A reservoir prediction method

Abstract

The present invention relates to the technical field of oil and gas exploration, and specifically relates to a reservoir prediction method. The method obtains seismic data and preprocesses the seismic data. Among them, the preprocessing includes converting the values of the seismic data into values between 0 and 1 to obtain a seismic image of seismic profile gray scale; making a training set and a test set for training a seismic feature autoencoder, and sending them into the seismic feature autoencoder for training and testing; sending the seismic image into the seismic feature autoencoder after training and testing to extract hidden features; obtaining the training set labels and training set inputs required by the thickness estimation depth network, and training the thickness estimation depth network based on the training set labels and training set inputs; inputting the hidden features into the trained thickness estimation depth network to predict the sandstone thickness at the remaining unknown positions of the seismic profile, and obtaining the sandstone thickness distribution of the entire profile.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas exploration, and particularly to a reservoir prediction method. Background Art

[0002] The distribution characteristics of each attribute value of seismic data at different time periods shown by seismic attribute slices. Using seismic attributes to predict the sandstone thickness distribution in the reservoir section is an important means of current 3D seismic comprehensive interpretation. In river and delta systems, thin-layer and thin-interbed sandstones are the main exploration targets, and the result of predicting the sandstone thickness through seismic attributes can provide geological guidance for further well placement.

[0003] Seismic attributes play an important role in reservoir prediction. In traditional sandstone thickness calculation, seismic sedimentology is mainly used to analyze seismic attributes; then, manually and qualitatively select the attribute slice that is most similar to the geological model; establish a fitting relationship between the finally selected attribute and the well-log sandstone thickness parameter, and use the attribute value beside the well to predict the sandstone thickness data trend; finally, guided by the predicted thickness trend, perform constraint correction according to the actual sandstone thickness obtained from well logging, and finally generate a sandstone thickness prediction map. However, the traditional method is not only affected by geological conditions, such as: affected by geological conditions such as interlayer conditions, the thickness of the sand body itself, and the reservoir location; at the same time, different selected seismic attributes cannot intuitively reflect the degree of sand body development in the study area and there is interference between adjacent layers. This makes the selected attributes may reflect the information of adjacent sedimentary units, resulting in poor well-seismic matching effect on the horizon slice; even in the area of well-seismic matching, the sand body thickness cannot be quantitatively described. Eventually, the prediction effect of sand body thickness using the selected seismic attributes is poor. Summary of the Invention

[0004] In order to solve the above technical problems existing in the prior art, the present invention provides a reservoir prediction method.

[0005] To achieve the above object, the embodiments of the present invention provide the following technical solutions:

[0006] In a first aspect, in an embodiment provided by the present invention, a reservoir prediction method is provided, and the method includes the following steps:

[0007] Obtain seismic data and preprocess the seismic data, wherein the preprocessing includes converting the value of the seismic data to between 0 and 1 to obtain a seismic image of seismic profile gray scale;

[0008] Make a training set and a test set for training a seismic feature autoencoder, and send them into the seismic feature autoencoder for training and testing;

[0009] Send the seismic image into the seismic feature autoencoder after training and testing to extract hidden features;

[0010] Obtain the training set labels and training set inputs required for the thickness estimation depth network, and train the thickness estimation depth network based on the training set labels and training set inputs;

[0011] Input the latent features into the trained thickness estimation depth network to predict the sandstone thickness at the remaining unknown positions of the seismic profile, and obtain the sandstone thickness distribution of the entire profile.

[0012] As a further solution of the present invention, the preprocessing of the seismic data is specifically to perform preprocessing along the root mean square amplitude value of the interlayer seismic data, and then normalize the root mean square amplitude value along the layer to [0,1].

[0013] As a further solution of the present invention, the objective function of the preprocessing is:

[0014]

[0015] where, X nor ∈R a×b is the value after preprocessing, X ∈ R a×b is the original seismic data, X min and X max are the minimum value and the maximum value in the seismic data respectively, and a and b represent the size of the seismic image respectively.

[0016] As a further solution of the present invention, the method for specifically making the training set and test set for training the seismic feature autoencoder is as follows:

[0017] S201. Adopt a sliding time window, and successively use each pixel point as the center on the seismic image, and intercept an image block x with a size of p×p at a sliding step of i , and fill the image block x i at the edge of the seismic image with zero padding for null values;

[0018] S202. Randomly divide the obtained image blocks into a training set and a test set, and train and test the seismic feature autoencoder based on the obtained training set and test set.

[0019] As a further solution of the present invention, the training completion criterion of the seismic feature autoencoder is to monitor the reconstruction effect through the reconstruction loss function. If the reconstruction loss function converges, the training is completed.

[0020] As a further solution of the present invention, the method for specifically extracting latent features by inputting the seismic image into the seismic feature autoencoder after training and testing is as follows:

[0021] Assume that the image block with a size of x i ∈R p×p becomes a size of x after being processed by the seismic feature autoencoderi '∈R 1 ×p ' vector, then the entire seismic image X nor After being processed by the seismic feature autoencoder f k All the hidden features are obtained and form a set Z.

[0022] As a further solution of the present invention, the formula for the seismic feature autoencoder to extract features is as follows:

[0023] Z = f k (X nor ),

[0024] where X nor ∈R a×b is the value after preprocessing, Z ∈ R n×p ' (where the size of n is n = a × b) is the set of all the obtained hidden features, and a and b respectively represent the size of the seismic image.

[0025] As a further solution of the present invention, the seismic feature autoencoder includes an encoder and a decoder;

[0026] The encoder is used to downsample the input seismic image to obtain hidden features;

[0027] The decoder is used to upsample the obtained hidden features and output the final reconstructed seismic data.

[0028] As a further solution of the present invention, the formula for the thickness estimation depth network is as follows:

[0029] W n×1 = Φ(Z),

[0030] where Φ represents the thickness estimation depth network, Z ∈ R n×p ' is all the hidden features extracted from the seismic image by the seismic feature autoencoder, and W n×1 ∈R n×1 is the predicted value of the thickness estimation depth network.

[0031] As a further solution of the present invention, among them, the training set label is the sandstone thickness value at the known position obtained by combining the seismic attribute data of the horizons at the known positions and well logging curves and lithology analysis;

[0032] The training set input is the hidden features at the known positions extracted by the seismic feature autoencoder.

[0033] The technical solution provided by the present invention has the following beneficial effects:

[0034] The method of the present invention replaces the traditional manual seismic interpretation. In the traditional method, the known sandstone thickness in the well is used, and under the constraint of optimized seismic attributes, the interpolation calculation of the sandstone thickness is carried out.

[0035] The method of the present invention combines artificial intelligence and geophysics, and uses a seismic feature autoencoder to extract all the implicit features of seismic data. The trained thickness estimation deep network is used for all the extracted implicit features to obtain the reservoir sandstone thickness.

[0036] In terms of quantitative interpretation, the present invention is based on the sandstone thickness value in the well and the implicit features of the seismic data extracted including seismic facies information, and applies the thickness estimation deep network to predict the sandstone thickness distribution of the target reservoir section, which can further provide geological guidance for well placement.

[0037] These aspects or other aspects of the present invention will be more clearly understood in the following description of the embodiments. It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and cannot limit the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other embodiments can be obtained based on these drawings without creative efforts.

[0039] Figure 1 It is a flowchart of a reservoir prediction method according to an embodiment of the present invention.

[0040] Figure 2 It is a specific flowchart of a reservoir prediction example based on the combination of a seismic feature autoencoder and a thickness estimation deep network according to an embodiment of the present invention.

[0041] Figure 3 It is the value of the root mean square amplitude attribute along the layer after preprocessing according to an embodiment of the present invention.

[0042] Figure 4 It is a schematic diagram of the process of making a training set and a test set according to an embodiment of the present invention.

[0043] Figure 5 It is a structural diagram of a seismic feature autoencoder according to an embodiment of the present invention.

[0044] Figure 6Visualization diagrams of the first four features of all the implicit features extracted in an embodiment of the present invention; (a) is the first feature of all the implicit features; (b) is the second feature of all the implicit features; (c) is the third feature of all the implicit features; (d) is the fourth feature of all the implicit features.

[0045] Figure 7 Structural diagram of the spatial channel squeeze-and-excitation module inside the thickness estimation deep network in an embodiment of the present invention.

[0046] Figure 8 Reservoir sandstone thickness prediction diagram in an embodiment of the present invention. Detailed implementation manners

[0047] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0048] The flowcharts shown in the accompanying drawings are only illustrative examples, and do not necessarily include all the contents and operations / steps, nor do they necessarily need to be executed in the described order. For example, some operations / steps can also be decomposed, combined, or partially merged, so the actual execution order may be changed according to the actual situation.

[0049] It should be understood that the terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the specification of the present invention and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an", and "the" are intended to include the plural forms.

[0050] Specifically, the embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0051] Please refer to Figure 1 and 2 , Figure 1 is a flowchart of a reservoir prediction method provided by an embodiment of the present invention. As shown in Figure 1 , this method is based on a seismic feature autoencoder and a thickness estimation deep network; this reservoir prediction method includes steps S10 to S50.

[0052] S10. Obtain seismic data and preprocess the seismic data. Among them, the preprocessing includes converting the values of the seismic data into the range between 0 and 1 to obtain a seismic image of the seismic profile gray level. In this way, reconstructing the seismic data by using the seismic feature autoencoder corresponds to reconstructing the seismic gray level image.

[0053] The seismic data is the root-mean-square amplitude data along the layers.

[0054] The preprocessing of the seismic data is specifically to preprocess the root-mean-square amplitude values of the seismic data along the layers, and then normalize the root-mean-square amplitude values along the layers to [0, 1]. The preprocessed root-mean-square amplitude values along the layers are as Figure 3 shown. The purpose of this is to reduce the difficulty of network training and improve the accuracy of reconstructing seismic data.

[0055] In the embodiment of the present invention, the objective function of the preprocessing is:

[0056]

[0057] where X nor ∈R a×b is the value after preprocessing, X ∈ R a×b is the original seismic data, X min and X max are respectively the minimum and maximum values in the seismic data, and a and b respectively represent the size of the seismic image.

[0058] S20. Prepare the training set and test set for training the seismic feature autoencoder, and send them into the seismic feature autoencoder for training and testing;

[0059] The specific method for preparing the training set and test set for training the seismic feature autoencoder is as follows:

[0060] As Figure 4 shown, S201. Adopt a sliding time window, and successively take an image block x with a size of p×p centered on each pixel point on the seismic image, and fill the image block x i at the edge of the seismic image with zero padding to fill the null value; where the p value is the side length value of the intercepted image block, and the side length interval between the corresponding sides of the first intercepted image block and the second image block is i the length of .

[0061] S202. Randomly divide the obtained image blocks into a training set and a test set, and train and test the seismic feature autoencoder based on the obtained training set and test set.

[0062] In the embodiment of the present invention, the training completion criterion of the seismic feature autoencoder is to monitor the reconstruction effect through the reconstruction loss function. If the reconstruction loss function converges, the training is completed.

[0063] Among them, the formula of the reconstruction loss function is as follows:

[0064] Lrec = ηL SmoothL1 + L MSE , (2)

[0065]

[0066]

[0067] wherein, L MSE refers to the root mean square error loss function, and L SmoothL1 refers to the SmoothL1 loss function. η ∈ (0, 1) is the weight of the root mean square error and the SmoothL1 loss function. x i refers to the input image patch of the seismic feature autoencoder, with a size of x i ∈ R p×p ; refers to the image patch reconstructed by the seismic feature autoencoder. n represents that there are n seismic image patches in total in the training set and the test set used for training the seismic feature autoencoder.

[0068] S30. Feed the seismic image into the seismic feature autoencoder after training and testing to extract the latent features.

[0069] In the embodiments of the present invention, the seismic image is the normalized root mean square amplitude data along the interlayer.

[0070] In the embodiments of the present invention, the seismic feature autoencoder extracts the latent features of the seismic image. The objective function for the seismic feature autoencoder to extract the latent variables is:

[0071] x i ' = f k (x i ), (5)

[0072] represents that the seismic image patch with a size of x i ∈ R p×p becomes a vector with a size of x i ' ∈ R 1×p ' after being processed by the encoder. Feed the image patch x i into the encoder f k of the seismic feature autoencoder to extract the latent features x i ' of the image patch. x i ' contains all the information of the image patch x i . In subsequent steps, replace x i with x i . The purpose of this part is: since the image patch is intercepted from the complete seismic image by a sliding time window, so as to understand an image patch x iBy understanding the process of size change after the encoder of the seismic feature autoencoder processes, one can understand the size change of the complete seismic image after being processed by the encoder of the seismic feature autoencoder in the subsequent steps.

[0073] In an embodiment of the present invention, the step of feeding the seismic image into the trained and tested seismic feature autoencoder to extract latent features is exemplified as follows:

[0074] Suppose an image patch of size x i ∈R p×p becomes a vector of size x' i ∈R 1 ×p ' after being processed by the seismic feature autoencoder. Then, after the entire seismic image X nor is processed by the seismic feature autoencoder f k , all latent features are obtained and form a set Z (the visualization of the first four latent features is shown as Figure 6 ).

[0075] In an embodiment of the present invention, the trained seismic feature autoencoder extracts all latent features along the interlayer root mean square amplitude value. The objective function for the seismic feature autoencoder to extract all latent variables is: From Equation 5, we can know the size change of an image patch after being encoded by the seismic feature autoencoder. Then, when we feed the complete seismic image X nor into the trained seismic feature autoencoder, we can obtain the abstract features Z of the complete seismic image. The size of the obtained Z is related to the size of x' i . For example, a complete seismic image of size X nor ∈R a×b can be divided into n = a × b image patches. After an image patch passes through the seismic feature autoencoder, its size becomes a vector of R' 1×p . Then, after n image patches pass through the seismic feature autoencoder, their sizes become n vectors of R' 1×p , that is: Z ∈ R' n×p .

[0076] Z = f k (X nor ), (6)

[0077] where X nor represents the entire interlayer root mean square amplitude data after preprocessing. After the encoder f k of the seismic feature autoencoder extracts features, all latent features are obtained to form a set Z ∈ R' n×p (where the size of n is n = a × b), and a and b respectively represent the sizes of the seismic image.

[0078] In an embodiment of the present invention, as Figure 5 shown, the seismic feature autoencoder includes an encoder and a decoder;

[0079] The encoder is used to downsample the input seismic image to obtain latent features;

[0080] The decoder is used to upsample the obtained latent features and output the final reconstructed seismic data. It is necessary to calculate the reconstruction error between the reconstructed seismic data and the corresponding input seismic data. During the training process of the seismic feature autoencoder, only when the reconstruction error gradually converges and stabilizes as the number of training times increases, can it be considered that the latent features extracted by the seismic feature autoencoder are effective and can fully contain the abstract features of the input seismic data.

[0081] In an embodiment of the present invention, the encoder f k includes a convolutional autoencoder f kcae and a variational autoencoder f kvae . The convolutional autoencoder f kcae is used to extract the latent feature Z CAE from the seismic image; the variational autoencoder f kvae is used to extract the latent feature Z VAE from the seismic image.

[0082] The latent feature Z kcae extracted by the convolutional autoencoder f CAE is weighted and summed with the latent feature Z kvae extracted by the variational autoencoder f VAE to obtain the final latent feature Z CAE+VAE .

[0083] The input of the decoder g k consists of two parts. The first part is Z CAE , and the second part is Z CAE+VAE . Z CAE and Z CAE+VAE are sent to the first decoding layer g k_1 of the decoder to obtain two decoding results g k_1 (Z CAE ) and g k_1 (Z CAE+VAE ). After the two decoding results are weighted and summed, they are sent together to the remaining decoding layers g k of the decoder to obtain the reconstructed seismic data The specific calculation process of a seismic picture block in the seismic feature autoencoder is as follows:

[0084] f k = f kcae + f kvae, (7)

[0085] Z CAE = f kcae (x),(8)

[0086] Z VAE = f kvae (x),(9)

[0087] Z CAE+VAE = 0.6 × Z CAE + 0.4 × Z VAE ,(10)

[0088]

[0089] S40. Obtain the training set labels and training set inputs required for the thickness estimation depth network, and train the thickness estimation depth network based on the training set labels and training set inputs (as Figure 2 , Figure 7 shown);

[0090] In the embodiments of the present invention, the formula of the thickness estimation depth network is as follows:

[0091] W n×1 = Φ(Z),

[0092] where Φ represents the thickness estimation depth network, Z ∈ R n×p ' is all the hidden features extracted from the seismic image by the seismic feature autoencoder, and W n×1 ∈ R n×1 is the predicted value of the thickness estimation depth network.

[0093] The training set label w i , is the sandstone thickness value at the known position obtained by combining the seismic attribute data of the horizon at the known position, well logging curves, and lithology analysis.

[0094] The training set input is the hidden features at the known position extracted by the seismic feature autoencoder.

[0095] Take the hidden features z i ∈ R 1×p ' extracted from the seismic image by the seismic feature autoencoder at the known position as the training set input of the thickness estimation depth network Φ, and the sandstone thickness value at the known position as the training set label w i ∈ R of the thickness estimation depth network for model training. Assume that after each hidden feature passes through all the pooling and convolutional operations Conv_pool, its hidden feature size becomes Get z' i , and after passing through the spatial channel squeeze-and-excitation module scSE, the feature size remains unchanged to get After passing through the first linear layer, the size of the hidden feature becomes obtaining After passing through the last linear layer, the hidden feature becomes a real number which is the predicted sandstone thickness value. All the hidden features are fed into the trained thickness estimation deep network to predict the sandstone thickness W at the remaining unknown positions of the seismic profile n×1 , obtaining the sandstone thickness of the entire profile The specific calculation process can be described as:

[0096] z' i = Conv_pool(z i ), (12)

[0097]

[0098]

[0099]

[0100] W n×1 = Φ(Z), (16)

[0101]

[0102] where Φ represents the thickness estimation deep network, Z ∈ R n×p ' are all the hidden features extracted from the seismic image by the seismic feature autoencoder, and W n×1 ∈ R n×1 is the predicted value of the thickness estimation deep network. The function represents changing the length size of from in to out, and the function represents changing the size of the input data W n×1 from R n×1 to R a×b , and obtaining the final sandstone thickness distribution map as Figure 8 shown.

[0103] S50. Input the hidden features into the trained thickness estimation deep network to predict the sandstone thickness at the remaining unknown positions of the seismic profile, and obtain the sandstone thickness distribution of the entire profile.

[0104] The method of the present invention replaces the traditional manual seismic interpretation. In traditional manual seismic interpretation, the known sandstone thickness in the well is used, and under the constraint of optimized seismic attributes, the interpolation calculation of the sandstone thickness is carried out. The method of the present invention combines artificial intelligence with geophysics, and uses a seismic feature autoencoder to extract all the implicit features of the seismic data. The trained thickness estimation deep network is used to obtain the reservoir sandstone thickness from all the extracted implicit features. In terms of quantitative interpretation, based on the sandstone thickness value in the well and the implicit features of the seismic data extracted including seismic facies information, the thickness estimation deep network is applied to predict the sandstone thickness distribution of the target reservoir section, which can further provide geological guidance for well placement.

[0105] Finally, the horizon data comes from the Niuzhuang Delta, which develops a delta slump turbidite sedimentary system mainly controlled by the axial east-west water flow. This horizon is undergoing a process in which the delta continuously advances from east to west and the lake basin continuously shrinks. Based on the geological background, it is considered that the upper and lower parts of this horizon are a delta slump turbidite sedimentary system, including seismic facies such as deep lake - semi-deep lake, shore - shallow lake, prodelta, delta front slope slump turbidite fan, distributary mouth bar, distal bar, etc. Slump turbidites often distribute in the paleotopographic lows in front of the main advancing direction of the delta, and the turbidite fans are distributed in a semi-circular shape outside the delta front. Combining the analysis of the sandstone thickness percentage obtained from well logging, the slump turbidite fans have many muddy interlayers and thin sandstone thickness. Since there is a correlation between seismic facies and sandstone thickness, the implicit features extracted by the seismic feature autoencoder can be regarded as a kind of seismic attribute containing seismic facies features. Therefore, the sandstone thickness distribution can be predicted based on the seismic attributes obtained from the implicit features extracted by the seismic feature autoencoder. Sandstone thickness prediction ( Figure 7 ) results show that the maximum value of the sandstone thickness distributes in the distal bar, followed by the distributary mouth bar, and the minimum value distributes in the deep lake - semi-deep lake, which is consistent with the delta sedimentation law.

[0106] It should be understood that although the above is described in a certain order, these steps are not necessarily executed in the above order successively. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover, a part of the steps in this embodiment may include multiple steps or multiple stages. These steps or stages are not necessarily executed at the same moment, but can be executed at different moments. The execution order of these steps or stages is not necessarily sequential, but can be executed alternately or in turn with at least a part of other steps or steps or stages in other steps.

[0107] It should be understood that, as used herein, unless the context clearly supports the exception, the singular form "a" is intended to also include the plural form. It should also be understood that the "and / or" used herein refers to any and all possible combinations of one or more of the associated listed items. The serial numbers of the disclosed embodiments of the present invention above are only for description and do not represent the advantages or disadvantages of the embodiments.

[0108] Those of ordinary skill in the art should understand that the discussion of any of the above embodiments is only exemplary and is not intended to imply that the scope of the disclosure of the embodiments of the present invention (including the claims) is limited to these examples; under the concept of the embodiments of the present invention, the technical features in the above embodiments or different embodiments can also be combined, and there are many other variations in different aspects of the embodiments of the present invention as above, which are not provided in detail for the sake of brevity. Therefore, any omission, modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the present invention shall be included within the protection scope of the embodiments of the present invention.

Claims

1. A reservoir prediction method, characterized in that, This method is based on a seismic feature autoencoder and a thickness estimation deep network; the method includes: Obtain seismic data and preprocess the seismic data. Among them, the preprocessing includes converting the values of the seismic data into values between 0 and 1 to obtain a seismic image of the seismic profile grayscale. Make a training set and a test set for training the seismic feature autoencoder, and send them into the seismic feature autoencoder for training and testing. Send the seismic image into the seismic feature autoencoder after training and testing to extract hidden features. Obtain the training set labels and training set inputs required by the thickness estimation deep network, and train the thickness estimation deep network based on the training set labels and training set inputs. Input the hidden features into the trained thickness estimation deep network to predict the sandstone thickness at the remaining unknown positions of the seismic profile, and obtain the sandstone thickness distribution of the entire profile. The seismic feature autoencoder includes an encoder and a decoder. The encoder f k includes a convolutional autoencoder f kcae and a variational autoencoder f kvae ; the convolutional autoencoder f kcae is used to extract the latent feature Z of the seismic image CAE ; the variational autoencoder f kvae is used to extract the latent feature Z of the seismic image VAE ; The latent feature Z extracted by passing through the convolutional autoencoder f kcae is weighted and summed with the latent feature Z extracted by passing through the variational autoencoder f CAE to obtain the final latent feature Z kvae ; VAE CAE+VAE ;​ The decoder g k has an input consisting of two parts. The first part is Z CAE , and the second part is Z CAE+VAE ; Z CAE and Z CAE+VAE are fed into the first decoding layer g k_1 of the decoder to obtain two decoding results g k_1 (Z CAE ) and g k_1 (Z CAE+VAE ). After weighted summation of the two decoding results, they are fed together into the remaining decoder g k of the decoder to obtain the reconstructed seismic data The specific calculation process of a seismic picture block in the seismic feature autoencoder is as follows: f k = f kcae + f kvae , (7) Z CAE = f kcae (x), (8) Z VAE = f kvae (x), (9) Z CAE+VAE = 0.6 × Z CAE + 0.4 × Z VAE , (10) 2. The reservoir prediction method according to claim 1, wherein The preprocessing of the seismic data is specifically to preprocess along the root mean square amplitude value of the interlayer seismic data, and then normalize the root mean square amplitude value along the layer to [0,1].

3. The reservoir prediction method according to claim 2, characterized in that The objective function of the preprocessing is as follows: Among them, X nor ∈R a×b is the value after preprocessing, X ∈ R a×b is the original seismic data, X min and X max are respectively the minimum value and the maximum value in the seismic data, and a and b respectively represent the size of the seismic image.

4. The reservoir prediction method according to claim 1, wherein, The specific method for making the training set and test set for training the seismic feature autoencoder is as follows: Using a sliding time window, centered on each pixel point in the seismic image in turn, and intercepting an image block x of size p×p with a sliding step size of ; and filling the image block x at the edge of the seismic image with null values by padding with zeros i ; i null values; Randomly divide all the obtained image blocks into a training set and a test set, and train and test the seismic feature autoencoder based on the obtained training set and test set.

5. The reservoir prediction method according to claim 1, characterized in that The training completion criterion of the seismic feature autoencoder is to monitor the reconstruction effect through the reconstruction loss function. If the reconstruction loss function converges, the training is completed.

6. The reservoir prediction method according to claim 1, wherein, The specific method for sending the seismic image into the seismic feature autoencoder after training and testing to extract hidden features is as follows: Assume the size is x i ∈R p×p The image patch after being processed by the seismic feature autoencoder becomes a vector of size x i '∈R 1×p' If so, for the entire seismic image X nor After being processed by the seismic feature autoencoder f k All the latent features are obtained and form a set Z 7. The reservoir prediction method according to claim 6, characterized in that The feature extraction formula of the seismic feature autoencoder is as follows: Z = f k (X nor ) Among them, X nor ∈R a×b is the value after preprocessing, Z ∈ R n×p' (where the size of n is n = a × b) is the set of all obtained latent features, and a and b respectively represent the size of the seismic image.

8. The reservoir prediction method according to claim 1, wherein The encoder is used to downsample the input seismic image to obtain hidden features. The decoder is used to upsample the obtained hidden features and output the final reconstructed seismic data.

9. The reservoir prediction method according to claim 1, wherein The formula of the thickness estimation deep network is as follows: W n×1 = Φ(Z), Among them, Φ represents the thickness estimation deep network, and Z ∈ R n×p' is all the latent features extracted from the seismic image by the seismic feature autoencoder, and W n×1 ∈ R n×1 is the predicted value of the thickness estimation deep network.

10. The reservoir prediction method according to claim 1, wherein Wherein The training set labels are the sandstone thickness values at the known positions obtained by combining the seismic attribute data of the horizons at the known positions, well logging curves, and lithology analysis. The training set inputs are the hidden features at the known positions extracted by the seismic feature autoencoder.

Images