Multi-modal data and knowledge hybrid-driven heterogeneous reservoir pressure prediction method

Abstract

The invention relates to a heterogeneous reservoir pressure prediction method based on multi-modal data and knowledge hybrid driving. The method comprises the following steps that a heterogeneous reservoir permeability field and pressure field multi-modal data set is established; establishing a deep neural network model for knowledge guidance, and training the deep neural network model based on the multi-modal data set; and inputting the acquired permeability field and spatio-temporal information into the trained deep neural network model to obtain the predicted oil reservoir pressure. According to the method, the requirement of a traditional deep learning method for the high-cost training data volume is remarkably reduced, the algorithm performance and prediction precision are improved, and the method has important significance for intelligent development of an oil field and improvement of the recovery ratio.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to a method for predicting heterogeneous reservoir pressure driven by a hybrid of multi-modal data and knowledge. Background Art

[0002] Reservoir pressure is a key factor affecting reservoir phase change behavior and fluid properties. Accurate acquisition of reservoir pressure is of great significance for understanding oil and gas migration laws, analyzing reservoir performance, and improving oilfield recovery. However, after decades of exploitation, most of my country's oil fields have entered the middle and late stages of development. The reservoirs are highly heterogeneous and the reservoir pressure distribution varies greatly, which brings great difficulties to reservoir pressure prediction.

[0003] In the process of oilfield development, the methods of obtaining reservoir pressure can be divided into direct measurement and indirect analysis. The direct measurement method uses permanent downhole pressure sensors to directly measure reservoir pressure, but this measurement method is costly and complicated to operate, and it is impossible to measure the pressure at all locations within the reservoir space. The indirect analysis method uses a reservoir numerical simulator to analyze reservoir pressure, which has become a commonly used reservoir pressure prediction method. However, this analysis method has a long simulation time and large calculation errors, and can usually only be used for reservoir pressure trend analysis. With the rapid development of artificial intelligence technology, some experts and scholars have chosen machine learning methods to solve the problem of reservoir pressure prediction, but these purely data-driven machine learning methods have not incorporated long-term accumulated empirical knowledge and physical laws, resulting in poor interpretability and generalization of the model. In addition, the rapidly growing parameter scale of the pure data-driven model increases its dependence on data, but it is usually difficult to obtain a large amount of high-fidelity reservoir pressure data in oil and gas production.

[0004] Based on the above situation, there is an urgent need for a new method to accurately predict the pressure of heterogeneous reservoirs. The new method combines the oil and gas seepage knowledge with the multimodal data of oilfield production to achieve accurate prediction of the pressure distribution of different heterogeneous reservoirs and solve the problems of high data dependence, poor generalization ability and unexplainableness of traditional pure data-driven methods. Summary of the invention

[0005] In view of the shortcomings of the existing technology, the present invention proposes a method for predicting heterogeneous reservoir pressure driven by a hybrid of multimodal data and knowledge. This method introduces oil and gas seepage knowledge into a deep neural network to guide the training process of the neural network, and integrates the multimodal data monitored during the oil field production process to achieve rapid and accurate prediction of heterogeneous reservoir pressure in small sample scenarios.

[0006] The technical solution adopted by the present invention to achieve the above-mentioned purpose is: a method for predicting heterogeneous reservoir pressure driven by multimodal data and knowledge, comprising the following steps:

[0007] Step 1: Establish a multimodal dataset of permeability and pressure fields of heterogeneous reservoirs;

[0008] Step 2: Establish a knowledge-guided deep neural network model and train the deep neural network model based on a multimodal dataset;

[0009] Step 3: Input the collected permeability field and spatiotemporal information into the trained deep neural network model to obtain the predicted reservoir pressure.

[0010] The step 1, establishing a multi-modal data set of permeability field and pressure field of a heterogeneous reservoir, comprises the following steps:

[0011] Step 1.1, obtaining permeability field and pressure field data of M different oil layers in a heterogeneous oil reservoir, wherein the permeability field and pressure field of each oil layer are a set of raw data, and the raw data of M groups of different oil layers constitute a training data set for training a deep neural network;

[0012] Among them, the permeability field is the image data containing the spatial distribution information of the permeability, and the pressure field is the time series data containing the temporal and spatial information; and all the pressure field data in the training data set are normalized;

[0013] Step 1.2, obtaining the permeability field data and pressure field data of Z different oil layers in the heterogeneous oil reservoir, the permeability and pressure of each oil layer are a set of raw data, and the Z sets of raw data of different oil layers constitute a test data set, which is used to test the performance of the trained neural network;

[0014] Among them, the permeability field is the image data containing the spatial distribution information of the permeability, the pressure field is the time series data containing the temporal and spatial information, and all the pressure field data in the test data set are normalized;

[0015] The training dataset and the test dataset constitute a multimodal dataset.

[0016] The step 2, establishing a knowledge-guided deep neural network model, and training the deep neural network model based on a multimodal data set, includes the following steps:

[0017] Step 2.1, establishing a knowledge-guided deep neural network model, including a convolutional neural network and a fully connected neural network; wherein the input of the convolutional neural network is the permeability field, and the features of the permeability field are extracted, and the extracted features are obtained after data fusion with the pressure field as the input of the fully connected neural network, and the output of the fully connected neural network is the reservoir pressure;

[0018] Step 2.2, embedding the oil and gas seepage knowledge into the loss function of the deep neural network model to guide the training process of the deep neural network model and accelerate the convergence of the model;

[0019] The oil and gas seepage knowledge includes a control equation, an initial condition equation and a boundary condition equation, and the corresponding loss functions are a control equation loss function, an initial condition loss function and a boundary condition loss function respectively;

[0020] Step 2.3: Input the normalized training data set into the deep neural network model to train the parameters of the deep neural network model.

[0021] The total loss function of the deep neural network model is as follows:

[0022]

[0023] Among them, the control equation loss function is:

[0024]

[0025] The initial condition loss function is:

[0026]

[0027] The boundary condition loss function is:

[0028]

[0029] The observed data loss function is:

[0030]

[0031] Among them, N pde represents the number of samples of the control equation, N ic Represents the number of samples of initial conditions, N bc The number of samples representing the boundary conditions, N data represents the number of samples of observed data, K(x,y) represents the permeability at any point (x,y) in the reservoir, S is the coefficient, p i Represents the predicted pressure value corresponding to the data matching point, N obs represents the number of observed samples; p ic,i represents the actual pressure value of the point corresponding to the initial condition, p bc,i Indicates the actual pressure value at the point corresponding to the boundary condition, p obs,i represents the actual pressure value corresponding to the data matching point, ε pde , ε ic , ε bc and ε data are the weight coefficients of each loss function term, and the adaptive weight updating method is used to update each weight coefficient.

[0032] The method of updating each weight by using an adaptive weight updating method comprises the following steps:

[0033] (1) For the deep neural network model driven by knowledge and data, initialize the adaptive weight coefficient set ε = {ε pde ,ε ic ,ε bc ,ε data};

[0034] (2) Establish a Gaussian probability model, whose mean is obtained by the reservoir pressure p and the adaptive weight set ε output by the deep neural network;

[0035] (3) Use K gradient descent to iteratively update the weight parameter ε in the Gaussian probability model and the deep neural network parameter θ;

[0036] When the adaptive weight update method does not converge, the total loss function Loss of the adaptive weight is defined based on the maximum likelihood estimation, and the Adam optimizer is used to adjust the adaptive weight ε and update the deep neural network parameters θ;

[0037] The optimization process ends when the adaptive weight update method converges.

[0038] The method of defining an adaptive weight loss function based on maximum likelihood estimation and using an Adam optimizer to adjust the adaptive weight ε and update the deep neural network parameter θ includes the following steps:

[0039] (a) Define the total loss function Loss of the deep neural network based on maximum likelihood estimation;

[0040] (b) Use the Adam optimizer to adjust the adaptive weight ε to maximize the probability of satisfying the constraints:

[0041] ε k+1 ←Adam(L(ε k θ k ; N)); k represents the kth iteration, k∈[1,K];

[0042] (c) Update the deep neural network parameters θ using the Adam optimizer:

[0043] θ k+1 ←Adam(L(ε k θ k ; N));

[0044] Get the optimal deep neural network model parameters θ * and the adaptive weight parameter ε * .

[0045] The loss function of the deep neural network is as follows:

[0046]

[0047] Where ε={ε pde ,ε ic ,ε bc ,ε data} is the adaptive weight of the loss function, θ is the deep neural network parameter; N pde Represents the number of samples of the control equation, N ic Represents the number of samples of initial conditions, N bc The number of samples representing the boundary conditions, N data Represents the number of samples of observed data, MSE pde , MSE bc , MSE ic , MSE data They represent the control equation loss function, boundary condition loss function, initial condition loss function, and observation data loss function respectively.

[0048] The step 2 is to establish a knowledge-guided deep neural network model, and after training the deep neural network model based on the multimodal data set, the deep neural network model is tested using the root mean square error and the square correlation coefficient to evaluate the model accuracy of the deep neural network model.

[0049] The heterogeneous reservoir pressure prediction system driven by multimodal data and knowledge includes:

[0050] A dataset construction module is used to build a multimodal dataset of permeability and pressure fields of heterogeneous reservoirs;

[0051] A deep neural network model training module is used to establish a knowledge-guided deep neural network model and train the deep neural network model based on a multimodal data set;

[0052] The pressure prediction module is used to input the collected permeability field and spatiotemporal information into the trained deep neural network model to obtain the predicted reservoir pressure.

[0053] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the heterogeneous reservoir pressure prediction method driven by a hybrid of multimodal data and knowledge is implemented.

[0054] The present invention has the following beneficial effects and advantages:

[0055] The present invention provides a method for predicting heterogeneous reservoir pressure driven by a hybrid of multimodal data and knowledge. Based on the oil and gas seepage theory, a deep neural network model driven by the fusion of prior knowledge and multimodal data is established to achieve efficient and accurate prediction of reservoir pressure in a small sample scenario. This method targets reservoir dynamic parameters that are expensive to acquire, and uses the oil and gas seepage theory as a guide to establish an accurate deep learning prediction model using only a small amount of observational data. This method significantly reduces the demand for high-cost training data for traditional deep learning methods, improves algorithm performance and prediction accuracy, and is of great significance for intelligent development of oil fields and improving oil recovery. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 This is a flow chart of a method for predicting pressure in a heterogeneous reservoir driven by a combination of multimodal data and knowledge according to the present invention;

[0057] Figure 2 This is a diagram of the heterogeneous reservoir pressure prediction result according to an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0059] A method for predicting pressure in heterogeneous reservoirs driven by multimodal data and knowledge. Figure 1 As shown, the following steps are included:

[0060] Step 1: Establish a multimodal dataset of permeability and pressure fields in a heterogeneous reservoir.

[0061] Step 1.1: Obtain the permeability field and pressure field data of M different oil layers in the heterogeneous reservoir. The permeability field and pressure field (including time, position, and pressure data) of each oil layer are a set of raw data. The M sets of raw data of different oil layers constitute the training data set for training the deep neural network. Among them, the permeability field is the image data containing the spatial distribution information of the permeability, and the pressure field is the time series data containing the spatiotemporal information, including time, position, and pressure data. All the pressure field data in the training data set are standardized using the maximum and minimum normalization method, and the formula is in is the standardized pressure data, P is the original pressure data, maxP is the maximum value in the original pressure data, and minP is the minimum value in the original pressure data.

[0062] Step 1.2: Obtain the permeability field data and pressure field data of Z different oil layers in the heterogeneous reservoir. The permeability and pressure of each oil layer are a set of raw data. The Z sets of raw data of different oil layers constitute the test data set, which is used to test the performance of the trained neural network. Among them, the permeability field is the image data containing the spatial distribution information of the permeability, and the pressure field is the time series data containing the spatiotemporal information. All the pressure field data in the test data set are normalized.

[0063] In this embodiment, a heterogeneous reservoir with different permeability field distributions is used, and the training data set selects M=30 groups of permeability field data and pressure field data of different oil layers; in order to verify the accuracy and scalability of the proposed knowledge and data hybrid driven heterogeneous reservoir pressure prediction method, the test data set selects Z=200 groups of permeability field data and pressure field data of different oil layers. Among them, a group of data in the data set contains multiple samples.

[0064] Step 2: Establish a knowledge-guided deep neural network model and train the established model using the standardized training data set in step 1.1.

[0065] Step 2.1: Establish a knowledge-guided deep neural network model, which includes a convolutional neural network and a fully connected neural network, wherein the convolutional neural network is used to extract image features of the permeability field. The extracted image features are fused with spatiotemporal information (including time and location) and used as the input of the fully connected neural network. The output of the model is the reservoir pressure. The convolutional neural network consists of an input layer, a convolutional layer, a fully connected layer and an output layer, wherein the number of convolutional layers is 6 and the number of fully connected layers is 1. The convolutional layers of the convolutional neural network model all use a 3×3 convolution kernel, and the activation function selects the Swish function. The fully connected neural network model consists of an input layer, a fully connected layer and an output layer, wherein the number of fully connected layers is 4. The activation function of the fully connected layer of the fully connected neural network model selects the Swish function. The formula of the Swish function is:

[0066] Swish(x)=x*Sigmoid(x)

[0067] Here, x is the input value.

[0068] Step 2.2: Embed the oil and gas seepage knowledge into the loss function of the neural network to guide the training process of the neural network and accelerate the convergence of the model. The embedded oil and gas seepage knowledge includes the control equation, the initial condition equation and the boundary condition equation. The corresponding loss functions are the control equation loss function, the initial condition loss function and the boundary condition loss function. The total loss function is:

[0069]

[0070] Among them, the control equation loss function is:

[0071]

[0072] The initial condition loss function is:

[0073]

[0074] The boundary condition loss function is:

[0075]

[0076] The observed data loss function is:

[0077]

[0078] N pde represents the number of samples of the control equation, N ic Represents the number of samples of initial conditions, N bc The number of samples representing the boundary conditions, N data represents the number of samples of observed data, K(x,y) represents the permeability at any point (x,y) in the reservoir, S is the coefficient, p i Represents the predicted pressure value corresponding to the data matching point, N obs represents the number of observed samples; o ic,i represents the actual pressure value of the point corresponding to the initial condition, p bc,i Indicates the actual pressure value at the point corresponding to the boundary condition, p obs,i represents the actual pressure value corresponding to the data matching point, ε pde , ε ic , ε bc and ε data are the weights of each loss function term, and the adaptive weight update method is used to update each weight term. Based on the Gaussian probability model and maximum likelihood estimation, the loss function of the deep neural network can be expressed as:

[0079]

[0080] Where ε={ε pde ,ε ic ,ε bc ,ε data} is the adaptive weight of the loss function. By minimizing the loss function L(ε; θ; N), the optimal deep neural network model parameter θ can be obtained. * and the adaptive weight parameter ε *

[0081] The optimization process of the weight adaptive update method:

[0082] (1) For a deep neural network driven by a mixture of knowledge and data, initialize the adaptive weight set ε = {ε pde ,ε ic ,ε bc ,ε data};

[0083] (2) Establish a Gaussian probability model whose mean is given by the output of the deep neural network and the adaptive weight set γ;

[0084] (3) Use K gradient descent iterations to update the weight parameters ε and deep neural network parameters θ:

[0085] When the algorithm does not converge:

[0086] (a) Define the adaptive weight loss function based on maximum likelihood estimation

[0087] (b) Use the Adam optimizer to adjust the adaptive weight ε to maximize the probability of satisfying the constraints:

[0088] ε k+1 ←Adam(L(ε k θ k ;N))

[0089] (c) Update the deep neural network parameters θ using the Adam optimizer:

[0090] θ k+1 ←Adam(L(ε k θ k ;N))

[0091] Get the best model parameters θ * and the adaptive weight parameter ε * ;

[0092] (4) The optimization process ends.

[0093] Step 2.3: Input the standardized training data set in step 1.1 into the deep neural network model to train the parameters of the neural network. The neural network parameters that need to be trained include the parameters of the convolutional neural network and the parameters of the fully connected neural network. The two network parameters are trained simultaneously, so that the fully connected network can not only extract the characteristics of the permeability field, but also include the distribution information of the pressure field.

[0094] Step 3: Input the standardized test data set in step 1.2 into the trained deep neural network model to predict reservoir pressure, using the root mean square error (RMSE) and squared correlation coefficient (R 2 ) is used to evaluate the prediction accuracy of the model, and the formula is as follows:

[0095]

[0096] Among them, N test is the number of samples in the test set, is the predicted value of the i-th sample in the test set, p i is the true value of the i-th sample in the test set, is the average value of samples in the test set.

[0097] Step 4: Input the collected permeability field and spatiotemporal information (including time and location) into the trained deep neural network model to obtain the predicted reservoir pressure.

[0098] In this example, the above data and model parameters are used to train the model to obtain the prediction results R for 200 test set data. 2 The average value is 99.63%, and the average RMSE value is 0.034MPa. The comparison results between the actual value and the predicted value are as follows: Figure 2 shown.

Claims

1. A method for predicting pressure in heterogeneous reservoirs driven by hybrid multimodal data and knowledge, characterized in that: The following steps are involved: Step 1: Establish a multimodal dataset of permeability and pressure fields of heterogeneous reservoirs; Step 2: Establish a knowledge-guided deep neural network model and train the deep neural network model based on a multimodal dataset; Step 3: Input the collected permeability field and spatiotemporal information into the trained deep neural network model to obtain the predicted reservoir pressure.

2. The method for predicting heterogeneous reservoir pressure driven by hybrid multimodal data and knowledge according to claim 1 is characterized in that: The step 1: establishing a multi-modal data set of permeability field and pressure field of a heterogeneous reservoir, comprises the following steps: Step 1.1, obtaining permeability field and pressure field data of M different oil layers in a heterogeneous oil reservoir, wherein the permeability field and pressure field of each oil layer are a set of raw data, and the raw data of M groups of different oil layers constitute a training data set for training a deep neural network; Among them, the permeability field is the image data containing the spatial distribution information of the permeability, and the pressure field is the time series data containing the temporal and spatial information; and all the pressure field data in the training data set are normalized; Step 1.2, obtaining the permeability field data and pressure field data of Z different oil layers in the heterogeneous oil reservoir, the permeability and pressure of each oil layer are a set of raw data, and the Z sets of raw data of different oil layers constitute a test data set, which is used to test the performance of the trained neural network; Among them, the permeability field is the image data containing the spatial distribution information of the permeability, the pressure field is the time series data containing the temporal and spatial information, and all the pressure field data in the test data set are normalized; The training dataset and the test dataset constitute a multimodal dataset.

3. The method for predicting heterogeneous reservoir pressure driven by hybrid multimodal data and knowledge according to claim 1 is characterized in that: The step 2, establishing a knowledge-guided deep neural network model, and training the deep neural network model based on a multimodal data set, includes the following steps: Step 2.1, establishing a knowledge-guided deep neural network model, including a convolutional neural network and a fully connected neural network; wherein the input of the convolutional neural network is the permeability field, and the features of the permeability field are extracted, and the extracted features are obtained after data fusion with the pressure field as the input of the fully connected neural network, and the output of the fully connected neural network is the reservoir pressure; Step 2.2, embedding the oil and gas seepage knowledge into the loss function of the deep neural network model to guide the training process of the deep neural network model and accelerate the convergence of the model; The oil and gas seepage knowledge includes a control equation, an initial condition equation and a boundary condition equation, and the corresponding loss functions are a control equation loss function, an initial condition loss function and a boundary condition loss function respectively; Step 2.3: Input the normalized training data set into the deep neural network model to train the parameters of the deep neural network model.

4. The method for predicting heterogeneous reservoir pressure driven by hybrid multimodal data and knowledge according to claim 3 is characterized in that: The total loss function of the deep neural network model is as follows: Among them, the control equation loss function is: The initial condition loss function is: The boundary condition loss function is: The observed data loss function is: Among them, N pde represents the number of samples of the control equation, N ic Represents the number of samples of initial conditions, N bc The number of samples representing the boundary conditions, N data represents the number of samples of observed data, K(x,y) represents the permeability at any point (x,y) in the reservoir, S is the coefficient, p i Represents the predicted pressure value corresponding to the data matching point, N obs represents the number of observed samples; [ ic,i represents the actual pressure value of the point corresponding to the initial condition, [ bc,i Indicates the actual pressure value at the point corresponding to the boundary condition, p obs,i represents the actual pressure value corresponding to the data matching point, ε pde , ε ic , ε bc and ε data are the weight coefficients of each loss function term, and the adaptive weight updating method is used to update each weight coefficient.

5. The method for predicting heterogeneous reservoir pressure driven by hybrid multimodal data and knowledge according to claim 4 is characterized in that: The method of updating each weight by using an adaptive weight updating method comprises the following steps: (1) For the deep neural network model driven by knowledge and data, initialize the adaptive weight coefficient set ε = {ε pde ,ε ic ,ε bc ,ε data }; (2) Establish a Gaussian probability model, whose mean is obtained by the reservoir pressure p and the adaptive weight set ε output by the deep neural network; (3) Use K gradient descent to iteratively update the weight parameter ε in the Gaussian probability model and the deep neural network parameter θ; When the adaptive weight update method does not converge, the total loss function Loss of the adaptive weight is defined based on the maximum likelihood estimation, and the Adam optimizer is used to adjust the adaptive weight ε and update the deep neural network parameters θ; The optimization process ends when the adaptive weight update method converges.

6. The method for predicting heterogeneous reservoir pressure driven by hybrid multimodal data and knowledge according to claim 5 is characterized in that: The method of defining an adaptive weight loss function based on maximum likelihood estimation and using an Adam optimizer to adjust the adaptive weight ε and update the deep neural network parameter θ includes the following steps: (a) Define the total loss function Loss of the deep neural network based on maximum likelihood estimation; (b) Use the Adam optimizer to adjust the adaptive weight ε to maximize the probability of satisfying the constraints: ε k+1 ←Adam(L(ε k θ k ; N)); k represents the kth iteration, k∈[1,K]; (c) Update the deep neural network parameters θ using the Adam optimizer: i k+1 ←Adam(L(e k ;θ k ;N)); Get the optimal deep neural network model parameters θ * and the adaptive weight parameter ε * .

7. The method for predicting heterogeneous reservoir pressure driven by hybrid multimodal data and knowledge according to claim 6 is characterized in that: The loss function of the deep neural network is as follows: Where ε={ε pde ,ε ic ,ε bc ,ε data } is the adaptive weight of the loss function, θ is the deep neural network parameter; N pde represents the number of samples of the control equation, N ic Represents the number of samples of initial conditions, N bc The number of samples representing the boundary conditions, N data Represents the number of samples of observed data, MSE pde , MSE bc , MSE ic , MSE data They represent the control equation loss function, boundary condition loss function, initial condition loss function, and observation data loss function respectively.

8. The method for predicting heterogeneous reservoir pressure driven by hybrid multimodal data and knowledge according to claim 1, characterized in that: The step 2 is to establish a knowledge-guided deep neural network model, and after training the deep neural network model based on the multimodal data set, the deep neural network model is tested using the root mean square error and the square correlation coefficient to evaluate the model accuracy of the deep neural network model.

9. A heterogeneous reservoir pressure prediction system driven by multimodal data and knowledge, characterized in that: include: A dataset construction module is used to build a multimodal dataset of permeability and pressure fields of heterogeneous reservoirs; A deep neural network model training module is used to establish a knowledge-guided deep neural network model and train the deep neural network model based on a multimodal data set; The pressure prediction module is used to input the collected permeability field and spatiotemporal information into the trained deep neural network model to obtain the predicted reservoir pressure.

10. A computer-readable storage medium, characterized in that: The storage medium stores a computer program, and when the computer program is executed by a processor, the method for predicting heterogeneous reservoir pressure driven by a hybrid of multimodal data and knowledge as described in any one of claims 1 to 8 is implemented.

Images