A method for predicting CO2 leakage after cap rock rupture, reservoir breakthrough for CO2 geological storage

Abstract

The present application belongs to the technical field of CO2 geological storage, and particularly relates to a method for predicting CO2 leakage after cap layer rupture and reservoir breakthrough in CO2 geological storage, wherein the present application establishes a CO2 storage model to simulate the storage process under different CO2 injection amounts, and establishes a correlation between the strain and pressure in the cap layer and the gas saturation in the cap layer and the gas saturation in the reservoir through machine learning, thereby proving the correlation between the CO2 injection amount and the strain and pressure response, indicating that the strain and pressure monitoring can track the CO2 injection process, provide protection for safe injection, and reduce the CO2 leakage risk caused by the geomechanical effect. The present application is convenient to operate, and after the machine learning is completed, the trained model is deployed to the production environment, and the discrimination of CO2 leakage can be completed only by processing the data monitored by the computer, so that the cap layer rupture and reservoir breakthrough problems can be found in advance and the corresponding positions can be determined.

Description

Technical Field

[0001] This invention relates to the field of CO2 geological storage technology, and is a method for predicting CO2 leakage after the rupture of the CO2 geological storage caprock and the breakthrough of the reservoir. Background Technology

[0002] CO2 is a greenhouse gas naturally present in the atmosphere and a byproduct of human activities. Although CO2's global warming potential is lower than that of other gases such as methane, its emissions make it a major contributor to global warming and ocean acidification. Therefore, it is essential to reduce CO2 emissions, and one important measure is CO2 sequestration. CO2 sequestration involves transporting CO2 gas to deep underground reservoirs through well injection or other means to reduce its emissions into the atmosphere, thus mitigating global warming.

[0003] Long-term CO2 sequestration may weaken the tensile strength of the caprock and reservoir, potentially leading to CO2 leakage. Large-scale CO2 leakage can have severe adverse consequences. These include: ground subsidence at the sequestration site, triggering microseismic events and disrupting the stability of the entire soil ecosystem; groundwater acidification, which, after interacting with soil water circulation, causes soil acidification and damages the soil environment; and CO2 overflowing onto the surface, failing to achieve the goal of reducing CO2 emissions. Therefore, monitoring the effectiveness of CO2 sequestration and providing early warning of CO2 leakage are crucial. Distributed acoustic sensing vertical seismic profiling (DAS) is a sensing system that uses optical fibers as the sensing element and signal medium. DAS VSP technology can monitor CO2 seismic reserves and holds great potential for reservoir monitoring.

[0004] Chinese patent document CN118425413A (202410254374.X) discloses a real-time monitoring method and system for carbon dioxide sequestration. It employs the deployment of monitoring sensors and detectors on the surface, in CO2 transport pipelines, and in reservoirs to acquire real-time data. The data is then compared with three preset thresholds to determine whether an alarm module is triggered. This patented monitoring method faces challenges such as high cost and difficult construction, and it also struggles with 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, proposing a CO2 storage reservoir model; based on the pressure values ​​and coordinates in the model, it is determined whether there is a leakage risk at the carbon dioxide storage site. This patent only simulates the reservoir and cannot fully consider the interaction and influence between the reservoir and the caprock in the CO2 geological storage system, neglecting the good sealing ability of the caprock on CO2 in the reservoir. Summary of the Invention

[0006] The main objective of this invention is to provide a method for predicting CO2 leakage after caprock rupture and reservoir breakthrough. It proposes a method for predicting CO2 leakage using machine learning, which can promptly detect caprock rupture and reservoir breakthrough phenomena and their specific locations, providing technical support for subsequent response measures.

[0007] The technical problem to be solved by this invention is achieved by the following technical solution: a method for predicting CO2 leakage after the rupture of the CO2 geological caprock and reservoir breakthrough, comprising the following steps:

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

[0009] The CO2 storage model includes an injection well.

[0010] A crack model is set within the cap layer, and the crack opening stress threshold is set.

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

[0012] S3. Extract strain and pressure data corresponding to different locations in the caprock at different injection times during CO2 injection, and simultaneously extract gas saturation data at the corresponding locations of the cover layer and reservoir closest to the extraction point in the caprock. Combine the strain and pressure data corresponding to different locations in the caprock extracted at the same injection time with the gas saturation data corresponding to the nearest location in the cover layer to form the first feature dataset, and combine the strain and pressure data corresponding to different locations in the caprock extracted at the same injection time with the gas saturation data corresponding to the nearest location in the reservoir to form the second feature dataset.

[0013] The first feature dataset and the second feature dataset are divided into training set, validation set and test set, respectively;

[0014] S4. Train machine learning models using the extracted first feature dataset and second feature dataset respectively, with strain and stress data as input data.

[0015] Using gas saturation as the criterion, if the gas saturation is greater than 0, it means that carbon dioxide is present, and the output is 1; if the gas saturation is equal to 0, it means that carbon dioxide is not present, and the output is 0.

[0016] S5. When the accuracy of the predicted values ​​of the test set reaches the set threshold, the threshold is set inside the machine learning model. The specific method is the existing technology, which will not be elaborated here. The trained machine learning model is obtained. The first feature dataset is used to train the caprock fracture prediction model, and the second feature dataset is used to train the wellbore breakthrough prediction model.

[0017] S6. Using the strain and pressure data collected by the DAS monitoring system, predict whether the caprock will rupture and whether the reservoir will break through, as well as their respective locations, through the caprock rupture prediction model and the wellbore breakthrough prediction model.

[0018] In a preferred embodiment of the present invention, in step S6, specifically, the strain and pressure data at various locations within the caprock are collected using a DAS monitoring system located inside the monitoring well or on the outer wall of the injection well. The data are then input into the caprock rupture prediction model. If the output is 1, then the caprock closest to the corresponding measurement location contains carbon dioxide, and the caprock has ruptured. If the output is 0, then the caprock closest to the corresponding measurement location does not contain carbon dioxide, and the caprock has not ruptured.

[0019] A DAS monitoring system is used to monitor the well or the outer wall of the injection well to collect strain and pressure data at various points in the caprock. The data is then input into the wellbore breakthrough prediction model. If the output is 1, the reservoir closest to the corresponding measurement location of the caprock contains carbon dioxide, and the reservoir has broken through. If the output is 0, the reservoir closest to the corresponding measurement location of the caprock does not contain carbon dioxide, and the reservoir has not broken through.

[0020] In actual prediction, the permeability data of the caprock can be input into the caprock fracture prediction model or wellbore breakthrough prediction model, and combined with the actual carbon dioxide injection volume and actual injection time for assistance, so as to improve the accuracy of the prediction.

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

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

[0023] The model's input layer receives standardized 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 non-linearity.

[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 built into 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 current technology.

[0027] 3) Data output: For discrete responses (classification tasks), the output layer uses a single neuron with a sigmoid activation function, suitable for binary classification problems, outputting 0 (no leakage) or 1 (leakage) as the predicted event result.

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

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

[0030] The input layer receives strain and pressure data after shape adjustment. The data shape is (batch_size, 1, 9), where 1 represents one dimension and 9 represents the number of features, including one layer of time, one layer of permeability, one layer of injection volume, three layers of strain, and three layers of pressure data. "Shape adjustment" is an existing technology that can be implemented with deep learning code. The purpose of shape adjustment is to match the code with the data; the above nine 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 non-linearity, and "same" padding is used to maintain the output size.

[0032] This is followed by a max pooling layer, which reduces the spatial size of features and increases the model's abstraction capabilities.

[0033] After passing through the flatten layer, the data is fed into two fully connected (dense) layers with 128 and 1 neurons respectively, to continue extracting features and performing non-linear transformations.

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

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

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

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

[0038] The input layer receives strain and pressure data after shape adjustment. The data shape is (batch_size, 1, 9), where 1 represents one dimension and 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 non-linearity, and "same" padding is used to maintain the output size.

[0040] This is followed by a max pooling layer, which reduces the spatial size of features and increases the model's abstraction capabilities.

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

[0042] Then comes the second max pooling layer;

[0043] After passing through the flatten layer, the data is fed into two fully connected (dense) layers with 128 and 1 neurons respectively, to continue extracting features and performing non-linear transformations.

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

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

[0046] In a preferred embodiment of this invention, the early stopping method requires that if the validation set loss value does not improve within 50 epochs, the model with the lowest previous loss value is saved. This 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 stopping is saved.

[0047] Preferably, the present invention further includes step S7: evaluating the performance of different machine learning models using accuracy, outputting a confusion matrix heatmap to intuitively display the prediction performance of different machine learning models on the test set, and selecting the machine learning model with higher accuracy.

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

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

[0050] The location of the injection well is selected based on the actual terrain, and monitoring wells are set up as needed. A DAS monitoring system is installed inside the monitoring well. The location of the monitoring well ensures that the corresponding prediction area of ​​the injection well is within the DAS monitoring range, taking into account factors such as signal-to-noise ratio. Alternatively, a DAS monitoring system can be installed directly outside the injection well.

[0051] Preferably, in step S1 of this invention, the caprock permeability model adopts the Barton-Bandis model. There are various permeability models, but the Barton-Bandis model is suitable for applications requiring consideration of 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 of this invention, both the caprock and reservoir are simulated using the Mohr-Coulomb criterion, while the caprock is simulated using the Drucker-Prager criterion. The Mohr-Coulomb criterion is applicable to cases considering compressive-shear failure, which can occur in both the caprock and reservoir; the Drucker-Prager criterion is applicable to a wider range of materials and stress states, and is suitable for caprocks.

[0053] The inventive concept of this invention is as follows: When carbon dioxide (CO2) is injected into deep saline aquifers, excessive overpressure may lead to caprock rupture and reservoir breach, resulting in CO2 leakage. Therefore, monitoring the geomechanical changes of the caprock and reservoir is of great significance for evaluating the stability of CO2 injection.

[0054] Research has revealed that if a carbon dioxide breakthrough occurs in the reservoir, the carbon dioxide will flow vertically along the outer side of the injection well, meaning it rises vertically along the well wall. Upon reaching the boundary between the caprock and reservoir, the caprock impedes the carbon dioxide, causing it to diffuse laterally along this boundary. This results in changes in pressure and strain at the caprock-reservoir boundary. Based on this phenomenon, this invention, through monitoring strain and pressure data within the caprock and utilizing machine learning, establishes a correlation between these data and the gas saturation of the reservoir below the caprock-reservoir boundary. This allows for the prediction of whether a carbon dioxide breakthrough will occur and the boundary location of subsequent carbon dioxide diffusion.

[0055] Furthermore, the caprock serves to seal the reservoir. However, cracks exist within the caprock. When the carbon dioxide pressure within the reservoir reaches a certain level, carbon dioxide enters the cracks, causing them to open and expand. If the caprock is penetrated by a crack, carbon dioxide will enter the cover layer, resulting in carbon dioxide leakage. Because carbon dioxide entering the cracks causes changes in pressure and strain within the caprock, the cracks typically extend along the path of least resistance. Therefore, the boundary between the caprock and cover layer, closest to the crack tip, usually ruptures first, leading to carbon dioxide leakage. Therefore, this invention, through monitoring strain and pressure data within the caprock and using machine learning, can establish a correlation between the strain and pressure data within the caprock and the gas saturation in the cover layer above the boundary between the caprock and cover layer, thereby predicting whether the caprock will rupture and the location of carbon dioxide leakage after rupture.

[0056] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention establishes a CO2 sequestration model to simulate the sequestration process under different CO2 injection volumes. It uses machine learning to establish a correlation between the strain and pressure in the caprock and the gas saturation in the caprock and the gas saturation in the reservoir, respectively, proving the correlation between CO2 injection volume and strain and pressure response. This shows that strain and pressure monitoring can track the CO2 injection process, provide a guarantee for safe injection, and reduce the risk of CO2 leakage caused by geomechanical effects.

[0057] 2. This invention proposes a novel data analysis method that uses machine learning to analyze DAS monitoring data, revealing anomalies in strain and pressure data in the caprock, thereby monitoring CO2 leakage from caprock rupture and reservoir breakthrough.

[0058] 3. Machine learning-based carbon dioxide leak monitoring methods can automatically and quickly monitor and indicate strain and pressure anomalies caused by caprock rupture and reservoir breakthrough, greatly reducing labor and time costs.

[0059] 4. This invention is easy to operate. After machine learning is completed, the trained model is deployed to the production environment. CO2 leakage can be identified simply by processing the data monitored by DAS through a computer. This enables early detection of caprock rupture and reservoir breakthrough problems, as well as determination of the corresponding locations, thereby enabling timely handling and ensuring production safety. Attached Figure Description

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

[0061] Figure 2 In this embodiment of the invention, the injection volume in the xz-6 plane is 2 × 10⁻⁶. 5 m 3 / d, Gas saturation distribution after 5 years of CO2 injection when reservoir permeability is 15mD;

[0062] Figure 3 Confusion matrix heatmaps for vertical MLP, CNN, and dual CNN test sets;

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

[0064] Figure 5 The injection volumes were 500, 1000, and 5000 m³, respectively. 3 Scatter plot of MLP longitudinal prediction and actual results at / d;

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

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

[0067] Figure 8 The injection volumes were 500, 1000, and 5000 m³, respectively. 3 Scatter plot of MLP horizontal prediction and actual results at / d;

[0068] Figure 9 The injection volume is 1×10 4 2×10 4 5×10 4 m 3 Scatter plot of MLP horizontal prediction and actual results at / d;

[0069] Figure 10 The injection volume is 1×10 5 1.5×10 5 2×10 5 m 3 Scatter plot of MLP horizontal prediction and actual results at / d;

[0070] Figure 11 The injection volumes were 500, 1000, and 5000 m³, respectively. 3 Scatter plot of CNN longitudinal prediction and actual results at / d;

[0071] Figure 12 The injection volume is 1×10 4 2×10 4 5×10 4 m 3 Scatter plot of CNN longitudinal prediction and actual results at / d;

[0072] Figure 13 The injection volume is 1×10 5 1.5×10 5 2×10 5 m 3 Scatter plot of CNN longitudinal prediction and actual results at / d;

[0073] Figure 14 The injection volumes were 500, 1000, and 5000 m³, respectively. 3 Scatter plot of CNN horizontal prediction and actual results at / d;

[0074] Figure 15 The injection volume is 1×10 4 2×10 4 5×10 4 m 3 Scatter plot of CNN horizontal prediction and actual results at / d;

[0075] Figure 16 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;

[0076] Figure 17 The injection volumes were 500, 1000, and 5000 m³, respectively. 3 Scatter plot of dual CNN longitudinal prediction and actual results at / d;

[0077] Figure 18 The injection volume is 1×10 4 2×10 4 5×10 4 m 3 Scatter plot of dual CNN longitudinal prediction and actual results at / d;

[0078] Figure 19 The injection volume is 1×10 5 1.5×10 5 2×10 5 m 3 Scatter plot of dual CNN longitudinal prediction and actual results at / d;

[0079] Figure 20 The injection volumes were 500, 1000, and 5000 m³, respectively. 3 Scatter plot of dual CNN lateral prediction and actual results at / d;

[0080] Figure 21 The injection volume is 1×10 4 2×10 4 5×10 4 m 3 Scatter plot of CNN horizontal prediction and actual results at / d;

[0081] Figure 22 The injection volume is 1×10 5 1.5×10 5 2×10 5 m 3 Scatter plot of CNN horizontal predictions versus actual results at / d. Detailed Implementation

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

[0083] This invention utilizes emerging, high-resolution, and low-cost distributed acoustic sensing (DAS) technology to develop a novel data analysis process for monitoring caprock leakage and reservoir breakthrough. Machine learning (ML) is applied to reveal leakage-related anomalies in DAS monitoring data during carbon dioxide injection. In this embodiment, a CO2 sequestration model is established, and synthetic datasets are obtained by varying the CO2 injection rate and reservoir permeability. The impact of these factors on caprock fracturing and reservoir breakthrough is then analyzed. This embodiment utilizes a multilayer perceptron (MLP), convolutional neural network (CNN), and dual convolutional neural network (Dual CNN) architecture to test the use of these synthetic strain and multi-level pressure data for rapid monitoring and prediction of CO2 leakage through the caprock and reservoir.

[0084] This invention is achieved through the following measures:

[0085] This invention utilizes a CO2 sequestration model established using CMG GEM. A monitoring well is installed 50m away from the injection well. By varying the CO2 injection rate and reservoir permeability, the strain and pressure response characteristics related to CO2 leakage are recorded by a low-frequency DAS within the monitoring well. A synthetic dataset is obtained, and machine learning (ML) is applied to reveal leakage-related anomalies in the DAS monitoring data during CO2 injection, enabling rapid monitoring and prediction of CO2 and thus providing early warning. The model comprises three parts from top to bottom: the caprock, the subsurface rock, and the reservoir. The reservoir is a porous rock layer that provides storage space for CO2; the caprock is a low-permeability rock layer above the reservoir, primarily preventing CO2 from migrating upwards to shallow aquifers or the surface; and the caprock provides an additional physical barrier to prevent CO2 from migrating upwards. Monitoring strain and pressure provides more comprehensive information about the subsurface environment. When CO2 leakage occurs, displacement changes lag behind stress changes, making strain monitoring more effective for early warning. The two monitoring datasets can be cross-validated, reducing the errors and uncertainties that may arise from relying on a single monitoring dataset. Based on strain and pressure values, combined with gas saturation, machine learning methods are used to determine whether carbon dioxide is leaking.

[0086] The model consists of 101 grids along the x-axis, each 10m long, totaling 1010m; 11 grids along the y-axis, each 10m long, totaling 110m; and 33 grids along the z-axis, each 10m long, totaling 330m. It comprises 10 caprock layers (100m total), 3 strata (30m total), and 20 reservoir layers (200m total). The model represents a closed saline aquifer where CO2 is injected into and stored, with no flow through the boundaries. The caprock is a brittle material with a high Young's modulus, and its permeability is modeled using the Barton-Bandis model. Initially, the permeability is essentially zero; once stress reaches a certain value, cracks open and rapidly propagate, leading to CO2 leakage.

[0087] This invention's storage model simulates 27 scenarios across 9 different daily CO2 injection rates and 3 different reservoir permeability levels. The strain and pressure responses generated in the reservoir vary depending on the amount of CO2 injected. A dataset is synthesized by extracting stress, strain, and gas saturation. This synthesized dataset is used as input data to develop and test the machine learning model. The ML testing algorithms include: Multilayer Perceptron (MLP), Convolutional Neural Network (CNN), and Dual CNN. This invention uses these ML methods to rapidly monitor CO2 leakage, considering a "discrete" response: a value of 1 indicates caprock rupture and CO2 leakage if CO2 is present in the caprock region {51, 6, 10}; a value of 1 indicates reservoir breakthrough if CO2 is present in the reservoir region {55, 6, 14}. Different CO2 injection rates and reservoir permeability result in different strains and pressures, and the caprock rupture and reservoir breakthrough scenarios also differ. Training the machine learning model with simulated data aims to find these correlations.

[0088] Specifically, such as Figure 1 As shown, the sequestration model consists of three parts from top to bottom: the overburden, the caprock, and the reservoir. It is a 3D model with 101, 11, and 33 grids in the x, y, and z directions, respectively. Each grid is represented by its xyz coordinates, where x is positive to the right, y is positive backward, and z is positive downward. Injection wells are drilled between {51, 6, 1} and {51, 6, 33}, and monitoring wells are drilled 40m to the right of the injection wells. Each grid in the model has a scale of 10m, totaling 1010m in the x direction, 110m in the y direction, and 330m in the z direction. The overburden has 10 grid layers, totaling 100m, the caprock has 3 grid layers, totaling 30m, and the reservoir has 20 grid layers, totaling 200m. The model represents a closed saline aquifer where CO2 is injected into and sequestered, with no flow passing through the boundaries.

[0089] The fractures in the caprock of the reservoir model are orthogonal in all directions at 10m intervals. The fracture model in this embodiment is based on the fracture model in the literature "ROCK FRACTURING DUE TO CO2 INJECTION". The remaining parts are not fractured. The compressibility of the caprock and reservoir differs, while other properties are the same. They are both simulated using the Mohr-Coulomb criterion, while the caprock is simulated using the Drucker-Prager criterion. The rock properties are shown in Table 1.

[0090] Table 1 Rock Properties

[0091]

[0092] The Barton-Bandis model was used for caprock permeability. The initial permeability was essentially zero. When the effective normal stress dropped below the crack opening stress threshold of 2000 kPa, cracks opened at the bottom of the caprock and rapidly expanded, extending in both horizontal and vertical directions until the caprock ruptured, resulting in CO2 leakage. The parameters of the Barton-Bandis model are shown in Table 2.

[0093] Table 2 Barton-Bandis model parameters

[0094]

[0095] This invention designs 27 schemes, with daily injection volumes of 500 m³ / d for reservoir permeability of 15 mD, 100 mD, and 500 mD, respectively. 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. See Table 3 for details.

[0096] Table 3 Scheme Parameters

[0097]

[0098] This invention performed calculations for the above 27 schemes, with an injection volume of 2×10 5 m 3 Taking a reservoir with a permeability of 15mD as an example, let's analyze this. Figure 2This is a gas saturation distribution map five years after CO2 injection. The map shows that after the caprock ruptured, CO2 entered the cover layer through the caprock, causing a CO2 reservoir breach. During this process, the strain and pressure of the caprock changed. This embodiment considers a method for measuring CO2 leakage: a "discrete" response. Any amount of CO2 reaching the cover layer area above the injection well location is considered a leak, with a value of 1, triggering an alarm. If CO2 is present in the reservoir {55, 6, 14}, the value is 1, indicating a reservoir breach and CO2 leakage, triggering an alarm. This invention extracts the strain, pressure, and gas saturation of the caprock {55, 6, 11}, {55, 6, 12}, and {55, 6, 13}, the gas saturation of the overburden {51, 6, 10}, and the gas saturation of the reservoir {55, 6, 14} from various schemes. This provides a set of several strain measurements and several pressure measurements. These data are used as a synthetic dataset to train an ML model, with the strain and pressure of the caprock {55, 6, 11}, {55, 6, 12}, and {55, 6, 13} as input data and the gas saturation of the caprock {55, 6, 11}, {55, 6, 12}, and {55, 6, 13} as reference values.

[0099] This invention collects and organizes data related to carbon dioxide sequestration leakage, extracting pressure and strain data from CMG simulations. Simultaneously, it ensures data reliability by removing outliers, missing values, and duplicate data to guarantee quality and consistency. Input and response data are randomly divided into training, validation, and test sets. The training set comprises 64%, the validation set 16%, and the test set 20%. The test set is used to evaluate the model's generalization ability during model selection and final evaluation, while the validation set is used for parameter tuning during model training.

[0100] This invention employs multilayer perceptron (MLP), convolutional neural network (CNN), and dual convolutional neural network (DualCNN) models to handle discrete task prediction of vertical caprock rupture leakage and lateral reservoir leakage. The detailed training process for each model is as follows:

[0101] 1. Multilayer Perceptron (MLP) Training

[0102] 1) Network Structure Design: We designed an MLP model consisting of fully connected (Dense) layers, which can extract features from the input data and make predictions. The input layer of the model receives standardized strain and stress data. The model contains two hidden layers: the first hidden layer has 64 neurons, and the second hidden layer has 32 neurons, both using 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 a single neuron with a sigmoid activation function, suitable for binary classification problems, outputting 0 (no leakage) or 1 (leakage) as the predicted event result.

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

[0106] 1) Network Structure Design: A CNN model containing one-dimensional convolutional layers (Conv1D) and max-pooling layers (MaxPooling1D) was designed to extract spatiotemporal features from the input data. The input layer receives strain and pressure data with adjusted shapes, in the form (batch_size, 1, 9), where 1 represents one dimension and 9 represents the number of features, including one layer of time data, one layer of permeability data, one layer of injection volume data, three layers of strain data, and three 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 non-linearity, and "same" padding is used to maintain the output size. This is followed by a max-pooling layer to reduce the spatial size of the features and increase the model's abstraction ability. After a flattening layer, the data is fed into two fully connected (dense) layers with 128 and 1 neuron respectively, continuing to extract features and perform non-linear transformations. The final fully connected layer uses a sigmoid activation function to suit the needs of binary classification problems.

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

[0109] 3. Training a Dual Convolutional Neural Network (DCNN)

[0110] 1) Network Structure Design: A dual CNN model containing one-dimensional convolutional layers (Conv1D) and max pooling layers (MaxPooling1D) was designed to extract spatiotemporal features from the input data. The input layer receives strain and pressure data after shape adjustment, with a data shape of (batch_size, 1, 9), where 9 represents the number of features (1 layer for time, 1 layer for permeability, 1 layer for injection volume, 3 layers for strain, and 3 layers for pressure).

[0111] 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 non-linearity, and "same" padding is used to maintain the output size. This is followed by a max-pooling layer to reduce the spatial size of the features and increase the model's abstraction ability. Then comes the second convolutional layer, also using 64 filters with a kernel size of 3, ReLU activation function, and "same" padding. This is followed by a second max-pooling layer. After a flattening layer, the data is fed into two fully connected (dense) layers with 128 and 1 neuron respectively, continuing to extract features and perform non-linear transformations. The final fully connected layer uses a sigmoid activation function to meet the requirements of binary classification problems.

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

[0113] This invention uses a pre-trained discrete response model to predict data on a pre-defined test set. This process is completely independent of the training process and aims to test the model's performance and its ability to predict unknown data. Performance is evaluated using accuracy, and the final output is a confusion matrix heatmap that visually displays the model's prediction performance on the test set.

[0114] Longitudinal caprock rupture and leakage: For longitudinal caprock rupture and leakage, the performance of prediction models trained with MLP, CNN, and dual CNN is not significantly different. Specifically, MLP achieves an accuracy of 99.9%, CNN 99.1%, and dual CNN 99.4%. The confusion matrices of each model are shown below. Figure 3 As shown, the three machine learning models can make accurate predictions for most situations.

[0115] Lateral wellbore breakthrough: For lateral wellbore breakthroughs, the performance of wellbore breakthrough prediction models trained with MLP, CNN, and dual CNN is not significantly different, but their performance is slightly better than that of the longitudinal models. Specifically, the accuracy of MLP is 99.8%, CNN is 99.4%, and dual CNN is 99.0%. The confusion matrices of each model are as follows: Figure 4 As shown, the three models can accurately predict most situations.

[0116] This invention provides leak warning and forecasting for injection schemes under different carbon dioxide injection volumes and caprock permeability. The interpretation and application of the warning and forecasting results can provide a scientific basis for practical operations, ensuring the safety and effectiveness of the injection scheme. The two-directional discrete prediction model primarily targets potential leak events during carbon dioxide injection, outputting 0 (no leak) or 1 (leak). By learning different characteristic parameters, the model can predict whether a leak exists. In the discrete scatter plot, the horizontal axis represents the simulated injection time, and the vertical axis represents the prediction result.

[0117] This embodiment takes a reservoir matrix permeability of 15md as an example. For different injection volumes, three different machine learning models, namely multilayer perceptron (MLP), convolutional neural network (CNN), and dual convolutional neural network (dual CNN), were used to predict and analyze vertical caprock fracture leakage and lateral reservoir breakthrough leakage. The following is a comprehensive analysis of these prediction results.

[0118] 1. Analysis of MLP Prediction Results

[0119] Analysis of Longitudinal Caprock Failure Prediction Results: Under the condition of a reservoir matrix permeability of 15 md, the longitudinal leakage situation was learned using MLP, and then the trained longitudinal caprock failure prediction model was called to verify it. The scatter plot of the prediction and actual results is shown below. Figure 5 , Figure 6 and Figure 7 As shown.

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

[0121] Analysis of lateral reservoir prediction results: Under the condition of a reservoir matrix permeability of 15 md, the lateral leakage situation was learned using MLP, and then the trained wellbore breakthrough prediction model was called to verify it. The scatter plot of the prediction and actual results is shown in the figure. Figure 8 , Figure 9 and Figure 10 As shown.

[0122] The overall accuracy of the wellbore breakthrough prediction model is above 99%, with an injection volume of 1×10⁻⁶. 4 1.5×10 5 m 3 The prediction accuracy for both cases is close to 100%.

[0123] 2. Analysis of CNN Prediction Results

[0124] Analysis of Vertical Caprock Failure Prediction Results: Under the condition of a reservoir matrix permeability of 15 md, the vertical leakage situation was learned using a CNN, and then the trained vertical caprock failure prediction model was used to verify it. The scatter plot of the prediction and actual results is shown below. Figure 11 , Figure 12 and Figure 13 As shown.

[0125] The prediction results of the CNN model are largely similar to those of the MLP model. The overall accuracy of the longitudinal prediction model is above 99%, with an injection size of 5000 and 1×10⁻⁶. 4 5×10 4 1.5×10 5 m 3 The prediction accuracy for the four scenarios is close to 100%.

[0126] Analysis of lateral reservoir prediction results: Under the condition of a reservoir matrix permeability of 15 md, a CNN was used to learn about lateral leakage, and then a pre-trained wellbore breakthrough prediction model was used to verify the results. The scatter plot of the prediction and actual results is shown below. Figure 14 , Figure 15 and Figure 16 As shown.

[0127] The overall accuracy of the wellbore breakthrough prediction model is above 99%, with an injection volume of 1×10⁻⁶. 4 2×10 4 5×10 4 1×10 5 m 3 The prediction accuracy for the four scenarios is close to 100%.

[0128] 3. Analysis of Dual CNN Prediction Results

[0129] Analysis of Vertical Caprock Failure Prediction Results: Under the condition of a reservoir matrix permeability of 15 md, a dual CNN was used to learn about vertical leakage, and then the trained vertical caprock failure prediction model was used to verify it. The scatter plot of the prediction and actual results is shown below. Figure 17 , Figure 18 and Figure 19 As shown.

[0130] Except for an injection volume of 500m 3 The simulation accuracy for / d is above 98%, and the overall accuracy of the vertical breakthrough prediction model is above 99%, with injection volumes of 1000 and 1×10. 4 m 3 The prediction accuracy for both cases is close to 100%.

[0131] Analysis of lateral reservoir prediction results: Under the condition of a reservoir matrix permeability of 15 md, a dual CNN was used to learn about lateral leakage, and then a pre-trained wellbore breakthrough prediction model was called to verify it. The scatter plot of the prediction and actual results is shown below. Figure 20 , Figure 21 and Figure 22 As shown.

[0132] The overall accuracy of the wellbore breakthrough prediction model is over 99%, with an injection volume of 5000m³. 3 The prediction accuracy for the / d scenario is close to 100%. This invention uses pre-trained discrete response models (including a vertical caprock fracture prediction model and a wellbore breakthrough prediction model) to evaluate and calculate the accuracy of the input data, thus validating the discrete response models. 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 is called before each training session. A scatter plot of the actual and predicted values ​​of the discrete response is drawn for the discrete model to visually demonstrate the model's classification effect. A confusion matrix is ​​also drawn to more intuitively show the number of true positives, false positives, true negatives, and false negatives predicted by the model. This work demonstrates that strain and pressure monitoring can be a powerful tool for tracking CO2 movement, ensuring safe injection operations and mitigating 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 caprock fracture, reducing the investment of manpower and time.

Claims

1. A method of predicting CO2 leakage after cap failure, reservoir breakthrough for CO2 geological storage, characterized in that, The method comprises the following steps: S1, establishing a CO2 storage model, the CO2 storage model being a closed saline aquifer, the CO2 storage model comprising, from top to bottom, a caprock, a caprock and a reservoir, and the rock properties of the caprock, the caprock and the reservoir being set respectively; An injection well is arranged in the CO2 storage model; A fracture model is arranged in the caprock, and a fracture opening stress threshold is set; S2, simulating the storage process under different CO2 injection amounts and different reservoir permeabilities; S3, extracting strain and pressure data corresponding to different positions in the caprock at different injection times during CO2 injection, and simultaneously extracting gas saturation data at positions in the caprock and the reservoir closest to the extraction points, respectively, to form a first feature data set and 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, training a machine learning model using the extracted first feature data set and second feature data set, with strain and pressure data as input data; Taking gas saturation as the judgment standard, if the gas saturation is greater than 0, that is, there is carbon dioxide, output 1; if the gas saturation is equal to 0, that is, there is no carbon dioxide, output 0; S5, when the accuracy of the test set prediction value reaches a set threshold, a trained machine learning model is obtained, a caprock rupture prediction model is trained from the first feature data set, and a wellbore breakthrough prediction model is trained from the second feature data set; S6, using the strain and pressure data collected by the DAS monitoring system in the caprock to predict whether the caprock is ruptured and the reservoir is broken through and the corresponding positions respectively through the caprock rupture prediction model and the wellbore breakthrough prediction model.

2. The method of predicting CO2 leakage after cap failure, reservoir breakthrough for CO2 geologic storage according to claim 1, characterized in that: In step S6, specifically, the DAS monitoring system on the inner wall of the monitoring well or the outer wall of the injection well is used to collect strain and pressure data at various positions in the caprock, input the caprock rupture prediction model, and output 1, indicating that the caprock closest to the corresponding measurement position in the caprock contains carbon dioxide, and the caprock has been ruptured. The output is 0, indicating that the caprock closest to the corresponding measurement position in the caprock does not contain carbon dioxide, and the caprock has not been ruptured. The DAS monitoring system on the inner wall of the monitoring well or the outer wall of the injection well is used to collect strain and pressure data at various positions in the caprock, input the wellbore breakthrough prediction model, and output 1, indicating that the reservoir closest to the corresponding measurement position in the caprock contains carbon dioxide, and the reservoir has been broken through. The output is 0, indicating that the reservoir closest to the corresponding measurement position in the caprock does not contain carbon dioxide, and the reservoir has not been broken through.

3. The method of predicting CO2 leakage after cap failure, reservoir breakthrough for CO2 geologic storage according to claim 1, characterized in that: In step S4, the machine learning model comprises an MLP model; 1) Design network structure: design an MLP model composed of fully connected layers, 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 using 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 sigmoid activation function, suitable for binary classification problems, outputting 0 or 1.

4. The method of claim 1, wherein the method further comprises: In step S4, the machine learning model includes a CNN model; 1) Network structure design: A CNN model containing one-dimensional convolutional layers and max-pooling layers is designed to extract spatio-temporal features from input data; The input layer receives the adjusted strain and pressure data, with a data shape of (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, and ReLU activation function is used to introduce nonlinearity, and "same" padding is used to maintain the output size; This is followed by a max-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 with 128 and 1 neurons respectively, to continue extracting features and performing nonlinear transformation; The last fully connected layer uses a sigmoid activation function to adapt to the needs of binary classification problems; 3) Model training: The discrete response model is trained for 100 epochs, and early stopping is set to prevent overfitting.

5. The method of claim 1, wherein: In step S4, the machine learning model includes a double CNN model; 1) Network structure design: A double CNN model containing one-dimensional convolutional layers and max-pooling layers is designed to extract spatio-temporal features from input data; The input layer receives the adjusted strain and pressure data, with a data shape of (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, and ReLU activation function is used to introduce nonlinearity, and "same" padding is used to maintain the output size; This is followed by a max-pooling layer to reduce the spatial size of the features and increase the abstraction ability of the model; Then comes the second convolutional layer, also using 64 filters with a kernel size of 3, ReLU activation function, and "same" padding; This is followed by a second max-pooling layer; After the flattening layer, the data is sent to two fully connected layers with 128 and 1 neurons respectively, to continue extracting features and performing nonlinear transformation; The last fully connected layer uses a sigmoid activation function to adapt to the needs of binary classification problems; 3) Model training: The discrete response model is trained for 100 epochs, and early stopping is set to prevent overfitting.

6. The method of claim 4 or 5, wherein the method further comprises: The early stopping method correspondingly requires: if the validation set loss value is not improved within 50 epochs, the model with the lowest loss value before saving is saved.

7. The method of claim 1, wherein the method further comprises: The method further comprises the following step S7: evaluating the performance of different machine learning models by using accuracy, outputting a confusion matrix heat map to intuitively display the prediction of different machine learning models on the test set, and selecting a machine learning model with higher accuracy.

8. The method for predicting CO2 geologic storage cap rock failure, reservoir breakthrough modeling of claim 1, wherein: In step S1, the crack model and the rock properties of the overburden layer, the cap rock and the reservoir are set according to the actual terrain; The opening stress threshold of the crack is set according to the rock type; The position of the injection well is selected according to the actual terrain, and a monitoring well is set according to the situation, and a DAS monitoring system is arranged in the monitoring well.

9. The method for predicting CO2 geologic storage cap rock failure, reservoir breakthrough modeling of claim 1, wherein: In step S1, the cap rock permeability model adopts the Barton-Bandis model.

10. The method for predicting CO2 geologic storage cap rock failure, reservoir breakthrough modeling of claim 1, wherein: In step S1, the cap rock and the reservoir are simulated by using the Mohr-Coulomb criterion, and the overburden layer is simulated by using the Drucker-Prager criterion.

Images