A domain decomposition-based multi-scale intelligent simulation method for geothermal development scenarios

Abstract

The present invention provides a domain decomposition-based multi-scale intelligent simulation method for geothermal development scenarios, belonging to the field of intelligent simulation. The method includes: obtaining geothermal resource data according to the geothermal development scenario, constructing a multi-scale simulation model, and using a deep neural network to obtain the deep feature representation of the multi-scale simulation model; performing feature fusion on the deep feature representation to obtain a common subspace; constructing an unsupervised clustering neural network model to perform expansion processing on the common subspace, and using a transfer learning algorithm to construct an inter-scale bridging model; using the inter-scale bridging model to obtain a domain decomposition-based multi-scale simulation model and performing simulation experiment verification, and building a simulation platform for geothermal development scenarios. The present invention solves the problems that existing simulation methods cannot fully analyze and utilize the deep features of different-scale simulation models themselves, and in the domain decomposition task, due to the simultaneous existence and overlapping of different-scale simulation regions, it is difficult for the simulation model to cross-scale transitions.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent simulation, and particularly relates to a domain decomposition-based multi-scale intelligent simulation method for geothermal development scenarios. Background Art

[0002] Geothermal resources widely exist in the earth's crust and are a clean, stable and renewable energy source. Geothermal development mainly includes exploration, drilling, resource extraction, utilization and management of geothermal resources. Geothermal modeling refers to the mathematical simulation of the formation, distribution, flow, extraction and utilization processes of geothermal resources to optimize development strategies and predict development effects. Traditional modeling methods use numerical methods such as the finite element method and finite difference method to solve key numerical values such as temperature and pressure distributions in the geothermal field. Physics knowledge-driven deep learning is an effective tool for computer simulation, but most existing methods are limited to single-scale scenarios described by simple models. During the process of deep geothermal exploitation, high-pressure fluids are injected underground through deep boreholes, forcing the micro-nano pores in the target rock layer to expand to several millimeters. If the finally formed giant fracture network is considered, the research scope will expand to several kilometers. However, observational and experimental results show that the influence of micro-damage accumulation in rocks on fracture behavior often cannot be accurately characterized by macroscopic continuous models. Therefore, how to develop a general "scale bridging" method for efficient communication of multi-scale information is the key to realizing multi-scale simulation by deep learning.

[0003] With the rapid development of artificial intelligence technology, the use of deep learning methods to achieve intelligent simulation of physical problems has received extensive attention from the academic and industrial communities. Currently, among many attempts to apply deep learning methods to multi-scale simulation, it mainly includes research on combining neural networks with traditional numerical methods, multi-scale model coupling methods based on data-driven or dynamic importance sampling, and work on applying transfer learning to simulation.

[0004] However, existing simulation methods simply rely on the knowledge-data comprehensive driving characteristics of deep learning methods and do not fully analyze and utilize the deep features of different scale simulation models themselves. At the same time, there is no method in the field of deep learning simulation to solve the problem that in the domain decomposition-based multi-scale simulation task, simulation regions of different scales will coexist and overlap with each other, resulting in difficulties in cross-scale transition of the simulation model, nor effectively applying multi-scale models independently to simulation research. Summary of the Invention

[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides a domain decomposition multi-scale intelligent simulation method for geothermal development scenarios, which solves the problem that the existing simulation methods are unable to fully analyze and utilize the deep characteristics of simulation models of different scales themselves, and in the domain decomposition multi-scale simulation tasks, simulation areas of different scales will exist at the same time and overlap with each other, thereby causing difficulties in cross-scale transition of simulation models.

[0006] In order to achieve the above objectives, the technical solution adopted by the present invention is: a domain decomposition multi-scale intelligent simulation method for geothermal development scenarios, comprising the following steps:

[0007] S1. According to the geothermal development scenario, geothermal resource data is obtained, and a multi-scale simulation model is constructed. The deep neural network is used to perform deep feature learning on the multi-scale simulation model to obtain the deep feature representation of the multi-scale simulation model;

[0008] S2. Based on the deep feature representation of the multi-scale simulation model, feature fusion is performed on the cross-scale features of the multi-scale simulation model to obtain a common subspace;

[0009] S3, build an unsupervised clustering neural network model, expand the common subspace, and use the transfer learning algorithm to build an inter-scale bridging model;

[0010] S4, using the inter-scale bridging model to perform domain decomposition transition on the multi-scale simulation model to obtain a domain decomposition multi-scale simulation model;

[0011] S5. Based on the physical process, the domain decomposition multi-scale simulation model is verified by simulation experiments, and the verified domain decomposition multi-scale simulation model is used to build a simulation platform for geothermal development scenarios to realize domain decomposition multi-scale intelligent simulation.

[0012] The beneficial effects of the present invention are as follows: the present invention utilizes physical information neural operators to fully analyze and utilize the deep features of simulation models of different scales themselves, thereby improving the pertinence and feasibility of the simulation of the present invention; by integrating deep feature representations, a multi-scale framework that takes into account both deep feature learning of simulation models and cross-scale feature fusion is realized, enriching the existing multi-scale simulation theory; and by constructing an inter-scale bridging model to perform domain decomposition on the multi-scale simulation model, further mining inter-scale features, achieving relative independence of macro-scale models and micro-scale models, and partially overlapping domain decomposition in the simulation domain, thereby improving the accuracy of the results of multi-scale intelligent simulation and the simulation efficiency.

[0013] Furthermore, the S1 comprises the following steps:

[0014] S101. Obtain geothermal resource data according to geothermal development scenarios;

[0015] S102. Use the prior knowledge of multiple scale models and geothermal resource data in the same physical process to pre-train the physics-informed neural operator, construct an initial multi-scale simulation model, and perform feature embedding on the initial multi-scale simulation model;

[0016] S103. According to the multi-scale simulation model with feature embedding, design a targeted representation learning model, and construct a deep neural network for learning and representation based on the deep features of the multi-scale simulation model;

[0017] S104. Use the deep neural network to learn the multi-scale simulation model with feature embedding, learn the feature space of the multi-scale simulation model, and obtain the deep feature representation of the multi-scale simulation model related to the geothermal reservoir.

[0018] Furthermore, the expression for performing feature embedding on the multi-scale simulation model is as follows:

[0019] ;

[0020] where represents the predicted value of the function represented by the neural network at the position y , represents the feature embedding of the input function , represents the feature embedding of the input coordinate y , q represents the number of basis functions selected in the target space, represents the neural network parameters; where , , represents q the n-dimensional Euclidean real space.

[0021] The beneficial effect of the above further solution is that the present invention eliminates the semantic gap by designing a targeted representation learning model, constructs a deep neural network to learn and represent the deep features of the pre-trained physics-informed neural operator simulation model, improves the correlation between the deep features and the geothermal reservoir, and the established simulation model feature space lays a solid foundation for the layering and domain decomposition coupling of the model.

[0022] Furthermore, the S2 includes the following steps:

[0023] S201. According to the deep feature representation of the multi-scale simulation model, use feature fusion to map the feature space of the multi-scale simulation model to the same common subspace;

[0024] S202. Use the topological manifold structure of the model data to balance the consistent structure and complementary information between the feature representations of simulation models at different scales in the common subspace, and obtain the common subspace with balanced feature representations.

[0025] The beneficial effects of the above further solution are as follows: By using the physical information operator method that develops multi-view feature fusion, the present invention realizes the mapping from the feature spaces of models at different scales to the common subspace, introduces the idea of topological manifolds, balances the consistency and complementarity between different view representations, improves the robustness of the present invention, and bridges the heterogeneous gap.

[0026] Furthermore, the S3 includes the following steps:

[0027] S301. Based on the K-means algorithm, construct an unsupervised clustering neural network model, and use the unsupervised clustering neural network model to mine and understand the manifold information of the inter-scale bridging model;

[0028] S302. Use the unsupervised clustering neural network model to expand the common subspace with balanced feature representations;

[0029] S303. Design a transductive transfer learning algorithm, perform knowledge transfer from the multi-scale simulation model to the inter-scale bridging model on the expanded common subspace, and construct the inter-scale bridging model according to the manifold information mined by the unsupervised clustering neural network model.

[0030] Furthermore, the expression of the unsupervised clustering neural network model is as follows:

[0031] ;

[0032] where, represents the unsupervised clustering neural network model, represents the i th cluster, represents the centroid of the i th cluster, x represents the sample point, represents the sample x and the geodesic distance between the centroid on the manifold.

[0033] Furthermore, the expression of using the inter-scale bridging model to perform domain decomposition transition on the multi-scale simulation model is as follows:

[0034] ;

[0035] ;

[0036] where, represents the transition path, represents a transition factor represents the low-dimensional representation of the macro-scale model in the common subspace represents the low-dimensional representation of the micro-scale model in the common subspace represents a bridging matrix represents bridging weight information represents the cluster center of the macro scale represents the cluster center of the micro scale

[0037] The beneficial effects of the above further solution are as follows: By constructing an unsupervised clustering neural network model, the present invention excavates the manifold information of the inter-scale bridging model, and realizes the use of the manifold information for the expansion of the feature common subspace; on the basis of further expanding the model feature common subspace, by using the expanded space, the present invention realizes the migration of knowledge from the known-scale simulation model to the inter-scale simulation model, solves the domain shift problem, and realizes the multi-scale simulation of complex physical systems in the form of domain decomposition of the present invention; the multi-scale intelligent simulation model realizes domain decomposition multi-scale simulation through the smooth transition of the inter-scale bridging model, and solves the problem of difficult cross-scale transition of the simulation model caused by the simultaneous existence and overlapping of simulation regions of different scales BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is the method flow chart of the present invention

[0039] Figure 2 is the overall structure diagram of the domain decomposition multi-scale intelligent simulation method for the geothermal development scenario in this embodiment DETAILED DESCRIPTION OF THE EMBODIMENTS

[0040] The following describes the specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions and creations using the concept of the present invention are within the scope of protection

[0041] Before describing this embodiment, the following terms are first explained

[0042] PINO: Physics-informed neural operator

[0043] K-means algorithm: K-means algorithm

[0044] Embodiment

[0045] As Figure 1As shown in the figure, the present invention provides a domain decomposition multi-scale intelligent simulation method for geothermal development scenarios, and its implementation method is as follows:

[0046] S1. According to the geothermal development scenario, obtain geothermal resource data, construct a multi-scale simulation model, and use a deep neural network to perform deep feature learning on the multi-scale simulation model to obtain a deep feature representation of the multi-scale simulation model. The specific steps are as follows:

[0047] S101. According to the geothermal development scenario, obtain geothermal resource data;

[0048] S102. Use the prior knowledge of the same physical process in multiple scale models and the geothermal resource data to pre-train the physics-informed neural operator, construct an initial multi-scale simulation model, and perform feature embedding on the initial multi-scale simulation model;

[0049] S103. According to the multi-scale simulation model with feature embedding, design a targeted representation learning model, and construct a deep neural network for learning and representation based on the deep features of the multi-scale simulation model according to the targeted representation learning model;

[0050] S104. Use the deep neural network to learn the multi-scale simulation model with feature embedding, learn the feature space of the multi-scale simulation model, and obtain the deep feature representation of the multi-scale simulation model related to the geothermal reservoir.

[0051] In this embodiment, the deep features of the multi-scale simulation model are related to the long-term evolution of the geothermal reservoir, the cross-scale correlation behavior, and the coupling relationship between different physical fields, etc.;

[0052] As Figure 2 shown, the physical models in the form of differential equations at different scales of the geothermal development scenario can be divided into: a macroscopic scale model (a continuous model based on the conservation laws of mass, momentum, energy, etc.), a mesoscopic scale model (a particle dynamics model simulating the mutual mechanical interaction between rock units), and a microscopic scale model (based on). To realize the learning and representation of the deep features of the multi-scale simulation model, use the prior knowledge of the physical models in the form of differential equations at different scales of the same physical process to pre-train the physics-informed neural operator and establish the simulation models at the corresponding scales; according to the process of the physics-informed neural operator model learning the parameterized partial differential equation solving operator, in the process of constructing the physics-informed neural operator, use the feature embedding based on Fourier features (RFF) to improve the efficiency of the method of the present invention. The calculation formula is as follows:

[0053] ;

[0054] Among them, represents the neural network The predicted value of the represented function at position y , represents the feature embedding of the input function , represents the feature embedding of the input coordinates y , q represents the number of basis functions selected in the target space, represents the parameters of the neural network ; among them, , , represents q d-dimensional Euclidean real space; based on automatic differentiation, a differential operator is constructed to minimize the loss function based on partial differential equations , the loss function based on the initial conditions and the loss function based on the boundary conditions to obtain the optimal parameters , completing the optimization of the physics-informed neural operator.

[0055] In this embodiment, the multi-scale simulation model of the present invention is constructed based on the physics-informed neural operator network; a targeted representation learning model for the PINO simulation model is designed, and a deep neural network for learning and representing the deep features of the model is constructed on the basis of eliminating the semantic gap, thereby learning the feature spaces of the PINO simulation model at multiple scales and obtaining the deep feature representation of the multi-scale simulation model related to the geothermal reservoir.

[0056] S2. According to the deep feature representation of the multi-scale simulation model, perform feature fusion on the cross-scale features of the multi-scale simulation model to obtain a common subspace. The specific steps are as follows:

[0057] S201. According to the deep feature representation of the multi-scale simulation model, use feature fusion to map the feature spaces of the multi-scale simulation model to the same common subspace;

[0058] S202. Use the topological manifold structure of the model data to balance the consistent structure and complementary information between the feature representations of different-scale simulation models in the common subspace, and obtain the common subspace with balanced feature representations.

[0059] In this embodiment, according to the analysis of the tensor-based canonical correlation analysis method (TCCA), a mapping from the feature space of the multi-scale model to the same common subspace is studied and constructed, and the preliminary fusion of the features of the multi-scale simulation model is realized; the overall method spans multiple scales of micro, meso, and macro, extracts the key features of each scale, and then integrates the extracted key features through a fusion mechanism to construct a simulation model with global consistency;

[0060] Specifically, as Figure 2As shown, based on the deep feature representation of the multi-scale simulation model, the learned feature representation of the multi-scale simulation model is used ( p = 1, 2, …, n; n represents the total number of views, which is the total number of different-scale models in this embodiment), calculate the covariance tensor of all views , which is used to discover the correlation information between all views. By adopting simplified calculation, it is transformed into an equivalent problem of maximizing multi-view canonical correlation, that is, to find a set of optimal rank-one tensors ( k = 1, 2, …, r; r represents the dimensionality of the feature after dimensionality reduction, which should be less than the minimum of the feature dimensionalities of different views), so that the covariance tensor can be expressed as a weighted sum of the rank-one tensors ; The calculation expression is as follows:

[0061] ;

[0062] wherein, represents the tensor product, represents; and integrate the rank-one tensors into a transformation matrix , and apply the transformation matrix to the feature representation , initially realizing the mapping from the original high-dimensional features to the low-dimensional common subspace;

[0063] Furthermore, considering that due to the different physical models used in the simulation models of different scales, there may be significant differences in the feature representations. Relying solely on tensor-based canonical correlation maximization to construct the common subspace cannot completely bridge the heterogeneous gap; after combining the research work on multi-view clustering, the idea of topological manifolds is introduced, and the topological manifold structure of the model data is used to balance the consistent structure and complementary information between the feature representations of simulation models of different scales, thereby improving the robustness of the PINO method for multi-view feature fusion and bridging the heterogeneous gap between the feature spaces of different views.

[0064] S3. Construct an unsupervised clustering neural network model to expand the common subspace, and use the transfer learning algorithm to construct an inter-scale bridging model. The specific steps are as follows:

[0065] S301. Based on the K-means algorithm, construct an unsupervised clustering neural network model, and use the unsupervised clustering neural network model to mine and understand the manifold information of the inter-scale bridging model;

[0066] S302. Use the unsupervised clustering neural network model to expand the common subspace with the balanced feature representation;

[0067] S303. Design a direct-push transfer learning algorithm to perform knowledge transfer from a multi-scale simulation model to an inter-scale bridging model on the expanded common subspace, and construct an inter-scale bridging model based on the manifold information mined by the unsupervised clustering neural network model.

[0068] In this embodiment, in constructing the inter-scale bridging simulation model, the direct-push transfer learning algorithm for the common subspace of model features is studied in terms of both expansion and utilization; for expansion, this embodiment uses a targeted unsupervised clustering algorithm based on the K-means algorithm to mine the manifold information of the inter-scale bridging model and apply this to the expansion of the common subspace of model features. The expression of the unsupervised clustering neural network model is as follows:

[0069] ;

[0070] where, represents the unsupervised clustering neural network model, represents the i th cluster, represents the centroid of the i th cluster, x represents the sample point, represents the sample and the geodesic distance between the centroid on the manifold;

[0071] After initially implementing the construction of the unsupervised clustering neural network model, improvements are made from the perspective of utilizing manifold information, and a direct-push transfer learning algorithm is designed to effectively transfer knowledge from the known-scale simulation model to the inter-scale bridging model using the expanded common subspace of model features, obtaining the inter-scale bridging model and realizing the multi-scale simulation of complex physical systems in the form of domain decomposition by the PINO method. The expression of the inter-scale bridging model is as follows:

[0072] ;

[0073] where, represents the bridging matrix, represents the clustering center at the macro scale, represents the clustering center at the micro scale.

[0074] S4. Use the inter-scale bridging model to perform domain decomposition transition on the multi-scale simulation model to obtain a domain-decomposed multi-scale simulation model.

[0075] In this embodiment, the multi-scale intelligent simulation model based on the neural operator network realizes domain-decomposed multi-scale simulation through the smooth transition of the inter-scale bridging model, obtaining a domain-decomposed multi-scale simulation model. The expression is as follows:

[0076] ;

[0077] Among them, and respectively represent the low-dimensional representations of the macroscopic-scale model and the microscopic-scale model in the common subspace. represents the transition path, which is used to define to smooth change, represents the transition factor, and , when at that time, , when at that time, , represents the bridging matrix, represents the bridging weight information.

[0078] S5. Based on the physical process, conduct simulation experiment verification on the domain decomposition multi-scale simulation model, and use the verified domain decomposition multi-scale simulation model to build a simulation platform for geothermal development scenarios to achieve domain decomposition multi-scale intelligent simulation.

[0079] In this embodiment, using the domain decomposition multi-scale simulation model, conduct simulation experiment verification based on typical physical processes, take the deep learning of the domain decomposition multi-scale simulation model as a multi-scale simulation tool, and build a simulation platform for geothermal development scenarios using the multi-scale simulation tool with verified performance to achieve domain decomposition multi-scale intelligent simulation.

[0080] In this embodiment, integrate the deep feature representation of the multi-scale simulation model and the scale-inter bridging model to develop the PINO method for domain decomposition multi-scale simulation.

[0081] This invention is described with reference to the flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to the embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, and the combination of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing devices generate a device for implementing the specified functions in Figure 1 one process or multiple processes and / or blocks Figure 1 one block or multiple blocks.

[0082] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means embodying the function specified in the flowchart Figure 1 a flowchart or multiple flowcharts and / or block Figure 1 a block or multiple blocks.

[0083] These computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, whereby the instructions executed on the computer or other programmable apparatus provide steps for implementing the function specified in the flowchart Figure 1 a flowchart or multiple flowcharts and / or block Figure 1 a block or multiple blocks.

Claims

1. A domain decomposition-based multi-scale intelligent simulation method for geothermal development scenarios, characterized in that It includes the following steps: S1. According to the geothermal development scenario, obtain geothermal resource data, construct a multi-scale simulation model, perform deep feature learning on the multi-scale simulation model using a deep neural network, and obtain the deep feature representation of the multi-scale simulation model; S2. According to the deep feature representation of the multi-scale simulation model, perform feature fusion on the cross-scale features of the multi-scale simulation model to obtain a common subspace; S3. Construct an unsupervised clustering neural network model to perform an expansion process on the common subspace, and use a transfer learning algorithm to construct an inter-scale bridging model. Specifically: S301. Based on the K-means algorithm, construct an unsupervised clustering neural network model, and use the unsupervised clustering neural network model to mine and understand the manifold information of the inter-scale bridging model; S302. Use the unsupervised clustering neural network model to expand the common subspace with the already balanced feature representation; S303. Design a transductive transfer learning algorithm, perform knowledge transfer from the multi-scale simulation model to the inter-scale bridging model on the expanded common subspace, and construct an inter-scale bridging model according to the manifold information mined by the unsupervised clustering neural network model; S4. Use the inter-scale bridging model to perform domain decomposition transition on the multi-scale simulation model to obtain a domain-decomposed multi-scale simulation model; S5. Based on the physical process, perform a simulation experiment verification on the domain-decomposed multi-scale simulation model, and use the verified domain-decomposed multi-scale simulation model to build a simulation platform for the geothermal development scenario to achieve domain-decomposed multi-scale intelligent simulation; The expression of the unsupervised clustering neural network model is as follows: Among them, represents an unsupervised clustering neural network model, represents the i th cluster, represents the i th centroid of the cluster, x represents a sample point, represents the sample x and the centroid on the geodesic distance on the manifold; The expression of using the inter-scale bridging model to perform domain decomposition transition on the multi-scale simulation model is as follows: Among them, represents the transition path, represents the transition factor, represents the low-dimensional representation of the macroscopic scale model in the common subspace, represents the low-dimensional representation of the microscopic scale model in the common subspace, represents the bridging matrix, represents the bridging weight information, represents the cluster center of the macroscopic scale, represents the cluster center of the microscopic scale.

2. The domain decomposition-based multi-scale intelligent simulation method for geothermal development scenarios according to claim 1, wherein The said S1 includes the following steps: S101. According to the geothermal development scenario, obtain geothermal resource data; S102. Use the prior knowledge of the same physical process in multiple scale models and the geothermal resource data to pre-train the physical information neural operator, construct an initial multi-scale simulation model, and perform feature embedding on the initial multi-scale simulation model; S103. According to the multi-scale simulation model with already embedded features, design a targeted representation learning model, and construct a deep neural network for learning and representation based on the deep features of the multi-scale simulation model; S104. Use the deep neural network to learn the multi-scale simulation model with already embedded features, learn the feature space of the multi-scale simulation model, and obtain the deep feature representation of the multi-scale simulation model related to the geothermal reservoir.

3. The domain decomposition-based multi-scale intelligent simulation method for geothermal development scenarios according to claim 2, wherein The expression of performing feature embedding on the multi-scale simulation model is as follows: Among them, represents the predicted value of the function represented by the neural network at the position y ; represents the feature embedding of the input function ; represents the feature embedding of the input coordinate y ; q represents the number of basis functions selected in the target space; represents the neural network parameters; among them, , , represents q dimensional Euclidean real space.

4. The domain decomposition-based multi-scale intelligent simulation method for geothermal development scenarios according to claim 1, wherein The said S2 includes the following steps: S201. According to the deep feature representation of the multi-scale simulation model, use feature fusion to map the feature space of the multi-scale simulation model to the same common subspace; S202. Use the topological manifold structure of the model data to balance the consistent structure and complementary information between the feature representations of different scale simulation models in the common subspace to obtain a common subspace with already balanced feature representation.

Images