Shale gas well casing deformation risk dynamic intelligent prediction method driven by multi-source data

Abstract

The invention relates to the technical field of shale gas intelligent wellbores, and particularly discloses a multi-source data-driven shale gas well casing deformation risk dynamic intelligent prediction method, which comprises the following steps of: collecting block data of a built shale gas well to construct an original data set; carrying out abnormal value processing, missing value filling, data normalization, one-hot coding and data enhancement on the original data set in sequence; a threshold coefficient is set for data screening, and a final sample data set is obtained; establishing a casing deformation intelligent prediction model based on the CBT model and the BLSTM model; the intelligent casing deformation prediction model is adopted to predict the shale gas well casing deformation risk; and correcting seismic attribute data by using a newly monitored micro-seismic event, and iterating the casing deformation intelligent prediction model by adopting an incremental learning method in combination with well drilling and completion data and casing deformation data of a new well. The dynamic and intelligent prediction method for the deformation of the casing pipe is constructed, and real-time and accurate prediction of the deformation of the fracturing casing pipe is achieved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent wellbore technology, specifically relating to a dynamic intelligent prediction method for casing deformation risk in shale gas wells driven by multi-source data. Background Technology

[0002] As a core barrier ensuring wellbore integrity, the safe operation of oil and gas well casing is crucial for oil and gas development. However, complex underground environments can easily induce casing deformation and failure. Casing deformation refers to the phenomenon where uneven stress on the wellbore caused by elastic and plastic deformation of the formation, activation of faults, changes in formation stress, and formation slippage, leading to casing deformation, bending, and damage. Such deformation poses significant risks: reduced pipe diameter hinders tool insertion, affecting production capacity; severe deformation damages seals, causing leaks and pollution; in extreme cases, it can crush the tubing string, rendering the entire well unusable and resulting in substantial safety and economic losses.

[0003] Current casing deformation prediction methods are mainly divided into three categories: First, methods based on mechanical analytical models, which derive analytical solutions by combining design loads with elastoplastic theory and simplified mechanical models. Second, numerical simulation methods, which establish finite element, finite difference, or discrete element models of the casing-cement sheath-formation system to simulate the actual load history. However, the mechanical analytical model method neglects the influence of formation heterogeneity, dynamic loads, and the coupling effect of the casing-cement sheath-formation, making it difficult to predict deformation under complex working conditions. Third, intelligent prediction methods, which currently mainly rely on seismic attribute data and fracturing-geological data, using machine learning algorithms to predict casing deformation under hydraulic fracturing conditions, but do not consider the combined use of seismic and engineering data, and the prediction effect is poor. Therefore, we propose a dynamic intelligent prediction method for shale gas well casing deformation risk driven by multi-source data. Summary of the Invention

[0004] The technical problem to be solved by this invention is to address the shortcomings of the prior art by providing a dynamic intelligent prediction method for casing deformation risk in shale gas wells driven by multi-source data. This method can accurately predict the location of casing deformation caused by changes in geostress and fault activation in shale gas wells and predict the casing deformation risk.

[0005] The technical solution adopted in this invention is: a multi-source data-driven dynamic intelligent prediction method for shale gas well casing deformation risk, comprising the following steps:

[0006] Step 1: Collect casing deformation data, drilling and completion data of deformed and undeformed casing sections from the established shale gas well blocks, and seismic attribute data obtained by monitoring microseismic events caused by oil and gas field fracturing and geological tectonic movements to construct the original dataset;

[0007] Step 2: Perform outlier processing on each data sample in the original dataset using the Isolation Forest algorithm and the Local Outlier Factor algorithm to obtain the initial dataset;

[0008] Step 3: Use the KNN algorithm to fill in the missing values ​​of each data sample in the initial dataset. After filling, normalize the data of each data sample to obtain the first-level sample dataset.

[0009] Step 4: Divide each data sample in the primary sample dataset into risk levels according to the amount of casing deformation, and classify the data samples of each risk level using one-hot encoding. After classification, the secondary sample dataset is obtained.

[0010] Step 5: For risk categories with a small number of data samples after classification in the secondary sample dataset, data augmentation is used to expand the data samples of the risk category to obtain the tertiary sample dataset;

[0011] Step 6: Filter the three-level sample dataset, set coefficient thresholds, calculate the maximum mutual information coefficient between each parameter and the sleeve deformation in each data sample, remove parameters whose maximum mutual information coefficient is less than the coefficient threshold, and obtain the final sample dataset.

[0012] Step 7: Randomly divide the final sample dataset into a training set and a test set at an 8:2 ratio, and establish a casing deformation intelligent prediction model based on the CBT model and the BLSTM model. The casing deformation intelligent prediction model includes an input layer, a CBT unit, a BLSTM unit, and an output layer. Input the training set into the casing deformation intelligent prediction model for training, and input the test set into the trained casing deformation intelligent prediction model for verification. After successful verification, the final casing deformation intelligent prediction model is obtained.

[0013] Step 8: Use the final intelligent prediction model for casing deformation to predict the risk of casing deformation in shale gas wells;

[0014] Step 9: By monitoring microseismic events caused by each fracturing of the new well and microseismic events caused by geological tectonic movements, the seismic attribute data of the affected strata area is obtained, which is the seismic attribute modification data. Combined with the drilling and completion data of the casing deformation section and the undeformed section of the new well and the casing deformation data of the new well, the intelligent prediction model of casing deformation is continuously iterated by the incremental learning method, so as to realize the dynamic prediction of the casing deformation risk of old and new shale gas wells in the region.

[0015] Preferably, the drilling and completion data for the deformed and undeformed sections of the casing in step 1 include: well number, geodetic coordinates, wellbore trajectory, fracturing fluid discharge, total fluid volume, pump pressure, number of clusters, length of the fracturing section, perforation density, and number of perforations.

[0016] The seismic attribute data includes ant volume, curvature volume, coherence volume, maximum likelihood volume, and chaotic volume;

[0017] The casing deformation data includes the location and amount of casing deformation.

[0018] Preferably, in step 2, the isolated forest algorithm performs outlier processing to identify and handle global outliers, including the following steps:

[0019] Step A1: Data sampling, randomly extracting samples from the original data to construct a single decision tree;

[0020] Step A2: Randomly select a feature, randomly select a split point within the value range of the feature, divide the sample into left and right branches according to the split point, and recursively repeat this process until a branch contains only 1 sample or the maximum depth of the tree is reached.

[0021] Step A3: Repeat steps A and B to construct multiple decision trees, forming an isolated forest;

[0022] Step A4: For each data sample, calculate its average path length across all decision trees, convert the average path length into anomaly score, and determine whether the data sample is an anomaly.

[0023] Preferably, the local outlier factor algorithm in step 2 performs outlier processing to identify and handle local outliers, including the following steps:

[0024] Step B1: Find the k nearest neighbors of each sample x and form the k nearest neighbor set of x;

[0025] Step B2: Calculate the reachability distance. For each nearest neighbor y of x, the reachability distance from x to y is the maximum value of the k nearest neighbor distance of y and the actual distance between x and y, to avoid the influence of extreme values.

[0026] Step B3: Calculate the local reachability density. The local reachability density of x is the reciprocal of the average reachability distance between x and all its k nearest neighbors. The larger the value of the reciprocal of the average reachability distance, the higher the local reachability density.

[0027] Step B4: Calculate the local outlier value: The local outlier value of x is the ratio of the average local reachability of its k nearest neighbors to the local reachability of x itself, reflecting the degree of density deviation of x relative to its neighbors; the larger the local outlier value, the greater the probability that the sample is an outlier.

[0028] Preferably, the learning rate, the number of neurons in the hidden layer, and the number of iterations in the intelligent prediction model for casing deformation described in step 7 are optimized using an adaptive particle swarm optimization algorithm.

[0029] Preferably, the incremental learning method in step 9 is a sample replay-based method, including the following steps:

[0030] Step C1: Obtain old data samples by randomly selecting representative data samples for each category from the data storage area of ​​the trained intelligent prediction model for casing deformation.

[0031] Step C2: Obtain new data samples based on the new seismic attribute parameters obtained in Step 9, the casing deformation section of the new well, the drilling and completion data of the undeformed section, and the casing deformation data of the new well;

[0032] Step C3: Obtain the merged data sample, randomly sample several old data samples, and merge them with all the new data samples into a training batch of data samples;

[0033] Step C4: Retrain the intelligent prediction model for casing deformation using the merged data samples;

[0034] Step C5: Update the data storage area by randomly selecting representative data samples from each category from the new data samples and storing them in the data storage area;

[0035] Step C6: Evaluate the performance of the intelligent prediction model for casing deformation on new data samples, old data samples, and merged data samples respectively. If the performance is poor, repeat steps C1 to C5 until the performance of the merged data samples is satisfactory.

[0036] Step C7: When the intelligent prediction model for casing deformation acquires new data samples, repeat steps C1 to C6 to achieve continuous iteration of the intelligent prediction model for casing deformation.

[0037] The beneficial effects of this invention are as follows: Based on multi-source data of seismic attribute parameters and drilling and completion engineering parameters, this invention constructs an intelligent prediction model for casing deformation through machine learning. This solves the problem that traditional methods cannot predict the risk of casing deformation in blocks before drilling and fracturing, as well as the dynamic prediction of casing deformation under formation disturbance caused by new well fracturing and microseismic events. Thus, it achieves dynamic prediction of one well for each well fracturing within the development area. Furthermore, by constructing the relationship between multi-source data and casing deformation risk through the intelligent prediction model, the invention provides a more comprehensive assessment of casing deformation risk from the perspective of integrated geological engineering, which is more in line with statistical principles, reduces the workload of calculation in the prediction process, and has extremely high accuracy and low cost, while also making the calculation results more reliable. Attached Figure Description

[0038] Figure 1 This is a flowchart of the present invention;

[0039] Figure 2 This is a block diagram of the intelligent prediction model for casing deformation of the present invention;

[0040] Figure 3 This is a schematic diagram illustrating the risk classification of sleeve deformation according to the present invention;

[0041] Figure 4 This is a schematic diagram illustrating the dynamic prediction of the intelligent prediction model for casing deformation according to the present invention. Detailed Implementation

[0042] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0043] Example

[0044] like Figure 1 , Figure 2 , Figure 3 and Figure 4 As shown in this embodiment, the multi-source data-driven dynamic intelligent prediction method for shale gas well casing deformation risk includes the following steps:

[0045] Step 1: Collect drilling and completion data, seismic attribute data, and casing deformation data of the existing shale gas well blocks to construct the original dataset;

[0046] The drilling and completion data for the deformed and undeformed sections of the casing include: well number, geodetic coordinates, wellbore trajectory, fracturing fluid discharge, total fluid volume, pump pressure, number of clusters, length of the fracturing section, perforation density, and number of perforations.

[0047] The seismic attribute data includes ant volume, curvature volume, coherence volume, maximum likelihood volume, and chaotic volume;

[0048] The casing deformation data includes the location and amount of casing deformation.

[0049] Step 2: Perform outlier processing on each data sample in the original dataset using the Isolation Forest algorithm and the Local Outlier Factor algorithm to obtain the initial dataset;

[0050] The isolated forest algorithm handles outliers by identifying and processing global outliers. Outliers are isolated points in the feature space, far from most normal samples. The principle behind outlier identification is to segment the data using randomly generated decision trees. Outliers reach the leaf nodes of the trees faster (shorter path length). Anomaly scores are calculated by voting on the path lengths of multiple decision trees; higher scores indicate a greater probability of being an outlier. The algorithm includes the following steps:

[0051] Step A1: Data sampling, randomly extract a small number of data samples from the original data to construct a single decision tree;

[0052] Step A2: Randomly select a feature, randomly select a split point within the value range of the feature, divide the sample into left and right branches according to the split point, and recursively repeat this process until a branch contains only one sample or the maximum depth of the tree is reached.

[0053] Step A3: Repeat steps A and B to construct multiple decision trees, forming an isolated forest;

[0054] Step A4: For each data sample, calculate its average path length across all decision trees, convert the average path length into anomaly score, and determine whether the data sample is an anomaly.

[0055] The Local Outlier Factor algorithm identifies and processes outliers by recognizing and handling local outliers. Outliers have a local density significantly lower than their neighbors. The principle behind outlier identification is to measure the density deviation between a sample point and its neighbors. If a point's density is significantly lower than its neighbors', then that point is likely an outlier. It can detect points that are not globally abnormal but are abnormal in a local context. Specifically, it includes the following steps:

[0056] Step B1: Find the k nearest neighbors of each data sample x and form the k nearest neighbor set of x;

[0057] Step B2: Calculate the reachability distance. For each nearest neighbor y of x, the reachability distance from x to y is the maximum value of the k nearest neighbor distance of y and the actual distance between x and y, to avoid the influence of extreme values.

[0058] Step B3: Calculate the local reachability density. The local reachability density of x is the reciprocal of the average reachability distance between x and all its k nearest neighbors. The larger the value of the reciprocal of the average reachability distance, the higher the local reachability density.

[0059] Step B4: Calculate the local outlier factor: The local outlier factor of x is the ratio of the average local reachability of its k nearest neighbors to the local reachability of x itself, reflecting the degree of density deviation of x relative to its neighbors; the larger the local outlier factor, the greater the probability that the data sample is an outlier.

[0060] The Isolation Forest algorithm mainly identifies outliers globally, while the Local Outlier Factor algorithm mainly identifies outliers locally. By combining the Isolation Forest and Local Outlier Factor methods to identify outliers, each data sample is subjected to outlier identification using both algorithms. If either the Isolation Forest or Local Outlier Factor algorithm identifies an outlier, it is considered an outlier. The data samples identified as outliers are then removed to obtain the initial dataset.

[0061] Step 3: Apply the KNN algorithm to each data sample in the initial dataset to impute missing values. The specific imputation process for missing values ​​is as follows:

[0062] Set the value of K in the KNN algorithm, train the original terrane variant dataset based on the KNN algorithm, and obtain the K nearest neighbor samples of each original terrane variant data sample, that is, obtain the trained KNN model.

[0063] Each data sample with missing parameter values ​​in the primary data sample is input into the trained KNN model to obtain K nearest neighbor samples for each data sample. The missing parameter value of the input parameter in each data sample is the average value of the corresponding input parameter in these K nearest neighbor samples.

[0064] For example: the s-th primary data sample containing T valid input parameters. = ( , , ,…, The component corresponding to the t-th input parameter in ) Missing, i.e. Given the missing parameter value of the t-th input parameter, the first-level data sample is obtained using the KNN algorithm. The K nearest neighbors of the sample: , , , ..., ,Right now = ( , , ,…, (k∈[1,K]), the missing parameter value of the t-th input parameter is calculated using the average of the t-th input parameter from the K nearest neighbor samples. :

[0065]

[0066] That is, to calculate the missing parameter value of the t-th input parameter in the s-th level data sample containing T valid input parameters;

[0067] To eliminate the difference in magnitude between different parameter data, the data samples after filling missing values ​​are normalized to obtain a primary sample dataset; specifically, the data is scaled to the interval [0,1], using the following formula:

[0068] In the formula, For the original data, For the normalized data, Indicates parameters The maximum value in, Indicates parameters The minimum value in;

[0069] Step 4: Classify the risk level of each data sample in the primary sample dataset according to the amount of casing deformation. Under different standard casing inner diameters, classify the casing deformation risk into three levels according to the amount of casing deformation s: low risk, medium risk and high risk.

[0070] Low risk: The transformer will not affect normal construction.

[0071] Medium risk: The casing affects normal construction and requires multiple replacements of small bridge plugs or special tools to complete the fracturing.

[0072] High risk: Severe casing deformation will affect normal construction and production, requiring well repair or pressure control, or abandonment of sections;

[0073] One-hot encoding was used to encode the three datasets of low risk, medium risk and high risk respectively, transforming them into low risk (1, 0, 0), medium risk (0, 1, 0) and high risk (0, 0, 1). Each risk level was labeled as a three-dimensional vector, which effectively transformed the category information into a numerical format that the machine learning model can understand, while ensuring the independence and equality between categories.

[0074] One-hot encoding was used to classify the data samples for each risk level, and a secondary sample dataset was obtained after classification.

[0075] Step 5: For risk categories with a small number of data samples after classification in the secondary sample dataset, data augmentation is used to expand the data samples of the risk category to obtain the tertiary sample dataset;

[0076] Data augmentation utilizes synthetic minority oversampling techniques to leverage the local similarity of minority class samples. Between any two similar minority class samples, new minority class samples are generated through linear interpolation. These new samples retain the characteristic distribution of the minority class while expanding the sample size, allowing the model to learn effective features of the minority class. Specifically, it includes the following steps:

[0077] Step D1: Determine the minority class sample set, let the minority class sample set be S. min The sample size is |S min |, The majority class sample set is S max ;

[0078] Step D2: Select the nearest neighbor parameter k, where k is a hyperparameter representing the number of nearest neighbors that each minority class sample needs to reference;

[0079] Step D3: Find the k nearest neighbors: for S min Each sample x in i In S minFind the k nearest samples among them and form a nearest neighbor set N. K (x i );

[0080] Step D4: Synthesize new samples from N K (x i One or more samples x are randomly selected from ) j (x) j ∈N K (x i And x j ≠x i Generate a random number λ in the interval [0,1], and synthesize a new sample x through linear interpolation. new ;

[0081] Step D5: Repeated expansion: For S min Repeat steps C3-C4 for all samples until the number of minority class samples and the number of majority class samples reach a preset balance ratio.

[0082] Step 6: Screen the three-level sample dataset, set a coefficient threshold, calculate the maximum mutual information coefficient between each parameter and casing deformation in each data sample, and remove parameters whose maximum mutual information coefficient is less than the coefficient threshold, i.e., parameters with weak correlation to casing deformation. In this embodiment, the well number, geodetic coordinates, and wellbore trajectory of the casing deformation well are collected. The cluster number, fracture section length, perforation density, number of holes, fracturing fluid discharge, total fracturing fluid volume, and pump pressure in the target well engineering parameters are correlated with the large-scale curvature, small-scale curvature, maximum likelihood volume, ant body, and chaotic body attribute parameters of the seismic data to obtain the mutual information coefficient between each pair of parameters. A threshold of 0.1 is set, and after calculation, parameters such as perforation density, number of holes, number of fracturing clusters, discharge, pump pressure, and large-scale curvature are removed. Specifically, this includes:

[0083] Seismic attribute data and drilling and completion data are selected as input parameters, denoted as X; the casing deformation risk level is selected as the output parameter, denoted as Y;

[0084] Input parameters are paired with an m×n grid. With output parameters Perform two-dimensional mesh partitioning, so that the value falls on the th The frequency of data points in the grid is used as The estimate makes it fall into the first place. The frequency of data points in a row as The estimate falls in the first The frequency of data points in a row as The estimate, namely:

[0085]

[0086] In the formula, express and The joint probability distribution, Indicates falling on the 1st The number of data points in the grid. Indicates falling on the 1st The number of data items in the cell;

[0087] Calculate input parameters With output parameters Mutual information between ,Right now:

[0088]

[0089] In the formula, express Marginal probability distribution, express The marginal probability distribution;

[0090] Calculate input parameters With output parameters Maximum mutual information coefficient between ,Right now:

[0091]

[0092] In the formula, This indicates that the condition is met. The maximum value function when n is n, F(n) is a function of n. This represents the smaller of m and n;

[0093] When the maximum mutual information coefficient is less than the coefficient threshold, the corresponding input parameters are removed to obtain the final combination of input parameters, which is the final sample dataset.

[0094] Step 7: Randomly divide the final sample dataset into training and test sets in an 8:2 ratio, and then use K-fold cross-validation to further divide the training set into training and validation datasets.

[0095] A casing deformation intelligent prediction model is established based on the CBT and BLSTM models. This model includes an input layer, CBT units, BLSTM units, and an output layer. The CBT units employ multi-scale convolutional kernels, batch normalization, and the Tanh activation function. Multi-scale convolutional kernels capture feature information at different scales. Batch normalization aims to stabilize the training process and accelerate model convergence. The Tanh activation function introduces a nonlinear function into the neural network. The CBT units can automatically learn meaningful local and global spatial features from the data. It can transform raw drilling and completion data, seismic data, fracturing parameters, and other data into high-level, abstract feature representations.

[0096] BLSTM units employ a bidirectional long short-term memory network, designed to capture past and future contextual information in sequence data;

[0097] The hyperparameters such as learning rate, number of neurons in the hidden layer, and number of iterations in the intelligent prediction model for casing deformation are optimized using an adaptive particle swarm optimization algorithm.

[0098] Population and particle initialization: particle swarm size N, maximum number of iterations T, initial particle position x i 0 Initial velocity v i 0 Update particle velocity; the formula for calculating the velocity of particle i in the d-th dimension during the (k+1)th iteration is as follows:

[0099]

[0100] In the formula, w is the inertia weight, c1 and c2 are learning factors, pk id,pbest is the historical best position of particle i in the d-th dimension in the k-th iteration, r1 and r2 are random numbers in the interval [0, 1], and pk id,gbest is the historical best position of the swarm in the d-th dimension in the i-th iteration. The particle position is updated. If the preset convergence condition is met or the maximum number of iterations is exceeded, the process stops; otherwise, the iteration process is repeated by updating the particle velocity. The formula for calculating the position of particle i in the d-th dimension in the (k+1)-th iteration is as follows:

[0101]

[0102] The training dataset in the training set is input into the intelligent prediction model of casing deformation for training, and then validated using the validation dataset in the training set. After validation, the test set is input into the intelligent prediction model of casing deformation for final verification. After passing the verification, the final intelligent prediction model of casing deformation is obtained.

[0103] Step 8: Use the final intelligent prediction model for casing deformation to predict the risk of downhole casing deformation in shale gas wells, and import the prediction result file into Petrel software to obtain the casing deformation risk volume of the block, which is divided into low (blue), medium (yellow), and high risk (red) according to color; perform horizontal and vertical profile cuts as needed to obtain the risk levels at different positions in the block, import the wellbore trajectory, and obtain the casing deformation risk points of different wells;

[0104] Step 9: By monitoring the microseismic events caused by each fracturing of the new well and the microseismic events caused by geological tectonic movements, obtain the seismic attribute data of the affected formation area, which is the seismic attribute modification data, specifically the seismic attribute modification data of the large-scale curvature of the coherence attribute, the small-scale curvature of the curvature attribute, the maximum likelihood volume, the ant volume, and the fracture-cavity parameter of the chaotic body attribute; combine the drilling and completion data of the deformed and undeformed sections of the casing of the new well and the casing deformation data of the new well, and use the incremental learning method to continuously iterate the intelligent prediction model for casing deformation, so as to realize the dynamic prediction of the casing deformation risks of old and new shale gas wells in the dynamic prediction area; the incremental learning method is a method based on sample replay, which can enable the established intelligent prediction model for casing deformation to have the ability of incremental learning, including the following steps:

[0105] Step C1: Obtain old data samples, and randomly select M representative data samples of each category from the data storage area of the trained intelligent prediction model for casing deformation;

[0106] Step C2: Obtain new data samples, based on the new seismic attribute parameters obtained in Step 9, the drilling and completion data of the deformed and undeformed sections of the casing of the new well, and the casing deformation data of the new well;

[0107] Step C3: Obtain combined data samples, randomly sample N (N < M) old data samples, and combine them with all the new data samples into a training batch of data samples;

[0108] Step C4: Use the combined data samples to train the intelligent prediction model for casing deformation again;

[0109] Step C5: Update the data storage area, randomly select P (P < M) representative data samples of each category from the new data samples and store them in the data storage area;

[0110] Step C6: Evaluate the performance of the intelligent prediction model for casing deformation on the new data samples, old data samples, and combined data samples respectively. If the performance is poor, repeat Steps C1 to C5 until the performance of the combined data samples is optimal;

[0111] Step C7: When the intelligent prediction model for casing deformation obtains new data samples, repeat Steps C1 to C6 to achieve continuous iteration of the intelligent prediction model for casing deformation.

[0112] Based on the intelligent model trained with existing historical data, whenever a new well is fractured or a microseismic event is detected, the model dynamically optimizes the intelligent prediction model of casing deformation by using the incremental learning capability of the intelligent prediction model of casing deformation, based on the new seismic attribute data of the formation area within the influence range of the microseismic event, combined with the drilling and completion data of the new well and the latest casing deformation data, and immediately realizes the intelligent prediction of casing deformation under the fracturing conditions of the next well.

[0113] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications and substitutions based on the technical solutions and inventive concepts provided by the present invention should be covered within the scope of protection of the present invention.

Claims

1. A multi-source data-driven dynamic intelligent prediction method for shale gas well casing deformation risk, characterized in that: Includes the following steps: Step 1: Collect casing deformation data, drilling and completion data of deformed and undeformed casing sections from the established shale gas well blocks, and seismic attribute data obtained by monitoring microseismic events caused by oil and gas field fracturing and geological tectonic movements to construct the original dataset; Step 2: Perform outlier processing on each data sample in the original dataset using the Isolation Forest algorithm and the Local Outlier Factor algorithm to obtain the initial dataset; Step 3: Use the KNN algorithm to fill in the missing values ​​of each data sample in the initial dataset. After filling, normalize the data of each data sample to obtain the first-level sample dataset. Step 4: Divide each data sample in the primary sample dataset into risk levels according to the amount of casing deformation, and classify the data samples of each risk level using one-hot encoding. After classification, the secondary sample dataset is obtained. Step 5: For risk categories with a small number of data samples after classification in the secondary sample dataset, data augmentation is used to expand the data samples of the risk category to obtain the tertiary sample dataset; Step 6: Filter the three-level sample dataset, set coefficient thresholds, calculate the maximum mutual information coefficient between each parameter and the sleeve deformation in each data sample, remove parameters whose maximum mutual information coefficient is less than the coefficient threshold, and obtain the final sample dataset. Step 7: Randomly divide the final sample dataset into a training set and a test set at an 8:2 ratio, and establish a casing deformation intelligent prediction model based on the CBT model and the BLSTM model. The casing deformation intelligent prediction model includes an input layer, a CBT unit, a BLSTM unit, and an output layer. Input the training set into the casing deformation intelligent prediction model for training, and input the test set into the trained casing deformation intelligent prediction model for verification. After successful verification, the final casing deformation intelligent prediction model is obtained. Step 8: Use the final intelligent prediction model for casing deformation to predict the risk of casing deformation in shale gas wells; Step 9: By monitoring microseismic events caused by each fracturing of the new well and microseismic events caused by geological tectonic movements, the seismic attribute data of the affected strata area is obtained, which is the seismic attribute modification data. Combined with the drilling and completion data of the casing deformation section and the undeformed section of the new well and the casing deformation data of the new well, the intelligent prediction model of casing deformation is continuously iterated by the incremental learning method, so as to realize the dynamic prediction of the casing deformation risk of old and new shale gas wells in the region.

2. The multi-source data-driven dynamic intelligent prediction method for shale gas well casing deformation risk according to claim 1, characterized in that: The drilling and completion data for the deformed and undeformed sections of the casing in Step 1 include: well number, geodetic coordinates, wellbore trajectory, fracturing fluid discharge, total fluid volume, pump pressure, number of clusters, length of the fracturing section, perforation density, and number of perforations. The seismic attribute data includes ant volume, curvature volume, coherence volume, maximum likelihood volume, and chaotic volume; The casing deformation data includes the location and amount of casing deformation.

3. The multi-source data-driven dynamic intelligent prediction method for shale gas well casing deformation risk according to claim 1, characterized in that: Step 2 of the Isolation Forest algorithm involves outlier handling to identify and process global outliers, including the following steps: Step A1: Data sampling, randomly extracting samples from the original data to construct a single decision tree; Step A2: Randomly select a feature, randomly select a split point within the value range of the feature, divide the sample into left and right branches according to the split point, and recursively repeat this process until a branch contains only 1 sample or the maximum depth of the tree is reached. Step A3: Repeat steps A and B to construct multiple decision trees, forming an isolated forest; Step A4: For each data sample, calculate its average path length across all decision trees, convert the average path length into anomaly score, and determine whether the data sample is an anomaly.

4. The multi-source data-driven dynamic intelligent prediction method for shale gas well casing deformation risk according to claim 1, characterized in that: Step 2, the local outlier factor algorithm, performs outlier processing to identify and handle local outliers, including the following steps: Step B1: Find the k nearest neighbors of each sample x and form the k nearest neighbor set of x; Step B2: Calculate the reachability distance. For each nearest neighbor y of x, the reachability distance from x to y is the maximum value of the k nearest neighbor distance of y and the actual distance between x and y, to avoid the influence of extreme values. Step B3: Calculate the local reachability density. The local reachability density of x is the reciprocal of the average reachability distance between x and all its k nearest neighbors. The larger the value of the reciprocal of the average reachability distance, the higher the local reachability density. Step B4: Calculate the local outlier value: The local outlier value of x is the ratio of the average local reachability of its k nearest neighbors to the local reachability of x itself, reflecting the degree of density deviation of x relative to its neighbors; the larger the local outlier value, the greater the probability that the sample is an outlier.

5. The multi-source data-driven dynamic intelligent prediction method for shale gas well casing deformation risk according to claim 1, characterized in that: In step 7, the learning rate, the number of neurons in the hidden layer, and the number of iterations in the intelligent prediction model for casing deformation are optimized using an adaptive particle swarm optimization algorithm.

6. The multi-source data-driven dynamic intelligent prediction method for shale gas well casing deformation risk according to claim 1, characterized in that: The incremental learning method described in step 9 is a sample replay-based method, which includes the following steps: Step C1: Obtain old data samples by randomly selecting representative data samples for each category from the data storage area of ​​the trained intelligent prediction model for casing deformation. Step C2: Obtain new data samples based on the new seismic attribute parameters obtained in Step 9, the casing deformation section of the new well, the drilling and completion data of the undeformed section, and the casing deformation data of the new well; Step C3: Obtain the merged data sample, randomly sample several old data samples, and merge them with all the new data samples into a training batch of data samples; Step C4: Retrain the intelligent prediction model for casing deformation using the merged data samples; Step C5: Update the data storage area by randomly selecting representative data samples from each category from the new data samples and storing them in the data storage area; Step C6: Evaluate the performance of the intelligent prediction model for casing deformation on new data samples, old data samples, and merged data samples respectively. If the performance is poor, repeat steps C1 to C5 until the performance of the merged data samples is satisfactory. Step C7: When the intelligent prediction model for casing deformation acquires new data samples, repeat steps C1 to C6 to achieve continuous iteration of the intelligent prediction model for casing deformation.

Images