CN106294540A - Multiple spot geological statistics modeling method based on p stable local sensitivity Hash retrieval Data Styles - Google Patents

Abstract

The invention discloses a kind of multiple spot geological statistics modeling method based on p stable local sensitivity Hash retrieval Data Styles, use multi-block technique framework to calculate the characteristic vector of Data Styles (event), then use p stable local sensitivity hash method that these characteristic vectors are hashed multiple Hash table.Extract the characteristic vector of the data event at waiting to estimate some during modeling and carry out p stable local sensitivity Hash calculation, obtain data event Hash barrel number, take out all of Data Styles in these Hash buckets and build target patterns data base, relatively data event and the manhatton distance of all Data Styles in target patterns data base, obtain most like data event.And SIMPAT algorithm is in simulation process, need to will wait to estimate at data event carry out Similarity measures with all Data Styles in pattern database.Comparing SIMPAT algorithm, present invention application local sensitivity Hash retrieval technique drastically increases the efficiency of Multiple-Point Geostatistics modeling algorithm.

Description

Multiple spot geological statistics based on p-stable local sensitivity Hash retrieval Data Styles is built Mould method

Technical field

The present invention relates to reservoir geologic modeling technical field, examine based on p-stable local sensitivity Hash in particular to one The multiple spot geological statistics modeling method (LSHSIM, Local Sensitive Hashing Simulation) of rope Data Styles.

Background technology

Based on Multiple-Point Geostatistics modeling algorithm, oil field Efficient Development had important supporting function.Arpat is 2003 In year Stanford Forecast Oil Reservoir Distribution center meeting, Multiple-Point Geostatistics algorithm SIMPAT based on pattern, algorithm SIMPAT are proposed Image reconstruction thought is incorporated in reservoir geologic modeling, and core concept is between coupling " Data Styles " and " data event " Similarity, detailed process is first to mate in priori geology library (i.e. pattern database) (to count with simulation Dai Gu region, work area According to event) the maximum reservoir pattern (Data Styles) of similarity measurement；Then this reservoir pattern (Data Styles) is covered and replace Simulation Dai Gu region, work area (data event).

Along with deepening continuously of oil field development, the precision of geological research is more and more higher, thus the yardstick of geological model more comes The least, the old filed of such as a lot of mid-later development phase, it is 10m × 10m that the precision of Geologic modeling reaches planar mesh, longitudinally 0.5m is the least.Current commonly used stochastic simulation technology sets up geological model, it usually needs set up multiple realization (such as 100 Individual), even four dimension module, therefore, the efficiency of Geologic modeling is increasingly becoming the hot issue of concern.Calculating focus is program generation Code need to take considerable time the part of operation, it is common that calculate performance bottleneck, the calculating focus of algorithm SIMPAT mainly around in The Similarity measures of all patterns of pattern database of matched data event and vast number, calculated performance becomes universal based on sample The bottleneck of formula modeling algorithm.Limit the application of algorithm SIMPAT.

Therefore, it is badly in need of studying a kind of speed-up computation improved method for SIMPAT algorithm, it is thus possible to efficiently based on sample Formula matching strategy carries out multiple spot geological statistics modeling.

Summary of the invention

Present invention aim at providing a kind of can improve based on pattern multiple spot geological statistics modeling algorithm computational efficiency Method (i.e. based on p-stable local sensitivity Hash retrieval Data Styles multiple spot geological statistics modeling method).Compare SIMPAT Algorithm, this algorithm is greatly improved the efficiency of Multiple-Point Geostatistics modeling algorithm.

LSHSIM algorithm of the present invention retrieves set of metadata of similar data pattern based on local sensitivity salted hash Salted, first adds up pattern data The multi-block technique internal variable value sum of the total data pattern in storehouse, is converted to multi-block technique the characteristic vector of Data Styles, enters And calculate the cryptographic Hash of characteristic vector based on p-stable local sensitivity hash algorithm, the data sample of identical for cryptographic Hash (similar) Formula is saved in identical Hash bucket, constitutes Data Styles Hash storehouse.First calculate based on p-stable local sensitivity Hash during simulation Method calculates the cryptographic Hash of data event, then retrieves the Data Styles of identical Hash bucket from Data Styles Hash storehouse, then calculates Relatively data event and the similarity of these Data Styles, finally find out most like Data Styles.The present invention is based on Hash Algorithm searches most like Data Styles, therefore greatly improves the efficiency of simulation.

For achieving the above object, what the present invention provided is a kind of based on p-stable local sensitivity Hash retrieval Data Styles Multiple spot geological statistics modeling method, comprises the following steps:

1) input training image TI, defines the size of simulated implementation R, the size of data model T；

2) arranging the size of multi-block technique BlockGrid, the parameter of input p-stable local sensitivity Hash, including Hash The wide w of bucket, Hash table quantity N；

3) scan training image TI with data model T, set up pattern database PatDB；

4) add up the multi-block technique block internal variable sum of total data pattern, obtain the characteristic vector of Data Styles, then Carry out p-stable local sensitivity Hash calculation, obtain Data Styles Hash storehouse PatLSHLib；

5) random walk is created according to simulated implementation R；

6) if there being non-analog node U in random walk, enter following step 7) path；

Otherwise enter following step 11) path；

7) the data event DataEvent at node U is extracted, the multi-block technique block internal variable sum of statistical data event, Obtain the characteristic vector of data event, carry out p-stable local sensitivity Hash calculation, obtain the cryptographic Hash of data event DevLSH；

8) from Data Styles Hash storehouse PatLSHLib, retrieval is identical with data event cryptographic Hash DevLSH Hash barrel number Data Styles, constitutes target data style library ANNPatDB；

9) the Data Styles Pat maximum with data event Dev Similarity value is searched from target data style library ANNPatDB；

10) cover by Data Styles Pat entirety and freeze in simulated implementation R the part at node U；Return above-mentioned steps 6)；

11) simulation terminates, and inputs simulated implementation R.

Preferably, described step 4) in, carry out piecemeal according to the Data Styles i.e. size of mesh opening of event, set up and divide Block grid, all nodal value sums in adding up each piece, formula is

Wherein, ICount, JCount are training image grid number of grids in both the horizontal and vertical directions, MCount, NCount are multi-block technique number of grids in both the horizontal and vertical directions, BlockGrid_m,nBe index [m, N] piecemeal in all nodal value sums, TIGrid_i,jBeing the nodal value indexing [i, j] in training image grid, sum represents and asks And calculating.

Preferably, described step 4) in, multi-block technique data being converted to after linear data structure just can be as The data input of p-stable local sensitivity hash algorithm, linear data structure index and the two dimensional data structure rope of multi-block technique The relation of drawing is

BlockGrid_i=BlockGrid_{M, n}, wherein i=m*NCount+n

Wherein i is the linear directory of multi-block technique, m and n is the 2-d index of multi-block technique.Table 1 is 4 data in Fig. 2 The multi-block technique statistical value of pattern (event), belongs to linear directory data structure.

Table 1 calculates the characteristic vector of Data Styles (event) based on Block Grid

Preferably, described step 4) in, there is the node not having variate-value in Data Styles, the change of training image Amount minima participates in statistical computation as dummy variable, and formula is

Wherein, TI is training image, min_TIBeing the variable minima of training image, var is the change of node in data event Value, if certain node of data event is empty (NULL), then replaces null value with the variable minima of training image and participates in calculating.

The algorithm core concept of the present invention

The present invention calculates the characteristic vector of Data Styles (event) based on multi-block technique framework, then uses p-stable office These characteristic vectors are hashed in multiple Hash table by portion's sensitive hash method.According to the definition of local sensitivity Hash, same In individual Hash table, similar Data Styles (event) is in the probability of same Hash bucket and is far longer than dissimilar Data Styles (event).During modeling, use identical parameters that characteristic vector and the p-stable local sensitivity Hash meter extracting data event is set Calculate its Hash barrel number in Hash table, take out all of Data Styles in these Hash buckets and be referred to as similar closest pattern Data base.Relatively data and the manhatton distance of all Data Styles in similar closest pattern database, selects Manhattan The Data Styles entirety covering data event that distance is minimum, completes this simulation.

The definition of local sensitivity Hash is that (D is space length tolerance, and P represents if a family of functions meets following condition Probability)

1) if distance D in space between 2 p and q (p, q)<r1, then P{h (p)==h (q) }>p1

2) if distance D in space between 2 p and q (p, q)>r2, then P{h (p)==h (q) }<p2

Under conditions of r1<r2, p1>p2 meaningful, then family of functions H is local sensitivity.

Local sensitivity hash method can make point close together be mapped to the general of same position by mapping calculation Rate is big, and the probability that distant point is mapped to same position is little.

Local sensitivity hash algorithm based on p-stable distribution utilizes the thought of p-stable to compose each characteristic vector v Give cryptographic Hash h.Owing to this hash function is local sensitivity, it is therefore assumed that two characteristic vectors v1 and v2 are very near, it The probability that will be mapped in identical bucket of cryptographic Hash the biggest.Hash function based on p-stable distribution is defined as

h_{A, b}(v): R^d→N (1)

Map characteristic vector v to the set of integers of a d dimension.

Having two stochastic variable a and b in hash function, wherein a is the random vector of a d dimension, and Euclidean distance is quantitative Characterizing the distance variable of the spacing of two points in d dimension space, the local sensitivity hash function in Euclidean distance is defined as

Wherein, a is an independent random vector obeying p-stable distribution, and w is that bucket is wide, b be in the range of [0, w] with Machine number.Fig. 1 has illustrated principle based on p-stable local sensitivity hash algorithm by one: have n=5 on two dimensional surface Individual yellow dots.Inquiry and the hithermost yellow dots of blue dot, traditional directory method is the Europe calculating blue dot with all yellow dots Formula distance is also ranked up, and takes the yellow dots that distance value is minimum, calculating time complexity O (n) of traditional directory method.Based on p- The local sensitivity hash algorithm of stable carries out projection according to given random vector a to all yellow dots and blue dot and calculates, Assuming that 3 different axis of projection X1, X2 and X3, each axis of projection is divided into multiple Hash bucket according to the wide w of bucket, and all of point is through projection After fall in different buckets, inquire about the yellow dots identical with blue dot barrel number.Example is positioned at the Huang of identical Hash bucket with blue dot Color dot has the point 1 in axis of projection X2, the point 2 in axis of projection X3, calculates all yellow dots in blue dot same Hash bucket Euclidean distance, takes the minimum point of distance value as Query Result.Contrast query time, traditional directory time complexity is O (5), Query time complexity O (2) based on p-stable local sensitivity Hash, improves search efficiency.

The present invention carries out piecemeal (Fig. 2) according to the size of mesh opening of Data Styles (event), sets up multi-block technique (Block Grid), all nodal value sums in adding up each piece (Block), formula is

Wherein ICount, JCount are training image grid number of grids in both the horizontal and vertical directions, MCount, NCount are multi-block technique number of grids in both the horizontal and vertical directions, BlockGrid_m,nBe index [m, N] piecemeal in all nodal value sums, TIGrid_i,jBeing the nodal value indexing [i, j] in training image grid, sum represents and asks And calculating.

Multi-block technique data being converted to just can be as p-stable local sensitivity hash algorithm after linear data structure Data input, and the linear data structure index of multi-block technique with two dimensional data structure index relative is

BlockGrid_i=BlockGrid_{M, n}, wherein i=m*NCount+n (4)

Wherein i is the linear directory of multi-block technique, m and n is the 2-d index of multi-block technique.Table 1 is 4 data in Fig. 2 The multi-block technique statistical value of pattern (event), belongs to linear directory data structure.

The Data Styles [a] of Fig. 1 is closely similar with the geometric shape of Data Styles [b], corresponding multi-block technique in respective table 1 Characteristic vector the most close.Otherwise Data Styles [a] is respectively provided with bigger difference with shape and the characteristic vector of Data Styles [c] Not.During simulation, needing to calculate the characteristic vector of data event, data event there may be the node not having variate-value.Now, The variable minima of training image participates in statistical computation as dummy variable, and formula is

The beneficial effects of the present invention is:

The present invention introduces local sensitivity Hash retrieval technique and models to multiple spot geological statistics, differentiates when greatly improving modeling Inquiry, for the computational efficiency of the Data Styles of replacement data event, enhances the practicality of multiple spot geological statistics modeling method.

Accompanying drawing explanation

Fig. 1 is searching principle schematic diagram based on p-stable local sensitivity hash algorithm；

Fig. 2 is the characteristic vector principle calculating Data Styles (event) based on Block Grid；

Fig. 3 is the flow chart of algorithm LSHSIM；

Fig. 4 is algorithm LSHSIM modeled example-type of sedimentary facies variable (classified variable),

In figure, a is training image, and b is simulated implementation figure；

Fig. 5 is algorithm LSHSIM modeled example-porosity categorical variable (continuous variable),

In figure, a is training image, and b is simulated implementation.

Detailed description of the invention

In order to preferably explain the present invention, it is further elucidated with the main contents of the present invention below in conjunction with specific embodiment, but Present disclosure is not limited solely to following example.

Before Multiple-Point Geostatistics modeling method LSHSIM based on local sensitivity Hash is described, the first relevant art of definition Language:

1, grid cell (C Cell): along the X direction, there is appointment long (ISize), wide in Y-direction and Z-direction (JSize), the rectangular cubic unit of high (KSize), grid cell stores concrete numerical value and represents its attribute.

2, grid body (G Grid): the three-dimensional structure being made up of a lot of grid cell C, in X-direction, Y-direction and Z side To dimension be I × J × K, essence is three-dimensional matrice.(i, j k) are meant that X-direction index is equal to equal to i, Y-direction index to G J, the Z-direction index grid cell equal to k.

3, training image (TI TrainImage): priori geologic concept model, uses grid body G_TIAs data medium, It it is the digital model that can state actual reservoir structure, geometric shape and distribution pattern thereof.

4, simulated implementation (R Realization): the model result of simulation, uses grid body G_RAs data medium, it is The digital model of actual reservoir structure, geometric shape and distribution pattern thereof can be stated.

5, data model (T Template): around center cell symmetrical structure body, uses grid body G_TAs data Carrier, is the local digital model that can state actual reservoir structure, combining form and distribution pattern thereof, based on pattern The basic structural unit of Multiple-Point Geostatistics modeling method.

6, Data Styles (Pat Pattern): the local space number obtained for framework scanning training image with data model According to assembly, use grid body G_PatAs data medium.

7, data event (Dev DataEvent): the local space obtained for framework scanning simulated implementation with data model Data set is fit, uses grid body G_DevAs data medium.

6, pattern database (PatDB Pattern Database): use data model scanning training image to obtain Take all Data Styles of this training image, referred to as pattern database.

7, block grid (BlockGrid Block Grid): calculate the block of Data Styles (event) based on piecemeal framework Interior all variate-value sums, are the characteristic vectors of Data Styles (event), are the inputs of p-stable local sensitivity Hash calculation Data, are the core information storehouses of LSHSIM modeling algorithm.

8, Data Styles Hash storehouse (PatLSHLib Pattern LSH Library): based on p-stable local sensitivity Salted hash Salted calculates the cryptographic Hash of Data Styles characteristic vector, and the Hash table of composition is referred to as Data Styles Hash storehouse.

9, similar neighbouring pattern database (ANNPatDB Approximate Nearest Neighbors Pattern Database): in modeling process, from Data Styles Hash storehouse, the Data Styles group identical with data event cryptographic Hash is retrieved The pattern database become.

Embodiment 1

Multiple spot geological statistics modeling method based on p-stable local sensitivity Hash retrieval Data Styles, is sedimentary facies The modeling of categorical variable, specifically comprises the following steps that

(1) with Fig. 4 a for training image TI, training image illustrates a sedimentary facies model, wherein comprises two kinds of depositions Microfacies, black for river course microfacies, white be interchannel gulf microfacies, size and the training image size of definition simulated implementation R Unanimously, data model T a size of 15 × 15；

(2) arrange multi-block technique BlockGrid a size of 5 × 5, then the characteristic vector dimension of training image is equal to 25, defeated Entering p-stable local sensitivity Hash parameter, wherein the wide w of Hash bucket is equal to 0.01, and Hash table quantity is equal to the dimension of characteristic vector 25；

(3) scan training image TI with data model T, set up pattern database PatDB；

(4) the multi-block technique internal variable sum of all Data Styles in statistics pattern database PatDB, obtains training image Characteristic vector, as input variable participate in p-stable local sensitivity Hash calculation, set up Data Styles Hash storehouse PatLSHLib；

(5) random walk is created according to simulated implementation R；

(6) if there being the node U not simulated in random walk, then step 7 is entered), otherwise enter step 11)；

(7) the data event DataEvent at node U place, employing step 4 are extracted) identical parametric statistics data event Characteristic vector, obtains cryptographic Hash DevLSH of data event；

(8) from all Data Styles that Data Styles Hash storehouse PatLSHLib retrieval is identical with DevLSH Hash barrel number, group Become target patterns data base ANNPatDB；

(9) from ANNPatDB, the Data Styles Pat maximum with data event Dev Similarity value is searched.(10) data are used Pattern Pat entirety covers and freezes in simulated implementation R the part at node U, returns step 6)；

(11) simulation terminates, and exports simulated implementation R.Fig. 4 b is that a stochastic simulation realizes.

Embodiment 2

The present embodiment is the modeling method of porosity categorical variable, and the operating procedure of the method is substantially the same manner as Example 1, Difference is:

The training image of input is Fig. 5 a, and Fig. 5 b is that a stochastic simulation of porosity categorical variable realizes.

Fig. 4 and Fig. 5 respectively show the inventive method in simulation phase model and the application of physical property model.Phase model belongs to Classified variable, Fig. 4 comprises two kinds of facies types, and Fig. 4 a is training image fluvial facies model, and white and black represent river respectively Road phase and interchannel gulf, Fig. 4 b is that a stochastic simulation realizes, and observes simulated implementation, and spatial distribution and the seriality in river course obtain Preferably reproduce.Fig. 5 is the example simulating continuous categorical variable, and Fig. 5 a is the porosity model that Fig. 4 fluvial facies model is corresponding, Fig. 5 b Being simulated implementation, the distribution of porosity feature of simulated implementation meets with training image very much.

Take in terms of two, comprehensively analyze the present invention advantage relative to traditional method from simulated time and internal memory (table 2) Part.Fig. 4 a is training image, and its dimension is 250 × 250, grid cell a size of 10m × 10m, and the dimension of data model sets It is set to 15 × 15,200 × 200 and 500 × 500 two dimensions of the selection of simulated implementation.See the calculating time of table 2, the present invention The calculating time well below the tradition multiple spot geological statistics algorithm such as SIMPAT and Filtersim, compare and improved later DisPat and PSCSIM algorithm also has bigger advantage.In terms of EMS memory occupation, new method committed memory 36MB, control well Internal memory is taken by Hash table processed.Consider time-consuming and internal memory, present invention greatly enhances computational efficiency.

Table 2LSHSIM calculates time-consuming and EMS memory occupation with other multiple spot geological statistics algorithms

Other unspecified part is prior art.Although above-described embodiment is made that detailed retouching to the present invention State, but its a part of embodiment that is only the present invention rather than all embodiment, people can also according to the present embodiment without Obtaining other embodiments under creative premise, these embodiments broadly fall into scope.

Claims (4)

1. a multiple spot geological statistics modeling method based on p-stable local sensitivity Hash retrieval Data Styles, its feature exists In, comprise the following steps:

1) input training image TI, defines the size of simulated implementation R, the size of data model T；

2) size of multi-block technique BlockGrid is set, inputs the parameter of p-stable local sensitivity Hash, wide including Hash bucket W, Hash table quantity N；

3) scan training image TI with data model T, set up pattern database PatDB；

4) add up the multi-block technique block internal variable sum of total data pattern, obtain the characteristic vector of Data Styles, then carry out P-stable local sensitivity Hash calculation, obtains Data Styles Hash storehouse PatLSHLib；

5) random walk is created according to simulated implementation R；

6) if there being non-analog node U in random walk, enter following step 7) path；

Otherwise enter following step 11) path；

7) extract the data event DataEvent at node U, the multi-block technique block internal variable sum of statistical data event, obtain The characteristic vector of data event, carries out p-stable local sensitivity Hash calculation, obtains cryptographic Hash DevLSH of data event；

8) from Data Styles Hash storehouse PatLSHLib, the data identical with data event cryptographic Hash DevLSH Hash barrel number are retrieved Pattern, constitutes target data style library ANNPatDB；

9) the Data Styles Pat maximum with data event Dev Similarity value is searched from target data style library ANNPatDB；

10) cover by Data Styles Pat entirety and freeze in simulated implementation R the part at node U；Return above-mentioned steps 6)；

11) simulation terminates, and inputs simulated implementation R.

Multiple spot geological statistics modeling based on p-stable local sensitivity Hash retrieval Data Styles the most according to claim 1 Method, it is characterised in that: described step 4) in, carry out piecemeal according to the Data Styles i.e. size of mesh opening of event, set up piecemeal net Lattice, all nodal value sums in adding up each piece, formula is

Wherein, ICount, JCount are training image grid number of grids in both the horizontal and vertical directions, MCount, NCount is multi-block technique number of grid in both the horizontal and vertical directions, BlockGrid_m,nIt it is the piecemeal of index [m, n] Interior all nodal value sums, TIGrid_i,jBeing the nodal value indexing [i, j] in training image grid, sum represents read group total.

Multiple spot geological statistics modeling based on p-stable local sensitivity Hash retrieval Data Styles the most according to claim 1 Method, it is characterised in that: described step 4) in, multi-block technique data being converted to just can be as p-after linear data structure The data input of stable local sensitivity hash algorithm, the linear data structure index of multi-block technique indexes with two dimensional data structure Relation is

BlockGrid_i=BlockGrid_{M, n}, wherein i=m*NCount+n

Wherein i is the linear directory of multi-block technique, m and n is the 2-d index of multi-block technique.

Multiple spot geological statistics based on p-stable local sensitivity Hash retrieval Data Styles the most according to claim 1 is built Mould method, it is characterised in that: described step 4) in, there is the node not having variate-value in Data Styles, the variable of training image Minima participates in statistical computation as dummy variable, and formula is

$\mathrm{var}=\left\{\begin{array}{ccc}{\min }_{TI}& if& var==NULL\\ var& if& var!=NULL\end{array}\right.$

Wherein, TI is training image, min_TIBeing the variable minima of training image, var is the variate-value of node in data event, If certain node of data event is empty (NULL), then replaces null value with the variable minima of training image and participate in calculating.