Dense), with 128 and 1 neurons respectively, to continue extracting features and performing nonlinear transformations.

[0044] The last fully connected layer uses the sigmoid activation function to meet the needs of the binary classification problem;

[0045] 3) Model training: The discrete response model was trained for 100 epochs, and the early stopping method was also set to prevent overfitting.

[0046] Preferably, the early stopping method of the present invention has the following requirements: if the validation set loss value does not improve within 50 epochs, the model with the lowest loss value before is saved. The above is an explanation of the early stopping method. If the loss value does not improve within 50 epochs, the simulation is stopped and the optimal model before the stop is saved.

[0047] Preferably, the present invention also includes step S7: using accuracy to evaluate the performance of different machine learning models, outputting a confusion matrix heat map to intuitively display the prediction status of different machine learning models in the test set, and selecting a machine learning model with higher accuracy.

[0048] Preferably, in step S1 of the present invention, the fracture model and the rock properties of the overburden, caprock and reservoir are set according to the actual terrain;

[0049] The opening stress threshold of the fracture is set according to the rock type;

[0050] The location of the injection well is selected according to the actual terrain, and the monitoring well is set according to the situation. The DAS monitoring system is installed in the monitoring well. The location of the monitoring well ensures that the corresponding predicted area of ​​the injection well is within the DAS monitoring range, and comprehensively considers factors such as the signal-to-noise ratio; in addition, the DAS monitoring system can also be directly installed outside the injection well.

[0051] Preferably, in step S1, the cap rock permeability model adopts the Barton-Bandis model. There are many permeability models, and the Barton-Bandis model is suitable for application scenarios that need to consider nonlinear shear characteristics of rock joints, hydraulic fracturing simulation, rock slope stability analysis, and dual medium models in reservoir simulation.

[0052] Preferably, in step S1, the caprock and the reservoir are both simulated using the Mohr-Coulomb criterion, and the overburden is simulated using the Drucker-Prager criterion. The Mohr-Coulomb criterion is applicable to the case where compression shear failure is considered, and both the caprock and the reservoir may experience compression shear failure; the Drucker-Prager criterion is applicable to a wider range of materials and stress states, and is applicable to the overburden.

[0053] The inventive concept of the present invention is that when carbon dioxide (CO2) is injected into a deep saline layer, excessive overpressure may cause the cap rock to rupture and the reservoir to break, resulting in CO2 leakage. Therefore, monitoring the geomechanical changes of the cap rock and reservoir is of great significance for evaluating the stability of CO2 injection.

[0054] After research, it was found that if there is a carbon dioxide breakthrough in the reservoir, the carbon dioxide will show a plume phenomenon along the outside of the injection well, that is, the carbon dioxide will rise vertically along the outer wall of the injection well. After reaching the junction of the cap rock and the reservoir, due to the obstruction of the cap rock, the carbon dioxide will further diffuse laterally along the junction of the cap rock and the reservoir. Therefore, the pressure and strain at the junction of the cap rock and the reservoir will change. Based on the above phenomenon, the present invention can establish a correlation between the strain and pressure data in the cap rock and the gas saturation of the reservoir below the junction of the cap rock and the reservoir through machine learning by monitoring the strain and pressure data in the cap rock, and then predict whether the carbon dioxide has broken through and the boundary position of the carbon dioxide diffusion after the breakthrough.

[0055] In addition, the cap layer plays a role in sealing the reservoir, and there are cracks in the cap layer. When the pressure of the carbon dioxide in the reservoir reaches a certain value, the carbon dioxide will enter the cracks, causing the cracks to open and further expand. When the cap layer is penetrated by the cracks, the carbon dioxide will enter the covering layer, i.e., causing carbon dioxide leakage. Since the entry of carbon dioxide into the cracks will cause the pressure and strain inside the cap layer to change, the extension direction of the cracks is usually in the direction of least resistance. Therefore, usually the junction of the cap layer and the covering layer is closest to the end of the crack and is first broken, and carbon dioxide leakage occurs. Therefore, the present invention can establish a correlation between the strain and pressure data in the cap layer and the gas saturation in the covering layer above the junction of the cap layer and the covering layer through machine learning by monitoring the strain and pressure data in the cap layer, and then predict whether the cap layer is broken and the leakage position of carbon dioxide after the break.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention establishes a CO2 storage model, simulates the storage process under different CO2 injection amounts, and uses machine learning to establish a correlation between the strain and pressure in the cap rock and the gas saturation in the overburden and the gas saturation in the reservoir, respectively, to prove the correlation between the CO2 injection amount and the strain and pressure response, indicating that strain and pressure monitoring can track the CO2 injection process, provide protection for safe injection, and reduce the risk of CO2 leakage caused by geomechanical effects.

[0057] 2. The present invention proposes a new data analysis method, which uses machine learning to analyze DAS monitoring data, revealing anomalies in strain and pressure data in the cap rock, and further monitoring CO2 leakage from cap rock rupture and reservoir breakthrough.

[0058] 3. The carbon dioxide leakage monitoring method based on machine learning can automatically and quickly monitor the strain and pressure anomalies caused by cap rock rupture and reservoir breakthrough, greatly reducing labor and time costs.

[0059] 4. The present invention is easy to operate. After machine learning is completed, the trained model is deployed in the production environment. The CO2 leakage can be identified by only processing the DAS monitoring data through a computer, so as to detect the cap rock rupture and reservoir breakthrough problems in advance and determine the corresponding location, thereby achieving timely processing and ensuring production safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 It is a structural schematic diagram of the sealing model of the present invention;

[0061] Figure 2 In the embodiment of the present invention, the injection amount of the xz-6 plane is 2×10 5 m 3 / d, gas saturation distribution diagram after 5 years of CO2 injection when the reservoir permeability is 15mD;

[0062] Figure 3 Confusion matrix heatmap for the longitudinal MLP, CNN, and dual-CNN test sets;

[0063] Figure 4 Confusion matrix heatmap for horizontal MLP, CNN, and dual CNN test sets;

[0064] Figure 5 The injection volume is 500, 1000, and 5000m 3 / d scatter plot of MLP longitudinal prediction and actual results;

[0065] Figure 6 The injection volume is 1×10 4 , 2×10 4 , 5×10 4 m 3 / d scatter plot of MLP longitudinal prediction and actual results;

[0066] Figure 7 The injection volume is 1×10 5 , 1.5×10 5 , 2×10 5 m 3 / d scatter plot of MLP longitudinal prediction and actual results;

[0067] Figure 8 The injection volume is 500, 1000, and 5000m 3 / d scatter plot of MLP horizontal prediction and actual results;

[0068] Fig. 9 The injection volume is 1×10 4 , 2×10 4 , 5×10 4 m 3 / d scatter plot of MLP horizontal prediction and actual results;

[0069] Fig.10 The injection volume is 1×10 5 , 1.5×10 5 , 2×10 5 m 3 / d scatter plot of MLP horizontal prediction and actual results;

[0070] Fig.11 The injection volume is 500, 1000, and 5000m 3 / d scatter plot of CNN longitudinal prediction and actual results;

[0071] Fig.12 The injection volume is 1×10 4 , 2×10 4 , 5×10 4 m 3 / d scatter plot of CNN longitudinal prediction and actual results;

[0072] Fig.13 The injection volume is 1×10 5 , 1.5×10 5 , 2×10 5 m 3 / d scatter plot of CNN longitudinal prediction and actual results;

[0073] Fig.14 The injection volume is 500, 1000, and 5000m 3 / d scatter plot of CNN horizontal prediction and actual results;

[0074] Fig.15 The injection volume is 1×10 4 , 2×10 4 , 5×10 4 m 3 / d scatter plot of CNN horizontal prediction and actual results;

[0075] Fig.16 The injection volume is 1×10 5 , 1.5×10 5 , 2×10 5 m 3 / d scatter plot of CNN horizontal prediction and actual results;

[0076] Fig.17 The injection volume is 500, 1000, and 5000m 3 / d scatter plot of the dual CNN longitudinal prediction and actual results;

[0077] Fig.18 The injection volume is 1×10 4 , 2×10 4 , 5×10 4 m 3 / d scatter plot of the dual CNN longitudinal prediction and actual results;

[0078] Fig.19 The injection volume is 1×10 5 , 1.5×10 5 , 2×10 5 m 3 / d scatter plot of the dual CNN longitudinal prediction and actual results;

[0079] Fig. 20 The injection volume is 500, 1000, and 5000m 3 / d scatter plot of double CNN horizontal prediction and actual results;

[0080] Fig.21 The injection volume is 1×10 4 , 2×10 4 , 5×10 4 m 3 / d scatter plot of CNN horizontal prediction and actual results;

[0081] Fig. 22 The injection volume is 1×10 5 , 1.5×10 5 , 2×10 5 m 3 Scatter plot of CNN horizontal prediction and actual results at / d. DETAILED DESCRIPTION

[0082] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings.

[0083] The present invention uses the emerging, high-resolution, low-cost distributed acoustic sensing (DAS) technology to develop a new data analysis process for monitoring cap rock leakage and reservoir breakthrough, and applies machine learning (ML) to reveal leakage-related anomalies in DAS monitoring data during carbon dioxide injection. The embodiment of the present invention establishes a CO2 storage model, obtains a synthetic data set by changing the CO2 injection amount and reservoir permeability, and analyzes and determines their impact on cap rock rupture and reservoir breakthrough. In this embodiment, a multi-layer perceptron (MLP), convolutional neural network (CNN) and dual convolutional neural network (Dual CNN) architecture test are used to use these synthetic strain and multi-level pressure data to quickly monitor and predict CO2 leakage through the cap rock and reservoir.

[0084] The present invention is achieved by the following measures:

[0085] The present invention uses the CO2 storage model established by CMG GEM, sets a monitoring well 50m away from the injection well, and changes the CO2 injection amount and reservoir permeability. The strain and pressure response characteristics related to CO2 leakage are recorded by the low-frequency DAS in the monitoring well. A synthetic data set is obtained, and ML is applied to reveal the anomalies related to leakage in the DAS monitoring data during the carbon dioxide injection process to quickly monitor and predict CO2, so as to provide early warning alarms. The model includes three parts from top to bottom: the cap layer, the cap layer and the reservoir. The reservoir is an underground porous rock that provides storage space for CO2; the cap layer is a low-permeability rock layer above the reservoir, and its main function is to prevent CO2 from migrating upward to the shallow aquifer or the surface. The cap layer refers to the rock layer above the reservoir and the cap layer, and its function is to provide an additional physical barrier to prevent CO2 from migrating upward. Monitoring strain and pressure can provide more comprehensive underground environmental information. When CO2 leaks, the displacement change lags behind the stress change, and the monitoring of strain can provide early warning. The two monitoring data can verify each other, reducing the errors and uncertainties that may be caused by single monitoring data. Based on the strain and pressure values, combined with the gas saturation, machine learning methods are used to determine whether carbon dioxide is leaking.

[0086] The model has 101 grids along the x direction, each grid is 10m, and the total length in the x direction is 1010m; there are 11 grids along the y direction, each grid is 10m, and the total length in the y direction is 110m; there are 33 grids along the z direction, each grid is 10m, and the total length in the z direction is 330m, including 10 layers of cover layer, totaling 100m, 3 layers of cap layer, totaling 30m, and 20 layers of reservoir layer, totaling 200m. The model is a closed saline layer, CO2 is injected into the reservoir and sealed, and there is no flow through the boundary; the cap layer is a brittle material with a high Young's modulus, and its permeability model adopts the Barton-Bandis model, with an initial permeability of basically 0. Once the stress reaches a certain value, the cracks will open and expand rapidly, and CO2 will leak.

[0087] The storage model of the present invention simulates 9 different daily CO2 injection amounts and 3 different reservoir permeabilities, with a total of 27 schemes. The strain and pressure response levels generated by different amounts of CO2 injected into the reservoir are also different. The data set is synthesized by extracting stress, strain and gas saturation, and the machine model is developed and tested using the synthesized data set as input data. The ML test algorithm includes: multi-layer perceptron (MLP), convolutional neural network (CNN) and dual convolutional neural network (Dual CNN). The present invention uses these ML methods to quickly monitor CO2 leakage, and considers the method of measuring CO2 leakage: "discrete" response, as long as CO2 exists in the cap layer {51, 6, 10} area, its value is 1, it is considered that the cap layer is broken and CO2 is leaking; as long as CO2 exists in the reservoir {55, 6, 14}, its value is 1, it is considered that the reservoir is broken and CO2 is leaking. The strains and pressures generated by different CO2 injection amounts and different reservoir permeabilities are also different, and the situations of cap layer rupture and reservoir breakthrough are also different. Using simulated data to train the machine learning model is to find their association.

[0088] Specifically, Figure 1 As shown in the figure, the storage model consists of three parts from top to bottom: the cover layer, the cap layer and the reservoir. It is a 3D model. The number of grids in the x, y and z directions of the model are 101, 11 and 33 respectively. Each grid is represented by its xyz coordinates, with the x direction being positive to the right, the y direction being positive backward and the z direction being positive downward. The injection wells are drilled from {51, 6, 1} to {51, 6, 33}, and the monitoring well is drilled 40m to the right of the injection well. The scale of each grid in the model is 10m, with a total of 1010m in the x direction, 110m in the y direction and 330m in the z direction. There are 10 layers of grids in the cover layer, totaling 100m, 3 layers of grids in the cap layer, totaling 30m, and 20 layers of grids in the reservoir, totaling 200m. The model is a closed saline layer, CO2 is injected into the reservoir and sealed, and no flow passes through the boundary.

[0089] The cracks in the caprock of the storage model are orthogonal at intervals of 10 m in all directions. The crack model in this embodiment is set using the crack model in the document "ROCK FRACTURING DUE TO CO2 INJECTION", and there are no cracks in the rest of the part. The caprock and reservoir have different compressibility, but the other properties are the same. They are both simulated using the Mohr-Coulomb criterion, while the overburden layer uses the Drucker-Prager criterion. The rock properties are shown in Table 1.

[0090] Table 1 Rock properties

[0091]

[0092] The cap rock permeability model adopts the Barton-Bandis model. The initial permeability is basically 0. When the effective normal stress drops below the crack opening stress threshold of 2000 kPa, the crack opens at the bottom of the cap rock and expands rapidly, extending in the horizontal and vertical directions until the cap rock breaks and CO2 leaks. The parameters of the Barton-Bandis model are shown in Table 2.

[0093] Table 2 Barton-Bandis model parameters

[0094]

[0095] The present invention designs 27 schemes, where the reservoir permeability is 15mD, 100mD, and 500mD, and the daily injection volume is: 500m 3 / d, 1000m 3 / d, 5000m 3 / d, 1×10 4 m 3 / d, 2×10 4 m 3 / d, 5×10 4 m 3 / d, 1×10 5 m 3 / d, 1.5×10 5 m 3 / d, 2×10 5 m 3 / d. The details are shown in Table 3.

[0096] Table 3 Scheme parameters

[0097]

[0098] The present invention has calculated the above 27 solutions respectively, taking the injection amount of 2×10 5 m 3 / d, and the reservoir permeability is 15mD as an example. Figure 2This is the gas saturation distribution diagram after CO2 injection for 5 years. It can be seen from the figure that after the cap layer ruptures, CO2 enters the overburden through the cap layer, and the CO2 reservoir breaks through and diffuses in the reservoir. In this process, the strain and pressure of the cap layer change. This embodiment considers the method of measuring CO2 leakage: "discrete" response, that is, any amount of CO2 reaching the overburden area above the cap layer at the injection well location is considered to be a leak, and its value is 1, and an alarm starts; as long as there is CO2 in the reservoir {55, 6, 14}, its value is 1, and it is considered that the reservoir has broken through and CO2 has leaked, and an alarm starts. The present invention extracts the strain, pressure and gas saturation of various cap layers {55, 6, 11}, {55, 6, 12}, {55, 6, 13}, the gas saturation of the overburden {51, 6, 10} and the gas saturation of the reservoir {55, 6, 14}, which provides a set of several strain measurement values ​​and several pressure measurement values. These data are used as synthetic data sets to train an ML model, wherein the strain and pressure of the cap layers {55, 6, 11}, {55, 6, 12}, {55, 6, 13} are used as input data, and the gas saturation of the cap layers {55, 6, 11}, {55, 6, 12}, {55, 6, 13} is used as a reference value.

[0099] The present invention collects and organizes data related to carbon dioxide storage leakage, and extracts pressure and strain data obtained by CMG simulation. At the same time, the data is ensured to be reliable, outliers, missing values ​​and duplicate data are cleaned up, and quality and consistency are guaranteed. The input data and response data are randomly divided into a training set, a validation set and a test set. The training set accounts for 64%, the validation set accounts for 16%, and the test set accounts for 20%. The test set is used to evaluate the generalization ability of the model in the model selection and final evaluation stages, while the validation set is used for parameter adjustment during the model training process.

[0100] The present invention uses multi-layer perceptron (MLP), convolutional neural network (CNN) and dual convolutional neural network (DualCNN) models to handle the discrete task prediction of vertical cap rock rupture leakage and lateral reservoir leakage. The following is the detailed training process of each model:

[0101] 1. Multilayer Perceptron (MLP) training

[0102] 1) Design of network structure: We designed an MLP model consisting of fully connected layers (Dense), which can extract features from input data and make predictions. The input layer of the model receives standardized strain and pressure data. The model contains two hidden layers, the first hidden layer has 64 neurons, and the second hidden layer has 32 neurons. Both use the ReLU activation function to introduce nonlinearity.

[0103] 2) Feature extraction: The first hidden layer uses 128 neurons to capture complex patterns in the input data. The second hidden layer uses 64 neurons to further refine the features.

[0104] 3) Data output: For discrete responses (classification tasks), the output layer uses one neuron with a sigmoid activation function, which is suitable for binary classification problems and outputs 0 (not leaked) or 1 (leaked) to predict the event outcome.

[0105] 2. Convolutional Neural Network (CNN) Training

[0106] 1) Design of network structure: A CNN model consisting of a one-dimensional convolutional layer (Conv1D) and a maximum pooling layer (MaxPooling1D) was designed to extract spatiotemporal features from the input data. The input layer receives the reshaped strain and pressure data, and the data shape is (batch_size, 1, 9), where 1 represents one dimension; 9 represents the number of features, including 1 layer of time, 1 layer of permeability, 1 layer of injection volume, 3 layers of strain, and 3 layers of pressure data;.

[0107] 2) Feature extraction: The first convolutional layer uses 64 filters with a kernel size of 3. The ReLU activation function is used to introduce nonlinearity and maintain the output size through "same" padding. This is followed by a maximum pooling layer to reduce the spatial size of the features and increase the abstraction ability of the model. After the flattening layer, the data is sent to two fully connected layers (Dense), with 128 and 1 neurons respectively, to continue extracting features and performing nonlinear transformations. The last fully connected layer uses the sigmoid activation function to meet the needs of the binary classification problem.

[0108] 3) Model training: The discrete response model was trained for 100 epochs. To prevent overfitting, an early stopping method was used. If the loss value of the model did not improve within 50 epochs, the model with the lowest loss was saved.

[0109] 3. Dual CNN training

[0110] 1) Design of network structure: A dual CNN model consisting of a one-dimensional convolutional layer (Conv1D) and a maximum pooling layer (MaxPooling1D) was designed to extract spatiotemporal features from the input data. The input layer receives the reshaped strain and pressure data, and the data shape is (batch_size, 1, 9), where 9 represents the number of features (1 layer of time, 1 layer of permeability, 1 layer of injection volume, 3 layers of strain, and 3 layers of pressure data).

[0111] 2) Feature extraction: The first convolutional layer of the model uses 64 filters, a kernel size of 3, and the ReLU activation function is used to introduce nonlinearity, and the output size is maintained through "same" padding. This is followed by a maximum pooling layer to reduce the spatial size of the features and increase the abstraction ability of the model. Then comes the second convolutional layer, which also uses 64 filters, a kernel size of 3, the ReLU activation function, and "same" padding. This is followed by a second maximum pooling layer. After the flattening layer (Flatten), the data is sent to two fully connected layers (Dense), with 128 and 1 neurons respectively, to continue extracting features and performing nonlinear transformations. The last fully connected layer uses the sigmoid activation function to meet the needs of the binary classification problem.

[0112] 3) Model training: The discrete response model was trained for 100 epochs. To prevent overfitting, the early stopping method was also set. If the validation set loss value did not improve within 50 epochs, the model with the lowest loss value was saved.

[0113] The present invention uses the trained discrete response model to predict the previously divided test set data. This process is completely independent of training and is intended to test the performance of the model and its predictive ability for unknown data. The accuracy is used to evaluate its performance, and the final output confusion matrix heat map intuitively displays the prediction of the model in the test set.

[0114] Longitudinal cover rupture leakage: For longitudinal cover rupture leakage, the performance of longitudinal cover rupture prediction models trained by MLP, CNN, and dual CNN is not much different. The accuracy of MLP is 99.9%, the accuracy of CNN is 99.1%, and the accuracy of dual CNN is 99.4%. The confusion matrix of each model is as follows Figure 3 As shown, the three machine learning models can make accurate predictions for most situations.

[0115] Horizontal wellbore breakthrough: For horizontal wellbore breakthrough, the performance of the wellbore breakthrough prediction models trained with MLP, CNN, and dual CNN is not much different, but the performance is slightly better than that of the vertical model. The accuracy of MLP is 99.8%, the accuracy of CNN is 99.4%, and the accuracy of dual CNN is 99.0%. The confusion matrix of each model is as follows Figure 4 As shown, the three models can make accurate predictions for most situations.

[0116] The present invention performs leakage warning and prediction for injection schemes under different carbon dioxide injection amounts and cap rock permeability. By interpreting and applying the warning and prediction results, a scientific basis can be provided for practical operations to ensure the safety and effectiveness of the injection scheme. The two-directional discrete prediction model mainly warns of leakage events that may occur during the carbon dioxide injection process, and the output is 0 (no leakage) or 1 (leakage). By learning different characteristic parameters, the model can predict whether there is a leak. In the discrete scatter plot, the horizontal axis represents the time of simulated injection, and the vertical axis represents the prediction result.

[0117] This embodiment takes the reservoir matrix permeability of 15md as an example, and uses three different machine learning models, namely multi-layer perceptron (MLP), convolutional neural network (CNN) and double convolutional neural network (double CNN), to predict and analyze the vertical cap rock rupture leakage and lateral reservoir breakthrough leakage for different injection rates. The following is a comprehensive analysis of these prediction results.

[0118] 1. Analysis of MLP prediction results

[0119] Analysis of the prediction results of vertical cap rock rupture: Under the condition that the reservoir matrix permeability is 15md, after using MLP to learn the vertical leakage situation, the trained vertical cap rock rupture prediction model is called to test it. The scatter plot of prediction and actual results is shown in the figure below. Figure 5 , Figure 6 and Figure 7 shown.

[0120] The overall accuracy of the longitudinal prediction model is above 99%, and the injection volume is 5000 and 1×10 4 , 1.5×10 5 , 2×10 5 m 3 / dThe prediction accuracy of the four situations is close to 100%.

[0121] Analysis of lateral reservoir prediction results: Under the condition that the reservoir matrix permeability is 15md, after using MLP to learn the lateral leakage situation, the trained wellbore breakthrough prediction model is called to test it. The scatter plot of prediction and actual results is shown in the figure Figure 8 , Fig. 9 and Fig.10 shown.

[0122] The overall accuracy of the wellbore breakthrough prediction model is above 99%, and the injection volume is 1×10 4 , 1.5×10 5 m 3 / dThe prediction accuracy of the two cases is close to 100%.

[0123] 2. Analysis of CNN prediction results

[0124] Analysis of the prediction results of vertical cap rock rupture: Under the condition that the reservoir matrix permeability is 15md, after learning the vertical leakage situation using CNN, the trained vertical cap rock rupture prediction model is called to test it. The scatter plot of prediction and actual results is shown in the figure below. Fig.11 , Fig.12 and Fig.13 shown.

[0125] The prediction results of the CNN model are similar to those of the MLP model. The overall accuracy of the longitudinal prediction model is above 99%, with an injection volume of 5000 and 1×10 4 , 5×10 4 , 1.5×10 5 m 3 / dThe prediction accuracy of the four situations is close to 100%.

[0126] Analysis of lateral reservoir prediction results: Under the condition that the reservoir matrix permeability is 15md, after using CNN to learn the lateral leakage situation, the trained wellbore breakthrough prediction model is called to test it. The scatter plot of prediction and actual results is shown in the figure Fig.14 , Fig.15 and Fig.16 shown.

[0127] The overall accuracy of the wellbore breakthrough prediction model is above 99%, and the injection volume is 1×10 4 , 2×10 4 , 5×10 4 , 1×10 5 m 3 / dThe prediction accuracy of the four situations is close to 100%.

[0128] 3. Analysis of dual CNN prediction results

[0129] Analysis of the prediction results of vertical cap rock rupture: Under the condition that the reservoir matrix permeability is 15md, after the dual CNN is used to learn the vertical leakage situation, the trained vertical cap rock rupture prediction model is called to test it. The scatter plot of the prediction and actual results is shown in the figure below. Fig.17 , Fig.18 and Fig.19 shown.

[0130] Injection volume 500m 3 / d simulation prediction accuracy is above 98%, and the overall accuracy of the vertical breakthrough prediction model is above 99%. 4 m 3 / dThe prediction accuracy of the two cases is close to 100%.

[0131] Analysis of lateral reservoir prediction results: Under the condition that the reservoir matrix permeability is 15md, after using dual CNN to learn the lateral leakage situation, the trained wellbore breakthrough prediction model is called to test it. The scatter plot of prediction and actual results is shown in the figure below. Fig. 20 , Fig.21 and Fig. 22 shown.

[0132] The overall accuracy of the wellbore breakthrough prediction model is above 99%, with an injection volume of 5000m 3 / d case prediction accuracy is close to 100%. The present invention uses a trained discrete response model (including a longitudinal cap rock rupture prediction model and a wellbore breakthrough prediction model) to evaluate the input data, calculate the accuracy, and verify the discrete response model. If the accuracy of the discrete response model exceeds a set threshold (e.g., 0.99), the model is saved as the best model. The best model will be called before each training starts. For the discrete model, a scatter plot of the actual value and predicted value of the discrete response is drawn to intuitively display the classification effect of the model. At the same time, a confusion matrix is ​​drawn to more intuitively display the number of true positive examples, false positive examples, true negative examples, and false negative examples predicted by the model. This work shows that strain and pressure monitoring may become a powerful tool for tracking CO2 movement, which can ensure safe injection operations and reduce the risk of CO2 leakage caused by geomechanical effects. This ML-based leakage monitoring method can automatically and quickly monitor strain and pressure anomalies indicating cap rock rupture, which can reduce manpower and time investment.

Claims

1. A method for predicting CO2 leakage after cap rock rupture and reservoir breakthrough in CO2 geological storage, characterized in that: The following steps are involved: S1. Establish a CO2 storage model. The CO2 storage model is a closed saline layer. The CO2 storage model includes a cover layer, a cap layer and a reservoir layer arranged in sequence from top to bottom, and the rock properties of the cover layer, the cap layer and the reservoir layer are set respectively; Injection wells are provided in the CO2 storage model; Setting a crack model in the caprock and setting the crack opening stress threshold; S2, simulate the storage process under different CO2 injection rates and different reservoir permeabilities; S3, extracting the strain and pressure data corresponding to different positions in the cap layer at different injection times during the CO2 injection process, and simultaneously extracting the gas saturation data at the corresponding positions of the cap layer and the reservoir closest to the extraction point of the cap layer, and combining the strain and pressure data corresponding to different positions of the cap layer extracted at the same injection time and the gas saturation data at the corresponding nearest position of the cap layer to form a first feature data set, and combining the strain and pressure data corresponding to different positions of the cap layer extracted at the same injection time and the gas saturation data at the corresponding nearest position of the reservoir to form a second feature data set; The first feature data set and the second feature data set are divided into a training set, a validation set and a test set respectively; S4, using the extracted first feature data set and the second feature data set to train the machine learning model respectively, with the strain and pressure data as input data; The gas saturation is used as the judgment standard. If the gas saturation is greater than 0, it means that carbon dioxide exists and the output is 1; if the gas saturation is equal to 0, it means that there is no carbon dioxide and the output is 0; S5. When the accuracy of the predicted value of the test set reaches the set threshold, a trained machine learning model is obtained. The cap rock rupture prediction model is obtained by training the first feature data set, and the wellbore breakthrough prediction model is obtained by training the second feature data set. S6. Using the strain and pressure data in the cap layer collected by the DAS monitoring system, the cap layer rupture prediction model and the wellbore breakthrough prediction model are used to predict whether the cap layer will rupture, whether the reservoir will break through, and their corresponding positions.

2. The method for predicting CO2 leakage after cap rock rupture and reservoir breakthrough of CO2 geological storage according to claim 1 is characterized by: In step S6, specifically, the DAS monitoring system in the monitoring well or on the outer wall of the injection well is used to collect strain and pressure data at various locations in the cap layer, and input into the cap layer rupture prediction model, and output 1, then the cap layer closest to the corresponding measurement position of the cap layer has carbon dioxide, and the cap layer has been ruptured; If the output is 0, there is no carbon dioxide in the cover layer closest to the corresponding measurement position of the cover layer, and the cover layer is not broken; The DAS monitoring system in the monitoring well or on the outer wall of the injection well is used to collect strain and pressure data at various locations in the caprock, and input them into the wellbore breakthrough prediction model. If the output is 1, then there is carbon dioxide in the reservoir closest to the corresponding measurement position of the caprock, and the reservoir has been breached; If the output is 0, there is no carbon dioxide in the reservoir closest to the corresponding measurement position of the caprock, and the reservoir has not been breached.

3. The method for predicting CO2 leakage after cap rock rupture and reservoir breakthrough of CO2 geological storage according to claim 1, characterized in that: In step S4, the machine learning model includes an MLP model; 1) Design the network structure: Design an MLP model consisting of fully connected layers that can extract features from input data and make predictions; The input layer of the model receives normalized strain and pressure data; The model contains two hidden layers. The first hidden layer has 64 neurons and the second hidden layer has 32 neurons. Both use the ReLU activation function to introduce nonlinearity. 2) Feature extraction: The first hidden layer uses 128 neurons to capture complex patterns in the input data; The second hidden layer uses 64 neurons to further refine the features; 3) Data output: For discrete responses, the output layer uses one neuron with a sigmoid activation function, which is suitable for binary classification problems and outputs 0 or 1.

4. The method for establishing a model for predicting cap rock rupture and reservoir breakthrough for CO2 geological storage according to claim 1, characterized in that: In step S4, the machine learning model includes a CNN model; 1) Design network structure: Design a CNN model consisting of a one-dimensional convolutional layer and a maximum pooling layer to extract spatiotemporal features from the input data; The input layer receives the reshaped strain and pressure data, and the data shape is (batch_size, 1, 9), where 1 represents one dimension; 9 represents the number of features, including 1 layer of time, 1 layer of permeability, 1 layer of injection volume, 3 layers of strain, and 3 layers of pressure data; 2) Feature extraction: The first convolutional layer uses 64 filters with a kernel size of 3. The ReLU activation function is used to introduce nonlinearity, and the output size is maintained by "same" padding; This is followed by a max pooling layer, which is used to reduce the spatial size of features and increase the abstraction ability of the model; After the flattening layer, the data is fed into two fully connected layers with 128 and 1 neurons respectively to continue extracting features and performing nonlinear transformations; The last fully connected layer uses the sigmoid activation function to meet the needs of the binary classification problem; 3) Model training: The discrete response model was trained for 100 epochs, and the early stopping method was set to prevent overfitting.

5. The method for establishing a model for predicting cap rock rupture and reservoir breakthrough of CO2 geological storage according to claim 1, characterized in that: In step S4, the machine learning model includes a dual CNN model; 1) Design network structure: Design a dual CNN model consisting of a one-dimensional convolutional layer and a maximum pooling layer to extract spatiotemporal features from the input data; The input layer receives the reshaped strain and pressure data, and the data shape is (batch_size, 1, 9), where 1 represents one dimension; 9 represents the number of features, including 1 layer of time, 1 layer of permeability, 1 layer of injection volume, 3 layers of strain, and 3 layers of pressure data; 2) Feature extraction: The first convolutional layer of the model uses 64 filters with a kernel size of 3. The ReLU activation function is used to introduce nonlinearity, and the output size is maintained by "same" padding; This is followed by a max pooling layer, which is used to reduce the spatial size of features and increase the abstraction ability of the model; Then comes the second convolutional layer, also using 64 filters, kernel size 3, ReLU activation function, and "same" padding; This is followed by a second max pooling layer; After the flattening layer, the data is fed into two fully connected layers with 128 and 1 neurons respectively to continue extracting features and performing nonlinear transformations; The last fully connected layer uses the sigmoid activation function to meet the needs of the binary classification problem; 3) Model training: The discrete response model was trained for 100 epochs, and the early stopping method was set to prevent overfitting.

6. The method for establishing a model for predicting cap rock rupture and reservoir breakthrough of CO2 geological storage according to claim 4 or 5, characterized in that: The corresponding requirement of the early stopping method is: if the validation set loss value does not improve within 50 epochs, then save the model with the lowest loss value.

7. The method for establishing a model for predicting cap rock rupture and reservoir breakthrough of CO2 geological storage according to claim 1, characterized in that: The method also includes step S7: using accuracy to evaluate the performance of different machine learning models, outputting a confusion matrix heat map to intuitively display the prediction status of different machine learning models in the test set, and selecting a machine learning model with higher accuracy.

8. The method for establishing a model for predicting cap rock rupture and reservoir breakthrough for CO2 geological storage according to claim 1, characterized in that: In step S1, the fracture model and the rock properties of the overburden, caprock and reservoir are set according to the actual terrain; The opening stress threshold of the fracture is set according to the rock type; The location of the injection well is selected according to the actual terrain, monitoring wells are set according to the situation, and a DAS monitoring system is installed in the monitoring wells.

9. The method for establishing a model for predicting cap rock rupture and reservoir breakthrough for CO2 geological storage according to claim 1, characterized in that: In step S1, the cap rock permeability model adopts the Barton-Bandis model.

10. The method for establishing a model for predicting cap rock rupture and reservoir breakthrough of CO2 geological storage according to claim 1, characterized in that: In step S1, the cap rock and reservoir are simulated using the Mohr-Coulomb criterion, and the overburden is simulated using the Drucker-Prager criterion.

Images