A reservoir parameter prediction method based on fused convolutional neural networks

Abstract

This invention relates to a reservoir parameter prediction method based on a fused convolutional neural network, belonging to the field of petroleum exploration and development technology. The invention performs convolution operations on seismic waveform data, seismic edge attribute data, and acoustic impedance inversion data near well logging data, respectively. These three convolution operations are then input into a fused convolutional network. By training this network, a relationship is established between the well logging data and the aforementioned seismic data. Using this relationship and the corresponding seismic data, reservoir parameters in areas lacking well logging data can be predicted. Therefore, this invention can estimate well logging data in areas lacking data using relevant seismic data, thereby enabling reservoir parameter prediction and avoiding the problem of inaccurate reservoir parameter prediction due to a lack of well logging data.

Description

Technical Field

[0001] This invention relates to a reservoir parameter prediction method based on a fused convolutional neural network, belonging to the field of petroleum exploration and development technology. Background Technology

[0002] Currently, reservoir parameter prediction is mainly achieved through pre-stack and post-stack inversion or extraction of conventional attributes such as seismic amplitude and frequency, directly or indirectly (by establishing a correlation between attributes and reservoir parameters). When rock physical parameters cannot distinguish lithology or establish a relationship between attributes and reservoir parameters, reservoir parameter prediction will not be possible.

[0003] In recent years, with the continuous development of artificial intelligence technology and its in-depth application in the field of oil exploration, a series of technologies and algorithms for reservoir characterization have emerged, such as CNN, RNN, DNN, Boltzmann machine, and GAN. Among them, the most popular is the reservoir characterization technology based on deep learning. This method predicts the lithology, physical properties, and hydrocarbon potential of reservoirs by training the network to learn the nonlinear relationship between reservoir parameters and seismic amplitude, frequency, waveform, etc., thereby solving the problem that previous methods could not make predictions because a clear relationship could not be established between wells and seismic data. Reservoir characterization technology based on popular deep learning algorithms is a type of technology based on big data. Its significant feature is that it requires a comprehensive and rich sample set. However, in the reality of oil and gas exploration, well data and the number of wells in many early exploration or low-grade study areas are very scarce and unevenly distributed. Existing popular deep learning reservoir characterization technologies do not have universality. Summary of the Invention

[0004] The purpose of this invention is to provide a reservoir parameter prediction method based on a fused convolutional neural network, so as to solve the problem that reservoir parameter prediction cannot be accurately achieved when there is limited well logging data.

[0005] To address the aforementioned technical problems, this invention provides a reservoir parameter prediction method based on a fused convolutional neural network, comprising the following steps:

[0006] 1) Acquire seismic data of the work area and wells in the work area that lack reservoir parameters. Determine the corresponding seismic data volume with each sampling point in the well as the center. The seismic data volume includes post-stack seismic data, seismic data edge attributes and wave impedance inversion data volume.

[0007] 2) The acquired seismic data volume is input into the trained fusion convolutional neural network model, which processes the input seismic data volume to obtain the corresponding well logging data, thereby enabling the prediction of reservoir parameters.

[0008] The fusion convolutional neural network model is used to perform convolution operations on the input post-stack seismic data, seismic data edge attributes, and wave impedance inversion data volume, and then fuse the results of each convolution operation to output the corresponding well logging data.

[0009] The training samples used in the training of the fused convolutional neural network model consist of well logging data with reactive reservoir parameters within the work area and their corresponding seismic data volumes.

[0010] This invention performs convolution operations on seismic waveform data, seismic edge attribute data, and acoustic impedance inversion data near well logging data, respectively. These three convolution operations are then input into a fused convolutional network. By training this network, a relationship is established between the well logging data and the aforementioned seismic data volumes. Using this relationship and the corresponding seismic data volumes, reservoir parameters in areas lacking well logging data can be predicted. Therefore, this invention can estimate well logging data in areas lacking data using relevant seismic data, thereby enabling reservoir parameter prediction and avoiding the problem of inaccurate reservoir parameter prediction due to a lack of well logging data.

[0011] Furthermore, the fused convolutional neural network model includes a convolution operation module, a combined convolution module, and a fully connected layer. The convolution operation module includes several convolutional layers for performing convolution operations on post-stack seismic data, seismic data edge attributes, and acoustic impedance inversion data volumes, respectively. The combined convolution module includes several combined convolutional layers for fusing the convolution operation results and transmitting the processed results to the fully connected layer after pooling. The fully connected layer is used to synthesize the pooling operation results to output predicted well logging data.

[0012] Furthermore, the logging data is natural gamma or porosity.

[0013] Furthermore, the edge attributes of the seismic data are dip angle, instantaneous phase, or seismic coherence volume.

[0014] Furthermore, the output of the fused convolutional neural network model is:

[0015]

[0016] Among them, Y c W represents the output of the c-th layer. c Let f represent the weight matrix from layer (c-1) to layer c, where f is the activation function and θ is the weight matrix. c Let K be the threshold vector for the c-th layer, where K represents 1, 2, and 3, which represent the post-stack seismic data, seismic data edge attributes, and wave impedance inversion data volume, respectively.

[0017] Furthermore, during the training process of the fused convolutional neural network model, its parameters are updated based on the error between the actual output and the expected output of the fused convolutional neural network. The parameter update amount of the fused convolutional neural network model is:

[0018]

[0019] ΔW=ηY c (1-Y c )Wξ c+1

[0020]

[0021] ξ c =Y c (1-Y c )Wξ c+1

[0022] Where, Δw c ΔW represents the weight update amount, and Δθ c η represents the threshold update amount, and η represents the learning rate.

[0023] Furthermore, to ensure that the maximum and minimum values ​​of the logging data are within the same range, the logging data in the training samples need to be standardized.

[0024] Furthermore, to ensure that the established relationship can accurately characterize the relationship between well logging data and seismic data volume, the seismic data volume corresponding to the well logging data is determined by taking the sampling point of the well logging data as the center and extracting the post-stack seismic data, seismic data edge attributes, and wave impedance inversion data volume within the wavelet wavefield range of two orthogonal directions passing through the sampling point. Attached Figure Description

[0025] Figure 1 This is a flowchart of the reservoir parameter prediction method based on fused convolutional neural networks of the present invention;

[0026] Figure 2-a These are the logging curves of multiple wells before standardization processing in the embodiments of the present invention;

[0027] Figure 2-b These are the well logging curves after multi-well standardization processing in the embodiments of the present invention;

[0028] Figure 3 This is a schematic diagram illustrating the construction of the sample set in this invention;

[0029] Figure 4 This is a diagram of the fused convolutional network model used in this invention;

[0030] Figure 5This is a lithological prediction profile obtained in the embodiments of the method of the present invention;

[0031] Figure 6 This is a property prediction profile obtained in an embodiment of the method of the present invention. Detailed Implementation

[0032] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0033] This invention performs convolution operations on seismic waveform data, seismic edge attribute data, and acoustic impedance inversion data near well logging data, respectively. These three convolution operations are then input into a fused convolutional network. By training this network, a relationship is established between well logging data and the three types of seismic data. Using this relationship and the corresponding three types of seismic data, predictions can be made about the well logging data. The predicted well logging data is then used to predict reservoir parameters. The implementation process of this method is as follows: Figure 1 As shown, the specific implementation process is as follows.

[0034] 1. Acquire seismic data and well logging data that can reflect reservoir parameters.

[0035] Reservoir parameters mainly include lithological parameters and physical properties. Research has shown that natural gamma ray spectroscopy is the logging data that reflects reservoir lithology, while porosity is the logging data that reflects reservoir physical properties. Therefore, the logging data obtained in this invention mainly includes natural gamma ray spectroscopy and porosity. After obtaining the logging data from each well in the study area, standardization is required to facilitate subsequent processing and ensure consistency in the logging data range across multiple wells, so that the maximum and minimum values ​​of the same logging data fall within the same range. For this embodiment, it is assumed that the obtained logging data is as follows... Figure 2-a As shown in the figure, after standardization, the corresponding well logging data is as follows: Figure 2-b As shown.

[0036] The acquired seismic data is processed, including extracting seismic data edge attributes and wave impedance inversion data volumes. Seismic data edge attributes are dominant attributes that reflect stratigraphic contact relationships, such as dip angle, instantaneous phase, and seismic coherence volume. Wave impedance inversion data volumes are obtained by inverting post-stack seismic data volumes.

[0037] 2. Samples are constructed based on well logging data, edge attributes of seismic data, wave impedance inversion data volume, and post-stack seismic data.

[0038] When constructing the sample, the sampling point P of the well logging data is taken as the center, such as... Figure 3As shown, n seismic waveform data, impedance inversion data, and seismic edge attribute data are extracted from two orthogonal azimuths passing through point P. The extraction length is the wavelet length. Through this process, sample data centered on the well logging data sampling point P can be established. That is, a sample includes a well logging data point and several seismic waveform data, impedance inversion data, and seismic edge attribute data within its local wavelet length range. The well logging data is either natural gamma or porosity.

[0039] 3. Establish a fusion convolutional network model and train the model using the constructed samples.

[0040] The fusion convolutional neural network model (hereinafter referred to as fusion convolutional network) established in this invention is as follows: Figure 4 As shown, it includes a convolution operation module, a combined convolution module, and a fully connected layer. The convolution operation module includes several convolution layers, which are used to perform convolution operations on the post-stack seismic data, seismic data edge attributes, and acoustic impedance inversion data volumes, respectively. The combined convolution module includes several combined convolution layers, which are used to fuse the convolution operation results and transmit the processed results to the fully connected layer after pooling. The fully connected layer is used to synthesize the pooling operation results to output the predicted well logging data.

[0041] Specifically, the convolution operation module includes three convolution kernels with sizes w1, w2, and w3, and network layers C1, C2, and C3, respectively. The convolution operation for seismic waveform data uses a model with a kernel size of w1 and a network layer count of C1; the convolution operation for wave impedance data uses a model with a kernel size of w2 and a network layer count of C2; and the convolution operation for seismic edge attribute data uses a model with a kernel size of w3 and a network layer count of C3. The specific convolution operation process is as follows:

[0042] The seismic waveform data (i.e., post-stack seismic data) is convolved and propagated forward along the network, as shown in the following formula:

[0043]

[0044] c = 1, 2, ..., C1 (1)

[0045] Among them, Y1 c This represents the output vector of the c-th layer. This represents the kernel size from layer c-1 to layer c.

[0046] The wave impedance data is convolved and propagated forward along the network, as shown in the following formula:

[0047]

[0048] c = 1, 2, ..., C2 (2)

[0049] in This represents the output vector of the c-th layer. This represents the kernel size from layer c-1 to layer c.

[0050] The earthquake edge properties are convolved and propagated forward along the network, as shown in the following formula:

[0051]

[0052] c = 1, 2, ..., C3 (3)

[0053] in This represents the output vector of the c-th layer. This represents the kernel size from layer c-1 to layer c. The combined convolution module processes the three convolution operations described above to obtain the output of the fused convolutional network.

[0054]

[0055] Among them, Y c W represents the output of the c-th layer. c Let f represent the weight matrix from layer c-1 to layer c, f be the activation function (to find the maximum value of the output of the c-th layer network), and θ be the value of the layer. c Let K be the threshold vector for the c-th layer, where K represents 1, 2, or 3.

[0056] The model is trained using the samples constructed in step 2. During training, the parameters of the fused convolutional network need to be adjusted by fusing the error between the actual output and the expected output of the convolutional network. The actual output Y of the fused convolutional network... c The expected output is for the predicted well logging data (e.g., natural gamma). The process of updating parameters for the fused convolutional network model in the sample data (such as natural gamma) is as follows:

[0057] 1) Calculate the error between the actual output and the expected output of the fused convolutional network:

[0058]

[0059] Where E c Y is the error value. c Indicates the actual output. This indicates the expected output.

[0060] 2) Calculate the parameter update amount of the entire fused convolutional neural network based on the error magnitude:

[0061]

[0062] Where Δw c ΔW represents the weight update amount, and Δθ cξ represents the threshold update amount, η represents the learning rate, and ξ represents the learning rate. c The expression is as follows:

[0063] ξ c =Y c (1-Y c )Wξ c+1 (7)

[0064] 3) Use the parameter update values ​​obtained in 2) to update the parameters of the fused convolutional network:

[0065]

[0066] 4) Repeat steps 1) to 3) until the error meets the requirements.

[0067] Through the above process, the fusion convolutional network model of the present invention can be trained.

[0068] 4. Use the trained model to predict reservoir parameters.

[0069] For areas lacking well logging data where reservoir parameter prediction is impossible, select certain points within the area as well logging data sampling points. Centered on these sampling points, within the wavelet wavelength range, according to... Figure 3 The corresponding seismic data volume is extracted in a certain way. The seismic data volume includes seismic waveform data, wave impedance inversion data volume and edge attribute data volume. The obtained seismic data volume is input into the trained fusion convolutional network model for learning to obtain the well logging data of the sampling point. Similarly, well logging data of multiple points can be obtained. Based on the obtained well logging data, the reservoir parameters can be predicted.

[0070] When the reservoir parameter to be predicted is a reservoir lithology parameter, the well logging data required is natural gamma. In this case, the sample used in step 3 for model training consists of natural gamma values ​​and corresponding seismic data volumes. When the reservoir parameter to be predicted is a reservoir lithology parameter, the well logging parameter required is porosity. In this case, the sample used in step 3 for model training consists of porosity values ​​and corresponding seismic data volumes.

[0071] The stratigraphic lithology profile of well WG2 predicted using this invention is as follows: Figure 5 As shown, the formation property profile of well WG2 predicted using this invention is as follows: Figure 6 As shown, the lithology and physical properties are consistent with the actual formation. Therefore, this invention establishes a fused convolutional network model, uses this model to determine the relationship between well logging data reflecting reservoir parameters and the corresponding seismic data volume, and based on the established relationship, well logging data can be estimated for wells lacking data, and then the reservoir parameters can be predicted based on the estimated well logging data.

Claims

1. A reservoir parameter prediction method based on a fused convolutional neural network, characterized in that, The method includes the following steps: 1) Obtain seismic data for the work area and wells in the work area that lack logging data reflecting reservoir parameters. Determine the corresponding seismic data volume with each sampling point in the well as the center. The seismic data volume includes post-stack seismic data, seismic data edge attributes, and wave impedance inversion data volume. The seismic data edge attributes are the dominant attributes that reflect the formation contact relationship. 2) The acquired seismic data volume is input into the trained fusion convolutional neural network model, which processes the input seismic data volume to obtain the corresponding well logging data, thereby enabling the prediction of reservoir parameters. The fusion convolutional neural network model described above is used to perform convolution operations on the input post-stack seismic data, seismic data edge attributes, and wave impedance inversion data volumes using the convolution operation module, and then fuses the results of each convolution operation to output the corresponding well logging data; the convolution operation module propagates forward along the network when performing convolution operations on each data volume in the seismic data volume; The training samples used in the training of the fused convolutional neural network model consist of well logging data reflecting reservoir parameters and their corresponding seismic data volumes within the work area.

2. The reservoir parameter prediction method based on a fused convolutional neural network according to claim 1, characterized in that, The fusion convolutional neural network model includes a convolution operation module, a combined convolution module, and a fully connected layer. The convolution operation module includes several convolutional layers for performing convolution operations on post-stack seismic data, seismic data edge attributes, and acoustic impedance inversion data volumes, respectively. The combined convolution module includes several combined convolutional layers for fusing the convolution operation results and transmitting the processed results to the fully connected layer after pooling. The fully connected layer is used to synthesize the pooling operation results to output predicted well logging data.

3. The reservoir parameter prediction method based on a fused convolutional neural network according to claim 1 or 2, characterized in that, The well logging data mentioned are natural gamma or porosity.

4. The reservoir parameter prediction method based on a fused convolutional neural network according to claim 1 or 2, characterized in that, The edge attributes of the earthquake data are dip angle, instantaneous phase, or seismic coherence volume.

5. The reservoir parameter prediction method based on a fused convolutional neural network according to claim 2, characterized in that, The output of the fused convolutional neural network model is: C = in, This represents the output of the c-th layer. Let f represent the weight matrix from layer (c-1) to layer c, and f be the activation function. Let c be the threshold vector of the c-th layer. 1, 2, and 3 represent post-stack seismic data, seismic data edge attributes, and wave impedance inversion data volume, respectively.

6. The reservoir parameter prediction method based on a fused convolutional neural network according to claim 5, characterized in that, During the training of the fusion convolutional neural network model, its parameters are updated based on the error between the actual output and the expected output of the fusion convolutional network. The parameter update amount of the fusion convolutional neural network model is: in, , Indicates the weight update amount. Indicates the threshold update amount. This represents the learning rate.

7. The reservoir parameter prediction method based on a fused convolutional neural network according to claim 1 or 2, characterized in that, The well logging data in the training samples need to be standardized.

8. The reservoir parameter prediction method based on a fused convolutional neural network according to claim 1 or 2, characterized in that, When determining the seismic data volume corresponding to the well logging data, it is done by taking the sampling point of the well logging data as the center and extracting the post-stack seismic data, seismic data edge attributes, and wave impedance inversion data volume within the wavelet wavefield range of two orthogonal directions passing through the sampling point.

Images