CN-121996974-A - Method for constructing hydrate reservoir fluid-solid production multi-field coupling prediction model

CN121996974ACN 121996974 ACN121996974 ACN 121996974ACN-121996974-A

Abstract

The invention relates to the technical field of intelligent exploitation of hydrates, in particular to a method for constructing a multi-field coupling prediction model of hydrate reservoir fluid solid output, which comprises the steps of establishing a coupling mathematical model covering hydrate decomposition, multiphase seepage and sediment constitutive relation, adopting an unstructured grid discrete reservoir and utilizing an alternate iterative algorithm to generate evolution data such as pore pressure, sand output and the like, then constructing a topological graph based on the unstructured grid, assigning a physical field state as an initial node characteristic, constructing a graph neural network, extracting local embedded characteristics fusing physical properties and a spatial structure through a graph convolution layer, constructing a long-term memory network processing characteristic sequence to generate a time sequence implicit vector representing a fluid solid migration state, and finally utilizing a loss function joint training network containing prediction errors and physical constraint terms to establish a nonlinear mapping and generate a prediction model with physical consistency.

Inventors

LEI GANG
LU HONGZHI
LIU TIANLE
LI XIAODONG
REN JIANFEI
Fahart Urakh

Assignees

广州南沙地大滨海研究院

Dates

Publication Date: 20260508
Application Date: 20260130

Claims (9)

1. The method for constructing the hydrate reservoir fluid-solid output multi-field coupling prediction model is characterized by comprising the following steps of: s1, establishing a coupling mathematical model covering hydrate decomposition, multiphase seepage, heat conduction and sediment constitutive relation, and adopting an unstructured grid discrete reservoir; S2, solving the coupling mathematical model by using an alternate iterative algorithm, and generating evolution data of pore pressure, temperature, saturation of each phase, solid displacement, corresponding accumulated gas yield and accumulated sand yield of grid nodes of the unstructured grid under continuous time steps; S3, constructing a topological graph based on the unstructured grid, mapping the grid nodes into graph nodes, mapping geometrical adjacency relations among the grid nodes into graph edges, and assigning the pore pressure, the temperature, the saturation of each phase and the solid displacement as initial node characteristics; S4, constructing a graph neural network, performing message transmission and aggregation on the initial node characteristics through a graph convolution layer, and extracting local embedded characteristics fusing physical attributes and spatial structures; S5, constructing a long-term and short-term memory network, receiving the time sequence of the local embedded feature, and generating a time sequence implicit vector representing the fluid-solid migration state through a gating mechanism; S6, utilizing a loss function containing a prediction error and a physical constraint term to jointly train the graph neural network and the long-term and short-term memory network, establishing nonlinear mapping of the time sequence implicit vector and the accumulated gas yield and the accumulated sand yield, and generating a prediction model.
2. The method for constructing a multi-field coupling prediction model for hydrate reservoir fluid production according to claim 1, wherein in step S1, the step of establishing a coupled mathematical model covering hydrate decomposition, multiphase seepage, heat conduction and sediment constitutive relations comprises: Defining hydrate decomposition rate influenced by saturation and contact area, and establishing mass source expression of solid-to-fluid phase conversion; combining the relative permeability and capillary pressure function to construct a multiphase fluid mass conservation partial differential equation set containing gravity and pressure gradients; Describing skeleton stress response by adopting an elastoplastic model, and constructing a fluid-solid output flux model by utilizing a plastic strain and flow rate combined criterion.
3. The method for constructing a hydrate reservoir fluid production multi-field coupling prediction model according to claim 1, wherein in step S1, the step of using an unstructured grid discrete reservoir comprises: unstructured discretization is carried out on the reservoir geometrical domain by utilizing a space subdivision algorithm, and local grid encryption is carried out in a region with a physical field gradient exceeding a preset threshold value; generating a spatial attribute field by using an interpolation algorithm based on geological logging data, and projecting continuous geological attributes to grid nodes of the unstructured grid; and converting the coupled mathematical model into an algebraic equation set by using a numerical discrete method, and processing the flux conservation of the unstructured grid interface.
4. The method for constructing a hydrate reservoir fluid production multi-field coupling prediction model according to claim 1, wherein in step S2, the step of solving the coupling mathematical model by using an alternate iterative algorithm includes: Setting a time stepping parameter, and constructing a nonlinear solver using the last time stepping solution as an initial value; the fluid and the temperature field are solved preferentially by adopting a segmentation iteration strategy, and then the parameters of the solid mechanics and the chemical field are updated; and calculating residual norms of each physical field among iterations, outputting state vectors when the residual meets preset tolerance, and advancing time steps.
5. The method for constructing a hydrate reservoir fluid production multi-field coupling prediction model according to claim 1, wherein in step S3, the step of constructing a topological graph based on the unstructured grid includes: combining the normalized multi-physical field state data with the grid nodes as vertexes to serve as an initial node attribute vector; establishing a connected edge based on the topological connection relation among the grid nodes to form an adjacent matrix for describing space connectivity; And calculating a weight coefficient based on the space distance and the conduction characteristic between the adjacent graph nodes, and representing the interaction strength between grids.
6. The method for constructing a hydrate reservoir fluid production multi-field coupling prediction model according to claim 1, wherein in step S4, the step of performing message passing and aggregation on the initial node features through a graph convolution layer includes: performing nonlinear transformation on the source node characteristics and the edge weights to generate a message vector containing local physical gradients; Calculating influence coefficients of the neighborhood nodes by using an attention mechanism, and weighting and converging message vectors on the adjacent edges; and fusing the aggregation characteristics and the historical characteristics thereof, and generating a new node embedded vector containing space topology information through an update network.
7. The method for constructing a multi-field coupling prediction model for fluid-solid production of a hydrate reservoir according to claim 1, wherein in step S5, the step of generating a time sequence implicit vector representing the fluid-solid migration state by a gating mechanism includes: Setting a cyclic structure comprising multiple gating units to process the time series of locally embedded features; selectively forgetting the historical hidden state by using a gating operator, and carrying out weighted fusion on the current input characteristics; and performing nonlinear mapping on the updated cell state, and outputting the time sequence implicit vector reflecting the evolution rule of the fluid-solid migration.
8. The method for constructing a multi-field coupling prediction model for the solid production of a hydrate reservoir according to claim 1, wherein in step S6, the step of training in combination by using a loss function including a prediction error and a physical constraint term comprises: Comparing the model predicted value with corresponding accumulated gas yield and accumulated sand yield in the evolution data, and constructing a basic loss item for representing the fitting precision of the data; Applying punishment to the difference value against the physical rule in the prediction sequence based on the physical monotonicity characteristics of the accumulated gas yield and the accumulated sand yield; and utilizing the self-adaptive weight to balance the basic loss term and the physical constraint term, and constructing a comprehensive objective function.
9. The method for constructing a multi-field coupling prediction model for the solid production of a hydrate reservoir according to claim 1, wherein in step S6, the step of establishing a nonlinear mapping between the timing implicit vector and the accumulated gas production and the accumulated sand production comprises: mapping the timing implication vector to a low-dimensional throughput space using a fully connected network; calculating the gradient of the comprehensive objective function relative to the trainable parameters of the model based on the automatic differential frame; And synchronously updating weights of the graph neural network and the long-term and short-term memory network according to the gradient by utilizing an adaptive optimization strategy.

Description

Method for constructing hydrate reservoir fluid-solid production multi-field coupling prediction model Technical Field The invention relates to the technical field of intelligent exploitation of hydrates, in particular to a method for constructing a multi-field coupling prediction model for the solid production of a hydrate reservoir. Background The natural gas hydrate is used as a clean substitute energy source with great potential, and the exploitation of the natural gas hydrate relates to multi-physical field strength coupling mechanisms such as phase change decomposition, multiphase seepage, heat conduction, sediment skeleton mechanical response and the like. The method has the advantages of accurately predicting the accumulated gas yield and the sand yield in the whole mining period, and has decisive significance for evaluating the capacity potential, making a scientific mining scheme and preventing and controlling the geological disasters of the shaft. Along with the development of artificial intelligence technology, the deep learning model is utilized to mine complex reservoir evolution data, and the deep learning model becomes a mainstream technology trend for replacing the traditional high-time-consumption numerical simulation and realizing high-efficiency capacity real-time prediction. However, the existing deep learning method is limited by a regular grid architecture, the irregular boundary of a reservoir and local encryption requirements are difficult to flexibly adapt, the pure data driving model lacks the constraint of physical conservation law, and non-monotonic output errors against physical facts often occur in long-period prediction, so that the generalization capability and the physical consistency of the model are insufficient. Disclosure of Invention In order to make up for the defects, the invention provides a method for constructing a multi-field coupling prediction model of hydrate reservoir fluid solid output, and aims to solve the problems that the existing method is difficult to adapt to unstructured grids and the prediction precision and consistency are insufficient due to lack of physical constraints. The invention provides a method for constructing a multi-field coupling prediction model for hydrate reservoir fluid and solid output, which comprises the following steps: s1, establishing a coupling mathematical model covering hydrate decomposition, multiphase seepage, heat conduction and sediment constitutive relation, and adopting an unstructured grid discrete reservoir; S2, solving the coupling mathematical model by using an alternate iterative algorithm, and generating evolution data of pore pressure, temperature, saturation of each phase, solid displacement, corresponding accumulated gas yield and accumulated sand yield of grid nodes of the unstructured grid under continuous time steps; S3, constructing a topological graph based on the unstructured grid, mapping the grid nodes into graph nodes, mapping geometrical adjacency relations among the grid nodes into graph edges, and assigning the pore pressure, the temperature, the saturation of each phase and the solid displacement as initial node characteristics; S4, constructing a graph neural network, performing message transmission and aggregation on the initial node characteristics through a graph convolution layer, and extracting local embedded characteristics fusing physical attributes and spatial structures; S5, constructing a long-term and short-term memory network, receiving the time sequence of the local embedded feature, and generating a time sequence implicit vector representing the fluid-solid migration state through a gating mechanism; S6, utilizing a loss function containing a prediction error and a physical constraint term to jointly train the graph neural network and the long-term and short-term memory network, establishing nonlinear mapping of the time sequence implicit vector and the accumulated gas yield and the accumulated sand yield, and generating a prediction model. Preferably, in step S1, the step of establishing a coupled mathematical model covering hydrate decomposition, multiphase percolation, thermal conduction and sediment constitutive relations includes: Defining hydrate decomposition rate influenced by saturation and contact area, and establishing mass source expression of solid-to-fluid phase conversion; combining the relative permeability and capillary pressure function to construct a multiphase fluid mass conservation partial differential equation set containing gravity and pressure gradients; Describing skeleton stress response by adopting an elastoplastic model, and constructing a fluid-solid output flux model by utilizing a plastic strain and flow rate combined criterion. Preferably, in step S1, the step of using an unstructured grid discrete reservoir includes: unstructured discretization is carried out on the reservoir geometrical domain by utilizing a space subdivision algorithm, and local