Artificial Intelligence Four-Dimensional Seismic Inversion Method Based on CNN Method

Abstract

The present invention discloses an artificial intelligence four-dimensional seismic inversion method based on the CNN method, comprising the following steps: establishing a basic reservoir velocity model and a monitored reservoir velocity model; performing forward modeling on the basic reservoir velocity model and the monitored reservoir velocity model; obtaining the true velocity change rate of the basic reservoir velocity model and the monitored reservoir velocity model; obtaining the synthetic seismic record difference between the synthetic seismic record of the monitored model and the synthetic seismic record of the basic model; thinning the synthetic seismic record difference and the true velocity change rate to construct a training set, obtaining the thinned synthetic seismic record and the thinned true velocity change rate; using the thinned synthetic seismic record as input and the thinned true velocity change rate as a label, and feeding them into a network for training to obtain a predicted velocity change rate. The present invention can accurately find remaining oil and gas reservoirs, improve the recovery rate, implement reservoir dynamic monitoring, and greatly improve work efficiency and inversion stability.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas exploration, and particularly to an artificial intelligence four-dimensional seismic inversion method based on the CNN method. Background Art

[0002] Four-dimensional seismic is a reservoir monitoring technology at the field scale. It is a three-dimensional seismic data volume collected repeatedly at different times under the same block and the same acquisition conditions. By comparing and analyzing the seismic data collected at different times, the part with differences is the part where the reservoir fluid changes. This technology can not only monitor the change of the underground fluid boundary position and the movement of the injected fluid, but also detect the dead oil area, guide the reasonable layout of oilfield development wells, predict the distribution of remaining oil, etc., and finally improve the recovery rate. By analyzing the differences in repeatedly collected seismic data to infer the change of reservoir physical properties and realizing the dynamic monitoring of the reservoir, monitoring the reservoir is an important means for old oilfields to increase recoverable reserves and improve the recovery rate.

[0003] The four-dimensional seismic inversion result based on the correlation method is extremely unstable and highly dependent on the size of the processing time window. Different inversion results will be presented as different time window lengths are selected by different processors. That is, the boundary of the change range of the longitudinal wave velocity of the seismic wave when the time shift is not 0 will move up and down with the time window length, so that the change range of the longitudinal wave velocity of the seismic wave in the formation fluid cannot be accurately judged. The four-dimensional seismic inversion result based on the correlation method can only roughly invert the change range of the reservoir velocity and cannot help the interpreter accurately determine the change rate of the longitudinal wave velocity of the seismic wave in the reservoir fluid. At the same time, it is also unable to well invert the very small velocity change values in the transverse and longitudinal directions of the longitudinal wave of the seismic wave in the formation fluid, that is, the resolution is not high. Summary of the Invention

[0004] The purpose of the present invention is to provide an artificial intelligence four-dimensional seismic inversion method based on the CNN method to solve the problems of low resolution and low accuracy of the existing prediction methods.

[0005] To achieve the above invention purpose, the technical solution adopted by the present invention is: an artificial intelligence four-dimensional seismic inversion method based on the CNN method, including the following steps:

[0006] Establish a basic reservoir velocity model and a monitored reservoir velocity model for four-dimensional seismic measurement;

[0007] Perform forward modeling on the basic reservoir velocity model and the monitored reservoir velocity model to obtain a synthetic seismic record of the monitored model and a synthetic seismic record of the basic model;

[0008] Obtain the true velocity change rate of the basic reservoir velocity model and the monitored reservoir velocity model;

[0009] Find the synthetic seismogram difference between the synthetic seismogram of the monitoring model and that of the basic model;

[0010] Thin out the synthetic seismogram difference and the true velocity change rate to construct a training set, obtaining the thinned synthetic seismogram difference and the thinned true velocity change rate;

[0011] Use the thinned synthetic seismogram difference as the input and the thinned true velocity change rate as the label, and feed them into the network for training to obtain the predicted velocity change rate.

[0012] Preferably, the basic reservoir velocity model and the monitoring reservoir velocity model collect seismic data of the same area at different times before and after in the same acquisition manner.

[0013] Preferably, the objective function of the true velocity change rate is:

[0014] ;

[0015] Among them, represents the depth-domain velocity model of the monitoring model, represents the depth-domain velocity model of the basic model, represents the velocity change rate.

[0016] Preferably, the objective function of the synthetic seismogram difference is:

[0017] ;

[0018] Among them, represents the synthetic seismogram of the monitoring model, represents the synthetic seismogram of the basic model, represents the synthetic seismogram difference.

[0019] Preferably, it also includes the error analysis of the true velocity change rate and the predicted velocity change rate, and the objective function of the error analysis is:

[0020] ;

[0021] Among them, represents the predicted velocity change rate, represents the true velocity change rate, represents the error value.

[0022] The beneficial effects of the present invention are mainly reflected in:

[0023] 1. The method of the present invention avoids solving various complex geophysical algorithms and the multi-solution problem of traditional four-dimensional seismic inversion, and is applicable to three-dimensional seismic data volumes collected repeatedly at different times under the same acquisition conditions in the same block.

[0024] 2. The present invention combines artificial intelligence with geophysics, directly establishing a non-linear relationship between the change rate of the propagation velocity of the longitudinal seismic wave in the reservoir fluid and the synthetic seismic record through the learning of the training set by a convolutional neural network, and performing four-dimensional seismic inversion through the non-linear relationship between the two to predict the change rate model of velocity.

[0025] 3. Compared with the traditional four-dimensional seismic inversion, the present invention can accurately find the remaining oil and gas reservoirs, improve the recovery rate, and implement reservoir dynamic monitoring, which can greatly improve the working efficiency and inversion stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is the flowchart of the design scheme of the present invention;

[0027] Figure 2 is the depth-domain velocity model of the established basic model;

[0028] Figure 3 is the depth-domain velocity model of the established monitoring model;

[0029] Figure 4 is the synthetic seismic record of the basic model;

[0030] Figure 5 is the synthetic seismic record of the monitoring model;

[0031] Figure 6 is the true velocity change rate model;

[0032] Figure 7 is the difference of the synthetic seismic records;

[0033] Figure 8 is the graph after thinning the difference of the synthetic seismic records;

[0034] Figure 9 is the graph after thinning the true velocity change rate;

[0035] Figure 10 is the error analysis graph of the prediction values of the convolutional neural network trained with training sets of different sizes. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0036] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0037] As Figures 1-10 shown, the artificial intelligence four-dimensional seismic inversion method based on the CNN method includes the following steps:

[0038] a. Establish a basic reservoir velocity model and a monitoring reservoir velocity model for 4D seismic survey; the basic reservoir velocity model and the monitoring reservoir velocity model are obtained by collecting seismic data of the same area at different times before and after with the same acquisition method. For example, the reservoir velocity model before production is the basic reservoir velocity model, and the reservoir velocity model after production is the monitoring reservoir velocity model, as Figures 2-3 shown.

[0039] b. Perform time-depth conversion on the basic reservoir velocity model in the depth domain and the monitoring reservoir velocity model in the depth domain respectively, and then use convolutional Ricker wavelet forward modeling to obtain the synthetic seismic record of the monitoring model and the synthetic seismic record of the basic model, as Figures 4-5 shown.

[0040] Among them, seismic forward modeling specifically is: using the velocity values of seismic waves collected by nodal instruments during underground propagation to establish an underground geological model, and according to the propagation principle of seismic waves in underground media, through mathematical methods, forward simulate and calculate the seismic record of the established geological model; the basis of seismic model forward modeling is that different rock layers have different velocities and densities, and the product of velocity and density is wave impedance; the difference in wave impedance will generate a reflection coefficient; assuming that the seismic wave is vertically incident, the reflection coefficient for normal incidence can be calculated:

[0041] ;

[0042] where R represents the reflection coefficient; , the wave impedance of adjacent two layers;

[0043] The seismic wavelet is also an important parameter in seismic model forward modeling; the actual seismic record from excitation, propagation to reception is equivalent to experiencing a filtering system. After a sharp pulse passes through the earth's filtering system, it becomes a pulse waveform with a certain duration, that is, the seismic wavelet , and the process of forward modeling the synthetic seismic record is the process of convolving the seismic wavelet with the reflection coefficient to obtain the amplitude of the reflected wave:

[0044] ;

[0045] where represents the seismic wavelet; represents the reflection coefficient; represents the synthetic seismic record;

[0046] In seismic model forward modeling, generally the convolutional Ricker wavelet is selected for forward modeling, and the results obtained are in good agreement with the actual seismic record.

[0047] c. Obtain the true velocity change rate of the base reservoir velocity model and the monitored reservoir velocity model; the objective function of the true velocity change rate is:

[0048] ;

[0049] where, represents the depth-domain velocity model of the monitoring model, represents the depth-domain velocity model of the base model, represents the velocity change rate.

[0050] d. Obtain the synthetic seismic record difference between the synthetic seismic record of the monitoring model and the synthetic seismic record of the base model; the objective function of the synthetic seismic record difference is:

[0051] ;

[0052] where, represents the synthetic seismic record of the monitoring model, represents the synthetic seismic record of the base model, represents the synthetic seismic record difference.

[0053] e. Decimate the synthetic seismic record difference and the true velocity change rate, and use the same decimation method to decimate the true velocity change rate to construct a training set, obtaining the decimated synthetic seismic record difference and the decimated true velocity change rate; for example: extract one seismic trace every 15 seismic traces in the horizontal direction, and a total of 200 seismic traces are extracted to obtain Figure 8 , Figure 9 .

[0054] f. Establish a one-to-one correspondence between the decimated synthetic seismic record difference and the true velocity change rate, use the decimated synthetic seismic record difference as the input and the decimated true velocity change rate as the label, and feed them into a convolutional neural network for training to obtain the predicted velocity change rate;

[0055] The model of the velocity change rate is specifically: calculate the magnitude of the velocity change rate of the fluid propagating in the reservoir before and after reservoir production. Let represent the velocity value at each position in the vertical direction within the depth range at a certain position on the surface before reservoir production (i.e., the base model), represent the velocity value at each position in the vertical direction within the depth range at the same position on the surface after reservoir production (i.e., the monitoring model). The formula is expressed as follows:

[0056] ;

[0057] ;

[0058] where, The depth-domain velocity model representing the monitoring model; The depth-domain velocity model representing the basic model; Represents the velocity magnitude; Represents the velocity change rate.

[0059] Furthermore, it also includes the error analysis of the true velocity change rate and the predicted velocity change rate. The objective function of the error analysis is:

[0060] ;

[0061] Among them, Represents the predicted velocity change rate, Represents the true velocity change rate, Represents the error value;

[0062] Specifically, the difference in synthetic seismic records obtained by thinning to 300 traces, 200 traces, 150 traces, 100 traces, 50 traces, and 25 traces and the true velocity change rate are fed into a convolutional neural network for training to obtain 6 different trained convolutional neural network models. These 6 models are used for prediction, and the error between the predicted velocity change rate and the true velocity change rate is calculated to obtain the error value, as Figure 9 shown.

[0063] The method of the present invention avoids solving various complex geophysical algorithms and the non-uniqueness of traditional four-dimensional seismic inversion, and is applicable to 3D seismic data volumes collected repeatedly at different times under the same block and the same acquisition conditions.

[0064] Moreover, the present invention combines artificial intelligence with geophysics, directly establishes the non-linear relationship between the propagation velocity change rate of seismic P-waves in reservoir fluids and synthetic seismic records through the learning of the training set by a convolutional neural network, and performs four-dimensional seismic inversion through the non-linear relationship between the two to predict the velocity change rate model.

[0065] It can be seen from Figure 10 that as the number of samples in the training set increases, the four-dimensional seismic inversion effect based on the convolutional neural network is more stable and has higher resolution. It can more accurately predict the magnitude of the propagation velocity change rate of fluids at each position in the reservoir before and after reservoir exploitation, and can more precisely predict the range boundary of the relative change in the position of fluids in the reservoir after reservoir exploitation.

[0066] Secondly, compared with traditional four-dimensional seismic inversion, the present invention can accurately find remaining oil and gas reservoirs, improve the recovery rate, implement reservoir dynamic monitoring, and greatly improve work efficiency and inversion stability.

[0067] It should be noted that, for the foregoing method embodiments, for the sake of simple description, they are all expressed as a series of action combinations. However, those skilled in the art should know that this application is not limited by the described action sequence, because according to this application, some steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and units involved are not necessarily essential to this application.

Claims

1. An artificial intelligence four-dimensional seismic inversion method based on the CNN method, characterized in that: Including the following steps: Establish a basic reservoir velocity model and a monitoring reservoir velocity model for four-dimensional seismic survey; the basic reservoir velocity model and the monitoring reservoir velocity model are obtained by collecting seismic data of the same area at different times before and after with the same acquisition method. The reservoir velocity model before exploitation is the basic reservoir velocity model, and the reservoir velocity model after exploitation is the monitoring reservoir velocity model; Perform forward modeling on the basic reservoir velocity model and the monitoring reservoir velocity model to obtain a synthetic seismic record of the monitoring model and a synthetic seismic record of the basic model. Among them, performing forward modeling on the basic reservoir velocity model and the monitoring reservoir velocity model to obtain a synthetic seismic record of the monitoring model and a synthetic seismic record of the basic model includes: performing time-depth conversion on the basic reservoir velocity model in the depth domain and the monitoring reservoir velocity model in the depth domain respectively, and then using convolutional Ricker wavelets for forward modeling to obtain a synthetic seismic record of the monitoring model and a synthetic seismic record of the basic model. Among them, the process of forward modeling the synthetic seismic record is the process of convolving the seismic wavelet with the reflection coefficient to obtain the amplitude of the reflected wave; Calculate the true velocity change rate of the basic reservoir velocity model and the monitoring reservoir velocity model; Calculate the difference in synthetic seismic records between the synthetic seismic record of the monitoring model and the synthetic seismic record of the basic model; Thin out the difference in synthetic seismic records and the true velocity change rate to construct a training set, obtaining the thinned difference in synthetic seismic records and the thinned true velocity change rate; Use the thinned difference in synthetic seismic records as the input and the thinned true velocity change rate as the label, and send them into the network for training to obtain the predicted velocity change rate.

2. The artificial intelligence four-dimensional seismic inversion method based on the CNN method according to claim 1, wherein: The basic reservoir velocity model and the monitoring reservoir velocity model are obtained by collecting seismic data of the same area at different times before and after with the same acquisition method.

3. The artificial intelligence four-dimensional seismic inversion method based on the CNN method according to claim 1, characterized in that: The objective function of the true velocity change rate is: ； Among them, represents the depth-domain velocity model of the monitoring model, represents the depth-domain velocity model of the basic model, represents the velocity change rate.

4. The artificial intelligence four-dimensional seismic inversion method based on the CNN method according to claim 1, characterized in that: The objective function of the difference in synthetic seismic records is: ； Among them, represents the synthetic seismogram of the monitoring model, represents the synthetic seismogram of the basic model, represents the difference in synthetic seismograms.

5. The artificial intelligence four-dimensional seismic inversion method based on the CNN method according to claim 1, wherein: It also includes error analysis of the true velocity change rate and the predicted velocity change rate, and the objective function of the error analysis is: ； Among them, represents the predicted speed change rate, represents the actual speed change rate, represents the error value.

Images