Well parameter interpolation method based on neural network
Through the neural network-based well parameter interpolation method, combined with Bayesian convolutional neural network and MC Dropout technology, the problems of uncertainty estimation and construction information processing in well parameter interpolation in the existing technology are solved, and the well parameter interpolation with high precision, automatic and non-reliance on construction information is achieved.
Patent Information
- Application Number
- CN202311433709.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-10-31
- Publication Date
- 2025-05-02
AI Technical Summary
The prior art is difficult to estimate uncertainty of the far-well position in well parameter interpolation, and the construction information needs to be processed to ensure the continuity of the interpolation path, resulting in a decrease in interpolation accuracy.
The well parameter interpolation method based on neural network is adopted to construct structural matrix and train neural networks to interpolate nearby positions, and the uncertainty of interpolation is estimated using Bayesian convolutional neural network and MC Dropout technology to eliminate interpolation with high uncertainty.
High-precision well parameter interpolation is realized, and the uncertainty of interpolation can be automatically calculated, which avoids dependence on construction information and improves the continuity and automation of interpolation.
Smart Images
Figure CN119919283A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of velocity modeling and seismic imaging in oil and gas exploration and development, and in particular relates to a well parameter interpolation method based on a neural network. Background Art
[0002] The conventional model domain well interpolation method first extracts the structural information from the offset image, and then interpolates the well parameters in the model domain along the structural strike. This method has two disadvantages: 1. It is difficult to estimate the uncertainty of the interpolation far away from the well location; 2. The extracted structural information needs to be processed to ensure the continuity of the interpolation path, which reduces the accuracy of the interpolation. Summary of the invention
[0003] The purpose of the present invention is to solve the above-mentioned problems existing in the prior art and to provide a well parameter interpolation method based on a neural network.
[0004] The present invention is implemented by the following technical scheme: A well parameter interpolation method based on a neural network comprises the following steps:
[0005] Construct a structure matrix S based on the offset image;
[0006] A training sample set is established by interpolating the adjacent positions of the well parameters according to the structure matrix S, and the neural network is trained using the training sample set; wherein each training sample uses a local offset image as input data, and uses the well position data and the interpolation data as label data;
[0007] The local offset image of the model domain position without well / interpolation parameters is input into the trained neural network, and the output data is used as the interpolation data of the neural network at the corresponding position;
[0008] According to the uncertainty threshold, the neural network interpolation with high uncertainty is eliminated, and the neural network interpolation with low uncertainty is retained;
[0009] Use the structure matrix to interpolate neighboring locations of the preserved low-uncertainty neural network interpolation.
[0010] Furthermore, the Bayesian convolutional neural network NN is used to perform MC Dropout neural network interpolation.
[0011] Furthermore, the uncertainty of the neural network interpolation is evaluated as follows:
[0012] A local offset image of the model domain location without well / interpolation parameters is input, and multiple neural network predictions are performed. The variance of each prediction result is statistically estimated to estimate the uncertainty: a small variance indicates low uncertainty, and a large variance indicates high uncertainty.
[0013] Furthermore, after using the structure matrix to interpolate the neighboring positions of the retained neural network with low uncertainty, the training sample set is expanded to retrain the neural network so that the positions with high uncertainty can be re-predicted.
[0014] Furthermore, the DTW algorithm is applied to all adjacent seismic records of the migration image, and the structural correspondence relations s extracted from the adjacent traces are combined to construct the structural matrix s.
[0015] Furthermore, the DTW algorithm is as follows:
[0016]
[0017] Among them, a is the error matrix of the adjacent tracks d1 and d2 in the offset image domain, and the variables z and Δz are the grid point numbers and differences in the depth domain. According to the error matrix, the distance matrix can be obtained:
[0018]
[0019] Where n is the number of grid points in the depth direction of the offset image. The construction path of w(Δz, z) is recorded as P(Δz, z).
[0020] According to the distance matrix, the structural information in the offset image can be extracted by reverse searching the path matrix P:
[0021] s(z-1)=P(s(z),z), for z=n,n-1,...,2 (3)
[0022] Among them, s is the structural correspondence between the two paths.
[0023] Furthermore, the structure matrix S is used to interpolate the well parameters at adjacent locations:
[0024] d[i+s[i]]=d w [i] (4)
[0025] Where i is the grid point number, d[i+u[i]] is the interpolation value outside the model domain, and d w is the reference value.
[0026] Compared with the prior art, the beneficial effects of the present invention include:
[0027] 1. The cognitive uncertainty of the neural network on the interpolation estimate is calculated while interpolating, and it has the characteristics of high interpolation accuracy, no dependence on the continuity of construction information, and high degree of automation.
[0028] 2. The present invention proposes to use a Bayesian neural network based on Monte Carlo Dropout, combined with a traditional dynamic programming method, to use the offset image structure information to perform nonlinear interpolation of well parameters in the well-free area of the model domain. This method uses dynamic programming to expand high-precision interpolation near the well, expands the training data set of the neural network, and uses the Bayesian neural network to estimate the uncertainty of interpolation, eliminates uncertain interpolation, and improves the interpolation accuracy. This method has the characteristics of high interpolation accuracy, continuity that does not rely on structural information, and high degree of automation. BRIEF DESCRIPTION OF THE DRAWINGS
[0029] Figure 1 Schematic diagram of nonlinear interpolation for Bayesian neural networks;
[0030] Figure 2 Schematic diagram of the input and output of a neural network. DETAILED DESCRIPTION
[0031] The present invention proposes to use a Bayesian neural network based on Monte Carlo Dropout, combined with a traditional dynamic programming method, and use the offset image structure information to perform nonlinear interpolation of well parameters in the well-free area of the model domain. This method uses dynamic programming to expand high-precision interpolation near the well, expands the training data set of the neural network, and uses the Bayesian neural network to estimate the uncertainty of interpolation, eliminates uncertain interpolation, and improves the interpolation accuracy. This method has the characteristics of high interpolation accuracy, continuity that does not rely on structural information, and high degree of automation.
[0032] A well parameter interpolation method based on a neural network comprises the following steps:
[0033] Construct a structure matrix S based on the offset image;
[0034] A training sample set is established by interpolating the adjacent positions of the well parameters according to the structure matrix S, and the neural network is trained using the training sample set; wherein each training sample uses a local offset image as input data, and uses the well position data and the interpolation data as label data;
[0035] The local offset image of the model domain position without well / interpolation parameters is input into the trained neural network, and the output data is used as the interpolation data of the neural network at the corresponding position;
[0036] According to the uncertainty threshold, the neural network interpolation with high uncertainty is eliminated, and the neural network interpolation with low uncertainty is retained;
[0037] Use the structure matrix to interpolate neighboring locations of the preserved low-uncertainty neural network interpolation.
[0038] Preferably, a Bayesian convolutional neural network (NN) is used to perform MC Dropout neural network interpolation.
[0039] Preferably, the uncertainty of the neural network interpolation is evaluated as follows:
[0040] A local offset image of the model domain location without well / interpolation parameters is input, and multiple neural network predictions are performed. The variance of each prediction result is statistically estimated to estimate the uncertainty: a small variance indicates low uncertainty, and a large variance indicates high uncertainty.
[0041] Preferably, after using the structure matrix to interpolate the adjacent positions of the retained low-uncertainty neural network interpolation, the training sample set is expanded to retrain the neural network so that the positions with high uncertainty before can be re-predicted.
[0042] Preferably, the DTW algorithm is applied to all adjacent seismic records of the migration image, and the structural correspondences s of the extracted adjacent traces are combined to construct the structural matrix s.
[0043] Preferably, the DTW algorithm is as follows:
[0044]
[0045] Among them, a is the error matrix of the adjacent tracks d1 and d2 in the offset image domain, and the variables z and Δz are the grid point numbers and differences in the depth domain. According to the error matrix, the distance matrix can be obtained:
[0046]
[0047] Where n is the number of grid points in the depth direction of the offset image. The construction path of w(Δz, z) is recorded as P(Δz, z).
[0048] According to the distance matrix, the structural information in the offset image can be extracted by reverse searching the path matrix P:
[0049] s(z-1)=P(s(z),z), for z=n,n-1,...,2 (3)
[0050] Among them, s is the structural correspondence between the two paths.
[0051] Preferably, the structure matrix s is used to interpolate the well parameters at adjacent locations:
[0052] d[i+s[i]-d w [i] (4)
[0053] Where i is the grid point number, d[i+u[i]] is the interpolation value outside the model domain, and d w is the reference value.
[0054] The present invention is further described in detail below in conjunction with the accompanying drawings:
[0055] like Figure 1 As shown, a Bayesian neural network based on Monte Carlo Dropout is used to perform nonlinear interpolation and estimate the uncertainty of the interpolation.
[0056] Step 1: Use the dynamic time warping (DTW) method to extract the structural direction information in the migration image. The DTW algorithm is as follows:
[0057]
[0058] Among them, a is the error matrix of the adjacent channels d1 and d2 in the offset image domain, and the variables z, Δ z is the depth domain grid point number and difference. According to the error matrix, the distance matrix can be:
[0059]
[0060] Where n is the number of grid points in the depth direction of the offset image. The construction path of w(Δz, z) is recorded as P(Δz, z).
[0061] According to the distance matrix, the structural information in the offset image can be extracted by reverse searching the path matrix P:
[0062] s(z-1)=P(s(z),z), for z=n,n-1,...,2 (3)
[0063] Where s is the structural correspondence between the two paths.
[0064] By applying the DTW algorithm to all adjacent seismic records of the migration image, s combinations of extracted adjacent traces can be constructed into a structure matrix s.
[0065] Step 2: Use the structure matrix S to interpolate the well parameters at adjacent locations:
[0066] d[i+s[i]]=d w [i] (4)
[0067] Where i is the grid point number, d[i+u[i]] is the interpolation value outside the model domain, and d w is the reference value (the value of the well or the interpolated location near the well).
[0068] This step greatly expands the training set samples for the next step of neural network training.
[0069] Step 3: Build a Bayesian convolutional neural network (NN) based on Monte Carlo Dropout. Train and establish a nonlinear regression relationship between the local offset image and the parameters at the well location and the adjacent location of the well after the second interpolation. Figure 2 shown.
[0070] The training labels (expected output of the model) are the well location data and interpolation data; the training model input is the local offset image corresponding to each label data.
[0071] Step 4: Use the trained NN to perform MC Dropout neural network interpolation at the model domain location without wells / interpolation parameters. Evaluate the uncertainty of the interpolation data and remove the interpolation with high uncertainty based on the uncertainty threshold. Input the local offset image of the model domain location without wells / interpolation parameters to the NN and output the corresponding interpolation data
[0072] Input a local offset image of the model domain position without well / interpolation parameters, and make multiple neural network predictions. Because the dropout function is used, the prediction results will be different each time. The variance of these results can be used to estimate uncertainty: small variance indicates low uncertainty, and large variance indicates high uncertainty.
[0073] Step 5: Use the structure matrix to extrapolate the retained low-uncertainty neural network interpolation. Extrapolation is to use the second step to interpolate the adjacent positions and repeat the third and fourth steps.
[0074] Because we did not keep all the neural network prediction interpolations in the fourth step, but only kept some interpolation results with low uncertainty; so we need to re-predict the positions with high uncertainty.
[0075] In the description of the present invention, it should be noted that, unless otherwise clearly specified and limited, the terms "connected" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.
[0076] In the description of the present invention, unless otherwise specified, the terms "upper", "lower", "left", "right", "inside", "outside", etc. indicate directions or positional relationships based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they cannot be understood as limitations on the present invention.
[0077] The above technical scheme is only a specific implementation method of the present invention. For those skilled in the art, it is easy to make various types of improvements or modifications based on the principles disclosed in the present invention, and it is not limited to the technical scheme described in the above specific embodiments of the present invention. Therefore, the above description is only preferred and does not have a restrictive meaning.
Claims
1. A well parameter interpolation method based on neural network, characterized in that: The following steps are involved: Construct a structure matrix S based on the offset image; A training sample set is established by interpolating the adjacent positions of the well parameters according to the structure matrix s, and the neural network is trained using the training sample set; wherein each training sample uses a local offset image as input data, and uses the well position data and the interpolation data as label data; The local offset image of the model domain position without well / interpolation parameters is input into the trained neural network, and the output data is used as the interpolation data of the neural network at the corresponding position; According to the uncertainty threshold, the neural network interpolation with high uncertainty is eliminated, and the neural network interpolation with low uncertainty is retained; Use the structure matrix to interpolate neighboring locations of the preserved low-uncertainty neural network interpolation.
2. The neural network-based well parameter interpolation method according to claim 1, characterized in that: The Bayesian convolutional neural network (NN) is used to perform MC Dropout neural network interpolation.
3. The neural network-based well parameter interpolation method according to claim 2, characterized in that: Evaluate the uncertainty of the neural network interpolation as follows: A local offset image of the model domain location without well / interpolation parameters is input, and multiple neural network predictions are performed. The variance of each prediction result is statistically estimated to estimate the uncertainty: a small variance indicates low uncertainty, and a large variance indicates high uncertainty.
4. The neural network-based well parameter interpolation method according to claim 1, characterized in that: After using the structure matrix to interpolate the neighboring locations of the retained neural network with low uncertainty, the training sample set is expanded to retrain the neural network so that it can re-predict the previously high uncertainty locations.
5. The neural network-based well parameter interpolation method according to claim 1, characterized in that: The DTW algorithm is applied to all adjacent seismic records of the migration image, and the structural correspondence relations s extracted from adjacent traces are combined to construct the structural matrix S.
6. The neural network-based well parameter interpolation method according to claim 5, characterized in that: The DTW algorithm is as follows: Among them, a is the error matrix of the adjacent tracks d1 and d2 in the offset image domain, and the variables z and Δz are the grid point numbers and differences in the depth domain. According to the error matrix, the distance matrix can be obtained: Where n is the number of grid points in the depth direction of the offset image. The construction path of w(Δz, z) is recorded as P(Δz, z). According to the distance matrix, the structural information in the offset image can be extracted by reverse searching the path matrix P: s(z 1)=P(s(z),z), for z=n,n 1,...,2 (3) Among them, s is the structural correspondence between the two paths.
7. The neural network-based well parameter interpolation method according to claim 5, characterized in that: Use the structure matrix S to interpolate well parameters at nearby locations: d[i+s[i]]=d w [i] (4) Where i is the grid point number, d[i+u[i]] is the interpolation value outside the model domain, and d w is the reference value.