This invention relates to a method for modeling a reservoir. More particularly, a method for modeling a 3D reservoir using multiple-point simulations with 2D training images.
In the oil and gas industry, flow simulations are widely used for predicting the reservoir performance, which in many cases is controlled by the distribution of reservoir properties, such as facies, porosity and permeability.
The variogram-based algorithms, e.g., SGSIM (Sequential Gaussian Simulation) and SISIM (Sequential Indicator Simulation), generate petrophysical distributions pixel-by-pixel through a prior variogram model accounting for the spatial continuity and are able to condition various types of data, such as well data, 2D or 3D trend information (Goovearts, 1997; Deutsch and Journel, 1998; Remy et al., 2009). However, these algorithms only reproduce up to 2-point statistics, histogram and variogram, which are not sufficient to generate complex geological features, such as lobes and channels.
The object-based or the Boolean algorithms (Haldorsen and Chang, 1986; Lantuejoul, 2002; Maharaja, 2008) can produce better geological patterns by dropping whole objects of given shapes into the simulation grid. These algorithms parameterize the objects according to the shape, size, anisotropy, sinuosity, and the interaction (erosion and overlap) with other objects. In order to match the desired facies proportions and to honor the conditioning data, an iterative process is utilized to remove, replace, and transform the previously dropped objects. However, the iterative process creates issues when the well spacing is smaller than the object size. This situation gets worse with dense wells, 2D or 3D soft data, and other exhaustive information.
The weakness of both variogram-based and object-based algorithms triggered the concept of multiple-point geostatistical simulations (mps) (Journel, 1992). With mps, the simulation proceeds pixel-wise along a random path which visits all the simulation nodes. Instead of borrowing statistics from a prior variogram model, the local conditional probabilities in mps are lifted as conditional proportions from a given training image (TI). The training image is a geological concept model depicting geological patterns such as the geometry, texture and distribution of objects deemed to prevail in the real world.
The SNESIM (Single Normal Equation Simulation) is the first practical mps algorithm developed to handle the categorical variables, such as lithofacies (Strebelle, 2000). Later, the FILTERSIM (Filter-based Simulation) algorithm was utilized to simulate the distribution of continuous variables, such as porosity and permeability (Zhang 2006).
In order to perform mps simulation to generate a 3D geological property, a 3D training image is required. Practically, the TI is obtained from outcrops, air photos, or even brush-painted by geologists, which are normally in 2D. How to generate a 3D TI with a 2D map and how to simulate 3D mps models from a 2D TI become challenging for geoscientists.
How to reconstitute a 3D structure from a 2D map is a topic of interest in various areas, such as image processing, material engineering, rock physics, and petroleum engineering. A number of methods have been proposed, which can be categorized into four groups: simple stack, statistics-based, process-based and mps-based.
The simple stack method combines a series of 2D sections into a 3D image. These 2D sections can be obtained directly from laboratory photographs (Dullien, 1992; Tomutsa and Radmilovic, 2003) or from X-ray computed tomography pictures (Dunsmuir et al., 1991; Fredrich 1999). In general, these methods require laborious operations and are very time consuming, hence are not suitable for the routine applications.
The statistics-based techniques describe the 3D models with some statistical measure, for instance the histogram and the 2-point correction functions. The statistics-based techniques then reconstruct the 2D model with respect to the statistical measures using stochastic procedures, such as simulated annealing (Yeong and Torquato, 1998), truncated Gaussian simulation (Biswal and Hilfer, 1999), and percolation system (Daian et al., 2004). The statistical measures could come from some empirical relations (Ioannidis et al., 1996) or be derived from a known 2D image (Quiblier, 1984). The main drawback with the statistics-based technique is the difficulty to reproduce the long range connectivity of interested variables.
The process-based algorithms reconstruct 3D porous medium by modeling its geological process (Bryant and Blunt, 1992; Biswal et al., 1999; Pilotti, 2000). This method is capable of reproducing long range connectivity for certain geological systems. However, process-based algorithms encounter difficulties when the sedimentation process becomes complex and/or involved irregular object shapes, for example the carbonate system. Moreover, the process-based training image is not stationary for mps simulation
In recent years, the mps algorithm for 3D image constructions using 2D maps as the input training images have been utilized. Okabe and Blunt (2005) proposed a method to use SNESIM algorithm for pore space reconstruction. Assuming the porous medium is isotropic, a 2D section image is used to provide the pore space patter in each X/Y/Z direction during the SNESIM simulations. Zhang et al. (2008, 2009) proposed another way to use the SNESIM algorithm for generating 3D pore space images. In this method, each horizontal layer is simulated in sequence. After simulation of one layer is complete, the training image is replaced with the simulation in that layer. Data is sampled with a predetermined template from the training image to the next simulation later. Finally, the SNESIM simulation is performed with the new TI and the sampled hard data. The newly proposed mps-based method can reproduce long range connectivity, but is exempted from the prior knowledge of geological process.
All of the methods discussed above focus on the micro scale, such as the 3D porous medium structure and material microstructures. Therefore, a need exists for a macro scale mps algorithm to generate 3D reservoir distributions (facies, porosity, permeability) with 2D training images.
In an embodiment, a method for modeling a reservoir includes: receiving and loading data; creating a 3D grid with a plurality of layers; generating a 2D grid for each layer in sequence; reconstructing or simulating a 2D image for the first layer; sampling data from the 2D image on the first layer; sampling data from the 2D grid for all other layers; setting sampled data as hard data; performing a filter based simulation to condition the hard data; and copying the filter based simulation from the 2D grid to the 3D grid.
In another embodiment, a method for modeling a reservoir includes: receiving and loading data; creating a 3D grid with a plurality of layers; generating a 2D grid for each layer in sequence; getting target facies proportions; sampling the data from the 2D grid, wherein the sampling is a point sampling, a geobody sampling or a hybrid sampling; performing a single normal equation simulation to condition the data; and copying the SNESIM based simulation from the 2D grid to the 3D grid.
The invention, together with further advantages thereof, may best be understood by reference to the following description taken in conjunction with the accompanying drawings in which:
a)-1(c) show various images according to one embodiment of the invention: (a) 2D training image; (b) good 3D training image; (c) poor 3D training image.
a)-3(b) depict a final 3D realization with random sampling, according to an embodiment of the invention: (a) original 3D realization; (b) de-noised realization.
a)-4(d) depict hard samples for data conditioning and simulated realization, according to an embodiment of the invention: (a) hard data sampled from layer 1; (b) 2D realization in layer 2; (c) hard data sampled from layer 2; (d) 2D realization in layer 3.
a)-6(b) depict a final 3D realization with histogram-based sampling, according to an embodiment of the invention: (a) original 3D realization; (b) de-noised realization.
a)-7(b) depict realizations with random sampling option and data mutations, according to an embodiment of the invention: (a) no mutation; (b) mutate 20% samples; (c) mutate 30% samples; (d) mutate 40% samples.
a)-8(d) depict realizations with random sampling option, according to an embodiment of the invention: (a) 100 samples; (b) 200 samples; (c) 300 samples; (d) 400 samples.
a)-10(b) depict realizations with regular grid sampling and data mutations, according to an embodiment of the invention: (a) no mutation; (b) mutate 30% samples.
a)-11(e) depict testing geobody sampling with two facies training image, according to an embodiment of the invention.
a)-13(b) depict a final 3D realization and one realization simulated directly with the given 2D two facies training image according to an embodiment of the invention: (a) realization with new workflow; (b) direct simulation with 2D TI.
a)-15(d) depict three final 3D realizations and one realization simulated directly with the given 2D three facies training image, according to an embodiment of the invention: (a) realization #1 with new workflow; (b) realization #2 with new workflow; (c) realization #3 with new workflow; (d) direct simulation with 2D TI.
a)-17(b) depict a final 3D realization and one realization simulated directly with the given 2D four facies training image, according to an embodiment of the invention: (a) realization with new workflow; (b) direct simulation with 2D TI.
Reference will now be made in detail to embodiments of the present invention, one or more examples of which are illustrated in the accompanying drawings. Each example is provided by way of explanation of the invention, not as a limitation of the invention. It will be apparent to those skilled in the art that various modifications and variations can be made in the present invention without departing from the scope or spirit of the invention. For instance, features illustrated or described as part of one embodiment can be used in another embodiment to yield a still further embodiment. Thus, it is intended that the present invention cover such modifications and variations that come within the scope of the appended claims and their equivalents.
The present invention focuses on a method to use the FILTERSIM algorithm for 3D continuous variable simulations using a 2D training image (TI) and a method for the use of the SNESIM algorithm to simulate 3D categorical facies with a 2D TI. The resulted 3D image should reproduce the geological features from the given 2D TI; should have reasonably good vertical continuities; and should have reasonably good vertical variations. For example,
The FILTERSIM algorithm (Zhang, 2006) first extracts all of the patterns from the given training image (TI) using a predefined template. As previously discussed, the training image is a geological concept model depicting geological patterns. The geological patterns are then grouped into different classes based on the filters. Finally the algorithm performs stochastic simulation using pattern recognition techniques. The simulated realization can be conditioned to various types of data, such as well hard data, soft probability or trend data, and azimuth and scaling factors. In general, the FILTERSIM algorithm requires a 3D TI for 3D simulations.
The method for 3D multiple point simulation uses 2D training images to simulate continuous variables with the FILTERSIM algorithm, shown in
Nx×Ny×Nz. The method can also be used to construct a 3D TI from any 2D maps, in that sense the size of the 2D TI must be Nx×Ny.
The method processes the 3D grid (G) from layer 1 to layer Nz in sequence. At each layer k ∈[1, Nz], a 2D simulation grid Gk (of size Nx×Ny) can be created and be used as the host for running the FILTERSIM simulation.
For the first layer (k=1), if the purpose of simulation is to reconstruct a 3D TI, then sample data (nk) from the given the TI is sampled, saved as hard data and run in the FILTERSIM simulation conditioning to the nk samples; otherwise, unconditional FILTERSIM simulations are run. For the remaining layers k ∈[2, Nz], first sample data (nk) data from the 2D grid Gk−1, save the samples as hard data in layer k, and then run the FILTERSIM simulation conditioning the nk new samples. The simulated realization in grid Gk can be post-processed to remove the simulation noise, for example using the Gaussian low pass filter. Finally, copy all temporary 2D simulations from grids Gk (k=1,2, . . . , Nz) to the 3D grid G to form a full 3D realization.
For the data sampling, various strategies can be applied, such as (1) fully random sampling, which does not consider the data values and patterns; (2) regular grid sampling, which samples every Nrx, Nry node in the X, Y direction, respectively; (3) stratified sampling, which divides the grid into Nw windows and randomly select nk/Nw nodes from each window; or (4) histogram-based sampling, which divides the histogram into Nint intervals and then randomly samples nk/Nint nodes for each interval.
If no vertical trend curve exists, then there is no preference as to which sampling algorithm is selected. However, when a vertical portion/mean curve exists, the histogram-based sampling is recommended. With the histogram-based sampling, nodes close to corresponding values from the vertical curve and sampled. The sampled hard data should be combined with the original user-supplied hard data to constrain FILTERSIM simulations.
Because the process borrows some conditioning data from the adjacent layer, some vertical continuity's between two successive layers are ensured. And because of the nature of stochastic simulation, there will be some variations between these two layers. However, when the number of sampled hard data becomes larger, the difference between two successive layers will get smaller. A certain portion of the sampled data should be mutated to allow for larger vertical variations. The operation can be data location shifting, data value modification, node dropout, or any combination thereof.
a) gives one final 3D realization with random sampling option and
Next, a vertical target mean curve was used to constrain the average probability for each layer. The input target means are shown in Table 1.
From
a) shows a 2D training image, in which the property values accumulate high along the channel centers and decreases gradually to zero towards the channel edges. The size of the training image is 240×240.
The 2D training image was used to construct a 3D training image of size 240×240×10. FILTERSIM was run with a 17×17 search template and a 5×5 patch template.
Next, the method was tested with the mutation concept.
The SNESIM algorithm (Strebelle, 2000) first scans the given TI for all possible patterns with a predefined search template and saves the scanned local simulation proportions into a search tree data structure. During simulation, the same search template is used to look for the local conditioning data in the simulation grid and its corresponding conditional probability. Similar to the FILTERSIM algorithm, it is not recommend to perform SNESIM simulations with a 2D TI.
The method uses 2D training images to simulate categorical variables with the SNESIM algorithm, shown in
The workflow processes the 3D grid (G) from layer 1 to layer Nz in sequence. At each layer k ∈[1, Nz], a 2D simulation grid (Gk) of size Nx×Ny should be created and will be used as the host for running the SNESIM simulation. For each layer, the target facies proportions can be the same as the global target, if there are no vertical proportion curves; otherwise, target facies proportions are derived from vertical proportion curves.
For the first layer (k=1), if the purpose of simulation is to reconstruct a 3D TI, then sample nk data from the given TI is sampled and the SNESIM simulation conditioning to the nk samples is run; otherwise, unconditional SNESIM simulations are run. For the remaining layers k(>1), first sample nk data from the 2D grid and then run the SNESIM simulation conditioning to the nk new samples. The simulated realization in grid Gk can be post-processed to remove the simulation noise. Finally, copy all temporary 2D simulations from grids Gk (k=1,2, . . . Nz) to the 3D grid G to constitute a full 3D realization.
Three methods are presented for data sampling: point sampling, geo-body sampling and hybrid sampling. The point sampling method, shown in
The geobody sampling method, shown in
The hybrid method, shown in
As previously discussed, the conditioning can either be well hard data or region data. Because well data allows for data relocation, the point sampling method sets the sample as well data for better conditioning. However, data relocation is less important with sampled geo-bodies, because multiple cells (with the same value) from a single geo-body may be relocated to the same simulation node. Hence, the geo-body sampling method relies on the region concept to provide conditioning data.
Point sampling works well to supply the sparse hard conditioning data to maintain the vertical connectivity's with reasonable vertical variations. However, with geo-body sampling the sampled hard data will be clustered as a set of connected geo-objects, which are normally dense. One concern with geobody sampling is whether or not there will be enough variations between two successive layers.
a) is a 2D two facies categorical training image representing fluvial channels;
Due to the current design of mps algorithms, it is difficult to reproduce long range channel continuities even with a very large search template for 3D mps simulations. However,
The 2D two facies channelized training image shown in
One final 3D realization using the method sown in
A three facies training image (of size 200×200) is shown in
A four facies training image in
One final 3D realization using the method is depicted in
G=3D grid
N=cell dimension
k=layer index
n=number of sample
r=regular grid
int=interval
w=windows
GB=geobody
Z=binary map
f=facies
In closing, it should be noted that the discussion of any reference is not an admission that it is prior art to the present invention, especially any reference that may have a publication date after the priority date of this application. At the same time, each and every claim below is hereby incorporated into this detailed description or specification as an additional embodiment of the present invention.
Although the systems and processes described herein have been described in detail, it should be understood that various changes, substitutions, and alterations can be made without departing from the spirit and scope of the invention as defined by the following claims. Those skilled in the art may be able to study the preferred embodiments and identify other ways to practice the invention that are not exactly as described herein. It is the intent of the inventors that variations and equivalents of the invention are within the scope of the claims while the description, abstract and drawings are not to be used to limit the scope of the invention. The invention is specifically intended to be as broad as the claims below and their equivalents.
The discussion of any reference is not an admission that it is prior art to the present invention, especially any reference that may have a publication data after the priority date of this application.
This application is a non-provisional application which claims benefit under 35 USC §119(e) to U.S. Provisional Application Ser. No. 61/716,050 filed Oct. 19, 2012, entitled “METHOD FOR MODELING A RESERVOIR USING 3D MULTIPLE-POINT SIMULATIONS WITH 2D TRAINING IMAGES,” which is incorporated herein in its entirety.
Number | Date | Country | |
---|---|---|---|
61716050 | Oct 2012 | US |