CN104573333A - Method for optimizing of model selection based on clustering analysis - Google Patents

Abstract

The invention discloses a method for optimizing of model selection based on clustering analysis. The method comprises the following steps of 1, establishing a plurality of three-dimensional quantitative geological models, calculating the attribute value of each grid node of each model, standardizing the attribute value, calculating the Euclidean distance between any two models, and obtaining a dissimilarity matrix for representing the difference of each model; 2, reducing the dimension of the dissimilarity matrix, and distinguishing the similarity of the model by vectors in a two-dimensional space; 3, clustering the model by a clustering analysis method, and selecting one or a plurality of models from each type for the research of numerical reservoir simulation; 4, comparing a well point attribute value histogram and a histogram of the selected model, and judging whether the selected model meets the geological concept or not; 5, comparing the reservoir calculated by the selected model and the P10, P50 and P90 reservoirs, and judging the representativeness of the selected model. The method has the characteristics that the simplicity and objectivity are realized, the application range is wide, and the method can be applied to the field of reservoir description.

Description

Cluster analysis is utilized to carry out the preferred method of model

Technical field

The present invention relates to reservoir description field, particularly relate to one and utilize cluster analysis to carry out the preferred method of model.

Background technology

Reservoir Stochastic Modeling technology results from phase early 1980s, and the application at present in oilfield prospecting developing practice is more and more extensive.Stochastic modeling can provide multiple equiprobable model realization, utilizes these realizations can carry out uncertainty assessment to reservoir.And in numerical reservoir simulation, consider and assess the cost, be usually merely able to carry out analog computation to limited several realizations, therefore must be optimized one or several from multiple model and carry out numerical simulation study.Conventional probabilistic model screening technique comprises arithmetic mean method, Geological Mode screening method, Method for Numerical, probabilistic reserves method, experimental design, Latin Hypercube Sampling and ranking method.Arithmetic mean ratio juris is that arithmetic mean is carried out in multiple realization, and using the averaging model that obtains as optimization model, its advantage is simple and fast, and shortcoming has smoothing effect, changes the Statistical Distribution Characteristics of reservoir heterogeneity and model; Geological Mode screening ratio juris is difference by contrasting between each model and Geological Mode, and therefrom select the model that degree of agreement is larger, the model obtained can meet geology conceptual schema preferably, but very consuming time and affect larger by subjective factor; Method for Numerical is by the method such as streamline simulation, history matching optimization model, and shortcoming is also consuming time many; And probabilistic reserves method, experimental design, Latin Hypercube Sampling and ranking method are all that to take geologic reserve as index preferred to carry out model, be not all suitable for the preferred of penetration rate model and phase model.

In view of the problems referred to above that existing model optimization techniques exists, urgently work out more simple, the objective model optimization techniques of one, be difficult to therefrom optimize the difficult problem in the industry that representative model carries out numerical reservoir simulation because of model One's name is legion to solve.

Summary of the invention

The object of the invention is the deficiency in order to overcome above-mentioned background technology, providing one to utilize cluster analysis to carry out the preferred method of model, there is feature simply objective and applied widely.

One provided by the invention utilizes cluster analysis to carry out the preferred method of model, comprise the following steps: step one, utilize stochastic modeling method to set up multiple 3-D quantitative geologic model, add up the property value of each grid node in each model and standardization is carried out to described property value, calculate the Euclidean distance between any two models, obtain the Dissimilarity matrix that characterizes difference between each model; Step 2, dimensionality reduction is carried out to obtained Dissimilarity matrix, realize the similarity distinguishing model in 2 dimension spaces with vector; Step 3, utilization clustering method carry out cluster to model, select one or several model and carry out numerical reservoir simulation research from each class; Step 4, contrast well point property value histogram and selected model histogram, judge whether selected model meets geologic concepts; Step 5, reserves that selected model is calculated with contrast with P10, P50 and P90 reserves that Monte Carlo Analogue Method obtains, judge that whether selected model representative.

In technique scheme, described step 3 comprises following process: 1) write Model tying code; 2) Cluster Validity Index is utilized to evaluate Clustering Effect.

In technique scheme, described step 3 the 1st) in item, with R language compilation Model tying code, and utilize k-means clustering algorithm to carry out cluster to model.

In technique scheme, described step 3 the 2nd) in item, adopt the validity of Dunn metrics evaluation cluster result.

In technique scheme, described step 3 the 2nd) in item, described Dunn index calculate formula is as follows: in formula, D (k) represents Dunn index value, and ci represents the i-th class data, and cj represents jth class data, and ck represents kth class data, and d (x, y) represents the distance between two data points.

The present invention utilizes cluster analysis to carry out the preferred method of model, there is following beneficial effect: with the difference between each model for foundation is classified to model, not only simple and quick, and more effective, be applicable to the preferred of each class models such as sedimentary facies, factor of porosity, permeability, water saturation.Because the model selected based on this method for optimizing has certain representativeness, the number of times of numerical simulation can be greatly reduced, efficiently solve and be difficult to therefrom to optimize representative model because of model One's name is legion and carry out this difficult problem of numerical reservoir simulation.

Accompanying drawing explanation

Fig. 1 is that the present invention utilizes cluster analysis to carry out the schematic flow sheet of the preferred method of model;

Fig. 2 is that the present invention utilizes cluster analysis to carry out step 3 the 1st in the preferred method of model) result schematic diagram after item cluster;

Fig. 3 is that the present invention utilizes cluster analysis to carry out step 4 well point property value and selected model Nogata comparison diagram in the preferred method of model;

Fig. 4 utilizes cluster analysis to carry out in the preferred method of model the comparison diagram of model selected by the step 5 reserves calculated and P10, P50 and P90 reserves obtained by Monte Carlo Analogue Method for the present invention.

Embodiment

Below in conjunction with drawings and Examples, the present invention is described in further detail, but this embodiment should not be construed as limitation of the present invention.

See Fig. 1, the present invention utilizes cluster analysis to carry out the preferred method of model, comprises the following steps:

Step one, stochastic modeling method is utilized to set up multiple 3-D quantitative geologic model, add up the property value of each grid node in each model and standardization is carried out to described property value, calculate the Euclidean distance between any two models, obtain the Dissimilarity matrix that characterizes difference between each model;

Step 2, dimensionality reduction is carried out to obtained Dissimilarity matrix, realize the similarity distinguishing model in 2 dimension spaces with vector;

Step 3, utilization clustering method carry out cluster to model, and from each class, select one or several model carry out numerical reservoir simulation research, this step comprises again following process:

1) write Model tying code, namely use R language compilation Model tying code, and utilize k-means clustering algorithm to carry out cluster to model;

2) utilize Cluster Validity Index to evaluate Clustering Effect, namely adopt the validity of Dunn metrics evaluation cluster result, described Dunn index calculate formula is as follows:

$D\left(k\right)=\frac{{\min }_{x\∈\mathrm{ci},y\∈\mathrm{cj}}d(x,y)}{\max \left(d(x,y)_{x,y\∈\mathrm{ck}}\right)},$

In formula, D (k) represents Dunn index value,

Ci represents the i-th class data,

Cj represents jth class data,

Ck represents kth class data,

D (x, y) represents the distance between two data points;

Step 4, contrast well point property value histogram and selected model histogram, judge whether selected model meets geologic concepts;

Step 5, reserves that selected model is calculated with contrast with P10, P50 and P90 reserves that Monte Carlo Analogue Method obtains, judge that whether selected model representative.

Below using the experimental data from the two sections of reservoirs in angle, West, WZ oil field as embodiment, research object is 50 penetration rate models that this area's sequential Gaussian simulation method is set up.

First stochastic simulation generates 50 penetration rate models, add up the penetration value certificate of each grid node in each model, reduced data is imported in SPSS (Statistical Productand Service Solutions, the data processing software of Chinese " statistical product and service solution " by name).Carry out standardization to data, then calculate the distance (difference namely between model) between each model with Euclidean distance function, computing formula is:

$\mathrm{dij}=\sqrt{{(x_{i1}-x_{j1})}^2+{(x_{i2}-x_{j2})}^2+...{(x_{\mathrm{in}}-x_{\mathrm{jn}})}^2}$

In formula, dij represents the distance between each model,

I and j represents two models respectively,

X _inand x _jnmodel respectively _iand model _jin the property value size of the n-th grid,

N is meshes number.

Table 1 is the Dissimilarity matrix (specifically seeing the following form) of difference between each model of sign of calculating, data in this matrix are first obtained by the quadratic sum, the then root of making even that calculate the permeability value difference of each grid node corresponding between any two models, and represented is the cumulative of the difference of every bit place permeability analog result.And in the geological research of reality, the size of each local permeability directly can affect the flow direction of fluid (namely permeability value shows that more greatly connectedness is better).So more accurate by the difference between the method expression penetration rate model, the model selected by this difference is also more representative.

Dissimilarity matrix between each model of table 1

Model	1	2	3	4	5	6	7	8	9	10
											1	0	16	97	13	153	151	123	117	151	115
2	16	0	124	121	152	150	119	109	149	129
											3	97	124	0	109	152	150	119	114	150	105
4	113	121	109	0	152	152	111	107	150	118
											5	153	152	152	152	0	154	153	152	154	153
6	151	150	150	150	152	0	150	151	153	150
											7	123	119	119	119	111	153	0	103	148	11
8	117	109	114	114	107	152	103	0	150	116
											9	151	149	150	150	150	154	148	150	0	149
10	115	129	105	105	118	153	111	116	149	0

Because higher-dimension can make the distinguishing limit between data thicken, bring difficulty to cluster, so all dimension-reduction treatment can be carried out to data before carrying out cluster analysis, and then realize the similarity distinguishing object in 2 dimension spaces with vector, thus simplify cluster process.

Table 2 utilizes MDS (Multidimensional Scaling, Chinese " multi-dimension analysis ", the module be contained in SPSS by name) technology to extract the coordinate figure of each model in a dimension and two-dimensions (specifically seeing the following form).Wherein, the dimension that on the low side-high level that a dimension refers to permeability high level is on the high side, the dimension that on the high side-high level that two-dimensions refers to permeability high level is on the low side.When carrying out cluster analysis, can directly with this two column data.

The coordinate figure of each model of table 2 in a dimension and two-dimensions

Utilize the coordinate figure of each model in a dimension and two-dimensions, use R language (a kind of language for statistical study, drawing and operating environment, belong to the software of a freedom of GNU and general public licence (General Pubilic License) system, free, open source) write correlative code and carry out Model tying research, and apply k-means clustering algorithm cluster carried out to model: the number k that first will determine cluster, due to k value size should 2 with between, wherein _nfor the number of all data points in data space, so k value size is here between 2 to 7, then chooses k value one by one in [2,7] interval and carry out cluster.

Fig. 2 is the result after cluster, comprises each cluster diagram from k=2 to k=7.Wherein, in figure, " Δ " represents cluster centre, circular expression model, and in figure, same class model is drawn a circle to approve together, the dimension that on the low side-high level that its transverse dimensions corresponds to permeability high level is on the high side and one dimension, the dimension i.e. two dimension that on the high side-high level that longitudinal dimension corresponds to permeability high level is on the low side.

Its classification results of effective cluster analysis should to reach in class closely, between class away from.Adopt the validity of Dunn metrics evaluation cluster result.This index uses the maximum distance between class and class to represent similar degree in the class, and use the minor increment between class and class to represent class inherited, the value of Dunn index is the business of the two simultaneously.Between the larger representation class of value of Dunn index and class, interval is far away, and Clustering Effect is better, and its computing formula is as follows:

$D\left(k\right)=\frac{{\min }_{x\∈\mathrm{ci},y\∈\mathrm{cj}}d(x,y)}{\max \left(d(x,y)_{x,y\∈\mathrm{ck}}\right)},$

In formula, D (k) represents Dunn index value,

Ci represents the i-th class data,

Cj represents jth class data,

Ck represents kth class data,

D (x, y) represents the distance between two data points;

Table 3 is k D values (specifically seeing the following form) when getting different value.Drawn by this table analysis, Clustering Effect when cluster numbers is 5 is best.

k	2	3	4	5	6	7
							D	0.765	0.8134	0.8457	0.8611	0.8588	0.8454

Utilize k-means clustering method, 50 original models have been divided into 5 large class (not shown)s.Wherein, in first kind model, be model 15 from the model that cluster centre is nearest; In Equations of The Second Kind model, be model 49 from the model that cluster centre is nearest; In 4th class model, be model 11 from the model that cluster centre is nearest; 3rd class and the 5th class all only have a model, are respectively model 2 and model 34.

Fig. 3 is the Nogata comparison diagram (in figure PERM and permeability, permeability) of the property value of well point permeability and the property value of selected model and model 15.Wherein, Fig. 3 .a is the histogram of well point permeability properties value, and Fig. 3 .b is the histogram of model 15 property value.Can find out that the result that model 15 is simulated and geologic concepts exist good consistance according to this figure.

Utilize cluster analysis to carry out the preferred method of model with the present invention and optimize corresponding factor of porosity, water saturation and NTG (Net-To-Gross ratio, net-gross ratio) model respectively.

Fig. 4 is the reserves cumulative probability distribution plan calculated with 50 models, P10 (optimistic geologic reserve), P50 (most probable geologic reserve) and P90 (conservative geologic reserve) reserves that in figure, rectangle expression Monte Carlo Analogue Method obtains, triangle represents the reserves calculated with selected 5 group models.The model that this figure explanation the present invention utilizes cluster analysis to carry out selected by the preferred method of model has good representativeness, namely originally need to do 50 simulations just getable result, only need do 5 simulations now just can obtain, greatly reduce the number of times of simulation, both simple and convenient, turn improve efficiency, the analog operation for numerical reservoir provides a kind of new thinking.

Obviously, those skilled in the art can carry out various change and modification to the present invention and not depart from the spirit and scope of the present invention.Like this, if these amendments of the present invention and modification belong within the scope of the claims in the present invention and equivalent technologies thereof, then the present invention is also intended to comprise these change and modification.

The content be not described in detail in this instructions belongs to the known prior art of professional and technical personnel in the field.

Claims (5)

1. utilize cluster analysis to carry out the preferred method of model, it is characterized in that: comprise the following steps:

Step one, stochastic modeling method is utilized to set up multiple 3-D quantitative geologic model, add up the property value of each grid node in each model and standardization is carried out to described property value, calculate the Euclidean distance between any two models, obtain the Dissimilarity matrix that characterizes difference between each model;

Step 2, dimensionality reduction is carried out to obtained Dissimilarity matrix, realize the similarity distinguishing model in 2 dimension spaces with vector;

Step 3, utilization clustering method carry out cluster to model, select one or several model and carry out numerical reservoir simulation research from each class;

Step 4, contrast well point property value histogram and selected model histogram, judge whether selected model meets geologic concepts;

Step 5, reserves that selected model is calculated with contrast with P10, P50 and P90 reserves that Monte Carlo Analogue Method obtains, judge that whether selected model representative.

2. according to claim 1ly utilize cluster analysis to carry out the preferred method of model, it is characterized in that: described step 3 comprises following process:

1) Model tying code is write;

2) Cluster Validity Index is utilized to evaluate Clustering Effect.

3. according to claim 2ly utilize cluster analysis to carry out the preferred method of model, it is characterized in that: described step 3 the 1st) in item, with R language compilation Model tying code, and utilize k-means clustering algorithm to carry out cluster to model.

4. according to claim 2ly utilize cluster analysis to carry out the preferred method of model, it is characterized in that: described step 3 the 2nd) in item, adopt the validity of Dunn metrics evaluation cluster result.

5. according to claim 4ly utilizing cluster analysis to carry out the preferred method of model, it is characterized in that: described step 3 the 2nd) in item, described Dunn index calculate formula is as follows:

$D\left(k\right)=\frac{\min d(x,y)_{x\∈\mathrm{ci},y\∈\mathrm{cj}}}{\max \left(d(x,y)_{x,y\∈\mathrm{ck}}\right)},$

In formula, D (k) represents Dunn index value,

Ci represents the i-th class data,

Cj represents jth class data,

Ck represents kth class data,

D (x, y) represents the distance between two data points.