Sunteți pe pagina 1din 5

A new reservoir prediction method: PCA value-weighted attribute optimization

Lifeng Liu*, Sam Zandong Sun, Haiyang Wang, Lab for Integration of Geology and Geophysics (LIGG), China University of Petroleum (Beijing), Haijun Yang, Jianfa Han, Bing Jing, Tarim Oilfield Co., CNPC
Summary This paper introduces the principle component attribute (PCA) technology into carbonate reservoir prediction, and a PCA value-weighted attribute optimization method is proposed here. As the calculation time-window of the PCA valueweighted attribute is the same (viz., free of the constraint of time-window scale), it can effectively avoid the negative effect of different time-windows on attribute optimization results. A test area is chosen from Tarim Basin, and the reservoir information of the known wells is treated as the training samples. The discriminant model of PCA valueweighted attribute and reservoir parameters are built. Through the retrospective test, the total fit rate between the attribute optimization results and the production results is up to 94.62%, proving that the discriminant capability of PCA value-weighted attribute for reservoir quality is significant. Finally, the proposed method is employed for the integration prediction of carbonate reservoir in test area. The prediction results finely exhibiting the distribution of favorable carbonate reservoir. This paper is aimed at providing a scientific and practical method for seismic attribute analysis and reservoir prediction. Introduction The attribute optimization technique is the most effective method to solve the problem of multi-solutions of seismic attributes (Chen, 1997; Leiphart, 2001; Dorrington, 2004). However, much attention is paid to the investigation of the optimization methods and the effectiveness of the seismic attribute itself is neglected. For instance, the results of seismic attributes analysis have a close relationship with the calculation time-window. For different attributes, the caculation time-windows are different due to the difference of the algorithms and the meanings. For instances, frequencydivision attributes require relatively larger time-window while the large time-window for the amplitude attributes will cause serious average effect. Therefore if these attributes with different time-windows are used for optimization, the reliability of the prediction results is queried. The principle of PCA PCA (principal components analysis) is one of the datadigging technologies in the multivariate statistic analysis. It mainly uses the dimension-reducing idea in mathematics. The idea is that through reducing the dimension of dataset to eliminate the correlation between different elements. Consequently, the problems in high dimension space are transformed into those in low dimension space, revealing the simple structure hiding behind the complex data and eliminating the noise interruption and redundancy. Finally, the instability of the calculation results caused by the correlation among the multi-variables can be avoided (Tipping, 1999; Gu, 2007). For one seismic trace, it can be taken as a linear dataset of waveforms. The original seismic data is a high-dimension array made up of n dimension waveform variables (Sun, 1999). Using PCA to reduce the dimension of data and calculate m waveforms (principle components), where n>m. Through the sum of linear weight for these principle components, we could realize the reconstruction for any seismic trace (shown in Figure 1). The most principal components and waveform features can be recognized from seismic data, and the seismic data could express the most of information in the primitive data as well as eliminate some random noise and interference. Finally, the principle components are orthogonal and independent to each other. Assuming seismic data X is a vector containing n columns (xi(t), i=1,2,,n; t=1,2,,t), its covariance is (1) C = XTX This covariance matrix must be a positive semidefinite symmetrical matrix which contains n rows and n columns. Conduct SVD decomposition to this matrix and we get (2) C = RR T The matrix is the eigenvalue of the symmetrical matrix, and R is the eigenvector corresponding to the eigenvalues. The orthoganal eigenvetor corresponding to the eigenvalue ii (sorted from big values to small ones) is the coefficient vector Ri corresponding to the ith principal element. The principal component P can be defined as, (3) P = XRT Each principal component is the linear combination of primitive seismic traces, and its dimension is the same to the primitive seismic trace. The ith column of the principal component P is the ith principal component, where the first column contains the most information of seismic data. The reconstruction to seismic trace is the inverse transformation of the process above, and (4) X = PR The variance contribution rate of the principal component P i is defined as i/j, In general the number (m) of the extracted principal component is needed to satisfy k/j>0.85. That is to say, the accumulative contribution rate of the former m principal components reaches to 85%, which could contain

2011 SEG

SEG San Antonio 2011 Annual Meeting

2024

PCA weight value attribute optimization


the most information of the primitive seismic data, where n>m, that is,
x11 x 21 # x t1 x12 " x1n P11 x 22 " x2 n = P21 # # # # xt 2 " xtn Pt1 P12 " P1m r11 r12 " r1n P22 " P2 m r21 r22 " r2 n # # # # # # # Pt 2 " Ptm rm1 rm 2 " rmn

(5)

Finally, through the sum of linear weight of the top 10 principle components, the reconstruction of seismic traces can be realized.

Figure 2: The time-window of seismic data used for PCA

Figure 1: The principal schematic of the principal component analysis (PCA) for seismic data

Reconstruction of seismic data A certain area in Tarim Basin is chosen as the testing area. According to the principal component theory, the analysis and test on the field seismic data is performed. The timewindow of the seismic data used for PCA ranges from 20ms to 160ms below the top of Lianglitage formation, as is shown in Figure 2. The seismic data used for testing contain 854532 traces, which can be taken as the array made up of 854532 waveform variables. The PCA method is employed here to transform the 854532 waveform variables into 854532 principal components (Y1, Y2,, Y28623), and each principal component is orthogonal and independent to each other. That is to say, every seismic trace could be represented by the sum of the linear weight of these principal components (shown in Figure 3). Calculate the eigenvalues and the accumulative contribution rate of these variance matrixes, as is shown in Table 1. The contribution rate of the first principal component is the greatest (65.28%), containing the most of the information of the primitive seismic trace, followed by the second principal component. The accumulative contribution rate of the top 10 principal components reaches to 95%, reflecting the most important waveform features.
Figure 3: The principal components through the decomposition of seismic traces in TZ45 area.

Application of PCA value-weighted attribute optimization method on carbonate reservoir prediction When using the principal components to reconstruct the seismic data, the value-weighted of each seismic trace can be taken as a new attribute, which is called PCA value-weighted attribute in this paper. The m row vectors in eigenvector matrix R in Equation 3 can be viewed as m PCA valueweighted attributes. Since the calculation time-window of the attribute is the same, it is not controlled by the window scale, thus effectively avoiding the effect of time-window on attribute optimization result. The PCA value-weighted

2011 SEG

SEG San Antonio 2011 Annual Meeting

2025

PCA weight value attribute optimization


attribute is introduced into the discriminant analysis, establishing the discriminant model of the attribute and reservoir parameters, further predicting the spatial distribution of reservoir. Firstly, considering the actual situation of reservoir development in the study area and the sensitivity of seismic data for reservoir, the carbonate reservoirs can be classified into three types: hydrocarbon layer, hydrocarbon-showing layer and non-reservoir layer. The three types of layers are taken as the classification to establish discriminant model, and the obtained PCA value-weighted attributes are taken as the input variables of the discriminant model. Then reintroduce the various indexes of given training samples into the built discriminant models. Through calculating the coincidence rate of reintroducing, thus discriminant ability of the model can be examined (shown in Figure 4). The figure shows that the total coincidence rate reaches to 94.62%, validating the effectiveness of the discriminant model. Finally, the PCA weight value attribute points are brought into the established discriminant model, determining the 3D spatial distribution of the carbonate reservoir in study area.
25000 25000
The predicted HC layer The predicted HC-showing layer The predicted Non-reservoir layer

seismic attribute optimization (Figure 5d), are not ideal. However, the PCA method was able to identify the location of the well clearly (Figure. 5e), proving the higher prediction accuracy of the PCA value-weighted attribute method compared with the conventional methods. Conventional attributes and PCA value-weighted attributes in the study area were extracted respectively for gain limestone member of Lianglitage formation. From the comparative analysis between them, the results of PCA value-weighted attribute optimization are consistent well with the oil/gas production situations of various wells (show in Figure 6). And the fit rate is up to 100%. Meanwhile, compared with the conventional attribute optimization results, its predictions show a strong heterogeneity, which accords with the actual geological conditions, thus more objectively reflecting the strong heterogeneity characteristics of carbonate reservoir. We use the three-dimensional visualization technology to sculpture the results of PCA value-weighted attribute optimization (show in Figure 7), tracking the carbonate reservoir in the vertical and lateral directions to reveal the favorable distribution and geometric shape of dissolution caved carbonate reservoirs in 3D space. Consequently, the connectivity between the reservoirs can be clear determined, further laying a foundation for the accurate classification of carbonate reservoir flow units and the formulation of development scenarios. Conclusions

22813

20000 20000 Samples number

15000 15000

The total coincidence Rate=94.62%

94.32%

10000 10000 3769 2398


94.56% 112

5000 5000

26

87

96.54%

48

367 1008

The drilling HC layer

The drilling HC-showing The drilling non-reservoir layer layer

Figure 4: The coincident rate of the predicted reservoir types compared with the drilling reservoir types

Reservoir integration prediction and results analysis Figure 5 shows the comparison of multi prediction result sections crossing Well TZ86. The core and FMI logging data shows that the main storage spaces of the well are dissolution pores, small-scale caves, and fractures. Dissolution pores and caves are connected by fractures, making the reservoir performance superior. For this type of reservoir, highyielding oil/gas flow can be obtained by acid fracturing. This type of reservoir has no large-scale dissolution caves and it is relatively in small scale and weak heterogeneity. So in the seismic profile, it does not show beaded strong reflection with strong heterogeneity, but shows the messy irregular reflection characteristics. Due to the limitations of low resolution for this type of reservoir, the common prediction means such as frequency-attenuation gradient (Figure. 5b), tuning-frequency energy (Figure. 5c) and conventional

Principal component analysis (PCA) is introduced into the discriminant analysis, and a PCA value-weighted attribute optimization method is proposed. Through the establishment of discriminant model of PCA value-weighted attributes and reservoir parameters, the spatial distribution of the reservoir is predicted. Because of the same calculation time window of the PCA attributes, the optimization is not controlled by timewindow scale, avoiding the negative effect of different time windows between different seismic attributes. Moreover, since the PCA value-weighted attributes are orthogonal, it can effectively avoid the instability of the optimization results. In the prediction of the carbonate rocks reservoir in the Tarim Basin, the total fit rate is up to 94.62% and the optimization results are consistent well with the actual oil/gas production situations of various wells. This method can still identify the relative small reservoirs which are not sensitive to the conventional attributes. To sum up, the PCA value-weighted attribute optimization method can significantly improve the reservoir prediction accuracy and finely sculpture the distribution of carbonate reservoir in 3D space. And it can be taken as a scientific and practical method for seismic attribute analysis and carbonate reservoir prediction.

2011 SEG

SEG San Antonio 2011 Annual Meeting

2026

PCA weight value attribute optimization

Figure 5: The compared results of Well TZ86 through a variety of forecasting methods: (a) Seismic data; (b) the gratitude of frequency attenuation; (c) frequency turning energy; (d) the optimization of the conventional attributes; (e) PCA value-weighted attribute optimization.

Figure 6: Comparison of reservoir prediction for gain limestone member of Lianglitage formation in study area: (a) The results by using conventional attribute optimization;(b) The results by using PCA value-weighted attribute optimization
ZG17 TZ49 TZ86

TZ452 ZG18 TZ451 TZ45 ZG16

TZ88

TZ63

Figure 7: The predicted spatial distribution of dissolution caved carbonate reservoirs for study area.

2011 SEG

SEG San Antonio 2011 Annual Meeting

2027

EDITED REFERENCES Note: This reference list is a copy-edited version of the reference list submitted by the author. Reference lists for the 2011 SEG Technical Program Expanded Abstracts have been copy edited so that references provided with the online metadata for each paper will achieve a high degree of linking to cited sources that appear on the Web. REFERENCES

Chen, Q., and S. Sidney, 1997, Seismic attribute technology for reservoir forecasting and monitoring: The Leading Edge, 16, 445456, doi:10.1190/1.1437657. Dorrington, K. P., and C. A. Link, 2004, Genetic algorithm /neural network approach to seismic attribute selection for well log prediction: Geophysics, 69, 212221, doi:10.1190/1.1649389. Gu, S. H., Y. S. Wang, and G. G. Wang, 2007, Application of principal component analysis model in data processing?: Journal of Surveying and Mapping, 24, no. 5, 38739. Leiphart, D. J., and B. S. Hart, 2001, Case history comparison of linear regression and a probabilistic neural network to predict porosity from 3D seismic attributes in lower brushy canyon channeled sandstones, southeast New Mexico: Geophysics, 66, 13491358, doi:10.1190/1.1487080. Sun, Z. D., 1999, Seismic methods for heavy oil reservoir monitoring and characterization: PhD thesis, University of Calgary. Tipping, M. E., and C. M. Bishop, 1999, Probabilistic principal component analysis: Journal of the Royal Statistical Society, 61, no. 3, 611622, doi:10.1111/1467-9868.00196.

2011 SEG

SEG San Antonio 2011 Annual Meeting

2028

S-ar putea să vă placă și