Pre-stack inversion method, system, storage medium and processor

Abstract

The embodiment of the application provides a pre-stack inversion method, system, storage medium and processor, and belongs to the technical field of oil and gas exploration and development. The method comprises the following steps: acquiring a pre-stack seismic data volume, an AVO attribute volume and a low-frequency model of an inversion target area, wherein the low-frequency model is a low-frequency part extracted from a supervised model obtained based on a plurality of logging data and a horizon interpolation method of the inversion target area; according to a predetermined rule, each two-dimensional profile of the pre-stack seismic data volume, the AVO attribute volume and the low-frequency model is intercepted respectively, and is input into a pre-trained two-dimensional network prediction model one by one, so that a two-dimensional target parameter profile about a target parameter at a corresponding position of each two-dimensional profile in an inversion model is obtained; and all obtained two-dimensional target parameter profiles are integrated according to a positional relationship to obtain the inversion model. The method can make the neural network model better learn the low-frequency trend of the target parameter and the geological body characteristics by using multi-source input, and improve the lateral continuity of the inversion result.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration and development technology, and specifically to a pre-stack inversion method, system, storage medium, and processor. Background Technology

[0002] Reservoir evaluation is a crucial step in reservoir research and an important aspect of oil and gas exploration and development. Current technologies typically employ pre-stack seismic inversion to extract rock elastic parameters from pre-stack seismic data, including velocity, density, Poisson's ratio, and Young's modulus, for reservoir lithology identification, fluid identification, and gas-bearing characterization.

[0003] Deep learning, as a data-driven approach, can learn the nonlinear mapping relationship between input features and corresponding labels using massive datasets. Existing deep learning-based pre-stack inversion methods all use a single-channel-to-single-channel mapping method, which may lead to lateral discontinuities in the inversion profile when the amount of well logging data is limited or the signal-to-noise ratio of seismic data is low.

[0004] Therefore, there is an urgent need for a multichannel inversion method based on deep learning and geological laws to guide data augmentation, in order to improve the lateral continuity of pre-stack inversion results. Summary of the Invention

[0005] The purpose of this invention is to provide a method that uses multi-source input to enable neural networks to better learn the low-frequency trends of target parameters and geological features, thereby improving the lateral continuity of inversion results.

[0006] To achieve the above objectives, embodiments of the present invention provide a pre-stack inversion method for obtaining an inversion model of target parameters, comprising:

[0007] The pre-stack seismic data volume, AVO attribute volume, and low-frequency model of the inversion target area are obtained. The low-frequency model is the low-frequency part extracted from the supervised model obtained from multiple well logging data and layer interpolation methods based on the inversion target area. The pre-stack seismic data volume, AVO attribute volume, and low-frequency model are all three-dimensional data volumes.

[0008] According to predetermined rules, each two-dimensional profile of the pre-stack seismic data volume, AVO attribute volume, and low-frequency model is extracted and input into a pre-trained two-dimensional network prediction model to obtain the two-dimensional target parameter profiles with respect to the target parameters at the corresponding locations of each two-dimensional profile in the inversion model; and

[0009] All the obtained two-dimensional target parameter profiles are synthesized into an inversion model according to their positional relationships.

[0010] Optionally, the prestack seismic data volume can be a three-dimensional data volume from three different perspectives, including prestack seismic data of time sampling points on multiple tracks. The multiple tracks include multiple main tracks and multiple connecting tracks, and each track of the multiple main tracks and multiple connecting tracks has multiple time sampling points.

[0011] Furthermore, the AVO attribute volume includes: the G attribute volume, the PG attribute volume, and the P+G attribute volume, and the G attribute volume, the PG attribute volume, and the P+G attribute volume are all calculated from pre-stack seismic data.

[0012] Optionally, three different angles of 3D data volume are available: 7 degrees, 19 degrees and 31 degrees.

[0013] Optionally, the two-dimensional network prediction model is a Unet network. The input of the Unet network first passes through a convolutional layer with a kernel of 21×5, then through three residual modules and a 1×1 convolutional layer, then through three residual modules and an upsampling layer, and finally through a three-channel convolutional layer with a kernel of 21×5 to obtain the predicted two-dimensional parameter profile. The main path of the residual module is two convolutional layers with kernels of 21×5, and the output is directly bypassed by skip connections to alleviate the gradient vanishing problem caused by the increase in the number of network layers.

[0014] Furthermore, during the training of the two-dimensional network prediction model, well logging data from the low-frequency model is used as the first supervision. The low-frequency model includes the first supervision weights for each curve in the three-dimensional pseudo-label, which are established based on the weight decay function.

[0015] Furthermore, the formula for the weight decay function is: Where α is the first supervisory weight, with a value ranging from 0.05 to 1, and x i Let x be the coordinate vector of each curve in the monitoring model, and let x be the coordinate vector of the well log closest to each curve.

[0016] Optionally, the weight decay function is an inverse distance weighting function, where the first supervisory weight is lower for locations farther from the well logging point.

[0017] Optionally, during the training of the two-dimensional network prediction model, synthetic seismic data is also used as a second supervisor. The weight of the first supervisor is lower and the weight of the second supervisor is higher the distance from the well logging point.

[0018] Furthermore, the calculation formula for synthetic seismic data is as follows:

[0019]

[0020]

[0021]

[0022]

[0023] Where E, σ, ρ and ΔE, Δσ, Δρ represent the average value and difference of Young's modulus, Poisson's ratio and density on both sides of the corresponding two-dimensional profile, respectively, θ represents the incident angle, and γ represents the ratio of transverse to longitudinal wave velocities.

[0024] Optionally, at any given location, the sum of the first supervision weight and the second supervision weight is 1.

[0025] Optionally, the seismic data in the pre-stack seismic data volume used to train the two-dimensional network prediction model can be synthetic seismic data or seismic experimental data.

[0026] Optionally, the target parameters are elastic parameters, including: Young's modulus, Poisson's ratio, and density.

[0027] Optionally, the pre-stack inversion method of this application is used to identify favorable reservoir regions from the inversion model according to predetermined identification criteria.

[0028] Furthermore, the predetermined identification criterion is: Young's modulus < 2.5 × 10⁻⁶. 7 N·m -2 It has a Poisson's ratio < 0.29 and a density < 2.55 g / cc.

[0029] Furthermore, the identified favorable reservoir areas are projected onto the HD layer of the channel to obtain the location of high-quality reservoirs within the channel.

[0030] On the other hand, the present invention provides a pre-stack inversion system, comprising: a two-dimensional network prediction model, a data acquisition module, a prediction module, and a synthesis module.

[0031] The data acquisition module acquires the pre-stack seismic data volume, AVO attribute volume, and low-frequency model of the inversion target area. The low-frequency model is the low-frequency part extracted from the supervised model obtained from multiple well logging data and layer interpolation methods based on the inversion target area. The pre-stack seismic data volume, AVO attribute volume, and low-frequency model are all three-dimensional data volumes.

[0032] The prediction module extracts each two-dimensional profile of the pre-stack seismic data volume, AVO attribute volume, and low-frequency model according to predetermined rules, and inputs them one by one into the pre-trained two-dimensional network prediction model to obtain the two-dimensional target parameter profile of the target parameter at the corresponding position of each two-dimensional profile in the inversion model.

[0033] The synthesis module synthesizes all the two-dimensional target parameter profiles obtained from the prediction module into an inversion model according to their positional relationships.

[0034] Optionally, the pre-stack inversion system of this application is used to identify favorable reservoir areas from the inversion model according to predetermined identification criteria, and to project the identified favorable reservoir areas onto the HD layers in the channel, thereby obtaining the location of high-quality reservoirs in the channel.

[0035] On the other hand, the present invention provides a machine-readable storage medium storing instructions that cause a machine to execute the pre-stack inversion method of the present application.

[0036] On the other hand, the present invention provides a processor, characterized in that it is used to run a program, wherein the program is run to execute the pre-stack inversion method of the present application.

[0037] The above technical solution employs various two-dimensional profiles, including pre-stack seismic data volumes, AVO attribute volumes, and low-frequency models, as multi-source inputs. These are fed into a pre-trained two-dimensional network prediction model to obtain two-dimensional target parameter profiles for each profile at the corresponding location. Finally, all two-dimensional target parameter profiles are synthesized into a complete inversion model according to their positional relationships. This allows for better learning of the low-frequency trends of the target parameters and geological body characteristics, improving the lateral continuity of the inversion results. The low-frequency model is extracted from the supervised model obtained using multiple well logging data and stratigraphic interpolation methods for the inversion target area. This enables prediction of target parameters based on well monitoring and multi-channel pre-stack seismic data, improving the accuracy of the prediction results.

[0038] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0039] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings:

[0040] Figure 1 This is a schematic diagram of a pre-stack inversion process based on a network model in the prior art;

[0041] Figure 2 This is a flowchart illustrating an embodiment of the pre-stack inversion method of this application;

[0042] Figure 3 yes Figure 2 A schematic diagram of the data flow in the embodiment;

[0043] Figure 4 yes Figure 2 The supervision model and the corresponding low-frequency model in the embodiments;

[0044] Figure 5 yes Figure 2 The pre-stack seismic data volume and AVO attribute volume in the embodiment;

[0045] Figure 6 yes Figure 2 The architecture diagram of the two-dimensional network prediction model in the embodiment;

[0046] Figure 7 yes Figure 2 Training curves of the two-dimensional network prediction model in the embodiment;

[0047] Figure 8 This is the well supervision weight distribution of the supervision model in another embodiment of the pre-stack inversion method of this application;

[0048] Figure 9 yes Figure 8 Comparison of prediction results for elasticity parameters between the example and the comparative example;

[0049] Figure 10 yes Figure 8 Comparison of prediction results for well logging curves between the example and comparative examples;

[0050] Figure 11 yes Figure 8 Comparison chart of prediction results for elasticity parameters using noiseless data in the example and comparative examples;

[0051] Figure 12 This is a pre-stack seismic data volume and AVO attribute volume obtained in another embodiment of the pre-stack inversion method of this application;

[0052] Figure 13 yes Figure 12 The supervision model and the corresponding low-frequency model in the embodiments;

[0053] Figure 14 yes Figure 13 Well supervision weight distribution in the supervision model;

[0054] Figure 15 yes Figure 12 Example of prediction results for elasticity parameters;

[0055] Figure 16 yes Figure 12 The example compares the predicted L6 curve results for blind wells with the corresponding L6 curves;

[0056] Figure 17 yes Figure 12 The example provides a cross-plot of elastic and physical property parameters from well logging curves; and

[0057] Figure 18 yes Figure 12 The example is used to delineate high-quality reservoirs and compare them with the verification results of well testing. Detailed Implementation

[0058] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the scope of the present invention.

[0059] Existing pre-stack inversion techniques based on deep learning neural network models are implemented using single-trace seismic data; that is, the input data consists of a single seismic trace curve and its attributes. In existing techniques, the neural network form is not fixed and can be CNN, RNN, or Transformer, etc. Figure 1 As shown, taking the U-Net network as an example, the input features are an n×(a+b) matrix, where n is the number of time-depth points, a represents the number of seismic data angles, and b represents the number of seismic attributes. The features are input into the neural network to obtain an end-to-end output, which is an n×c matrix, where n is the number of time-depth points and c represents the number of features to be predicted. Based on this, model-driven or data-driven forward models can be added to alleviate the problem of limited well data. However, the above-mentioned existing technologies are based only on a single seismic trace curve and its attributes, containing limited information, often resulting in lateral discontinuities in the profile of the inversion model.

[0060] This invention provides a pre-stack inversion method that uses a two-dimensional profile of a multi-channel pre-stack seismic data volume as input. Simultaneously, it inputs a two-dimensional profile of an AVO attribute volume calculated from the pre-stack seismic data, and a two-dimensional profile of a low-frequency model extracted from a supervised model based on multiple well logging data and stratigraphic interpolation methods for the inversion target area, into a two-dimensional network prediction model. This enables the two-dimensional network prediction model to learn the low-frequency trends of the target parameters and geological features reflected in the multi-source inputs. The pre-stack inversion method of this invention is used to obtain an inversion model of the target parameters, improving the lateral continuity of the inversion results.

[0061] The following combination Figure 2-3 This embodiment describes the specific implementation method. The process of this embodiment is as follows: Figure 2 As shown, it includes the following steps:

[0062] Step 1: Obtain the pre-stack seismic data volume, AVO attribute volume, and low-frequency model of the inversion target area. The low-frequency model is the low-frequency part extracted from the supervised model obtained from multiple well logging data and layer interpolation methods based on the inversion target area. The pre-stack seismic data volume, AVO attribute volume, and low-frequency model are all three-dimensional data volumes.

[0063] Step 2: According to the predetermined rules, extract each two-dimensional profile of the pre-stack seismic data volume, AVO attribute volume, and low-frequency model, and input them one by one into the pre-trained two-dimensional network prediction model to obtain the two-dimensional target parameter profile of the target parameter at the corresponding position of each two-dimensional profile in the inversion model.

[0064] Step 3: Combine all the obtained two-dimensional target parameter profiles into an inversion model according to their positional relationships.

[0065] AVO (Adaptive Volatility) attribute is a characterization property for the variation of energy with angle in pre-stack data, which can indicate geological bodies such as river channels. Seismic data is a finite-band data volume, lacking low-frequency and high-frequency components. In this embodiment, a three-dimensional data volume model for supervision is obtained through well logging and stratigraphic interpolation, and its low-frequency component is extracted as part of the multi-source input. At the same time, the high-frequency component or the supervision model itself is used as the first supervision. Therefore, using pre-stack seismic data, AVO attribute, and low-frequency model as multi-source inputs allows the neural network to better learn the low-frequency trends of elastic parameters and geological body characteristics.

[0066] In this embodiment, the purpose of step 1 is to obtain the multi-source inputs of the two-dimensional network prediction model. For example... Figure 3 As shown, the U-Net2D network is a two-dimensional network prediction model. Its multi-source inputs include angle-seismic data, AVO attributes calculated based on the angle-seismic data, and E_15hz, Poi_15hz, and Rho_15hz data volumes. The angle-seismic data consists of pre-stack seismic data volumes at three different angles, including pre-stack seismic data from time-sampling points along multiple tracks. The multiple tracks include main tracks and connecting tracks, each with multiple time-sampling points. The AVO attribute volume is shown below. Figure 3 The data includes: G attribute volume, PG attribute volume, and P+G attribute volume. It should be noted that each of the G, PG, and P+G attribute volumes is calculated using angular seismic data. Specifically, the E_15hz, Poi_15hz, and Rho_15hz data volumes are low-frequency models of elastic parameters related to Young's modulus, Poisson's ratio, and density, extracted from a supervised model obtained based on multiple well logging data and stratigraphic interpolation methods for the inverted target area. Therefore, in step 1, the two-dimensional network prediction model has nine inputs: three angular seismic data volumes (e.g., 7°, 19°, and 31° three-dimensional data volumes), three AVO attribute volumes (G, PG, and P+G attribute volumes, respectively), and three low-frequency models (E_15hz, Poi_15hz, and Rho_15hz data volumes, respectively).

[0067] In addition, it should be noted that Figure 3The well + interpolation high-frequency model shown is the high-frequency component extracted from the above-mentioned supervised model and is used as the first supervision of the prediction results. In some implementations, based on the similarity of well paths, curves closer to the well are given higher confidence in the interpolation model, thereby increasing the well supervision weight in the above-mentioned supervised model and better supervising the prediction results.

[0068] To reduce the computational load of concurrent data, a two-dimensional network prediction model is used in step 2. This requires that all the multi-source inputs be two-dimensional data. Therefore, the nine inputs obtained in step 1 are sliced ​​to obtain a set of nine two-dimensional profiles. Each set of nine two-dimensional profiles is then input into the two-dimensional network prediction model U-Net2D. After processing by the U-Net2D network model, a two-dimensional target parameter profile with respect to the target parameters is obtained at the corresponding position of each two-dimensional profile in the inversion model. For example... Figure 3 The target parameters output by the U-Net2D network model shown are elastic parameters, namely Young's modulus, Poisson's ratio, and density. However, this invention is not limited to these parameters, and those skilled in the art can select target parameters according to actual needs.

[0069] It should be noted that the two-dimensional target parameter profile output in step 2 corresponds to the position of the input two-dimensional profile, that is, the position of the output two-dimensional target parameter profile in the inversion model is the same as the position of the input set of 9 two-dimensional profiles.

[0070] like Figure 3 As shown, in some implementations, synthetic seismic data obtained through the YPD approximation formula is also used as seismic forward modeling supervision, that is, synthetic seismic data is used as a second supervision.

[0071] Furthermore, the supervision weight of the second supervision can be set, with the first supervision weight being lower and the second supervision weight being higher for locations farther away from the well logging.

[0072] The formula for calculating synthetic seismic data is as follows:

[0073]

[0074]

[0075]

[0076]

[0077] In the formula, E, σ, ρ and ΔE, Δσ, Δρ represent the average value and difference of Young's modulus, Poisson's ratio and density on both sides of the corresponding two-dimensional profile, respectively, θ represents the incident angle, and γ represents the ratio of transverse to longitudinal wave velocities.

[0078] Furthermore, it can be set that at any given location point, the sum of the first monitoring weight (i.e., well monitoring weight) and the second monitoring weight (i.e., seismic monitoring weight) is 1. That is, at locations with high well monitoring confidence, the well monitoring weight is set higher while the seismic monitoring weight is set lower; conversely, at locations with low well monitoring confidence, the well monitoring weight is set lower while the seismic monitoring weight is set higher.

[0079] To further illustrate the pre-stack inversion method of the present invention and demonstrate its technical advantages, the following is combined with... Figures 4 to 11 The specific implementation process of another embodiment of the present invention is described in detail below. In this embodiment, numerical model data with a river system and tight sandstone data from the actual work area are used. This data is artificially synthesized, and its purpose is to verify the effectiveness of the pre-stack inversion method of the present invention. To save space, this embodiment only describes the differences from the previous embodiment.

[0080] In this embodiment, the following is adopted: Figure 6 The Res-Unet two-dimensional network prediction model shown is similar to the encoding of an autoencoder (see...). Figure 6 (left path) and decoding (see Figure 6 The network model's input features are first encoded through a 21×5 convolutional layer, then encoded through three residual modules and a 1×1 convolutional layer. The features are then decoded through three residual modules and an upsampling layer, and finally, a three-channel convolutional layer (21×5 kernel) is used to obtain a two-dimensional profile of the predicted target parameters. In this embodiment, the target parameters are Young's modulus, Poisson's ratio, and density. Figure 6 In the ResBlock, the residual module is defined. The main path of the residual module consists of two convolutional layers with 21×5 kernels. It skips the convolutional layers and outputs directly through skip connections. This method of learning the residual between the input and output can alleviate the gradient vanishing problem caused by the increase in the number of network layers.

[0081] In this embodiment, the inversion target area is a tight sandstone work area with a river system. Well logging data from the inversion target area are subjected to stratigraphic interpolation to obtain... Figure 4 As shown in (a)-(c), supervised models reflecting Young's modulus, Poisson's ratio, and density characteristics are used to extract data from these models. Figure 4 As shown in (d)-(f), these are low-frequency models reflecting Young's modulus, Poisson's ratio, and density characteristics, respectively. These low-frequency models will be used as input to the aforementioned Res-Unet network model, while the supervisory model will serve as the first supervisor for that model.

[0082] In this embodiment, seismic data volumes with angles of 7 degrees, 19 degrees, and 31 degrees are calculated according to the YPD approximation formula, as follows: Figure 5As shown in (a)-(c), with 10% random noise added, the seismic data volume after adding noise is as follows: Figure 5 As shown in (d)-(f). Based on the noisy seismic data volume, the corresponding AVO attribute volumes can be calculated, namely the G attribute, PG attribute, and P+G attribute, as shown in... Figure 5 As shown in (g)-(i), eight pseudo-wells (Well1 to Well8) are randomly selected from the three-dimensional data volume. The Young's modulus, Poisson's ratio, and density curves corresponding to these pseudo-wells are used for training and testing the two-dimensional network prediction model. Figure 5 As shown in (a)-(i), 7 wells are used as training wells (blue), and 1 well is used as a test well (red, Well6). The selection of the pseudo-label supervision model for well and layer interpolation is... Figure 4 The three-dimensional data volumes shown in (a)-(c) utilize a weight decay formula to assign higher well monitoring weights to pseudo-labels closer to the well. This weight formula can be expressed as:

[0083]

[0084] Where α represents the well monitoring weight, with a value ranging from 0.05 to 1. That is, when the monitoring model curve coincides with the well curve, the monitoring weight is 1; when the monitoring model curve is far from the well curve, the minimum monitoring weight is set to 0.05. i Let represent the coordinate vector of each curve in the supervised model, and let x represent the coordinate vector of the nearest well to that curve. The supervised weights of each curve in the supervised model can be calculated using this weight formula, such as... Figure 8 The figure shows the weight distribution of the calculated three-dimensional volume. It can be seen that the well location has the highest weight. As the distance from the curve to the well increases, the weight decreases rapidly. When the curve is far enough away from the well, the weight decreases to the minimum value of 0.05. The constraint effect of the well on this location is the weakest, and it is constrained only by seismic data.

[0085] In some implementations, an inverse distance weighting function is also used as the weight decay function, where the first supervisory weight is lower for locations farther away from the logging point.

[0086] In this embodiment, the following loss function is used during the training of the two-dimensional network prediction model:

[0087] loss=α×wellloss+(1-α)×seiloss

[0088] wellloss=(E pre -E h ) 2 +(Poi pre -Poi h ) 2 +(Rhopre -Rho h ) 2

[0089] seiloss=(F(E pre Poi pre Rho pre )-Sei) 2 ,

[0090] Where loss, wellloss, and seiloss represent the total network loss, well monitoring loss, and seismic monitoring loss, respectively. pre E h Poi pre Poi h Rho pre With Rho h These represent the predicted results for Young's modulus, Poisson's ratio, and density, along with high-frequency pseudo-labels. α represents the weight of well monitoring, and F represents the forward modeling function that converts the elastic parameters into seismic data. The forward modeling function can be composed of the YPD approximation formula and a convolution model.

[0091] In this embodiment, both the input features and labels of the two-dimensional network prediction model are normalized. The normalization formula can be expressed as:

[0092]

[0093] in, X represents the normalized data, and X represents the data to be normalized. max This represents the maximum value of the actual data.

[0094] In this embodiment, the learning rate of the two-dimensional network prediction model is set to 0.0001, the training iterations are 2000, and the Adam adaptive moment estimator optimizer is used to train the network. As the number of training iterations increases, the loss curve gradually converges, and the network training is complete when the curve stabilizes and no longer decreases. Figure 7 The training curves of the training process of the two-dimensional network prediction model in this embodiment are shown.

[0095] Will Figure 4-5 The three-dimensional data volume is sliced ​​according to predetermined rules to obtain each two-dimensional profile of the corresponding data volume. These profiles are then input into a trained two-dimensional network prediction model to obtain two-dimensional target parameter profiles for Young's modulus, Poisson's ratio, and density. Finally, all the obtained two-dimensional target parameter profiles are synthesized according to their positional relationships to form inversion models reflecting Young's modulus, Poisson's ratio, and density parameters, respectively, as shown below. Figure 9As shown in (a)-(c), the correlation coefficient and relative error between the predicted Young's modulus and the actual results are 91.04% and 7.13%, respectively; the correlation coefficient and relative error between the predicted Poisson's ratio and the actual results are 92.15% and 4.29%, respectively; and the correlation coefficient and relative error between the predicted density and the actual results are 94.39% and 1.15%, respectively. The actual results are artificially synthesized data used to verify the effectiveness of this embodiment. The comparison between the prediction results and the actual results of this embodiment demonstrates the effectiveness of the pre-stack inversion method and the two-dimensional network prediction model of this invention.

[0096] Figure 9 (d)-(f) show the inversion models of Young's modulus, Poisson's ratio, and density parameters obtained by the existing semi-supervised 1DCNN prediction model. The correlation coefficient and relative error between the predicted Young's modulus and the actual results are 71.3% and 12.9%, respectively; the correlation coefficient and relative error between the predicted Poisson's ratio and the actual results are 56.66% and 9.53%, respectively; and the correlation coefficient and relative error between the predicted density and the actual results are 65.94% and 2.82%, respectively. Both the elasticity parameter results predicted by this invention and the results of the semi-supervised 1DCNN can identify river channels. However, the prediction results of this invention are better at distinguishing between river channels and non-river channels, and are smoother and more continuous laterally. In contrast, the results of the semi-supervised 1DCNN are blurry in distinguishing river boundaries, and the lateral variation trend is more jittery. Compared with the results of this embodiment, it can be seen that the pre-stack inversion method and the two-dimensional network prediction model of this invention are more effective than the existing technology.

[0097] In this embodiment, the prediction results for the location of test wells were further compared. For example... Figure 10 As shown, the yellow line represents the actual elastic parameter curve, the blue line represents the elastic parameter curve predicted by this invention, and the green line represents the elastic parameter curve predicted by the semi-supervised 1DCNN. The correlation coefficient and relative error of the Young's modulus prediction results of this invention are 98.88% and 3.13%, respectively; the correlation coefficient and relative error of the Poisson's ratio prediction results are 99.64% and 1.35%, respectively; and the correlation coefficient and relative error of the density prediction results are 99.62% and 0.32%, respectively. The correlation coefficient and relative error of the Young's modulus prediction results of the semi-supervised 1DCNN algorithm are 97.36% and 5.32%, respectively; the correlation coefficient and relative error of the Poisson's ratio prediction results are 96.67% and 5.47%, respectively; and the correlation coefficient and relative error of the density prediction results are 94.9% and 1.24%, respectively. The test results demonstrate that the elastic parameter results inverted by this invention have better lateral continuity, and the prediction accuracy is improved compared to the single-channel intelligent inversion method.

[0098] In this embodiment, the noise resistance of this embodiment is further compared with that of the prior art by testing with noise-free data based on data containing 10% noise. Figure 11 (a)-(c) represent the prediction results of this invention using noiseless seismic data. Figure 11 (d)-(f) show the prediction results of a semi-supervised 1DCNN. (Comparison) Figure 9 and Figure 11 It was found that both methods achieved high accuracy in inversion under noise-free conditions. However, since semi-supervised 1DCNN is a single-channel inversion method, its lateral continuity is slightly lower than that of the method described in this invention. When 10% noise was added to the input features, the inversion results of both methods changed. However, the prediction results of this invention were less affected by noise, and the inversion results still clearly depicted the river channel with good lateral continuity. In contrast, the inversion results of 1DCNN were significantly affected by noise, resulting in weakened differentiation of river channel boundaries and drastic lateral changes. Therefore, the pre-stack inversion method of this embodiment has better noise resistance.

[0099] To verify the inversion effect of this invention on actual work areas, this application also provides an embodiment for a tight sandstone actual work area. In this embodiment, the pre-stack seismic data main line has 906 channels, the connecting lines have 896 channels, each seismic data has 320 time sampling points, the channel spacing is 12.5m, and the time sampling interval is 1ms. The sedimentary environment of the reservoir is fluvial-lacustrine facies deposition, with well-developed channel sand bodies. Vertically, multiple phases of channel sand bodies are superimposed on each other, resulting in strong reservoir heterogeneity and discontinuous seismic phase axes. There are a total of 8 logging wells in the work area (named L1-L8). The seismic data synthesized from the 8 wells after well-seismic calibration has a high degree of matching with the corresponding well-side seismic channels, with an average correlation coefficient of 0.7.

[0100] like Figure 12 As shown in (a)-(c), by partially stacking pre-stack seismic data, stacked data volumes at near, mid, and far angles of 7 degrees, 19 degrees, and 31 degrees are obtained. The work area contains three stratigraphic levels, from top to bottom: c1, HD, and c2. HD is the river channel stratigraphic level as interpreted by experts, which is used for subsequent verification of results.

[0101] The AVO properties of a three-dimensional volume can be calculated using pre-stack seismic data, such as... Figure 12 (d)-(f) show the three-dimensional AVO attribute volumes for the G attribute, PG attribute, and P+G attribute, respectively.

[0102] Seven well logs were selected as training wells, and well L6 was used as a blind well to test the model's effectiveness. Figure 13This indicates that a low-frequency model (0-15 Hz) and a supervised model (0-80 Hz) of the three-dimensional volume can be obtained through interpolation of training well data and layers c1 and c2. The supervised model and training well data are used to calculate the well supervision weights for each curve in the supervised model based on a weight decay formula, such as... Figure 14 As shown. The input features of the two-dimensional network prediction model are multi-angle superimposed data, AVO attributes, and low-frequency models, and the number of training iterations is set to 2000. The trained network is then generalized to the three-dimensional work area, yielding the prediction results for Young's modulus, Poisson's ratio, and density, as shown. Figure 15 As shown, the prediction results in this embodiment differ significantly from the supervised model. Although the network training is supervised by pseudo-labels obtained through interpolation, the network only learns the elastic parameter characteristics of well logging and well-side data due to the limitation of well supervision weights. Curves far from well logging locations are minimally constrained by well data, and network parameters are primarily updated through matching seismic data.

[0103] Use such as Figure 16 The blind well L6 shown was tested against a two-dimensional network prediction model. The prediction results showed good agreement with the actual results, with relative errors of 7.04%, 7.62%, and 0.88%, respectively. For example... Figure 17 Cross-plot analysis of the logging curves shown reveals a correlation between the network's inversion parameters and physical property parameters. Since Young's modulus is negatively correlated with porosity, Poisson's ratio is positively correlated with clay content, and density is negatively correlated with gas saturation, the inverted Young's modulus, Poisson's ratio, and density can be used to delineate... Figure 18 The favorable reservoir region shown is one with a Young's modulus less than 2.5 × 10⁻⁶. 7 N·m -2 The Poisson's ratio is less than 0.29, and the density is less than 2.55 g / cc. The identified favorable reservoirs are projected onto the HD layer of the channel to obtain the locations of high-quality reservoirs within the channel. This is verified using well test production data from wells L2, L3, and L6. The identified high-quality reservoir locations and well test production show a good correlation, demonstrating the effectiveness of this invention in applying real-world data.

[0104] This invention provides a pre-stack inversion system, comprising: a two-dimensional network prediction model, a data acquisition module, a prediction module, and a synthesis module.

[0105] The data acquisition module acquires the pre-stack seismic data volume, AVO attribute volume, and low-frequency model of the inversion target area. The low-frequency model is the low-frequency part extracted from the supervised model obtained from multiple well logging data and layer interpolation methods based on the inversion target area. The pre-stack seismic data volume, AVO attribute volume, and low-frequency model are all three-dimensional data volumes.

[0106] The prediction module extracts each two-dimensional profile of the pre-stack seismic data volume, AVO attribute volume, and low-frequency model according to predetermined rules, and inputs them one by one into the pre-trained two-dimensional network prediction model to obtain the two-dimensional target parameter profile of the target parameter at the corresponding position of each two-dimensional profile in the inversion model.

[0107] The synthesis module synthesizes all the two-dimensional target parameter profiles obtained from the prediction module into an inversion model according to their positional relationships.

[0108] In some embodiments, the pre-stack inversion system of this application is used to identify favorable reservoir regions from the inversion model according to predetermined identification criteria, and to project the identified favorable reservoir regions onto the HD layers of the channel, thereby obtaining the location of high-quality reservoirs in the channel.

[0109] Preferred embodiments of the pre-stack inversion system of the present invention can be found in the above-described embodiments of the pre-stack inversion method, and will not be repeated here.

[0110] This invention provides a pre-stack inversion device, including a processor and a memory. The two-dimensional network prediction model, data acquisition module, prediction module, and synthesis module are all stored as program units in the memory, and the processor executes the program units stored in the memory to realize the corresponding functions.

[0111] The processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and adjusting kernel parameters can improve the lateral continuity of the inversion results.

[0112] The memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0113] This invention provides a storage medium storing a program that, when executed by a processor, implements the pre-stack inversion method of this application.

[0114] This invention provides a processor for running a program, wherein the program executes the pre-stack inversion method of this application.

[0115] This invention provides a device including a processor, a memory, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the pre-stack inversion method steps of this application. The device described herein can be a server, PC, PAD, mobile phone, etc.

[0116] This application also provides a computer program product, which, when executed on a data processing device, is adapted to perform the program that initializes the pre-stack inversion method steps of this application.

[0117] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0118] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0120] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0121] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0122] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0123] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0124] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0125] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A pre-stack inversion method for obtaining an inversion model of target parameters, comprising: The pre-stack seismic data volume, AVO attribute volume, and low-frequency model of the inversion target area are obtained. The low-frequency model is the low-frequency part extracted from the supervised model obtained from multiple well logging data and layer interpolation methods based on the inversion target area. The pre-stack seismic data volume, the AVO attribute volume, and the low-frequency model are all three-dimensional data volumes. According to predetermined rules, each two-dimensional profile of the pre-stack seismic data volume, the AVO attribute volume, and the low-frequency model is extracted and input into a pre-trained two-dimensional network prediction model to obtain a two-dimensional target parameter profile with respect to the target parameter at the corresponding location of each two-dimensional profile in the inversion model; and All the obtained two-dimensional target parameter profiles are synthesized into the inversion model according to their positional relationships. During the training process of the two-dimensional network prediction model, well logging data from the supervised model is used as the first supervision. The supervised model includes establishing the first supervision weights for each curve in the three-dimensional pseudo-label based on the weight decay function. The formula for the weight decay function is: Where α is the first supervisory weight, and its value ranges from 0.05 to 1. Let be the coordinate vector of each curve in the supervised model. The coordinate vector of the well log closest to each of the aforementioned curves. During the training of the two-dimensional network prediction model, synthetic seismic data is also used as a second supervisor. The weight of the first supervisor decreases and the weight of the second supervisor increases the further away the location is from the well logging data. The calculation formula for the synthetic seismic data is as follows: ， Among them, E, , and , , θ represents the average value and difference of Young's modulus, Poisson's ratio and density on both sides of the corresponding two-dimensional profile, respectively, θ represents the incident angle, and γ represents the ratio of transverse to longitudinal wave velocities.

2. The pre-stack inversion method according to claim 1, characterized in that, The pre-stack seismic data volume consists of three-dimensional data volumes from three different perspectives, including pre-stack seismic data from time sampling points on multiple survey lines. The multiple survey lines include multiple main survey lines and multiple connecting survey lines, and each of the multiple main survey lines and the multiple connecting survey lines has multiple time sampling points.

3. The pre-stack inversion method according to claim 2, characterized in that, The AVO attribute body includes: the G attribute body, the PG attribute body, and the P+G attribute body, and the G attribute body, the PG attribute body, and the P+G attribute body are all calculated from the pre-stack seismic data.

4. The pre-stack inversion method according to claim 2, characterized in that, The three-dimensional data volumes at the three different angles are 7 degrees, 19 degrees and 31 degrees respectively.

5. The pre-stack inversion method according to claim 1, characterized in that, The two-dimensional network prediction model is a Unet network. The input of the Unet network first passes through a convolutional layer with a kernel of 21×5, then through three residual modules and a 1×1 convolutional layer, then through three residual modules and an upsampling layer, and finally through a three-channel convolutional layer with a kernel of 21×5 to obtain the predicted two-dimensional target parameter profile. The main path of the residual module is two convolutional layers with kernels of 21×5, and the output is directly bypassed by skip connections to alleviate the gradient vanishing problem caused by the increase in the number of network layers.

6. The pre-stack inversion method according to claim 1, characterized in that, The weight decay function is an inverse distance weighting function, where the first supervisory weight is lower for locations farther from the well logging point.

7. The pre-stack inversion method according to claim 1, characterized in that, At any given location, the sum of the first supervision weight and the second supervision weight is 1.

8. The pre-stack inversion method according to claim 1, characterized in that, The target parameters are elastic parameters, including Young's modulus, Poisson's ratio, and density.

9. The pre-stack inversion method according to claim 8, used to identify favorable reservoir regions from the inversion model according to predetermined identification criteria.

10. The pre-stack inversion method according to claim 9, characterized in that, The predetermined identification criterion is: Young's modulus < 2.5 × 10⁻⁶. 7 N•m -2 It has a Poisson's ratio < 0.29 and a density < 2.55 g / cc.

11. The pre-stack inversion method according to claim 9, characterized in that, The identified favorable reservoir regions are projected onto the HD layer of the channel to obtain the location of high-quality reservoirs within the channel.

12. A pre-stack inversion system for obtaining an inversion model of target parameters, comprising: Two-dimensional network prediction model; The data acquisition module acquires the pre-stack seismic data volume, AVO attribute volume, and low-frequency model of the inversion target area. The low-frequency model is the low-frequency part extracted from the supervised model obtained from multiple well logging data and layer interpolation methods based on the inversion target area. The pre-stack seismic data volume, the AVO attribute volume, and the low-frequency model are all three-dimensional data volumes. The prediction module, according to predetermined rules, extracts each two-dimensional profile of the pre-stack seismic data volume, the AVO attribute volume, and the low-frequency model, and inputs them one by one into a pre-trained two-dimensional network prediction model to obtain a two-dimensional target parameter profile with respect to the target parameter at the corresponding position of each two-dimensional profile in the inversion model; and The synthesis module synthesizes all the two-dimensional target parameter profiles obtained by the prediction module into the inversion model according to their positional relationships. During the training process of the two-dimensional network prediction model, well logging data from the supervised model is used as the first supervision. The supervised model includes establishing the first supervision weights for each curve in the three-dimensional pseudo-label based on the weight decay function. The formula for the weight decay function is: Where α is the first supervisory weight, and its value ranges from 0.05 to 1. Let be the coordinate vector of each curve in the supervised model. The coordinate vector of the well log closest to each of the aforementioned curves. During the training of the two-dimensional network prediction model, synthetic seismic data is also used as a second supervisor. The weight of the first supervisor decreases and the weight of the second supervisor increases the further away the location is from the well logging data. The calculation formula for the synthetic seismic data is as follows: ， Among them, E, , and , , θ represents the average value and difference of Young's modulus, Poisson's ratio and density on both sides of the corresponding two-dimensional profile, respectively, θ represents the incident angle, and γ represents the ratio of transverse to longitudinal wave velocities.

13. The pre-stack inversion system according to claim 12, used to identify favorable reservoir regions from the inversion model according to predetermined identification criteria, and project the identified favorable reservoir regions onto the HD layer of the channel to obtain the location of high-quality reservoirs in the channel.

14. A machine-readable storage medium storing instructions for causing a machine to perform the pre-stack inversion method as described in any one of claims 1-11.

15. A processor, characterized in that, Used to run a program, wherein the program is run to execute: the pre-stack inversion method as described in any one of claims 1-11.

Images