Shear wave velocity prediction method based on deep learning

Abstract

The application provides a shear wave velocity prediction method based on deep learning, comprising the following steps: step 1, obtaining a longitudinal wave velocity curve Vp, a neutron porosity curve CNL, a density curve DEN, a porosity curve POR, a shale content curve SH and a corresponding measured shear wave velocity curve Vs; step 2, pre-processing the logging curves; step 3, constructing a shear wave velocity prediction model; step 4, continuously modifying the model parameters until the model stability and prediction accuracy meet the requirements; and step 5, predicting the shear wave velocity of a well to be predicted by using the shear wave velocity prediction model. The shear wave velocity prediction method based on deep learning can accurately predict the logging shear wave velocity, can replace the empirical formula method and the rock physics modeling method, can accurately and quickly predict the logging shear wave velocity by using the logging curves and logging interpretation results, has strong generalization ability and has practical popularization and application significance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil exploration, and particularly relates to a shear wave velocity prediction method based on deep learning. BACKGROUND

[0002] It is a basic prerequisite to use pre-stack AVO analysis technology and pre-stack inversion technology to predict reservoir and reservoir fluid by using shear wave velocity logging data. However, in the actual exploration and development production process of oilfields, due to cost problems and other factors, shear wave logging data is rarely collected. This leads to great difficulties in the application of geophysical technology based on pre-stack seismic data, because it is restricted by shear wave logging data. In this case, the estimation of shear wave velocity using conventional logging data has become a hot research topic, and a large number of scholars have carried out research on shear wave velocity estimation methods. At present, there are mainly two methods for shear wave velocity estimation: empirical formula method and rock physics modeling method. The empirical formula method is simple and easy to use, but the estimation accuracy is not high and is limited by the region; while the rock physics modeling method has high estimation accuracy, but the process is complex, the parameters are numerous, and the operation is difficult, so the efficiency is not high in practical application. Therefore, how to quickly and accurately predict shear wave velocity is a technical problem to be solved.

[0003] In the Chinese patent application with the application number CN202010434118.0, a shear wave velocity prediction method and device are involved, which comprises: acquiring conventional logging curves and known well shear wave velocity curves; determining preferred curves according to the conventional logging curves and the known well shear wave velocity curves; performing normalization processing on the preferred curves to determine normalized preferred curves; establishing a deep feedforward neural network model combined with the conventional logging curves, training the known well shear wave velocity to determine a shear wave velocity prediction model; inputting the normalized preferred curves into the shear wave velocity prediction model to determine the shear wave velocity.

[0004] In the Chinese patent application with the application number CN202010321843.7, a shear wave velocity prediction method, device and equipment are involved, wherein the method comprises: acquiring to-be-predicted data, wherein the to-be-predicted data includes seismic records, seismic wavelets, P-wave velocities and formation densities; inputting the to-be-predicted data into a target shear wave velocity prediction model to obtain a predicted value of the shear wave velocity; wherein the target shear wave prediction model is obtained by unsupervised pre-training of a target neural network structure constructed by using a convolutional neural network structure and a gated recurrent neural network structure.

[0005] In the Chinese patent application with the application number: CN202110214077.9, a reservoir parameter determination method, device, electronic equipment and storage medium are involved, wherein the method comprises: calculating an initial formation elastic parameter body according to a constructed target rock physical model and obtained e lan interpretation data of a target area; determining a target formation elastic parameter body through a pre-stack elastic parameter inversion method based on obtained seismic data of the target area and the initial elastic parameter body; constructing an inverse function model of the target rock physical model based on a neural network learning method; wherein the inverse function model is used to convert the formation elastic parameter body into reservoir parameters; inputting the target formation elastic parameter body into the inverse function model, and determining the reservoir parameters of the target area according to the output of the inverse function model.

[0006] The above prior art is quite different from the present application, and cannot solve the technical problems we want to solve. Therefore, we have invented a new deep learning-based shear wave velocity prediction method. SUMMARY

[0007] The purpose of the present application is to provide a deep learning-based shear wave velocity prediction method that replaces the empirical formula method and rock physical modeling method to improve the accuracy and efficiency of shear wave velocity estimation.

[0008] The purpose of the present application can be achieved by the following technical measures: a deep learning-based shear wave velocity prediction method, comprising:

[0009] Step 1, obtaining a longitudinal wave velocity curve Vp, a neutron porosity curve CNL, a density curve DEN, a porosity curve POR, a shale content curve SH and a corresponding measured shear wave velocity curve Vs;

[0010] Step 2, preprocessing the logging curves;

[0011] Step 3, constructing a shear wave velocity prediction model;

[0012] Step 4, continuously modifying the model parameters until the model stability and prediction accuracy meet the requirements;

[0013] Step 5, using the shear wave velocity prediction model to predict the shear wave velocity of the well to be predicted.

[0014] The purpose of the present application can also be achieved by the following technical measures:

[0015] In step 1, wells with measured shear wave velocity in the study area and adjacent areas are collected, and these wells have longitudinal wave velocity curve Vp, neutron porosity curve CNL, density curve DEN, porosity curve POR, and shale content curve SH at the same time; wherein Vp, CNL, and DEN are measured curves, and POR and SH are obtained by well logging interpretation.

[0016] In step 2, the logging curves are preprocessed, including wellbore collapse correction, abnormal value processing, and multi-well consistency processing, to obtain a set of logging curves with high quality and good inter-well comparison.

[0017] In step 2, the wellbore collapse correction and abnormal value processing of the logging curves use a multivariate regression method, and a regression formula of a normal curve segment is established and applied to the correction of the wellbore collapse segment and abnormal values.

[0018] In step 2, the multi-well consistency processing is performed on the Vp curve to eliminate inconsistencies between different wells and system errors caused by different instruments and different time measurements.

[0019] In step 3, a deep feedforward neural network DFNN algorithm is used to establish a nonlinear relationship between the longitudinal wave velocity curve Vp, the neutron porosity curve CNL, the density curve DEN, the porosity curve POR, the shale content curve SH, and the shear wave velocity curve Vp, and a shear wave velocity prediction model is constructed.

[0020] In step 3, a part of the wells are selected as training wells, and a deep feedforward neural network is used to establish a nonlinear mapping relationship between the longitudinal wave velocity curve Vp, the neutron porosity curve CNL, the density curve DEN, the porosity curve POR, the shale content curve SH, and the shear wave velocity curve Vs, and a shear wave velocity prediction model is constructed; the DFNN network contains 2 hidden layers, each containing 3 neurons, and during model training, the input layer is the feature space of the training samples, i.e., the reservoir parameters used to predict the shear wave velocity, and the output layer is the shear wave velocity.

[0021] In step 4, the neural network parameters are further adjusted according to the error between the predicted shear wave velocity and the measured shear wave velocity on the verification well, and the most suitable neural network parameters are continuously iterated and optimized to complete the final logging shear wave velocity prediction model construction.

[0022] In step 5, the logging shear wave velocity prediction model is directly applied to the target area to be predicted well, and by inputting the longitudinal wave velocity curve Vp, the neutron porosity curve CNL, the density curve DEN, the porosity curve POR, and the shale content curve SH, the shear wave velocity Vs is directly obtained, and the shear wave velocity prediction is completed.

[0023] The deep learning-based shear wave velocity prediction method of the present invention, based on the construction of the relationship between P-wave velocity and S-wave velocity, adds more parameters (logging curves) that affect the relationship between P-wave velocity and S-wave velocity, such as CNL, DEN, POR, SH, etc., making the S-wave velocity prediction more reasonable and reliable. On the one hand, compared with the traditional empirical formula method using linear fitting, the method of the present invention, by using a deep neural network algorithm with strong nonlinear expression capabilities, can fully explore the relationship between the P-wave velocity curve Vp, the neutron porosity curve CNL, the density curve DEN, the porosity curve POR, the shale content curve SH and the S-wave velocity Vs, thereby improving the accuracy of the prediction results. On the other hand, compared with the rock physics modeling method, the method of the present invention can achieve a rapid output of S-wave velocity Vs from the input P-wave velocity curve Vp, the neutron porosity curve CNL, the density curve DEN, the porosity curve POR, and the shale content curve SH. The process is simple, ensuring accuracy while greatly improving the efficiency of S-wave velocity prediction.

[0024] In summary, the shear wave velocity prediction method based on deep learning can replace the empirical formula method and rock physics modeling method, realize the rapid and accurate prediction of well logging shear wave velocity, significantly improve work efficiency, and this method has strong generalization ability and has practical significance for promotion and application. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 Flowchart of a specific embodiment of the shear wave velocity prediction method based on deep learning of the present invention;

[0026] Figure 2 5 reservoir parameter curves for predicting shear wave velocity in a specific embodiment of the present invention;

[0027] Figure 3 Schematic diagram of a deep feedforward neural network model in one embodiment of the present invention;

[0028] Figure 4 A reservoir parameter curve, a measured shear wave velocity curve, and a predicted shear wave velocity curve on a training well in a specific embodiment of the present invention;

[0029] Figure 5 This is a diagram of a reservoir parameter curve, a measured shear wave velocity curve, and a predicted shear wave velocity curve for verifying a well in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0030] It should be noted that the following detailed descriptions are exemplary and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0031] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations and / or combinations thereof.

[0032] The shear wave velocity prediction method based on deep learning of the present invention includes the following steps:

[0033] Step 1: Obtain the P-wave velocity curve Vp, the neutron porosity curve CNL, the density curve DEN, the porosity curve POR, the shale content curve SH, and the corresponding measured S-wave velocity curve Vs;

[0034] Wells with measured shear wave velocity in the study area and adjacent areas are collected. These wells should also have a compressional wave velocity curve Vp, a neutron porosity curve CNL, a density curve DEN, a porosity curve POR, and a shale content curve SH. Among them, the three curves Vp, CNL, and DEN are measured curves, and the two curves POR and SH are generally obtained through well logging interpretation.

[0035] Step 2: pre-process the logging curves, mainly including outlier processing and inter-well consistency processing;

[0036] Perform preprocessing on well logging curves. This primarily includes borehole collapse correction, outlier processing, and multi-well consistency processing. Well logging curves typically use multivariate regression to correct for borehole collapse and outlier processing. By establishing a regression formula for normal curve segments, this formula is applied to correct for wellbore collapse segments and outliers. Multi-well consistency processing typically operates on Vp curves to eliminate inconsistencies in units between wells, as well as systematic errors caused by different instruments and measurements at different times. The ultimate goal is to obtain a set of high-quality logging curves that can be easily compared across wells.

[0037] Step 3: Use the deep feedforward neural network (DFNN) algorithm to establish a nonlinear relationship between the P-wave velocity curve (Vp), the neutron porosity curve (CNL), the density curve (DEN), the porosity curve (POR), the shale content curve (SH), and the S-wave velocity curve (Vp), and construct a S-wave velocity prediction model.

[0038] A selection of wells was selected as training wells. A deep feedforward neural network (DFNN) was used to establish a nonlinear mapping relationship between five well logging curves—the compressional wave velocity curve (Vp), the neutron porosity curve (CNL), the density curve (DEN), the porosity curve (POR), and the shale content curve (SH)—and the shear wave velocity curve (Vs). This led to the construction of a shear wave velocity prediction model. The DFNN network employed consisted of two hidden layers, each containing three neurons. During model training, the input layer was the feature space of the training samples, i.e., the reservoir parameters used to predict shear wave velocity, and the output layer was the shear wave velocity.

[0039] Step 4: Modify the model parameters through continuous iteration until the model stability and prediction accuracy meet the requirements;

[0040] The neural network parameters are further adjusted based on the error between the predicted shear wave velocity and the measured shear wave velocity on the verification well. The most appropriate neural network parameters are continuously optimized through iterations to complete the construction of the final well logging shear wave velocity prediction model.

[0041] Step 5: Use the shear wave velocity prediction model to predict the shear wave velocity of the well to be predicted.

[0042] The well logging shear wave velocity prediction model is directly applied to the wells to be predicted in the target area. By inputting five well logging curves, including the compressional wave velocity curve Vp, the neutron porosity curve CNL, the density curve DEN, the porosity curve POR, and the shale content curve SH, the shear wave velocity Vs is directly obtained to complete the shear wave velocity prediction.

[0043] The following are several specific embodiments of the present invention.

[0044] Example 1

[0045] like Figure 1 As shown, Figure 1 This is a flow chart of the shear wave velocity prediction method based on deep learning of the present invention. Step 101, collect wells with measured shear wave velocity in the study area and adjacent areas, and these wells must also have a compressional wave velocity curve Vp, a neutron porosity curve CNL, a density curve DEN, a porosity curve POR, and a shale content curve SH; wherein the three curves Vp, CNL, and DEN are measured curves, and the two curves POR and SH are generally obtained through well logging interpretation. Figure 2 shown.

[0046] Step 102, pre-process the well logging curves. Mainly includes borehole collapse correction, outlier processing, multi-well consistency processing, etc. The borehole collapse correction and outlier processing of the well logging curves generally adopts the method of multiple regression, and the regression formula of the normal curve segment is established, which is applied to the correction of the borehole collapse segment and the outlier; the multi-well consistency processing is generally operated on the Vp curve, and the inconsistency between different wells is eliminated, and the system error caused by different instruments and different time measurements is eliminated, etc. The final purpose is to obtain a group of well logging curves with high quality and good comparison between wells. As shown in Figure 3 .

[0047] Step 103, use the well curves processed in step 102 to select a part of wells as training wells, and establish a depth neural network learning model between the five well logging curves of longitudinal wave velocity curve Vp, neutron porosity curve CNL, density curve DEN, porosity curve POR and shale content curve SH and transverse wave velocity curve Vs by using deep feedforward neural network. The DFNN network used contains 2 hidden layers, each hidden layer contains 3 neurons, as shown in Figure 3 . When the model is trained, the input layer is the feature space of the training sample, that is, the reservoir parameters used to predict the transverse wave velocity, which is a 5-dimensional column vector [x1, x2, x3, x4, x5] T ; for each input data and the output result after the neuron, the Xavier initialization method is used to randomly generate the corresponding weight coefficient w and bias term b Figure 1 . The activation function of each neuron selects the Sigmoid function, and its expression is:

[0048]

[0049] Where: x is the input data.

[0050] In the process of forward propagation, all input layer data are input to each neuron in the first hidden layer in the form of weighted average, then the input data of the three neurons neu1, neu2 and neu3 in the first hidden layer are respectively represented as:

[0051] z1=w(x1,1)*x1+w(x2,1)*x2+w(x3,1)*x3+w(x4,1)*x4+w(x5,1)*x5+b1

[0052] z2=w(x1,2)*x1+w(x2,2)*x2+w(x3,2)*x3+w(x4,2)*x4+w(x5,2)*x5+b2

[0053] z3=w(x1,3)*x1+w(x2,3)*x2+w(x3,3)*x3+w(x4,3)*x4+w(x5,3)*x5+b3

[0054] Among them: z1, z2, z3 are the input data of the three neurons neu1, neu2 and neu3 in the first hidden layer respectively.

[0055] After the nonlinear transformation of the logistic function, the outputs of the three neurons in the first hidden layer are f1(z1), f2(z2), and f3(z3). The weighted average of these is used as the input of the neurons in the next hidden layer. Therefore, the input data of the three neurons in the second hidden layer are:

[0056] z4=w (1，4) *f1(z1)+w (2，4) *f2(z2)+w (3，4) *f3(z3)+b4

[0057] z5=w (1，5) *f1(z1)+w (2，5) *f2(z2)+w (3，5) *f3(z3)+b5

[0058] z6=w (1，6) *f1(z1)+w (2，6) *f2(z2)+w (3，6) *f3(z3)+b6

[0059] Among them: z4, z5, z6 are the input data of the three neurons neu4, neu5 and neu6 in the second hidden layer respectively.

[0060] Similarly, the output results of the three neurons in the second hidden layer are: f4(z4), f5(z5), and f6(z6). Their weighted average sum is used as the input of the output layer neurons, that is:

[0061] z7=w (4，7) *f4(z4)+w (5，7) *f5(z5)+w (6，7) *f6(z6)+b7

[0062] Among them: z7 is the input data of neuron neu7.

[0063] After passing through the output layer, the final prediction result f7(z7) can be obtained.

[0064] The final result of the neural network output is based on the initialized w and b. In order to achieve the least squares error between the final output result and the measured data, it is necessary to construct a function with w and b as independent variables:

[0065]

[0066] in: is the output result, and y is the measured data.

[0067] The conjugate gradient method is used for numerical optimization, and the iteration is stopped when the number of iterations reaches the set number.

[0068] like Figure 4 shown.

[0069] In step 104, the deep learning model trained in step 103 is applied to the verification well, and the neural network parameters are further adjusted according to the error between the predicted shear wave velocity and the measured shear wave velocity in the verification well. The most appropriate neural network parameters are continuously iterated to complete the construction of the final well logging shear wave velocity prediction model. Figure 5 shown.

[0070] Step 105 : directly apply the well logging shear wave velocity prediction model completed in step 104 to the wells to be predicted in the target area to complete the well logging shear wave velocity prediction.

[0071] Example 2:

[0072] Step 101: Collect wells with measured shear wave velocity in the study area and adjacent areas. These wells must also have a P-wave velocity curve Vp, a porosity curve POR, and a shale content curve SH. The P-wave velocity curve Vp is the measured curve, and the POR and SH curves are generally obtained through well logging interpretation.

[0073] Step 102 pre-processes the well logging curves. This primarily includes borehole collapse correction, outlier processing, and multi-well consistency processing. Well logging curve borehole collapse correction and outlier processing generally utilize multivariate regression methods. By establishing a regression formula for normal curve segments, this formula is applied to correct for wellbore collapse segments and outliers. Multi-well consistency processing typically operates on the Vp curve to eliminate inconsistencies in units between wells, as well as systematic errors caused by different instruments and measurements at different times. The ultimate goal is to obtain a set of high-quality well logging curves that can be easily compared across wells.

[0074] In step 103, using the well curves processed in step 102, a portion of wells are selected as training wells. A deep neural network learning model is established between the three well logging curves, namely the compressional wave velocity curve Vp, the porosity curve POR, and the shale content curve SH, and the shear wave velocity curve Vs. The DFNN network used contains two hidden layers, each of which contains three neurons, such as Figure 3 When training the model, the input layer is the feature space of the training samples, that is, the reservoir parameters used to predict the shear wave velocity, using a 5-dimensional column vector [x1, x2, x3, x4, x5] T For each input data and the output after passing through the neuron, the study uses the Xavier initialization method to randomly generate the corresponding weight coefficient w and bias term b Figure 1 The activation function of each neuron is chosen to use the Sigmoid function, whose expression is:

[0075]

[0076] Where: x is the input data.

[0077] During the forward propagation process, all input layer data are input to each neuron in the first hidden layer in the form of weighted average. Then the input data of the three neurons neu1, neu2 and neu3 in the first hidden layer are expressed as:

[0078] z1=w(x1,1)*x1+w(x2,1)*x2+w(x3,1)*x3+w(x4,1)*x4+w(x5,1)*x5+b1

[0079] z2=w(x1,2)*x1+w(x2,2)*x2+w(x3,2)*x3+w(x4,2)*x4+w(x5,2)*x5+b2

[0080] z3=w(x1,3)*x1+w(x2,3)*x2+w(x3,3)*x3+w(x4,3)*x4+w(x5,3)*x5+b3

[0081] Among them: z1, z2, z3 are the input data of the three neurons neu1, neu2 and neu3 in the first hidden layer respectively.

[0082] After the nonlinear transformation of the logistic function, the outputs of the three neurons in the first hidden layer are f1(z1), f2(z2), and f3(z3). The weighted average of these is used as the input of the neurons in the next hidden layer. Therefore, the input data of the three neurons in the second hidden layer are:

[0083] z4=w(1，4) *f1(z1)+w (2，4) *f2(z2)+w (3，4) *f3(z3)+b4

[0084] z5=w( 1，5) *f1(z1)+w (2，5) *f2(z2)+w (3，5) *f3(z3)+b5

[0085] z6=w (1 ,6)*f1(z1)+w (2，6) *f2(z2)+w (3，6) *f3(z3)+b6

[0086] wherein: z4, z5, z6 are input data of three neurons neu4, neu5 and neu6 in the second hidden layer respectively.

[0087] Similarly, the output results of the three neurons in the second hidden layer are f4(z4), f5(z5), f6(z6) respectively, and their weighted average sum is the input of the output layer neuron, that is:

[0088] z7=w (4，7) *f4(z4)+w (5，7) *f5(z5)+w (6，7) *f6(z6)+b7

[0089] wherein: z7 is the input data of neuron neu7.

[0090] After the output layer, the final prediction result f7(z7) can be obtained.

[0091] The final result of the neural network output is based on the initialized w and b, in order to make the final output result and the measured data reach the least square error, a function with w and b as independent variables needs to be constructed:

[0092]

[0093] wherein: is the output result, and y is the measured data.

[0094] The conjugate gradient method is used for numerical optimization, and when the number of iterations reaches the set number of times, the iteration is stopped.

[0095] In step 104, the deep learning model trained in step 103 is applied to the verification well, and the neural network parameters are further adjusted according to the error between the predicted shear wave velocity and the measured shear wave velocity in the verification well. The most appropriate neural network parameters are continuously iterated to complete the construction of the final well logging shear wave velocity prediction model.

[0096] Step 105 : directly apply the well logging shear wave velocity prediction model completed in step 104 to the wells to be predicted in the target area to complete the well logging shear wave velocity prediction.

[0097] Example 3:

[0098] Step 101: Collect wells with measured shear wave velocity in the study area and adjacent areas, and these wells must also have a compressional wave velocity curve Vp, a neutron porosity curve CNL, and a density curve DEN.

[0099] Step 102 pre-processes the well logging curves. This primarily includes borehole collapse correction, outlier processing, and multi-well consistency processing. Well logging curve borehole collapse correction and outlier processing generally utilize multivariate regression methods. By establishing a regression formula for normal curve segments, this formula is applied to correct for wellbore collapse segments and outliers. Multi-well consistency processing typically operates on the Vp curve to eliminate inconsistencies in units between wells, as well as systematic errors caused by different instruments and measurements at different times. The ultimate goal is to obtain a set of high-quality well logging curves that can be easily compared across wells.

[0100] In step 103, a portion of the well logs processed in step 102 is selected as training wells. A deep neural network learning model is established between the three well logs (P-wave velocity curve Vp, neutron porosity curve CNL, density curve DEN) and the S-wave velocity curve Vs using a deep feedforward neural network. The DFNN network used contains two hidden layers, each containing three neurons, such as Figure 3 When training the model, the input layer is the feature space of the training samples, that is, the reservoir parameters used to predict the shear wave velocity, using a 5-dimensional column vector [x1, x2, x3, x4, x5] T For each input data and the output after passing through the neuron, the study uses the Xavier initialization method to randomly generate the corresponding weight coefficient w and bias term b Figure 1 The activation function of each neuron is chosen to use the Sigmoid function, whose expression is:

[0101]

[0102] Where: x is the input data.

[0103] During the forward propagation, all the input layer data are inputted to each neuron in the first hidden layer in the form of weighted average, and the input data of the three neurons neu1, neu2 and neu3 in the first hidden layer are respectively represented as:

[0104] z1 = w(x1, 1) * x1 + w(x2, 1) * x2 + w(x3, 1) * x3 + w(x4, 1) * x4 + w(x5, 1) * x5 + b1

[0105] z2 = w(x1, 2) * x1 + w(x2, 2) * x2 + w(x3, 2) * x3 + w(x4, 2) * x4 + w(x5, 2) * x5 + b2

[0106] z3 = w(x1, 3) * x1 + w(x2, 3) * x2 + w(x3, 3) * x3 + w(x4, 3) * x4 + w(x5, 3) * x5 + b3

[0107] Wherein, z1, z2, z3 are respectively the input data of the three neurons neu1, neu2 and neu3 in the first hidden layer.

[0108] After the nonlinear conversion processing of the logic function, the output results of the three neurons in the first hidden layer are respectively f1(z1), f2(z2), f3(z3), and the weighted average of them is taken as the input of the neurons in the next hidden layer. Therefore, the input data of the three neurons in the second hidden layer are respectively:

[0109] z4 = w (1，4) *f1(z1) + w (2，4) *f2(z2) + w (3，4) *f3(z3) + b4

[0110] z5 = w (1，5) *f1(z1) + w (2，5) *f2(z2) + w (3，5) *f3(z3) + b5

[0111] z6 = w (1，6) *f1(z1) + w (2，6) *f2(z2) + w (3，6) *f3(z3) + b6

[0112] Wherein, z4, z5, z6 are respectively the input data of the three neurons neu4, neu5 and neu6 in the second hidden layer.

[0113] Similarly, the output results of the three neurons in the second hidden layer are respectively f4(z4), f5(z5), f6(z6), and the weighted average of them is taken as the input of the neurons in the output layer, i.e.

[0114] z7 = w (4，7) *f4(z4) + w (5，7) *f5(z5) + w (6，7) *f6(z6) + b7

[0115] Wherein: z7 is the input data of neuron neu7.

[0116] After the output layer, the final prediction result f7(z7) can be obtained.

[0117] The final result of the neural network output is based on the initialized w and b, in order to make the final output result and the measured data reach the least square error, a function with w and b as independent variables needs to be constructed:

[0118]

[0119] Wherein: is the output result, y is the measured data.

[0120] The conjugate gradient method is used for numerical optimization, and when the number of iterations reaches the set number of times, the iteration is stopped.

[0121] Step 104, the deep learning model trained in step 103 is applied to the verification well, and the neural network parameters are further adjusted according to the error between the predicted shear wave velocity and the measured shear wave velocity on the verification well, and the neural network parameters are iterated and optimized to complete the final logging shear wave velocity prediction model.

[0122] Step 105, the logging shear wave velocity prediction model completed in step 104 is directly applied to the target area to be predicted well to complete the logging shear wave velocity prediction.

[0123] Finally, it should be pointed out that: the above only describes the preferred embodiments of the present application and is not used to limit the present application, although the present application has been described in detail with reference to the foregoing embodiments, for those skilled in the art, the technical solutions recorded in the foregoing embodiments can be modified, or some technical features can be replaced. Any modification, equivalent replacement, improvement, etc. within the spirit and principles of the present application shall be included in the protection scope of the present application.

[0124] In addition to the technical features described in the specification, they are known to those skilled in the art.

Claims

1. A shear wave velocity prediction method based on deep learning, characterized in that: The deep learning-based shear wave velocity prediction method includes: Step 1: Obtain the P-wave velocity curve Vp, the neutron porosity curve CNL, the density curve DEN, the porosity curve POR, the shale content curve SH, and the corresponding measured S-wave velocity curve Vs; Step 2, preprocessing the well logging curve; Step 3, constructing a shear wave velocity prediction model; Step 4: Modify the model parameters through continuous iteration until the model stability and prediction accuracy meet the requirements; Step 5: using the shear wave velocity prediction model to predict the shear wave velocity of the well to be predicted; In step 2, the well logging curves are pre-processed, including wellbore collapse correction, outlier processing, and multi-well consistency processing, to obtain a set of well logging curves with high quality and good inter-well comparability; Multi-well consistency processing operates on the Vp curve to eliminate the inconsistency of units between different wells and the systematic errors caused by different instruments and different time measurements; In step 3, the deep feedforward neural network (DFNN) algorithm is used to establish a nonlinear relationship between the P-wave velocity curve Vp, the neutron porosity curve CNL, the density curve DEN, the porosity curve POR, the shale content curve SH, and the S-wave velocity curve Vp, and a S-wave velocity prediction model is constructed. In step 3, a number of wells are selected as training wells. A deep feedforward neural network is used to establish a nonlinear mapping relationship between the five well logging curves (P-wave velocity curve Vp, neutron porosity curve CNL, density curve DEN, porosity curve POR, and shale content curve SH) and the shear wave velocity curve Vs, thereby constructing a shear wave velocity prediction model. The DFNN network used contains two hidden layers, each containing three neurons. During model training, the input layer is the feature space of the training samples, that is, the reservoir parameters used to predict the shear wave velocity, and the output layer is the shear wave velocity. In step 4, the neural network parameters are further adjusted based on the error between the predicted shear wave velocity and the measured shear wave velocity in the verification well, and the most appropriate neural network parameters are continuously iterated to complete the construction of the final well logging shear wave velocity prediction model.

2. The shear wave velocity prediction method based on deep learning according to claim 1, characterized in that: In step 1, wells with measured shear wave velocity in the study area and adjacent areas are collected. These wells must also have a compressional wave velocity curve Vp, a neutron porosity curve CNL, a density curve DEN, a porosity curve POR, and a shale content curve SH. The three curves Vp, CNL, and DEN are measured curves, while the two curves POR and SH are obtained through well logging interpretation.

3. The shear wave velocity prediction method based on deep learning according to claim 1, characterized in that: In step 2, the borehole collapse correction and outlier processing of the logging curve adopt the multivariate regression method. By establishing the regression formula of the normal curve segment, it is applied to the correction of the borehole collapse segment and outlier value.

4. The shear wave velocity prediction method based on deep learning according to claim 1, characterized in that: In step 5, the well logging shear wave velocity prediction model is directly applied to the wells to be predicted in the target area. By inputting five well logging curves, namely the compressional wave velocity curve Vp, the neutron porosity curve CNL, the density curve DEN, the porosity curve POR, and the shale content curve SH, the shear wave velocity Vs is directly obtained to complete the shear wave velocity prediction.

Images