Crack prediction method and system based on adaptive feature selection

Abstract

This invention provides a fracture prediction method and system based on adaptive feature selection, relating to the field of oil and gas geological exploration technology. The method includes: acquiring multi-source geological data and extracting a fracture development feature set; calculating the feature contribution based on geomechanical constraints and dynamically selecting a target feature subset; establishing a fracture development probability mapping relationship using the target feature subset and calibration data; obtaining the prediction deviation by comparing the prediction results with actual data, and adjusting the geomechanical constraints accordingly and adaptively optimizing the feature selection to update the target feature subset; finally, generating a corrected fracture development probability distribution based on the updated feature subset. This invention achieves dynamic optimization of feature selection, improving the accuracy and reliability of fracture prediction.

Description

Technical Field

[0001] This invention relates to the field of oil and gas geological exploration technology, and in particular to a fracture prediction method and system based on adaptive feature selection. Background Technology

[0002] In the field of oil and gas exploration and development, accurately predicting the development location and distribution patterns of underground reservoir fractures is crucial for optimizing drilling deployment and improving oil and gas recovery. In existing technologies, fracture prediction based on multi-source geological data is a common practice. These methods typically rely on multi-scale data such as seismic, well logging, and geological structures, extracting a series of features related to fracture development to construct a feature set. Subsequently, machine learning or statistical models are used to establish a mapping relationship from these features to fracture development probabilities, ultimately generating a fracture probability distribution map to guide exploration decisions.

[0003] Features extracted from multi-source data are high-dimensional and numerous, with complex correlations and redundancies among them. Indiscriminately using all features not only significantly increases the computational complexity of the model but also easily introduces noise, leading to overfitting and reducing its generalization ability in unknown regions. Feature selection often relies on general statistical criteria or fixed expert experience, lacking a dynamic response to the specific geomechanical background of the reservoir. The formation of reservoir fractures is controlled by specific geomechanical conditions, such as the in-situ stress field and rock mechanical properties, which vary significantly across different regions and depths. Fixed feature selection strategies struggle to adaptively optimize feature subsets, preventing the model from accurately capturing the key driving factors controlling fracture development in the current target area. This results in significant deviations between predicted results and actual fracture distribution revealed by drilling, leading to insufficient predictive reliability. Summary of the Invention

[0004] The present invention provides a crack prediction method and system based on adaptive feature selection, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a crack prediction method based on adaptive feature selection, comprising: Acquire multi-source geological data of the target reservoir area and perform multi-scale decomposition to extract fracture development feature set of fracture development characteristics; Based on the geomechanical constraints of reservoir fracture development, the contribution of each feature in the fracture development feature set to fracture prediction is calculated, and the fracture development feature set is dynamically selected according to the contribution to obtain the target feature subset. Using the target feature subset as input variables and combined with the calibration data of known fracture development locations, a mapping relationship from the target feature subset to the fracture development probability is established. The mapping relationship outputs the fracture development probability distribution of each spatial location within the target reservoir area. The prediction deviation is obtained by comparing the fracture development probability distribution with the fracture distribution data revealed by actual drilling. The geomechanical constraints are adjusted based on the prediction deviation, and the composition of the target feature subset is updated. The feedback adjustment is performed by quantifying the contribution rate of different geomechanical constraints to the prediction error and adaptively optimizing the dynamic selection to obtain the updated target feature subset. Based on the updated target feature subset, a new mapping relationship is established from the updated target feature subset to the crack development probability, and a corrected crack development probability distribution is generated as the crack prediction result.

[0006] Multi-source geological data of the target reservoir area were acquired and decomposed at multiple scales. The fracture development feature set, including fracture development characteristics, was extracted. In the multi-source geological data, the seismic attribute data is decomposed in the time and frequency domain to extract the seismic waveform features at different frequency scales. The well logging data is vertically layered and horizontally interpolated to extract the physical property parameter variation features that reflect the pore structure and mechanical properties of the rock. The geological structure data is inverted by tectonic stress field to extract the tectonic geometric features and stress distribution features that reflect the degree of tectonic deformation and stress concentration as tectonic stress features. Based on the stratigraphic calibration information of the target reservoir area, the seismic waveform features, the physical property parameter variation features, and the tectonic stress features are transformed into a unified geological stratigraphic coordinate system. The vertical depth correspondence between different data sources is established through the stratigraphic calibration information, and the seismic waveform features, physical property parameter variation features, and tectonic stress features are vertically aligned. After vertical alignment, based on the spatial transformation relationship between well location coordinates and seismic network coordinates, the seismic waveform characteristics, the physical property parameter variation characteristics, and the tectonic stress characteristics are laterally spatially interpolated and uniformly mapped to the same spatial grid nodes to form the spatially registered fracture development feature set.

[0007] Based on the geomechanical constraints of reservoir fracture development, the contribution of each feature in the fracture development feature set to fracture prediction is calculated. The fracture development feature set is then dynamically selected based on these contributions to obtain a target feature subset, which includes: The fracture development mechanism parameters in the geomechanical constraints of reservoir fracture development are extracted as constraint parameters. The physical coupling strength of any two features in the fracture development feature set during the fracture formation process is calculated based on the constraint parameters as the correlation weight. A correlation weight matrix is ​​constructed based on all correlation weights. Each feature in the fracture development feature set is weighted by the corresponding weight value in the association weight matrix to calculate its contribution to fracture prediction under the geomechanical constraints. The fracture development density is obtained by calculating the ratio of the number of calibrated fracture locations in the calibration data of known fracture development locations to the total number of spatial grid nodes in the target reservoir area. A contribution threshold is set based on the fracture development density, and features with a contribution higher than the contribution threshold are retained. Redundancy checks are performed on the retained features. The contribution difference between any two retained features is calculated. When the contribution difference is less than a preset proportion of the contribution threshold, the sum of the correlation weights between any two features and the remaining features in the crack development feature set is calculated using the correlation weight matrix. Features whose sum of correlation weights is greater than the sum of correlation weights of the other feature are retained to form the target feature subset.

[0008] Using the target feature subset as input variables and combining it with calibration data of known fracture development locations, a mapping relationship from the target feature subset to fracture development probabilities is established. This mapping relationship outputs the fracture development probability distribution at various spatial locations within the target reservoir region, including: The feature values ​​of all features in the target feature subset at each spatial grid node in the target reservoir area are extracted as feature vectors. The fracture development status of each calibrated location in the calibration data of the known fracture development location is marked as a training label. The feature vectors and the training labels constitute a training sample set. The training sample set is nonlinearly mapped and a crack development probability prediction value is output. The parameters of the nonlinear mapping are adjusted by minimizing the deviation between the crack development probability prediction value and the training label, and a mapping relationship from the target feature subset to the crack development probability is established. The feature vectors of all spatial grid nodes in the target reservoir area are input into the mapping relationship. The mapping relationship outputs the corresponding fracture development probability value for each spatial grid node. The fracture development probability values ​​of all spatial grid nodes are organized in the spatial coordinate system of the target reservoir area to form the fracture development probability distribution of each spatial location in the target reservoir area.

[0009] The geomechanical constraints are adjusted based on the prediction deviation, and the composition of the target feature subset is updated. This adjustment quantifies the contribution rate of different geomechanical constraints to the prediction error, and adaptively optimizes the dynamic selection to obtain the updated target feature subset, which includes: The spatial position of the crack development probability distribution is compared with the calibration data of the known crack development location. For each spatial grid node, the difference between the predicted crack development probability value and the actual calibrated crack development state is calculated as the prediction error. The prediction error of all spatial grid nodes is organized in the spatial coordinate system to obtain the spatial distribution of the prediction deviation. For each deviation location in the spatial distribution of the prediction deviation, the geomechanical constraint parameter value corresponding to the deviation location is extracted, and the correlation strength between the geomechanical constraint parameter value and the prediction error at the deviation location is calculated as the contribution rate of the geomechanical constraint to the prediction error. Based on the contribution rate, the geomechanical constraints that have the greatest impact on the prediction error are identified. A parameter space search is performed on the fracture development mechanism parameters in the identified geomechanical constraints. By calculating the contribution trend of each feature in the fracture development feature set to fracture prediction under different parameter combinations, a parameter combination that makes the contribution distribution negatively correlated with the spatial distribution of the prediction deviation is selected. Based on the selected parameter combination, the dynamic selection is performed to obtain the updated target feature subset.

[0010] Based on the updated target feature subset, a new mapping relationship is established from the updated target feature subset to the crack development probability, and a corrected crack development probability distribution is generated as the crack prediction result, including: The updated feature vector is obtained by extracting the feature values ​​of all features in the updated target feature subset at each spatial grid node in the target reservoir area, keeping the training labels in the calibration data of the known fracture development location unchanged, and reconstructing the updated training sample set with the updated feature vector and the training labels. The updated training sample set is nonlinearly mapped and the updated crack development probability prediction value is output. The residual between the updated crack development probability prediction value and the training label is calculated and compared with the prediction deviation before the feedback adjustment. When the residual is less than the prediction deviation before the feedback adjustment, the feedback adjustment is confirmed to be effective and the updated target feature subset is fixed. The parameters of the nonlinear mapping are readjusted to establish the updated mapping relationship from the updated target feature subset to the crack development probability. The updated mapping relationship outputs a corresponding corrected fracture development probability value for each spatial grid node. The corrected fracture development probability values ​​of all spatial grid nodes are organized in the spatial coordinate system of the target reservoir region to form the corrected fracture development probability distribution. The corrected fracture development probability distribution is used as the fracture prediction result.

[0011] A second aspect of the present invention provides a crack prediction system based on adaptive feature selection, comprising: The feature extraction unit is used to acquire multi-source geological data of the target reservoir area and perform multi-scale decomposition to extract a fracture development feature set of fracture development characteristics. A dynamic selection unit is used to calculate the contribution of each feature in the fracture development feature set to fracture prediction based on the geomechanical constraints of reservoir fracture development, and to dynamically select the fracture development feature set according to the contribution to obtain a target feature subset. The mapping establishment unit is used to establish a mapping relationship from the target feature subset to the fracture development probability by using the target feature subset as an input variable and combining it with the calibration data of known fracture development locations. The mapping relationship outputs the fracture development probability distribution of each spatial location in the target reservoir area. The feedback adjustment unit is used to compare the fracture development probability distribution with the fracture distribution data revealed by actual drilling to obtain the prediction deviation, adjust the geomechanical constraints based on the prediction deviation, and update the composition of the target feature subset. The feedback adjustment quantifies the contribution rate of different geomechanical constraints to the prediction error and adaptively optimizes the dynamic selection to obtain the updated target feature subset. The result correction unit is used to re-establish the mapping relationship from the updated target feature subset to the crack development probability based on the updated target feature subset, and generate a corrected crack development probability distribution as the crack prediction result.

[0012] A third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0013] Fourth aspect of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0014] The beneficial effects of this application are as follows: This method significantly improves the accuracy and reliability of fracture prediction. By acquiring multi-source geological data and performing multi-scale decomposition, it can comprehensively capture fracture development information at different scales, from macroscopic structure to microscopic lithology, and construct an information-rich and dimensionally complete initial feature set, laying a solid data foundation for subsequent accurate prediction.

[0015] By calculating the contribution of each feature based on geomechanical constraints and dynamically selecting features, the system effectively eliminates features with weak or redundant correlation to fracture development, focusing on a subset of target features that have a core indicative role in prediction. This process reduces data dimensionality and model complexity, avoids the risk of overfitting, and enhances the physical interpretability of the model, making the prediction results more consistent with geomechanical laws.

[0016] By establishing a mapping relationship between a subset of target features and known calibration data, a spatially continuous probability distribution of fracture development can be directly output. This probabilistic result not only indicates favorable zones for fracture development but also quantifies the likelihood of development, providing a refined decision-making basis for drilling deployment and development planning, and overcoming the limitations of overly absolute predictions from traditional binary classification.

[0017] The core advantage of this method lies in introducing a feedback adjustment mechanism based on prediction bias. By comparing predicted probabilities with actual drilling data, the contribution of different geomechanical constraints to prediction errors is quantitatively analyzed, and the feature selection process and the constraints themselves are adaptively optimized accordingly. This closed-loop optimization process enables the model to continuously learn and evolve from new exploration data, dynamically correcting its understanding of geological patterns, thereby continuously improving its extrapolation and generalization capabilities under new blocks or complex geological conditions.

[0018] Finally, a corrected fracture prediction result is generated based on the updated target feature subset. This result integrates multi-source data, geological prior knowledge, and actual verification information. Its accuracy and reliability have been iteratively enhanced, enabling it to more realistically reflect the spatial distribution pattern of underground fractures and significantly reduce the uncertainty of exploration and development decisions. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the crack prediction method based on adaptive feature selection according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the multi-source geological data processing and feature extraction process. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0022] Figure 1 This is a flowchart illustrating the crack prediction method based on adaptive feature selection according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes: Acquire multi-source geological data of the target reservoir area and perform multi-scale decomposition to extract fracture development feature set of fracture development characteristics; Based on the geomechanical constraints of reservoir fracture development, the contribution of each feature in the fracture development feature set to fracture prediction is calculated, and the fracture development feature set is dynamically selected according to the contribution to obtain the target feature subset. Using the target feature subset as input variables and combined with the calibration data of known fracture development locations, a mapping relationship from the target feature subset to the fracture development probability is established. The mapping relationship outputs the fracture development probability distribution of each spatial location within the target reservoir area. The prediction deviation is obtained by comparing the fracture development probability distribution with the fracture distribution data revealed by actual drilling. The geomechanical constraints are adjusted based on the prediction deviation, and the composition of the target feature subset is updated. The feedback adjustment is performed by quantifying the contribution rate of different geomechanical constraints to the prediction error and adaptively optimizing the dynamic selection to obtain the updated target feature subset. Based on the updated target feature subset, a new mapping relationship is established from the updated target feature subset to the crack development probability, and a corrected crack development probability distribution is generated as the crack prediction result.

[0023] In one optional implementation, multi-source geological data of the target reservoir area is acquired and decomposed at multiple scales. The fracture development feature set extracted includes: In the multi-source geological data, the seismic attribute data is decomposed in the time and frequency domain to extract the seismic waveform features at different frequency scales. The well logging data is vertically layered and horizontally interpolated to extract the physical property parameter variation features that reflect the pore structure and mechanical properties of the rock. The geological structure data is inverted by tectonic stress field to extract the tectonic geometric features and stress distribution features that reflect the degree of tectonic deformation and stress concentration as tectonic stress features. Based on the stratigraphic calibration information of the target reservoir area, the seismic waveform features, the physical property parameter variation features, and the tectonic stress features are transformed into a unified geological stratigraphic coordinate system. The vertical depth correspondence between different data sources is established through the stratigraphic calibration information, and the seismic waveform features, physical property parameter variation features, and tectonic stress features are vertically aligned. After vertical alignment, based on the spatial transformation relationship between well location coordinates and seismic network coordinates, the seismic waveform characteristics, the physical property parameter variation characteristics, and the tectonic stress characteristics are laterally spatially interpolated and uniformly mapped to the same spatial grid nodes to form the spatially registered fracture development feature set.

[0024] like Figure 2 As shown, the method includes: The seismic attribute data acquired for the target reservoir region encompasses multi-dimensional seismic response information, including amplitude, frequency, and phase. When performing time-frequency domain decomposition on the seismic attribute data, wavelet transform time-frequency analysis is employed to decompose the original seismic signal into different frequency components. The decomposition frequency range is set to 10Hz to 80Hz, with a frequency interval of 5Hz. Instantaneous amplitude, instantaneous frequency, and instantaneous phase at each frequency scale are extracted using a sliding time window. These parameters form the seismic waveform feature vector. High-frequency components reflect the fine-grained response of thin-layer fractures, while low-frequency components reflect the overall trend of the thick-layer background, forming a multi-scale seismic waveform feature set.

[0025] During well logging data processing, the acquired raw logging curves include physical property measurements such as natural gamma ray, sonic transit time, density, and resistivity. Vertical stratification, based on abrupt changes in the logging curves and geological stratification standards, divides continuous logging data into several geological units. Within each unit, rock physical property parameters such as porosity, permeability, Young's modulus, and Poisson's ratio are extracted. Lateral interpolation uses Kriging interpolation or radial basis function methods to extrapolate the distribution of physical property parameters between well locations based on logging data from drilled wells, forming a characteristic of physical property parameter changes reflecting variations in pore structure and rock mechanical properties. Abnormal changes in elastic parameters often indicate fracture development zones; areas with decreased Young's modulus and increased Poisson's ratio typically correspond to fracture-rich areas.

[0026] The processing of geological structural data involves structural geometric information such as fault location, fold morphology, and dip angle variations. Structural stress field inversion, based on tectonic deformation theory, utilizes the finite element method (FEM) numerical simulation, inputting the regional tectonic evolution history and boundary stress conditions to calculate the three-dimensional stress tensor distribution within the reservoir. Extracted structural geometric features include parameters such as fault density, radius of curvature, and formation dip gradient; stress distribution features include indicators such as the direction of maximum principal stress, differential stress value, and stress concentration factor. High curvature regions and stress concentration zones are locations where fractures are predominantly developed; these tectonic stress features serve as key input parameters for fracture prediction.

[0027] In the data fusion phase, stratigraphic calibration information is provided by seismic interpretation and well logging data, establishing a depth-time conversion relationship between time-domain seismic data and depth-domain geological data. Seismic waveform characteristics are converted from time coordinates to depth coordinates using time-depth curves, ensuring that seismic characteristics at different frequency scales correspond to the same geological stratigraphic interface with well logging physical parameters. The vertical alignment process employs a dynamic time warping algorithm to eliminate depth deviations caused by velocity variations, ensuring accurate matching of seismic waveform characteristics, physical parameter variation characteristics, and tectonic stress characteristics across vertical geological stratigraphic levels.

[0028] The corresponding nodes of well location coordinates in the 3D seismic grid are determined. Inverse distance weighting or co-kriging methods are used to extend the discrete well point feature values ​​to all grid nodes in the study area. The spatial grid is set according to the seismic sampling interval, typically with a horizontal grid step size of 25 to 50 meters and a vertical grid step size of 2 to 5 meters. After lateral interpolation, seismic waveform characteristics, physical property parameter variation characteristics, and tectonic stress characteristics have clear values ​​at each spatial grid node, forming a spatially registered multidimensional feature data volume. This feature set contains dozens to hundreds of feature variables, providing a data foundation for subsequent feature selection and fracture prediction.

[0029] In one optional implementation, based on the geomechanical constraints of reservoir fracture development, the contribution of each feature in the fracture development feature set to fracture prediction is calculated. The fracture development feature set is then dynamically selected based on the contribution to obtain a target feature subset, including: The fracture development mechanism parameters in the geomechanical constraints of reservoir fracture development are extracted as constraint parameters. The physical coupling strength of any two features in the fracture development feature set during the fracture formation process is calculated based on the constraint parameters as the correlation weight. A correlation weight matrix is ​​constructed based on all correlation weights. Each feature in the fracture development feature set is weighted by the corresponding weight value in the association weight matrix to calculate its contribution to fracture prediction under the geomechanical constraints. The fracture development density is obtained by calculating the ratio of the number of calibrated fracture locations in the calibration data of known fracture development locations to the total number of spatial grid nodes in the target reservoir area. A contribution threshold is set based on the fracture development density, and features with a contribution higher than the contribution threshold are retained. Redundancy checks are performed on the retained features. The contribution difference between any two retained features is calculated. When the contribution difference is less than a preset proportion of the contribution threshold, the sum of the correlation weights between any two features and the remaining features in the crack development feature set is calculated using the correlation weight matrix. Features whose sum of correlation weights is greater than the sum of correlation weights of the other feature are retained to form the target feature subset.

[0030] Parameters related to fracture development mechanisms are extracted from the geomechanical constraints of reservoir fracture development. These parameters include tectonic stress field parameters, rock mechanics parameters, and fracture formation kinetic parameters. Specifically, tectonic stress field parameters include the directions of maximum and minimum principal stresses and differential stress values, obtained through comprehensive calculations using seismic inversion data and well logging data. Rock mechanics parameters include Young's modulus, Poisson's ratio, and rock brittleness index, extracted from well logging curves or determined through core experiments. The fracture formation kinetic parameters include strain rate and energy release rate, reflecting the dynamic process of fracture propagation.

[0031] Based on constraint parameters, the physical coupling strength between any two features in the fracture development feature set during the fracture formation process is calculated. This physical coupling strength is quantified by analyzing the degree of interaction between the two features during the geomechanical process. For example, the coupling strength between curvature attribute and stress concentration is calculated by analyzing the influence of curvature changes on local stress field disturbances; the coupling strength between density attribute and elastic modulus is quantified using rock physics relationships. The physical coupling strength between each pair of features is used as a correlation weight to construct a correlation weight matrix. This correlation weight matrix is ​​a symmetric matrix with a dimension equal to the number of features in the fracture development feature set. The element in the i-th row and j-th column of the matrix represents the correlation weight between the i-th and j-th features.

[0032] The statistical properties of the feature itself are weighted by the corresponding weight values ​​in the association weight matrix. Specifically, for the i-th feature, all weight values ​​of the feature in the i-th row of the association weight matrix are extracted, and the sum of the products of these weight values ​​and the response intensity of the corresponding feature at the known crack location is calculated. After normalization, the contribution value of the feature is obtained.

[0033] During feature selection, the fracture development density is calculated by the ratio of the number of identified fracture locations to the total number of spatial grid nodes in the target reservoir area. When the total number of grid nodes is 10,000 and the number of identified fracture locations is 150, the fracture development density is 0.015. A contribution threshold is set based on the product of the fracture development density and an empirical coefficient. The empirical coefficient typically ranges from 3 to 5; in this embodiment, it is set to 4, resulting in a contribution threshold of 0.06. Features with a contribution higher than this threshold are retained for further testing.

[0034] A redundancy check is performed on the retained features. The contribution difference between any two features is calculated. If the contribution difference between two features is less than a preset proportion of the contribution threshold, the two features are considered redundant. The preset proportion is usually set to 0.1 to 0.2, and in this embodiment, it is set to 0.15, meaning that the redundancy check is triggered when the contribution difference is less than 0.009. At this time, the row vectors corresponding to the two features are extracted from the correlation weight matrix, and the sum of all elements in each row vector is calculated. This sum represents the sum of the correlation weights of the feature with the remaining features in the crack development feature set. The sum of the correlation weights of the two features is compared, and the feature with the larger value is retained, while the feature with the smaller value is removed. After the above dynamic selection and redundancy check, a target feature subset is finally formed. This subset contains feature combinations that have a high contribution to crack prediction and strong information complementarity.

[0035] In one optional implementation, using the target feature subset as an input variable and combining it with calibration data of known fracture development locations, a mapping relationship from the target feature subset to fracture development probabilities is established. The mapping relationship outputs the fracture development probability distribution at each spatial location within the target reservoir region, including: The feature values ​​of all features in the target feature subset at each spatial grid node in the target reservoir area are extracted as feature vectors. The fracture development status of each calibrated location in the calibration data of the known fracture development location is marked as a training label. The feature vectors and the training labels constitute a training sample set. The training sample set is nonlinearly mapped and a crack development probability prediction value is output. The parameters of the nonlinear mapping are adjusted by minimizing the deviation between the crack development probability prediction value and the training label, and a mapping relationship from the target feature subset to the crack development probability is established. The feature vectors of all spatial grid nodes in the target reservoir area are input into the mapping relationship. The mapping relationship outputs the corresponding fracture development probability value for each spatial grid node. The fracture development probability values ​​of all spatial grid nodes are organized in the spatial coordinate system of the target reservoir area to form the fracture development probability distribution of each spatial location in the target reservoir area.

[0036] A three-dimensional spatial grid model was established for the target reservoir area, with the grid node spacing set to 10 to 50 meters based on the reservoir's heterogeneity. At each spatial grid node, the values ​​of various feature parameters contained in the target feature subset were read. The target feature subset includes multiple types of features such as seismic curvature attributes, coherence attributes, stress field parameters, and rock mechanics parameters. For a specific grid node, assuming the target feature subset contains 12 features, the values ​​of these 12 features at that node were extracted and arranged into a 12-dimensional feature vector according to the feature number order. The same extraction operation was performed on all grid nodes within the reservoir area to obtain a complete set of feature vectors.

[0037] Training labels are constructed based on actual drilling core observations and imaging logging data. At the calibrated locations, fracture development is quantified according to the number of fractures per unit length. A fracture linear density exceeding 3 fractures per meter is marked as 1, indicating fracture development; a fracture linear density below 0.5 fractures per meter is marked as 0, indicating no development; and a density between these two is marked as 0.5, indicating moderate development. The spatial coordinates of the calibrated locations are combined with their corresponding feature vectors and fracture development status labels to form a single training sample. Typically, 50 to 200 wells within the target reservoir area provide calibration data, forming a training sample set containing hundreds to thousands of training samples.

[0038] A neural network is used to implement the nonlinear mapping. The network consists of an input layer, two hidden layers, and an output layer. The number of neurons in the input layer equals the dimension of the feature vector. The first hidden layer has 64 neurons, the second hidden layer has 32 neurons, and the output layer contains one neuron that outputs the predicted crack development probability. The hidden layers use the ReLU activation function, and the output layer uses the Sigmoid activation function to limit the output value to between 0 and 1. The training sample set is divided into a training set and a validation set in an 8:2 ratio. During training, the feature vectors of the training set are input into the network, and the mean squared error between the predicted crack development probability and the corresponding training label is calculated. The Adam optimization algorithm is used to adjust the weight and bias parameters of each layer of the network, with a learning rate of 0.001 and 300-500 iterations. The validation set is used to monitor the training process. Training stops when the error on the validation set no longer decreases for 20 consecutive iterations to prevent overfitting. After training, the network parameters are fixed, forming a deterministic mapping relationship from the target feature subset to the crack development probability.

[0039] During the prediction phase, all spatial grid nodes within the target reservoir region are traversed. For each node, its feature vector is extracted and input into the established mapping relationship. The mapping relationship outputs the fracture development probability value for that node. This probability value reflects the likelihood of fractures occurring at that location under current geological conditions; the closer the value is to 1, the higher the probability of fracture development. The probability values ​​of all grid nodes are organized according to their spatial positions in the three-dimensional coordinate system of the target reservoir region to form a three-dimensional probability data volume. This data volume forms a planar distribution map of fracture development probability on a plane and displays the vertical variation of fracture development probability on a cross-section, comprehensively characterizing the fracture development probability distribution features at various spatial locations within the target reservoir region.

[0040] In one optional implementation, the geomechanical constraints are adjusted based on the prediction deviation, and the composition of the target feature subset is updated. The feedback adjustment quantifies the contribution rate of different geomechanical constraints to the prediction error, and adaptively optimizes the dynamic selection to obtain the updated target feature subset, which includes: The spatial position of the crack development probability distribution is compared with the calibration data of the known crack development location. For each spatial grid node, the difference between the predicted crack development probability value and the actual calibrated crack development state is calculated as the prediction error. The prediction error of all spatial grid nodes is organized in the spatial coordinate system to obtain the spatial distribution of the prediction deviation. For each deviation location in the spatial distribution of the prediction deviation, the geomechanical constraint parameter value corresponding to the deviation location is extracted, and the correlation strength between the geomechanical constraint parameter value and the prediction error at the deviation location is calculated as the contribution rate of the geomechanical constraint to the prediction error. Based on the contribution rate, the geomechanical constraints that have the greatest impact on the prediction error are identified. A parameter space search is performed on the fracture development mechanism parameters in the identified geomechanical constraints. By calculating the contribution trend of each feature in the fracture development feature set to fracture prediction under different parameter combinations, a parameter combination that makes the contribution distribution negatively correlated with the spatial distribution of the prediction deviation is selected. Based on the selected parameter combination, the dynamic selection is performed to obtain the updated target feature subset.

[0041] The predicted fracture development probability distribution is compared spatially with the actual drilling calibration data one by one. A three-dimensional spatial grid is established for the study area, and each grid node stores a predicted fracture development probability value. The actual fracture development state of that node is obtained from the calibration data; this state is represented in binary form, with developed fractures marked as 1 and no fractures marked as 0. The difference between the predicted probability value and the actual state value is calculated; this difference is the prediction error for that node. This calculation is performed on all grid nodes, and the prediction error of each node is organized according to its three-dimensional spatial coordinates to form a spatial distribution data volume of the prediction deviation.

[0042] For each grid location with a deviation in the spatial distribution of prediction errors, the corresponding geomechanical constraint parameters are extracted. These parameters include stress field parameters, rock mechanics parameters, and tectonic curvature at that location. The correlation coefficient between the value of each type of geomechanical constraint parameter and the prediction error at that location is calculated. The absolute value of the correlation coefficient represents the strength of the association between the constraint and the prediction error. The correlation strengths of all deviation locations are statistically summed and normalized to obtain the contribution rate of each geomechanical constraint to the overall prediction error.

[0043] Based on the calculated contribution rates, geomechanical constraints exceeding a set threshold are identified in descending order. For the identified key constraints, mechanical mechanism parameters controlling fracture development, such as stress difference coefficients and fracture angle parameters, are extracted. A parameter space is constructed within the reasonable physical range of these parameters, and multiple parameter combinations are generated using a grid search method. For each parameter combination, the process of calculating the contribution of each feature in the fracture development feature set based on geomechanical constraints is re-executed to obtain the feature contribution distribution under that combination. Spatial correlation analysis is performed between the feature contribution distribution and the spatial distribution of prediction deviation, and the spatial correlation coefficient is calculated. The parameter combination that results in a negative correlation coefficient with the largest absolute value is selected. The feature contribution distribution corresponding to this combination shows an inverse relationship with the prediction deviation in space, effectively compensating for prediction errors. The selected parameter combination is used to re-execute the dynamic feature selection process, and features with updated contribution rates higher than the selection threshold constitute the updated target feature subset, achieving adaptive optimization and adjustment of the feature set.

[0044] In one optional implementation, based on the updated target feature subset, a new mapping relationship is established from the updated target feature subset to the crack development probability, and a corrected crack development probability distribution is generated as the crack prediction result, including: The updated feature vector is obtained by extracting the feature values ​​of all features in the updated target feature subset at each spatial grid node in the target reservoir area, keeping the training labels in the calibration data of the known fracture development location unchanged, and reconstructing the updated training sample set with the updated feature vector and the training labels. The updated training sample set is nonlinearly mapped and the updated crack development probability prediction value is output. The residual between the updated crack development probability prediction value and the training label is calculated and compared with the prediction deviation before the feedback adjustment. When the residual is less than the prediction deviation before the feedback adjustment, the feedback adjustment is confirmed to be effective and the updated target feature subset is fixed. The parameters of the nonlinear mapping are readjusted to establish the updated mapping relationship from the updated target feature subset to the crack development probability. The updated mapping relationship outputs a corresponding corrected fracture development probability value for each spatial grid node. The corrected fracture development probability values ​​of all spatial grid nodes are organized in the spatial coordinate system of the target reservoir region to form the corrected fracture development probability distribution. The corrected fracture development probability distribution is used as the fracture prediction result.

[0045] To establish a mapping relationship for the updated target feature subset, it is necessary to reconstruct the numerical representation of this feature subset in a three-dimensional spatial grid. Specifically, assume that the target reservoir region has been divided into M × N × L spatial grid nodes, each node corresponding to three-dimensional coordinates (x, y, y). i y j z k For each of the p features contained in the updated target feature subset, its feature value is extracted at each grid node to form a p-dimensional updated feature vector. For example, if the updated target feature subset includes features such as curvature attribute, stress difference coefficient, and fault distance index, then the specific values ​​of these features are extracted at node (x1, y1, z1) to form the feature vector of that node. For calibration data with known fracture development locations, its training label remains a binary identifier of the original fracture development state, i.e., fracture development location is marked as 1, and non-development location is marked as 0. The updated feature vectors of all nodes corresponding to the calibration locations are paired with their training labels to form an updated training sample set.

[0046] The updated training sample set is nonlinearly mapped using machine learning algorithms capable of characterizing complex nonlinear relationships, such as support vector machines, random forests, or neural networks. The feature vector of each training sample is input into the mapping model, which outputs a predicted crack development probability value for that sample location, ranging from 0 to 1. The residuals between the predicted values ​​and training labels for all training samples are calculated, using root mean square error or mean absolute error as the residual metric. The calculated residual value is compared numerically with the prediction deviation before the feedback adjustment. If the residual is less than the prediction deviation before the feedback adjustment and the decrease exceeds a preset threshold (e.g., 10%), the feedback adjustment is considered effective. At this point, the composition of the updated target feature subset is fixed, and no further feature additions or deletions are performed.

[0047] After confirming the validity of the feature subset, the internal parameters of the nonlinear mapping are readjusted to optimize the fitting effect. For neural network algorithms, hyperparameters such as the number of hidden layer nodes, learning rate, and number of iterations are adjusted; for tree model algorithms, parameters such as tree depth and number of leaf node samples are adjusted. Cross-validation is used to determine the optimal parameter combination to avoid overfitting. After parameter adjustment, the mapping model is retrained using the entire updated training sample set to establish an updated mapping relationship from the updated target feature subset to the crack development probability. This mapping relationship has the ability to convert feature vectors at any spatial location into predicted crack development probabilities.

[0048] Using the updated mapping relationship after training, prediction is performed for each spatial grid node within the target reservoir region. The values ​​of each feature in the updated target feature subset at each node are extracted to form the node's feature vector, which is then input into the mapping model to obtain the corresponding corrected fracture development probability value. The probability values ​​for the entire region are calculated by traversing all M × N × L grid nodes. The probability values ​​of each node are then assigned according to their spatial coordinates (x, y, y). i y j z k The data is organized to form a three-dimensional probability distribution data volume. The value of each voxel in this data volume represents the probability of fracture development at that location; the closer the value is to 1, the higher the probability of fracture development. This three-dimensional probability distribution data volume is used as the corrected fracture development probability distribution, which is the final fracture prediction result and can be used for reservoir evaluation and well location deployment decisions.

[0049] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0050] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0051] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A crack prediction method based on adaptive feature selection, characterized in that, include: Acquire multi-source geological data of the target reservoir area and perform multi-scale decomposition to extract fracture development feature set of fracture development characteristics; Based on the geomechanical constraints of reservoir fracture development, the contribution of each feature in the fracture development feature set to fracture prediction is calculated, and the fracture development feature set is dynamically selected according to the contribution to obtain the target feature subset. Using the target feature subset as input variables and combined with the calibration data of known fracture development locations, a mapping relationship from the target feature subset to the fracture development probability is established. The mapping relationship outputs the fracture development probability distribution of each spatial location within the target reservoir area. The prediction deviation is obtained by comparing the fracture development probability distribution with the fracture distribution data revealed by actual drilling. The geomechanical constraints are adjusted based on the prediction deviation, and the composition of the target feature subset is updated. The feedback adjustment is performed by quantifying the contribution rate of different geomechanical constraints to the prediction error and adaptively optimizing the dynamic selection to obtain the updated target feature subset. Based on the updated target feature subset, a new mapping relationship is established from the updated target feature subset to the crack development probability, and a corrected crack development probability distribution is generated as the crack prediction result.

2. The method according to claim 1, characterized in that, Multi-source geological data of the target reservoir area were acquired and decomposed at multiple scales. The fracture development feature set, including fracture development characteristics, was extracted. In the multi-source geological data, the seismic attribute data is decomposed in the time and frequency domain to extract the seismic waveform features at different frequency scales. The well logging data is vertically layered and horizontally interpolated to extract the physical property parameter variation features that reflect the pore structure and mechanical properties of the rock. The geological structure data is inverted by tectonic stress field to extract the tectonic geometric features and stress distribution features that reflect the degree of tectonic deformation and stress concentration as tectonic stress features. Based on the stratigraphic calibration information of the target reservoir area, the seismic waveform features, the physical property parameter variation features, and the tectonic stress features are transformed into a unified geological stratigraphic coordinate system. The vertical depth correspondence between different data sources is established through the stratigraphic calibration information, and the seismic waveform features, physical property parameter variation features, and tectonic stress features are vertically aligned. After vertical alignment, based on the spatial transformation relationship between well location coordinates and seismic network coordinates, the seismic waveform characteristics, the physical property parameter variation characteristics, and the tectonic stress characteristics are laterally spatially interpolated and uniformly mapped to the same spatial grid nodes to form the spatially registered fracture development feature set.

3. The method according to claim 1, characterized in that, Based on the geomechanical constraints of reservoir fracture development, the contribution of each feature in the fracture development feature set to fracture prediction is calculated. The fracture development feature set is then dynamically selected based on these contributions to obtain a target feature subset, which includes: The fracture development mechanism parameters in the geomechanical constraints of reservoir fracture development are extracted as constraint parameters. The physical coupling strength of any two features in the fracture development feature set during the fracture formation process is calculated based on the constraint parameters as the correlation weight. A correlation weight matrix is ​​constructed based on all correlation weights. Each feature in the fracture development feature set is weighted by the corresponding weight value in the association weight matrix to calculate its contribution to fracture prediction under the geomechanical constraints. The fracture development density is obtained by calculating the ratio of the number of calibrated fracture locations in the calibration data of known fracture development locations to the total number of spatial grid nodes in the target reservoir area. A contribution threshold is set based on the fracture development density, and features with a contribution higher than the contribution threshold are retained. Redundancy checks are performed on the retained features. The contribution difference between any two retained features is calculated. When the contribution difference is less than a preset proportion of the contribution threshold, the sum of the correlation weights between any two features and the remaining features in the crack development feature set is calculated using the correlation weight matrix. Features whose sum of correlation weights is greater than the sum of correlation weights of the other feature are retained to form the target feature subset.

4. The method according to claim 1, characterized in that, Using the target feature subset as input variables and combining it with calibration data of known fracture development locations, a mapping relationship from the target feature subset to fracture development probabilities is established. This mapping relationship outputs the fracture development probability distribution at various spatial locations within the target reservoir region, including: The feature values ​​of all features in the target feature subset at each spatial grid node in the target reservoir area are extracted as feature vectors. The fracture development status of each calibrated location in the calibration data of the known fracture development location is marked as a training label. The feature vectors and the training labels constitute a training sample set. The training sample set is nonlinearly mapped and a crack development probability prediction value is output. The parameters of the nonlinear mapping are adjusted by minimizing the deviation between the crack development probability prediction value and the training label, and a mapping relationship from the target feature subset to the crack development probability is established. The feature vectors of all spatial grid nodes in the target reservoir area are input into the mapping relationship. The mapping relationship outputs the corresponding fracture development probability value for each spatial grid node. The fracture development probability values ​​of all spatial grid nodes are organized in the spatial coordinate system of the target reservoir area to form the fracture development probability distribution of each spatial location in the target reservoir area.

5. The method according to claim 1, characterized in that, The geomechanical constraints are adjusted based on the prediction deviation, and the composition of the target feature subset is updated. This adjustment quantifies the contribution rate of different geomechanical constraints to the prediction error, and adaptively optimizes the dynamic selection to obtain the updated target feature subset, which includes: The spatial position of the crack development probability distribution is compared with the calibration data of the known crack development location. For each spatial grid node, the difference between the predicted crack development probability value and the actual calibrated crack development state is calculated as the prediction error. The prediction error of all spatial grid nodes is organized in the spatial coordinate system to obtain the spatial distribution of the prediction deviation. For each deviation location in the spatial distribution of the prediction deviation, the geomechanical constraint parameter value corresponding to the deviation location is extracted, and the correlation strength between the geomechanical constraint parameter value and the prediction error at the deviation location is calculated as the contribution rate of the geomechanical constraint to the prediction error. Based on the contribution rate, the geomechanical constraints that have the greatest impact on the prediction error are identified. A parameter space search is performed on the fracture development mechanism parameters in the identified geomechanical constraints. By calculating the contribution trend of each feature in the fracture development feature set to fracture prediction under different parameter combinations, a parameter combination that makes the contribution distribution negatively correlated with the spatial distribution of the prediction deviation is selected. Based on the selected parameter combination, the dynamic selection is performed to obtain the updated target feature subset.

6. The method according to claim 1, characterized in that, Based on the updated target feature subset, a new mapping relationship is established from the updated target feature subset to the crack development probability, and a corrected crack development probability distribution is generated as the crack prediction result, including: The updated feature vector is obtained by extracting the feature values ​​of all features in the updated target feature subset at each spatial grid node in the target reservoir area, keeping the training labels in the calibration data of the known fracture development location unchanged, and reconstructing the updated training sample set with the updated feature vector and the training labels. The updated training sample set is nonlinearly mapped and the updated crack development probability prediction value is output. The residual between the updated crack development probability prediction value and the training label is calculated and compared with the prediction deviation before the feedback adjustment. When the residual is less than the prediction deviation before the feedback adjustment, the feedback adjustment is confirmed to be effective and the updated target feature subset is fixed. The parameters of the nonlinear mapping are readjusted to establish the updated mapping relationship from the updated target feature subset to the crack development probability. The updated mapping relationship outputs a corresponding corrected fracture development probability value for each spatial grid node. The corrected fracture development probability values ​​of all spatial grid nodes are organized in the spatial coordinate system of the target reservoir region to form the corrected fracture development probability distribution. The corrected fracture development probability distribution is used as the fracture prediction result.

7. A crack prediction system based on adaptive feature selection, used to implement the method as described in any one of claims 1-6, characterized in that, include: The feature extraction unit is used to acquire multi-source geological data of the target reservoir area and perform multi-scale decomposition to extract a fracture development feature set of fracture development characteristics. A dynamic selection unit is used to calculate the contribution of each feature in the fracture development feature set to fracture prediction based on the geomechanical constraints of reservoir fracture development, and to dynamically select the fracture development feature set according to the contribution to obtain a target feature subset. The mapping establishment unit is used to establish a mapping relationship from the target feature subset to the fracture development probability by using the target feature subset as an input variable and combining it with the calibration data of known fracture development locations. The mapping relationship outputs the fracture development probability distribution of each spatial location in the target reservoir area. The feedback adjustment unit is used to compare the fracture development probability distribution with the fracture distribution data revealed by actual drilling to obtain the prediction deviation, adjust the geomechanical constraints based on the prediction deviation, and update the composition of the target feature subset. The feedback adjustment quantifies the contribution rate of different geomechanical constraints to the prediction error and adaptively optimizes the dynamic selection to obtain the updated target feature subset. The result correction unit is used to re-establish the mapping relationship from the updated target feature subset to the crack development probability based on the updated target feature subset, and generate a corrected crack development probability distribution as the crack prediction result.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images