Novel method for constructing oil reservoir simulation agent model based on transfer learning

Abstract

The invention discloses a new method for constructing an oil reservoir simulation agent model based on transfer learning, and relates to the technical field of petroleum engineering, and the method comprises the steps: building a fine-scale oil reservoir model; constructing a high-fidelity data set; carrying out scale upgrading and coarsening on the oil reservoir model; constructing a low-fidelity data set; constructing a deep learning model; training a pre-training model by using the low-fidelity data; finely tuning the pre-training model by using the high-fidelity data; and predicting output parameters. According to the method, it can be guaranteed that under the condition that a small amount of high-fidelity training data is used, oil reservoir multi-term flow under the variable well control condition can be accurately predicted, dependence on the high-fidelity training data in the agent model building process is effectively reduced, and the agent model building time is shortened on the premise that prediction precision is not affected. According to the method, the embedded module is introduced to process the time-varying well control parameters, the problem that matrix element parameters corresponding to well positions are difficult to effectively diffuse to the whole output result through finite times of convolution is solved, and the global influence of the well control parameters on the output result of the proxy model is improved.

Description

Technical Field

[0001] The present invention relates to the field of petroleum engineering technology, and in particular to a new method for constructing an oil reservoir simulation agent model based on transfer learning. Background Art

[0002] Accurately characterizing underground fluid flow is crucial for the development of oil, gas, and groundwater resources. Due to the complexity of geological structures and multiphase flows of oil, gas, and water, high-precision numerical simulation techniques are usually required to characterize underground fluid flow. In practical engineering problems, numerical simulations often require expensive computing resources. Recent studies have shown that proxy models built based on deep learning methods can effectively replace the numerical simulation process. Proxy models can be used to solve geological uncertainty quantification problems and predict the evolution of flow fields under different geological models. They can also be used to deal with production optimization problems and predict the evolution of flow fields under time-varying well control.

[0003] Existing methods usually require a large amount of high-fidelity training data to build proxy models, and the generation of these training data itself consumes a lot of computing resources and takes a long time to calculate, which limits the application of proxy models in practical engineering problems. Transfer learning methods can use existing knowledge to accelerate the learning process of new tasks and reduce the time consumption of training data generation and model training. In the process of building proxy models, the processing of well control information is often represented by a single-channel two-dimensional or three-dimensional matrix and used as the input feature of the deep learning model, where the matrix elements corresponding to the production wells are assigned negative well control values, while the injection wells are assigned positive well control values, and the remaining matrix elements are set to zero. The well control information input matrix is ​​a sparse matrix. In the deep learning model, it is difficult to effectively diffuse the matrix element parameters corresponding to the well position to the entire output result after a limited number of convolutions.

[0004] Therefore, it is urgent to study a new method to construct a reservoir simulation agent model based on transfer learning. Summary of the invention

[0005] To solve the above technical problems, the present invention discloses a new method for constructing a reservoir simulation proxy model based on transfer learning, which can reduce the dependence on high-fidelity training data during the proxy model construction process and reduce the proxy model construction time without affecting the prediction accuracy.

[0006] To achieve the above object, the present invention adopts the following technical solutions:

[0007] A new method for constructing a reservoir simulation proxy model based on transfer learning includes the following steps:

[0008] S1. Establish fine-scale reservoir model;

[0009] S2, build a high-fidelity dataset;

[0010] S3, scale-up and coarsening of reservoir model;

[0011] S4, build low-fidelity dataset;

[0012] S5. Build a deep learning model;

[0013] S6. Train the pre-trained model with low-fidelity data;

[0014] S7. Fine-tune the pre-trained model with high-fidelity data;

[0015] S8. Predict output parameters.

[0016] Optionally, in step S1, a fine-scale reservoir model is constructed according to actual geological conditions using reservoir numerical simulation software. The original fine-scale reservoir simulation is a Cartesian grid model established according to actual data, and the permeability field is generated by a sequential Gaussian simulation method.

[0017] Optionally, in step S2, the step of constructing a high-fidelity data set specifically includes: randomly perturbing well control parameters, using numerical simulation software to perform a small amount of simulation calculations on the fine-scale reservoir model, and generating high-fidelity input and output parameters required for model training; according to the iterative mechanism of the previous and next time steps in the numerical simulation solution process, selecting the permeability distribution, well location and well control distribution, and the pressure and saturation distribution at the current moment as the input parameters of the model, and the pressure and saturation distribution at the next moment as the output parameters of the model.

[0018] Optionally, in step S3, the step of scaling up and coarsening the reservoir model specifically includes: using scaling means to perform network coarsening on the fine-scale reservoir model to obtain a coarse-scale reservoir model.

[0019] Optionally, in step S4, the step of constructing a low-fidelity data set specifically includes: randomly perturbing well control parameters, performing a large number of simulation calculations on the coarse-scale reservoir model using numerical simulation software, and generating low-fidelity input and output parameters required for model training.

[0020] Optionally, in step S5, the step of constructing a deep learning model specifically includes: establishing a deep learning model based on a U-Net neural network, and improving the basic U-Net network to add an embedding module to process time-varying well control parameters, in the embedding module, the well control parameters are replicated and convolved and embedded in the deep learning model; in the constructed deep learning model, the input parameters are processed by the encoding module and enter the latent space, and are fused with the well control parameters input by the embedding module in the latent space, and the output parameters are reconstructed by the decoding module to obtain the final prediction result.

[0021] Optionally, in step S6, the step of training the pre-trained model with low-fidelity data specifically includes: using the large amount of low-fidelity training data obtained in step S4 to train the deep learning model established in step S5, setting hyperparameters in the process, and using an optimization algorithm to minimize the loss function, so that the deep learning model can establish a mapping relationship between low-fidelity input and output, and obtain a trained pre-trained model.

[0022] Optionally, in step S7, the step of fine-tuning the pre-trained model with high-fidelity data specifically includes: fine-tuning the pre-trained model obtained in step S6 using a small amount of high-fidelity training data obtained in step S2, and the fine-tuning process adopts step-by-step fine-tuning, first fine-tuning the input, output, decoding and embedding modules, and then fine-tuning all parameters, so that the model becomes a proxy model capable of characterizing fine-scale reservoir fluid flow. The deep learning model parameters of the pre-trained model constructed using low-fidelity data are expressed as Among them, θ input ,θ enc ,θ emb ,θ dec ,θ output Represent the network parameters of the input, encoding, embedding, decoding and output modules respectively, and l represents low fidelity; after preliminary fine-tuning, where h represents high fidelity; finally, after fine-tuning all parameters,

[0023] Optionally, in step S8, the step of predicting the output parameters specifically includes: inputting the new input parameters into the proxy model obtained in step S7, and the model gives corresponding calculation results to achieve rapid prediction of the output parameters. The proxy model calculation process is expressed as: In the formula represents the reservoir pressure and saturation distribution at time n+1, u n+1 is the well control measure at time n+1, is the pressure and saturation distribution at time n, is the surrogate model, and θ is an adjustable parameter in the surrogate model.

[0024] The beneficial effects of the present invention are:

[0025] (1) The present invention defines a new method for constructing a reservoir simulation proxy model based on transfer learning combined with multi-fidelity training data. This method can ensure accurate prediction of multiple reservoir flows under variable well control conditions while using a small amount of high-fidelity training data, effectively reducing the dependence on high-fidelity training data during the proxy model construction process, and reducing the proxy model construction time without affecting the prediction accuracy.

[0026] (2) The invention proposes to introduce an embedded module to process time-varying well control parameters, which solves the problem that it is difficult to effectively diffuse the matrix element parameters corresponding to the well locations to the entire output result after a finite number of convolutions, and effectively improves the global impact of the well control parameters on the output results of the proxy model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 Simulate permeability and well location maps for fine-scale reservoirs;

[0028] Figure 2 This is the EU-Net neural network architecture diagram proposed in the present invention;

[0029] Figure 3 This is a structural diagram of the embedded module proposed in the present invention;

[0030] Figure 4 Schematic diagram of the transfer learning training method based on a multi-fidelity training data set proposed in the present invention;

[0031] Figure 5 It is a schematic diagram of the prediction process of the proxy model proposed in the present invention;

[0032] Figure 6 The results of the water saturation field and pressure field predicted by the numerical simulation and the proxy model invented in this paper and the absolute error distribution diagram;

[0033] Figure 7 It is a flow chart of the method for constructing a proxy model based on transfer learning combined with multi-fidelity training data of the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0035] A new method for constructing reservoir simulation proxy models based on transfer learning, such as Figure 7 As shown, the following steps are included:

[0036] S1. Establish fine-scale reservoir model

[0037] A fine-scale reservoir model was constructed using a reservoir simulation numerical simulator. The fine-scale reservoir model contains 200×200 networks, each with a grid size of 18×18×10ft, 1 injection well, and 3 production wells. All wells in the reservoir are controlled by bottom hole pressure. The initial oil saturation and water saturation of the reservoir are 0.8 and 0.2, respectively. The porosity is a constant of 0.15. The initial reservoir pressure is 6000psi. The distribution of permeability and well location is shown in the figure. Figure 1 shown.

[0038] S2. Build a high-fidelity dataset

[0039] 100 groups of well control parameters are randomly generated, each group generates 20 random well control parameters, corresponding to 20 time steps, and each time step is simulated for 30 days. The bottom hole pressure of the injection well ranges from 8000 to 9500 psi, and the bottom hole pressure of the production well ranges from 3500 to 5000 psi.

[0040] For the seepage problem of oil and water two-phase underground fluid, the basic seepage differential equation is expressed as: in, represents the divergence operator, represents the gradient operator, j represents the phase state including oil and water, ρ j represents the j-phase density, μ j represents the j-phase viscosity, k represents the porous medium permeability, k rj represents the relative permeability of phase j, g represents the gravitational acceleration, z represents the depth, and p j represents the j-phase pressure, q j represents the source and sink phases, φ represents the porosity of the porous medium, S j Represents the saturation of phase j. The seepage differential equation is solved by numerical simulation. According to the iterative mechanism of the previous and next time steps in the numerical simulation solution process, the permeability distribution, well location and well control distribution, pressure and saturation distribution at the current moment are selected as the input parameters of the model, and the pressure and saturation distribution at the next moment are selected as the output parameters of the model.

[0041] S3. Upscaling and roughening the reservoir model

[0042] Grid-based upscaling was adopted. After the upscaling, the reservoir model contained 40×40 grids, each with a size of 90×90×10ft. Other settings were the same as those of the fine-scale reservoir model.

[0043] S4. Constructing low-fidelity datasets

[0044] 400 sets of well control parameters are randomly generated, 20 random well control parameters are generated each time, corresponding to 20 time steps, and each time step is simulated for 30 days. The selection of pressure range and input and output parameters is consistent with step S2.

[0045] S5. Build a deep learning model

[0046] A deep learning model is established based on the U-Net neural network, and an embedded module is added to the basic U-Net network to process time-varying well control parameters. The schematic diagram of the improved EU-Net is shown in the figure below. Figure 2 In the embedding module, the well control parameters are replicated and convolved into the deep learning model. The schematic diagram of the embedding module is shown in Figure 3 shown.

[0047] In the constructed deep learning module, the input parameters are processed by the encoding module and enter the latent space. The tensor generated by the encoding module is represented as In the latent space, the well control parameters are integrated with the embedding module input. w Well simulation problem for a well, well control information at a single moment After replication and convolution, the output is x e ×y e ×n c Tensor, processed well control information And the output E of the encoding module is integrated into a dimension x e ×y e ×(n e +n c ), which is a tensor of well control information and other input information after convolution operation. The output parameters are reconstructed by the decoding module to obtain the final prediction result.

[0048] S6. Training pre-trained models with low-fidelity data

[0049] The deep learning model established in step S5 is trained using a large amount of low-fidelity training data obtained in step S4, and the hyperparameters in the process are set. The optimization algorithm is used to minimize the loss function so that the deep learning model can establish a mapping relationship between low-fidelity input and output. During the model training process, the optimization algorithm uses the Adam algorithm, the learning rate is set to 0.0001, and the number of iterations is set to 300.

[0050] S7. Fine-tune the pre-trained model with high-fidelity data

[0051] The pre-trained model obtained in step S6 is fine-tuned using a small amount of high-fidelity training data obtained in step S1, so that the model can characterize the fine-scale reservoir fluid flow. The fine-tuning process adopts step-by-step fine-tuning, first fine-tuning the input, output, decoding and embedding modules, and then fine-tuning all parameters. The fine-tuning process is as follows: Figure 5 shown.

[0052] For the pre-trained model built using low-fidelity data, the deep learning model parameters are expressed as Among them, θ input ,θ enc ,θ emb ,θ dec ,θ output Represent the network parameters of the input, encoding, embedding, decoding and output modules respectively, and l represents low fidelity; after preliminary fine-tuning, where h represents high fidelity; finally, after fine-tuning all parameters,

[0053] S8. Prediction output parameters

[0054] The new input parameters are input into the proxy model obtained in step S7. The model can give the corresponding calculation results and realize the rapid prediction of the output parameters. The prediction process is as follows: Figure 4 The output results of the model prediction process are used as input parameters for the next moment, and continuous prediction of time-varying well control pressure and saturation is achieved through continuous iteration.

[0055] Figure 6 The pressure and saturation distribution (P,S) calculated using the numerical simulator for a randomly generated well control parameter in this example are shown. w ) and the pressure and saturation calculated using the proxy model Distribution diagram and error distribution diagram between the two It can be seen that the proxy model constructed using the present invention can accurately predict the flow of reservoir fluid under variable well control conditions. In addition, from the efficiency comparison, compared with the fine-scale reservoir simulation, a single simulation takes 97.23 seconds, and for the coarse-scale reservoir model, a single simulation only takes 9.42 seconds. Using multi-fidelity data to build a model will greatly reduce the dependence on fine-scale reservoir simulation and reduce the time consumption caused by training data generation during model construction, indicating that the present invention accelerates the construction process of the proxy model.

[0056] In view of the problem that conventional reservoir simulation proxy models require a large number of fine-scale reservoir simulations, the present invention adopts a method of constructing proxy models by combining transfer learning with multi-fidelity training data, thereby reducing the reliance on fine-scale reservoir simulations during the proxy model construction process. In view of the time-varying well control problem, an innovative method of using embedded module processing is proposed to increase the prediction accuracy of the model. At the same time, compared with reservoir simulation proxy models constructed by other methods, the prediction accuracy and computational efficiency of the present invention are significantly superior. The present invention can establish a proxy model that accurately predicts reservoir fluid flow and reduce the time cost during the proxy model construction process, providing technical support for the application of proxy models in practical engineering problems.

[0057] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A new method for constructing a reservoir simulation agent model based on transfer learning, characterized in that: The steps include: S1. Establish fine-scale reservoir model; S2, build a high-fidelity dataset; S3, scale-up and coarsening of reservoir model; S4, build low-fidelity dataset; S5. Build a deep learning model; S6. Train the pre-trained model with low-fidelity data; S7. Fine-tune the pre-trained model with high-fidelity data; S8. Predict output parameters.

2. A new method for constructing a reservoir simulation agent model based on transfer learning as claimed in claim 1, characterized in that: In step S2, the steps of constructing a high-fidelity data set specifically include: randomly perturbing well control parameters, using numerical simulation software to perform a small amount of simulation calculations on the fine-scale reservoir model, and generating high-fidelity input and output parameters required for model training; according to the iterative mechanism of the previous and next time steps in the numerical simulation solution process, the permeability distribution, well location and well control distribution, and the pressure and saturation distribution at the current moment are selected as the input parameters of the model, and the pressure and saturation distribution at the next moment are selected as the output parameters of the model.

3. A new method for constructing a reservoir simulation agent model based on transfer learning as claimed in claim 2, characterized in that: In step S3, the step of scaling up and coarsening the reservoir model specifically includes: using scaling means to perform network coarsening on the fine-scale reservoir model to obtain a coarse-scale reservoir model.

4. A new method for constructing a reservoir simulation agent model based on transfer learning as claimed in claim 3, characterized in that: In step S4, the step of constructing a low-fidelity data set specifically includes: randomly perturbing well control parameters, using numerical simulation software to perform a large number of simulation calculations on the coarse-scale reservoir model, and generating low-fidelity input and output parameters required for model training.

5. A new method for constructing a reservoir simulation agent model based on transfer learning as claimed in claim 4, characterized in that: In step S5, the step of constructing a deep learning model specifically includes: establishing a deep learning model based on a U-Net neural network, and improving the basic U-Net network to add an embedding module to process time-varying well control parameters. In the embedding module, the well control parameters are replicated and convolved and embedded in the deep learning model; in the constructed deep learning model, the input parameters are processed by the encoding module and enter the latent space, and are fused with the well control parameters input by the embedding module in the latent space, and the output parameters are reconstructed by the decoding module to obtain the final prediction result.

6. A new method for constructing a reservoir simulation agent model based on transfer learning as claimed in claim 5, characterized in that: In step S6, the step of training the pre-trained model with low-fidelity data specifically includes: using a large amount of low-fidelity training data obtained in step S4 to train the deep learning model established in step S5, setting hyperparameters in the process, and using an optimization algorithm to minimize the loss function, so that the deep learning model can establish a mapping relationship between low-fidelity input and output, and obtain a trained pre-trained model.

7. A new method for constructing a reservoir simulation agent model based on transfer learning as claimed in claim 6, characterized in that: In step S7, the step of fine-tuning the pre-trained model with high-fidelity data specifically includes: fine-tuning the pre-trained model obtained in step S6 using a small amount of high-fidelity training data obtained in step S2, and the deep learning model parameters of the pre-trained model are expressed as Among them, θ input ,θ enc ,θ emb ,θ dec ,θ output They represent the network parameters of the input, encoding, embedding, decoding and output modules respectively, l represents low fidelity, and the fine-tuning process adopts step-by-step fine-tuning. First, the input, output, decoding and embedding modules are fine-tuned. After preliminary fine-tuning, Among them, h represents high fidelity. Finally, after fine-tuning all parameters, The model becomes a proxy model capable of representing fine-scale reservoir fluid flow.

8. A new method for constructing a reservoir simulation agent model based on transfer learning as claimed in claim 7, characterized in that: In step S8, the step of predicting the output parameters specifically includes: inputting the new input parameters into the proxy model obtained in step S7, and the model gives the corresponding calculation results to achieve rapid prediction of the output parameters; the proxy model calculation process is expressed as: In the formula, represents the reservoir pressure and saturation distribution at time n+1, u n+1 is the well control measure at time n+1, is the pressure and saturation distribution at time n, g~ is the proxy model, and θ is the adjustable parameter in the proxy model.

Images