Figure 6- 105 Spatial Distribution Characteristics of 3D Visualized Sculpture Sand Bodies
The development of seismic attribute technology has experienced direct hydrocarbon detection, bright spots, flat spots and dark spots in the 1960s, instantaneous attributes (instantaneous frequency, amplitude and phase) in the 1970s, multi-attribute analysis and multi-dimensional attributes (such as dip, azimuth and coherence) in the 1980s, and three-dimensional continuous attributes in the 1990s. The azimuth function AVO(AVAZ) and 3D 3DVCM, namely three-dimensional voxel coherence, are developed. Practice has proved that VCM seismic attribute analysis technology provides an important new method for seismic interpretation and oil and gas exploration. After several ups and downs, we seek experience in success, learn lessons from failure and move towards a more mature stage of development. With the introduction of new knowledge such as mathematics and information science, seismic attributes extracted from seismic data are more and more abundant. There are more than 60 seismic attributes related to reflection time, amplitude, frequency, phase and absorption attenuation, and new attributes are still emerging. In addition to extracting attributes from frequency spectrum, autocorrelation function, complex trace analysis and linear prediction, fractal and wavelet transform have also been used to extract attributes from data time windows in recent years. The emergence of a large number of new seismic attributes has caused the popularity of multi-attribute joint inversion analysis techniques, such as cluster analysis, neural network and covariance. The rapid development and popularization of 3D seismic exploration provides favorable conditions for the development of seismic attribute technology. The wide application of seismic stratigraphy and sequence stratigraphy enriches the connotation of seismic attribute technology, which can be used to identify seismic facies. At present, the wide demand for reservoir prediction, reservoir description and reservoir monitoring shows a new development prospect for seismic attribute technology.
Three-dimensional earthquakes have increased the amount of seismic data by several orders of magnitude. Contemporary powerful computer capabilities, man-machine interactive interpretation workstations and advanced 3D visualization interpretation technology provide conditions for extracting and analyzing 3D seismic attributes and describing reservoirs. People's evaluation of 3D seismic attribute technology is that 3D seismic can generate 3D seismic attribute bodies with different spatial models, and then transform them into seismic facies bodies close to geological models, thus explaining the difficult-to-identify structural, stratigraphic and lithologic problems, and pushing 3D seismic interpretation to the comprehensive, quantitative and practical direction of structural, reservoir, lithologic, physical properties, oil and gas bearing, seismic, drilling and logging data. Although seismic attribute technology is an important means to predict reservoirs and describe the characteristics of oil and gas reservoirs, its potential risks must be fully recognized in use. Therefore, many methods (such as geostatistics, neural network, multiple regression, model verification, etc.) should be used to predict reservoir characteristics by using three-dimensional seismic attribute volume. Correlate logging with seismic data, and then distribute more accurate reservoir characteristics around the well to the whole survey area.
There are two ways to classify seismic attributes: one is "geometric attributes and physical attributes" proposed by Taner and others in the mid-1990s. Geometric attribute refers to the geometric shape of seismic horizon (such as dip angle, azimuth angle and curvature, etc.). ); Physical properties refer to the kinematic and dynamic properties of seismic waves, mainly including amplitude, frequency, phase, waveform, velocity and attenuation. The other is the "pre-stack and post-stack attribute" proposed by Brown 1996. Prestack attributes refer to seismic information extracted from prestack seismic data. With the development of data processing technology and the increase of the proportion of interpretation processing, more and more attention has been paid to the study of extracting attributes from prestack data, such as AVO, normal moveout, P-wave layer velocity, envelope and its derivatives, instantaneous information (amplitude, frequency, phase acceleration, Q factor, bandwidth), main frequency, normalized amplitude, wave impedance and so on. These attributes are related to the geological structure, stratigraphic structure, lithology, physical properties, hydrocarbon content and absorption attenuation of the fault zone strata. For lateral reservoir prediction and reservoir description, the most important and basic attributes are amplitude, velocity and wave impedance. On the basis of the above classification, 1997 SEG annual meeting put forward statistical attribute, which is a comprehensive new attribute derived from various seismic attributes under the condition that statistical methods are related to the same geological attribute.
In short, the so-called seismic attributes are some seismic information or characteristic parameters measured or calculated according to seismic data, and its calculation and analysis methods include cluster analysis, multivariate statistical analysis, neural network inversion, pattern recognition and well data constrained inversion and other modern new technologies. Therefore, the technology of seismic attribute extraction and analysis has made great progress.
Since the 1980s, China Offshore Oil Corporation has cooperated with foreign countries, and on the basis of introducing, digesting and absorbing foreign inversion processing technologies such as three instants and wave impedance, combined with the needs of reservoir and oil and gas reservoir description, it has successively carried out multi-seismic information oil and gas detection. Fuzzy mathematical clustering, judgment and quantitative analysis of oil and gas detection information; Interactive seismic velocity simulation, interval velocity scanning time-distance curve fitting inversion; Calculation of thin reservoir thickness (inversion of thin layer thickness by main amplitude and main frequency); Automatic interpretation of microseismic facies (fuzzy dynamic clustering combined with pattern recognition to explain microseismic facies); Reservoir thickness constraint inversion; Inverse iteration seismic velocity simulation of wave equation: hydrocarbon detection, such as pattern recognition of various seismic characteristic parameters, such as autocorrelation, autoregression and power spectrum of seismic signals; According to Biot's two-phase medium theory, the seismic stress and strain are calculated, and the functional relationship between them and formation physical parameters is established. Although the above seismic attribute inversion technology has made some achievements in offshore oil and gas exploration and development, lateral reservoir prediction and oil and gas reservoir description, it is still difficult to solve because of the multi-solution and constraint verification of the inversion results. Therefore, it is not surprising that the predicted results of lithology, physical properties and hydrocarbon content are inconsistent with the actual drilling.
With the development of offshore oil and gas exploration and development, seismic reservoir prediction and oil and gas description are required to be more accurate and reliable, so as to improve the success rate of drilling, reduce risks and effectively improve the social and economic benefits of exploration and development. Therefore, since 2000, on the basis of introduction, digestion, absorption and innovation, it has been practiced to directly invert reservoir physical properties (porosity, permeability and saturation) and oil-gas bearing property according to seismic multi-attributes under the constraint of drilling data, and the inversion results are verified and iteratively corrected by interactive verification of drilling data, so as to ensure the reliability of the advanced seismic multi-attribute inversion reservoir parameters technology.
I. Technical principle
This technology is a series of supporting technologies consisting of multi-channel geostatistics technology, neural network regression analysis and optimization technology, neural network modeling and pattern recognition inversion technology and well data interactive verification inversion technology. Among them, neural network technology is the core, which includes the following three kinds of neural networks.
(A) the general neural network
It is assumed that the distribution law of the intersection of seismic attributes and their corresponding well data geological attributes is linear. The mathematical model of linear intersection regression simulation is as follows.
Single attribute mathematical model of 1. earthquake
There is a linear correlation between the attribute of geological target and its corresponding single seismic attribute beside the well, such as amplitude. The mathematical model of linear regression fitting is:
High-tech and Practice of Offshore Oil in China
Where: y is the geological property of logging curve; X is the seismic attribute; A and b are unknown coefficients.
Use the minimum variance formula:
High-tech and Practice of Offshore Oil in China
Let the minimum square difference E=O, and the values of a and b coefficients can be solved by equation (6-46)4.
Substituting Equation (6-45) can fit the geostatistical optimal linear correlation between seismic attributes near the well and the scattered distribution of logging attributes. As can be seen from Figure 6- 106, the fluctuation of scattered points relative to the regression fitting line is uneven and large, which shows that the geostatistical regression fitting result of this method is not ideal.
2. Linear regression mathematical model of earthquake multi-attribute intersection.
Figure 6- 106 Crossplot of Logging Curve Target and Seismic Attribute
Figure 6- 107 Combination of simultaneous sampling values of seismic multi-attributes and logging curves
As can be seen from Figure 6- 107, assuming that there are three seismic attribute functions S 1(t), S2(t), S3(t), a geological target attribute G(t) and four unknown weighting coefficients W0, W 1, W2 and W3, the following linear equations can be established.
High-tech and Practice of Offshore Oil in China
The minimum variance formula is:
High-tech and Practice of Offshore Oil in China
Let E=O, and the following four unknown coefficient values can be obtained.
High-tech and Practice of Offshore Oil in China
3. Regression calculation of linear equation of seismic multi-attribute convolution operator.
As can be seen from Figure 6- 108, the frequency of logging records (geological attributes) is much higher than that of seismic records, so the correlation processing results between them are not ideal. Therefore, a well uses multiple seismic attribute sampling points to record geological attribute points, as shown in Figure 6- 109.
Fig. 6- 108 frequency comparison between logging curve and seismic attribute curve
Figure 6- 109 Operator Convolution Combination of Seismic Multi-attribute Multi-points and Single Sample Value of Logging Curve
As can be seen from the figure, the operator composed of five sampling points of each seismic attribute is just close to the seismic wavelet, so the linear regression equation of its convolution operator is:
High-tech and Practice of Offshore Oil in China
The minimum mean square error is e, and let E=0, which can be obtained by the following formula:
High-tech and Practice of Offshore Oil in China
Find out the values of unknown weighting coefficients W0, W 1, W2 and W3, and substitute them into equation (6-49) to complete the regression simulation of the intersection line between G and Si.
In a word, from the above mathematical model, it can be seen that seismic multi-attribute inversion of reservoir parameters is a multi-channel geological calculation technology in mathematics, which includes all methods to predict variables related to reservoir lithology, physical properties and hydrocarbon content by using multiple variables.
Multilayer neural network
Figure 6- 1 10 Composition of Multilayer Neural Network
As can be seen from Figure 6- 1 10, the multilayer neural network is composed of an input layer, an output layer and one or more intermediate result layers, and each layer is composed of several nodes, and each node has a weight, which determines the output result of the output layer. The input layer inputs seismic attributes, so its number of points depends on the number of weighting coefficients of linear correlation regression equation between geological attributes and seismic attributes. If convolution operator is used, the number of points of its effective attribute will increase with the increase of operator length. If the operator length is 3, each attribute will be sampled three times, and the corresponding sampling time is -65438.
Because the predicted geological properties are relatively simple, such as porosity or density, the output layer has only one node, and an intermediate layer composed of three nodes is used according to experience.
Determine the optimal weight coefficient of nodes through trial processing, such as single sample data series {S 1, S2, S3, G}, where S 52 and S3 are different seismic attributes, such as amplitude, frequency and wave impedance; G is the geological attribute recorded by a well, such as porosity or 〓 degree, etc. Because there are many seismic records corresponding to the existing well records in the time window, there are also many samples to be tested.
Data operation adopts nonlinear algorithm. The conventional method is to calculate the measured target logging record and the seismic attribute prediction target logging record by the least square difference algorithm, so that the measured target logging curve is similar to the logging I line corresponding to the seismic attribute prediction target with minimum error. Back propagation method, that is, gradient descent method, is usually used to solve the problem. In modern times, the * * * yoke descent method 〓 simulated annealing method is used to accelerate the convergence of operation and avoid the phenomenon of non-convergence.
Fig. 6-11shows the intersection scatter distribution of single seismic attribute and geological attribute of logging curve. Because it has seismic properties, its input layer has only one node. It is exactly the same as the scatter distribution in Figure 6- 106, but its regression curve is nonlinear regression fitting, which makes the regression curve closer to the scatter concentration. It improves its accuracy, but in low seismic attributes, due to the scattered jump of its intersection point, the fitting curve appears unstable big jump anomaly, because the neural network always tries to fit the curve closer to the intersection point of all attributes.
Fig. 6-11nonlinear regression curve
(3) Probabilistic statistical neural network
It uses the same data (including experimental data) as the multi-layer neural network. For example, suppose that the data series of seismic records in the analysis time window recorded by all wells are:
High-tech and Practice of Offshore Oil in China
N in the sequence is the number of samples processed; S 1, S2 and S3 are three seismic attributes; G is the geological attribute such as logging corresponding to seismic attribute.
For the given data of training neural network, neural network assumes that each new output logging curve value can be written as a linear combination of logging curve values in training data. For the new data sample X and the seismic attribute values S 1j, S2j and S3j, the relationship is as follows:
High-tech and Practice of Offshore Oil in China
The calculation formula of new logging curve value G(x) is:
High-tech and Practice of Offshore Oil in China
Among them:
High-tech and Practice of Offshore Oil in China
D(x, xi) is the distance between the input point and each training point xi; Measured by the ratio of seismic attributes to σj in multi-dimensional space domain, σj of each seismic attribute can be different.
Equations (6-52) and (6-53) describe the application of probabilistic statistical neural networks. The training of the network consists of the best set of (σj) smoothing parameters, and the criterion for determining these parameters is that the network should have the lowest effective error.
Then the prediction error of training (processing) data is defined by the following formula:
High-tech and Practice of Offshore Oil in China
It can be seen from Equation (6-54) that the prediction error E depends on the value of σ j, and the nonlinear * * * yoke gradient algorithm can minimize E and find the value of σ j. ..
The fitting curve calculated by the above probabilistic statistical neural network is shown in Figure 6- 1 12.
Fig. 6- 1 12 regression curve of probabilistic statistical neural network
As can be seen from Figure 6- 1 12, it overcomes the instability phenomenon of abnormal point jump in the fitting curve in Figure 6- 1 1, and is the best one in the fitting curve. However, due to the large amount of data used, advanced and complex mathematical model and heavy workload, large computer supporting equipment is needed. Therefore, probabilistic statistical neural network is also called mathematical neural network.
In a word, the key technology of multi-attribute comprehensive seismic prediction of reservoir characteristics is to extract a large number of reservoir geological characteristics carried by seismic data and optimize seismic attributes related to reservoir geological characteristics by using advanced technologies such as multi-channel geological statistics, optimization of regression analysis, pattern recognition, interactive verification of well constraints, accelerated convergence of operation, and optimal correlation curve fitting of intersection points.
Second, the realization method of technology
(A) the implementation of technical concepts
After the unknown is predicted by the known, the reliability of the unknown is predicted by the known verification. That is to say, the correlation between various seismic attributes (such as amplitude, frequency, phase, trace integration, time, etc.) from the well point. ) and geological properties (such as density, porosity, impedance, permeability, etc.). ) and train the neural network to invert the optimized cross-well multi-seismic attributes into predicted reservoir parameters. Then, the reliability of predicted reservoir parameters is verified by reservoir parameters not included in geostatistics wells. For those that do not meet the error requirements, the predicted reservoir parameters should be iteratively revised to meet the error requirements.
(II) Technical processing flow and steps
See figure 6- 1 13 for the processing flow.
As can be seen from Figure 6- 1 13, the specific method steps for its implementation are as follows.
A. Input logging target curves for correction and deep time conversion (such as density and porosity curves, etc.). );
B. Input pure wave and fidelity seismic data volume (such as stacking and migration pure wave data volume);
C, selecting seismic traces beside wells, calculating seismic attributes of required wells, correctly corresponding the calculated seismic attributes with interval positions of logging target curves by using synthetic seismic records, and then determining the correlation between logging target curves and multi-seismic attributes by using multi-channel geostatistics technology of neural networks and analysis of regression fitting curves of their intersection distribution laws;
Figure 6- 1 13 Technical Processing Flow of Seismic Multi-attribute Inversion of Reservoir Parameters
D, through stepwise regression analysis, determining the sequence and number of seismic attributes with maximum or optimized values related to the logging target curve;
E. Based on multi-well multi-channel geostatistics and probabilistic statistical neural network technology, neural network is trained with various correlation model curves between logging target curves and multi-seismic attributes, which lays a foundation for multi-attribute inversion prediction of cross-well seismic logging target curves;
F, inputting the cross-well seismic data volume into a neural network, carrying out pattern recognition on the seismic data volume by using the number and sequence combination of the selected optimal seismic attributes, and carrying out inversion prediction on the reservoir parameters;
G, interactively verifying the reliability of the optimized number and sequence combination of seismic attributes and the predicted reservoir parameters derived from the geological demand data volume, namely the final processing result, by using the logging target curve calculated by the nonparametric model;
H. If the residual can't meet the requirements of inversion accuracy, iteratively correct it by modifying the model and seismic attribute parameters until the accuracy requirements are met, and then output and display the processing results, providing a basis for the interpretation of reservoir lithology, physical properties and hydrocarbon content.
Third, application examples and effects
(1) General situation and geological tasks of the work area
Nanpu 35-2 Oilfield is located in the west of Shijiutuo Uplift in Bohai Sea, belonging to buried hill drape anticline structure, which is affected by faults in the later stage. 1996 drilling Nanpu 35-2- 1 Well discovered a 55m oil layer in the lower member of Minghuazhen Formation. DuriNg the period of 1997, based on the reinterpretation of 2D seismic data, three wells were drilled successively. Except for these three wells, all other wells encountered oil and gas reservoirs of Nm and ng. It is a fluvial reservoir with good physical properties and good reservoir-cap combination. It is a multi-oil-water layer combination system and a structural lithologic heavy oil reservoir cut by faults. On the basis of 220km2 high-resolution 3D seismic data collected by 1997 and processed by 1998, four more wells were drilled. In reservoir description and interpretation, it is found that 3D seismic wave impedance data volume can not reliably predict the variation and distribution range of reservoir thickness, and it is difficult to distinguish sandstone and mudstone because of the poor correlation between acoustic logging curve and sandstone and mudstone profile, and even errors occur. The actual drilling proves that the predicted reservoir thickness is quite different from the actual drilling reservoir thickness. For example, the predicted thickness of the zero oil group 10 sand layer in Well 7 is less than 1.3m, while the actually drilled thickness is 9m, with a difference of 7.7m Therefore, in order to meet the needs of oilfield development, in 2006, 5438+0 used the above technology to invert the reservoir parameters of high-resolution 3D migration pure wave data volume, which improved the accuracy of reservoir parameter prediction and provided a reliable basis for oilfield development.
(2) Analysis of drilling reservoir parameters.
Using the neural network multi-channel geostatistics, crossplot regression and other technologies, the crossplots of gamma and acoustic wave, wave impedance and gamma, and density of 8 wells in the work area are made and analyzed, so as to determine the sand and mud strata that can distinguish the main reservoir-cap combinations in the construction area. From Figure 6- 1 14, it can be seen that sandstone has low gamma impedance and mudstone has high gamma impedance, but there is an indistinguishable overlap between their wave impedances, indicating that reservoir parameters cannot be accurately predicted by using interval velocity and impedance of acoustic logging as prediction targets.
Fig. 6- 1 14 gamma and acoustic impedance crossplot of Nanpu 35-2- 1, 2, 6 and 7 wells.
As can be seen from Figure 6- 1 15, the intersection of gamma and density is distributed in a straight line, that is, low gamma corresponds to low density and sandstone reservoirs. Therefore, taking density logging curve as reservoir prediction target can effectively and accurately predict reservoir parameters. Through the above processing test analysis, it is determined that the geological target of inversion processing is density data volume. After interpretation and well constraint calibration, the reservoir prediction and sand body tracking in the work area are completed.
Figure 6- 1 157 Gamma and Density Crossplot
(3) Multi-attribute analysis of seismic data volume
Using the above processing technology, the input well density curve and seismic data volume are processed by seismic multi-attribute inversion density data volume. Firstly, the program calculates the multi-attributes of each seismic trace near the well, establishes the objective function between it and the well density curve, and determines which seismic attribute is the best by step-by-step recursive method; The nonlinear relationship between well density and seismic effective attributes is calculated by probability statistical neural network. The reliability of the correlation model between well density and seismic attributes is verified by interactive verification method. From the error analysis, seismic attribute list and Figure 6- 1 16, it can be seen that with the increase of seismic attributes, the average error of all wells (off-line) decreases compared with the verification error of verification wells (on-line). Therefore, the density data volume of seismic data volume is inversed by using the above four seismic attributes, and good sandstone reservoir prediction results are obtained.
Figure 6- 1 16 Error Analysis Diagram
Table 6- 10 seismic attribute list
Similarly, if the geological target curve of the input well is the reservoir physical parameters, such as porosity, permeability and saturation. Using the above technology, the seismic data volume can be inverted into porosity, permeability and other data volumes. That is, it is inverted into a predicted geological target corresponding to the example geological target. Numbers 1, 2, 3, 4 ... 10 in Figure 6- 16 are the same as those in Table 6- 10, 1+2 and 1+2 respectively.
(4) Effect analysis
A. As can be seen from Figure 6- 1 17, the inversion wave impedance of Well Nanpu 35-2- 1 and 7 is compared with the profile of density-connected well, and the distribution law of high and low wave impedance layers on the wave impedance profile is poor, so it is difficult to track and explain the reservoir correlation reliably, while the distribution law of low density layers on the corresponding density profile is good, with layered distribution characteristics and easy to be reliable. Therefore, this density data volume inversion effectively solves the problem of poor reservoir prediction accuracy of previous wave impedance data inversion, and provides a reliable basis for reservoir description of Nanpu 35-2 reservoir.
Fig. 6- 1 17 Comparison between seismic attribute inversion wave impedance and relative density correlation profile of Nanpu 35-2- 1 Well 7.
The actual drilling results of well b. 1 and well 7 prove that the reservoir interpretation results of multi-attribute inversion density body are accurate. For example, the actual drilling thickness of Zero Oil Group 1 No.0 Sand Layer1Well is 1.3m, and the actual drilling thickness of Well No.7 is 9m. The data of the two wells are compared and analyzed, and the sand layers of the two wells are not connected. It can be seen from the relative density profile in Figure 6- 1 17 that the predicted reservoir thickness of zero oil group 10 sand layer, 1 well and 7 well and the unconnected sand layer between the two wells are consistent with the drilling results.
Figure 6- 1 18 Relative Density Profile of Well Nanpu 35-2-6 and Well 10.
C For well 10, the multi-attribute inversion data volume constrained by the well is not involved, and the actual relative density curve encountered is similar to the relative density profile predicted by the well location, as shown in Figure 6- 1 18. As can be seen from the figure, the correlation between the measured and predicted relative density curves is obvious and satisfactory.
D As can be seen from Figure 6- 1 19, the measured density curves of 8 wells in the work area are similar to the predicted density curves. This shows that the seismic multi-attribute inversion density data volume is completed under the control of multi-well constraints and needs to be verified by wells. Therefore, this technology effectively solves the problem of multiple solutions and reliability of inversion results, which is its unique effect.
E. It can be seen from the slice figure 6- 120 along the layer of oil group II cut from the density data that the distribution pattern of sand layer of oil group II clearly reflects the sedimentary characteristics of river facies such as netted distributary channel, crevasse fan and sand bar. This result shows the effect of reservoir prediction, which is obviously consistent with the sedimentary characteristics of Tertiary fluvial facies and the actual drilling results in the work area.
F The porosity profile of 1 well cut from the porosity data volume of multi-attribute inversion shows that the porosity curve of 1 well has a good correlation with the porosity profile predicted by seismic multi-attribute inversion. This example shows that the effect of seismic multi-attribute inversion of reservoir physical properties can not be achieved in the past.
Figure 6- 1 19 Multi-attribute Analysis Diagram
Figure 6- 120 Reservoir Distribution Prediction
G. Seismic multi-attribute reservoir parameter inversion technology has made brilliant achievements in adjustment while drilling, such as the adjustment of 23 wells in Qinhuangdao 32-6 Oilfield, which greatly reduced the reservoir description results of ODP in offshore Dongfang 1- 1, Wenchang 13, Qinhuangdao 32-6 and Nanbao B35-2 oil and gas fields.
In a word, this technology has played an important role in fine reservoir description of oil and gas reservoirs. With the deepening of oil and gas exploration and development, this technology will certainly make greater contributions to the fine reservoir description of oil and gas reservoirs in practice.
Figure 6- 12 1 Jingnanbao 35-2- 1: 1 100 million gap profile