| The coupled multi-physical field models such as free flow-porous media flow are widely used in fluid calculation and simulation for reservoir simulation.How to accurately predict the physical fields by using the existing physical information of the models has been one of the great importance problems in fluid mechanics calculations.The reservoir production capacity is affected by multiple factors such as geology and engineering,and conventional production prediction methods can hardly reflect its real production characteristics.Most of the existing prediction models are calculated based on existing equations or formulas,which mostly lack the application and correction of actual data.In this thesis,deep learning is applied to the coupled free flow-porous media flow model and shale reservoir production prediction model,an efficient numerical simulation method is established for the coupled multi-physical field model,basic theoretical analysis is given and supported,and numerical simulation is performed.In addition,starting from the practical data,the influence of parameters on reservoir capacity is analyzed,and the future capacity is predicted based on the existing data to guide the actual production planning and provide a reference for the capacity prediction of shale reservoirs in China.The main work and contributions of this thesis are as follows:1)Deep learning method for Stokes model of incompressible flowFor the steady-state and unsteady-state forms of the Stokes equation for incompressible flow in the free flow region,a deep learning method is proposed to solve and improve the theoretical analysis,which consolidates the mathematical basis of deep neural networks to solve such problems.Firstly,a deep learning algorithm for the generalized constant Stokes equation is proposed based on the deep Galerkin method,and the theoretical basis of deep neural networks for solving computational hydrodynamics equations is discussed,extending the experimental calculations to high dimensions and achieving competitive results with traditional methods.Secondly,the generalized constant deep learning algorithm is extended to the non-stationary case to design deep learning algorithms capable of handling time-contained models,compiling and designing network structures rationally,learning complex mapping relations,and achieving non-stationary nonlinear predictions.2)Deep learning method for coupled free flow-porous media flow modelFor the free-flow-porous medium flow coupling model,the deep learning algorithm to deal with the complex multi-physical field coupling model is proposed and the theoretical analysis is improved.Firstly,CDNNs(coupled deep neural networks)are proposed for solving the time-varying coupled Stokes-Darcy problem.The method is mesh-free and can be solved independently in parallel,the complex interface conditions are reasonably compiled into the deep neural network to constrain the approximate solution,and the custom loss function ensures the physical properties and energy conservation of the numerical solution.The relevant fundamental theoretical analysis proves the ability of the method to solve the coupled problem,and the numerical experiments further verify the accuracy and effectiveness of the proposed method.Secondly,the free flow-porous medium flow coupling model Stokes Darcy-Forchheimer system is well studied.The relationships and interactions between each physical quantity are considered by CDNNs,convergence analysis is performed to derive relevant theoretical results,and the validity of the conclusions is verified by numerical experiments.Finally,a deep learning method for the solution of steady and unsteady shale oil models are studied,and a mesh-free parallel algorithm is designed to study the shale oil model by reasonably compiling the cross-interface conditions,and numerical experimental simulations are conducted to further illustrate the processing capability and applicability of the deep learning method for such complex coupled models.3)Deep learning method for coupled stochastic free-flow-porous medium flow modelFor the coupled free-flow-porous medium flow model with random field,a deep learning algorithm is proposed to deal with the coupled multi-physical field model with random field,and the theoretical analysis is improved and numerical experimental simulation is carried out.Statistical methods are combined with deep learning methods to study the evolution of physical quantities in the free fluid flow and permeability regions in uncertain systems.Specifically,the uncertain stochastic system is converted into a countably many deterministic system by Monte Carlo method,and the variables of each deterministic system are solved independently and in parallel by combining the CDNNs method.The experiments demonstrate that combining statistical methods with deep neural network methods to study the coupled stochastic multi-physics field model can effectively overcome the limitations of grid segmentation of traditional methods,improve the solution accuracy and efficiency,and provide a heuristic reference for related problems.4)Deep learning prediction method for shale reservoir capacityA deep learning-based prediction method is proposed and numerical experiments are conducted for the shale reservoir time series and the reference capacity data.Firstly,based on the shale oil production data from seven major shale oil and gas producing regions in the United States,we explore the feasibility and applicability of the deep learning method LSTM and the statistical method ARIMA for establishing the capacity prediction model for shale oil time-series data,find the optimal parameters,design the model to process the time-series data,reasonably divide the training set and the prediction set,verify the model performance and accuracy,and further improve the performance of accurately predicting the future time,and then establish the long-and short-term shale reservoir capacity prediction model based on time-series data.Then,based on 14 production wells in Xinjiang J oil field,we study the influence of geological reservoir engineering and other influencing factors on shale oil production capacity,and make model optimization based on various prediction methods to predict the production capacity data with reference.In summary,this thesis makes full use of fluid mechanics,machine learning,numerical simulation and other multidisciplinary theories and techniques to systematically study the deep learning solution method of the coupled free-flow-porous media flow multi-physical field model based on theoretical analysis,establishes a coupled deep neural network for solving the coupled multi-physical field model and a stochastic deep learning algorithm framework for the stochastic coupled model,and establishes a deep learning method for solving the high-dimensional the coupled free-flow-porous media flow.Based on deep learning to predict shale reservoir capacity data,the deep learning method is compared with the statistical method to find the optimal model to predict the time-series capacity data;by comparing multiple prediction model methods to predict the capacity data with reference,the model is improved and optimized to enhance the accuracy and efficiency.A model based on the deep learning method for predicting time-series and parameter-containing capacity data is established to further provide reference and guidance for the research of shale oil capacity prediction in China. |
