CN-121069524-B - Geomechanical modeling method, system and medium based on artificial intelligence

CN121069524BCN 121069524 BCN121069524 BCN 121069524BCN-121069524-B

Abstract

The application relates to a geomechanical modeling method, a geomechanical modeling system and a geomechanical modeling medium based on artificial intelligence, which comprise the steps of carrying out space-time alignment and feature decoupling on continuous field data and discrete structure data to generate continuous-discrete mixed feature tensors with unified space-time references, constructing a multi-scale physical coupling mixed neural network, realizing cross-scale bidirectional interaction through a self-adaptive physical interface to obtain a dynamic coupling geomechanical prediction model, carrying out parameter optimization on the geomechanical model by adopting a physical-data dual-driven dynamic training strategy, carrying out stratum parameter on-line inversion on the basis of real-time monitoring data and the trained geomechanical prediction model to generate an uncertainty quantization result, mapping the uncertainty quantization result to a three-dimensional geological model, and outputting a visual risk thermodynamic diagram. The application not only realizes the real-time updating and uncertainty evaluation of the geological parameters, but also provides a quantization basis for engineering decision-making in an intuitive visual mode, and enhances the scientificity and reliability of the decision-making.

Inventors

XIONG JIAN
LIU XINYU
LIU XIANGJUN
QIN JIANHUA
LIANG LIXI
WAN YOUWEI
DING YI
WEI XIAOCHEN

Assignees

西南石油大学

Dates

Publication Date: 20260505
Application Date: 20250801

Claims (9)

1. The geomechanical modeling method based on artificial intelligence is characterized by comprising the following steps of: S101, performing space-time alignment and feature decoupling processing on the acquired continuous field data and discrete structure data to generate continuous-discrete mixed feature tensors with unified space-time references; s102, constructing a multi-scale physical coupling hybrid neural network based on the continuous-discrete hybrid feature tensor, and realizing cross-scale bidirectional interaction through a self-adaptive physical interface to obtain a dynamically coupled geomechanical prediction model, wherein the hybrid neural network comprises a macroscopic continuous field modeling sub-network based on a field map convolution network and a microscopic crack dynamic sub-network based on a topology enhancement map network; S103, performing parameter optimization on the geomechanical prediction model by adopting a physical-data dual-drive dynamic training strategy to obtain a trained geomechanical prediction model; s104, carrying out formation parameter on-line inversion based on real-time monitoring data and a trained geomechanical prediction model through a variation self-encoder to generate an uncertainty quantization result, wherein the uncertainty quantization result comprises the fracture probability and the stress concentration standard deviation of spatial distribution; S105, mapping an uncertainty quantification result to a three-dimensional geological model, and outputting a visual risk thermodynamic diagram, wherein the three-dimensional geological model comprises a stratum interface topological structure and lithology distribution parameters; The step S103 includes the steps of: Acquiring a training data set; Performing constraint processing of a momentum conservation equation on macroscopic stress field distribution to generate a first physical loss term, performing constraint processing of a mass conservation equation on a microscopic crack opening change path to generate a second physical loss term, and combining the first physical loss term and the second physical loss term to obtain physical conservation loss; Comparing the macroscopic stress field distribution with the mean square error processing of the actual measurement data of the historical underground strain gauge to generate a first data loss item, comparing the Hausdorff distance processing of the microscopic crack opening change path and the historical microseismic positioning result data to generate a second data loss item, and combining the first data loss item and the second data loss item to obtain data matching loss; Performing inverse proportion adjustment on the physical loss weight of the physical conservation loss and the data loss weight of the data matching loss based on the training period number to obtain a dynamic weight; And superposing the physical conservation loss and the data matching loss according to dynamic weights to generate a total loss function, and carrying out iterative optimization on parameters of the geomechanical prediction model by adopting a self-adaptive learning rate strategy until the total loss function meets a preset convergence condition, and outputting the trained geomechanical prediction model.
2. The artificial intelligence based geomechanical modeling method of claim 1, wherein the continuous field data comprises seismic inversion velocity field data and InSAR deformation field data, and the discrete structure data comprises fracture topology map, fault point cloud data and microseismic event point cloud data.
3. The geomechanical modeling method based on artificial intelligence according to claim 2, wherein S101 comprises the steps of: Performing time stamp resampling on the seismic wave inversion speed field data, and performing moving average filtering on the InSAR deformation field data to obtain time-aligned corrected continuous field data; Performing Delaunay triangulation processing on the fault point cloud data, and constructing a three-dimensional unstructured grid to obtain corrected fault point cloud data; performing density clustering on the microseismic event point cloud data by adopting a clustering algorithm, and deleting outliers to obtain corrected microseismic event point cloud data; Based on the corrected fault point cloud data and the corrected microseismic event point cloud data, obtaining corrected discrete structure data matched with the corrected continuous field data space reference through space interpolation gridding processing; According to the corrected continuous field data, a double-channel residual convolution neural network is adopted to extract strain energy density gradient characteristics; performing feature extraction on the corrected discrete structure data by adopting a graph neural network to obtain fracture connectivity features; and performing channel splicing and normalization processing according to the strain energy density gradient characteristics and the fracture connectivity characteristics to obtain the continuous-discrete mixed characteristic tensor.
4. The geomechanical modeling method based on artificial intelligence according to claim 1, wherein S102 comprises the steps of: Inputting continuous field features in the continuous-discrete mixed feature tensor into a field graph convolution network, embedding a discrete form of a preset momentum conservation equation into the field graph convolution network to serve as a physical residual layer, carrying out convolution processing on the continuous field features under physical constraint, and outputting macroscopic stress field distribution to obtain a macroscopic continuous field modeling sub-network; inputting discrete structural features in the continuous-discrete mixed feature tensor into a topological enhancement graph network, defining a crack propagation criterion function in the topological enhancement graph network, calculating a crack opening change path according to the stress transfer relation between the crack node attribute and the adjacent node, and outputting a micro-crack dynamic parameter to obtain a micro-crack dynamic sub-network; Based on a self-adaptive physical interface, the micro-crack dynamic sub-network and the macroscopic continuous field modeling sub-network are interacted; inputting dynamic parameters of the micro cracks into a rigidity degradation model, calculating local rigidity degradation coefficients, feeding the local rigidity degradation coefficients back to a grid rigidity matrix of a field map convolution network, and updating parameters of an elastic matrix to obtain a geomechanical prediction model.
5. The artificial intelligence based geomechanical modeling method of claim 1, wherein the first physical loss term is generated by calculating a deviation between a model predicted macroscopic stress field and a conservation of momentum equation; generating a second physical loss term by calculating the deviation between the crack opening change path predicted by the model and the mass conservation equation; Comparing the macroscopic stress field predicted by the model with the actual measurement data of the historical underground strain gauge at the same spatial position and time point, calculating the average value of square differences between the macroscopic stress field predicted by the model and the actual measurement data of the historical underground strain gauge, and generating a first data loss item; and calculating the Hausdorff distance between the crack opening change path predicted by the model and the historical microseismic positioning result data, and generating a second data loss term.
6. The artificial intelligence based geomechanical modeling method of claim 1, wherein S104 comprises the steps of: Acquiring real-time monitoring data, inputting the real-time monitoring data into a variation self-encoder for parameter inversion processing, and generating posterior probability distribution of stratum parameters, wherein the real-time monitoring data comprises underground manometer monitoring data and microseismic event monitoring data, and the stratum parameters comprise rock elastic modulus, fracture toughness and pore pressure; Randomly sampling a probability density function of posterior probability distribution to generate a rock elastic modulus sampling value, a fracture toughness sampling value and a pore pressure sampling value which meet the distribution characteristics of the posterior probability distribution; Combining the rock elastic modulus sampling value, the fracture toughness sampling value and the pore pressure sampling value according to the space grid unit of the trained geomechanical prediction model to generate a plurality of groups of stratum parameter combinations; and (3) inputting a plurality of groups of stratum parameters into a trained geomechanical prediction model for fracture probability statistical treatment, and outputting an uncertainty quantization result.
7. The artificial intelligence based geomechanical modeling method of any of claims 1-6, wherein S105 comprises the steps of: constructing a three-dimensional geological model based on the space grid units in the modified discrete structure data; Mapping the cracking probability into a transparency channel by adopting a gradient sensitive rendering technology, and mapping the standard deviation of the stress concentration into the intensity of a red channel to obtain an initial thermodynamic diagram; Mapping the energy release rate in the microseismic event point cloud data into the point cloud size, and performing space alignment superposition on the initial thermodynamic diagram and the microseismic event point cloud with the size attribute to generate a dynamic risk thermodynamic diagram; and carrying out edge detection processing on the dynamic risk thermodynamic diagram according to a preset risk threshold value to generate an avoidance area boundary polygon, and overlapping the avoidance area boundary polygon to the three-dimensional geological model to output a visual risk thermodynamic diagram.
8. An artificial intelligence based geomechanical modeling system for implementing an artificial intelligence based geomechanical modeling method according to any of claims 1-7, comprising: the space-time alignment and feature decoupling module is used for performing space-time alignment and feature decoupling processing on the acquired continuous field data and discrete structure data to generate continuous-discrete mixed feature tensors with unified space-time references; the multi-scale neural network construction module is used for constructing a multi-scale physical coupling hybrid neural network based on the continuous-discrete hybrid characteristic tensor, realizing cross-scale bidirectional interaction through a self-adaptive physical interface and obtaining a dynamically coupled geomechanical prediction model; The physical-data driven dynamic training module is used for optimizing parameters of the geomechanical model by adopting a physical-data double-driven dynamic training strategy to obtain a trained geomechanical prediction model The parameter inversion and risk visualization module is used for carrying out stratum parameter on-line inversion based on real-time monitoring data and a trained geomechanical prediction model through the variation self-encoder to generate an uncertainty quantification result, mapping the uncertainty quantification result to a three-dimensional geological model and outputting a visualized risk thermodynamic diagram.
9. A computer readable storage medium storing a computer program, characterized in that the computer program when executed by a processor implements the steps of an artificial intelligence based geomechanical modeling method according to any of claims 1-7.

Description

Geomechanical modeling method, system and medium based on artificial intelligence Technical Field The invention relates to the technical field of geomechanics, in particular to a geomechanical modeling method, system and medium based on artificial intelligence. Background Geomechanical modeling is a core technology in the fields of oil and gas exploration, geothermal development, geological disaster early warning and the like, and can predict the rock breaking risk under engineering activities such as hydraulic fracturing by dynamically simulating stratum stress, strain and cracks. However, the conventional geomechanical modeling method often relies on geomechanical theory and numerical simulation technology such as finite element and discrete element methods to perform geomodeling, and has many defects in treating the coupling problem of complex geological structures and multiple physical fields. On one hand, the method is difficult to effectively integrate multi-source heterogeneous geological data, such as continuous stress field data and seepage field data and discrete fault, crack and other structural data, so that deviation exists between a model and actual geological conditions. On the other hand, the method adopts traditional numerical simulation to simulate, wherein physical factors can be considered by embedding a physical equation, but the method has high calculation complexity, relies on manually set boundary conditions, and is difficult to adapt to real-time monitoring data for dynamic updating. And pure data driven models such as deep learning models, while being able to fit data efficiently, lack constraints on the laws of physical conservation, thereby causing the predicted outcome to be distorted in physical consistency. Disclosure of Invention The invention provides a geomechanical modeling method, a geomechanical modeling system and a geomechanical modeling medium based on artificial intelligence for solving the technical problems. The invention is realized by the following technical scheme: The application provides a geomechanical modeling method based on artificial intelligence, which comprises the following steps: S101, performing space-time alignment and feature decoupling processing on the acquired continuous field data and discrete structure data to generate continuous-discrete mixed feature tensors with unified space-time references; s102, constructing a multi-scale physical coupling hybrid neural network based on the continuous-discrete hybrid feature tensor, and realizing cross-scale bidirectional interaction through a self-adaptive physical interface to obtain a dynamically coupled geomechanical prediction model, wherein the hybrid neural network comprises a macroscopic continuous field modeling sub-network based on a field map convolution network and a microscopic crack dynamic sub-network based on a topology enhancement map network; s103, performing parameter optimization on the geomechanical model by adopting a physical-data dual-drive dynamic training strategy to obtain a trained geomechanical prediction model; s104, carrying out formation parameter on-line inversion based on real-time monitoring data and a trained geomechanical prediction model through a variation self-encoder to generate an uncertainty quantization result, wherein the uncertainty quantization result comprises the fracture probability and the stress concentration standard deviation of spatial distribution; and S105, mapping the uncertainty quantification result to a three-dimensional geological model, and outputting a visual risk thermodynamic diagram, wherein the three-dimensional geological model comprises a stratum interface topological structure and lithology distribution parameters. The continuous field data comprise seismic wave inversion speed field data and InSAR deformation field data, and the discrete structure data comprise a fracture topological graph, fault point cloud data and microseismic event point cloud data. Optionally, the step S101 includes the following steps: Performing time stamp resampling on the seismic wave inversion speed field data, and performing moving average filtering on the InSAR deformation field data to obtain time-aligned corrected continuous field data; Performing Delaunay triangulation processing on the fault point cloud data, and constructing a three-dimensional unstructured grid to obtain corrected fault point cloud data; performing density clustering on the microseismic event point cloud data by adopting a clustering algorithm, and deleting outliers to obtain corrected microseismic event point cloud data; Based on the corrected fault point cloud data and the corrected microseismic event point cloud data, obtaining corrected discrete structure data matched with the corrected continuous field data space reference through space interpolation gridding processing; According to the corrected continuous field data, a double-channel residual convolution neural network is ad