A multi-view clustering method for oil well fracturing selection based on multi-order nearest neighbor cycle graph constraints

Abstract

The present invention relates to the technical field of oil and gas development, and particularly relates to a multi-view clustering well selection method for oil well fracturing based on multi-order nearest neighbor cycle graph constraints. The method includes obtaining dynamic and static parameters affecting the fracturing effect of oil wells and normalizing the dynamic and static parameters; dividing the multi-view data set according to the sources of the dynamic and static parameters; adding low-level feature extractors to each view; extracting high-level semantic features of each view by using the self-attention mechanism, and calculating the embedded features with information interaction in a weighted average manner; the decoder reconstructs all data, including the missing views; constructing a multi-order fusion nearest neighbor graph; calculating the loss function and updating the network parameters by using the stochastic gradient descent algorithm; using the embedded features finally obtained by the model to perform K-means clustering to complete the well selection for oil well fracturing measures. The method provided by the present invention applies deep multi-view clustering to the field of well selection for oil well fracturing measures, providing an important reference for well selection for oil well fracturing measures.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas development, and particularly relates to a multi-view clustering well selection method for oil well fracturing based on multi-order near-neighbor cycle graph constraints. Background Art

[0002] With the rapid exploitation of oil resources in China, the main oilfields discovered currently have entered a depletion period to a certain extent. To alleviate the shortage of oil and gas supply and increase the oilfield production, the use of fracturing technology helps to solve the problem of exploring the potential of remaining oil wells. Well selection for fracturing measures is an essential and indispensable link in the secondary exploitation project of oil wells. It is a systematic and comprehensive analysis and evaluation of the geological, reservoir, and rock mechanics properties of the oil and gas reservoirs to be fractured before fracturing design and construction. The results directly affect whether the reservoir stimulation measures are economical and effective. With the extension of the oilfield development time and the increase in the number of wells with repeated fracturing measures, the difficulty of well selection is increasing, and the fracturing oil production increase effect is gradually decreasing. Selecting wells with the potential for further measures from these low-production wells has become a difficult problem.

[0003] In recent years, scholars at home and abroad have proposed various well selection methods for fracturing measures. The commonly used methods include well testing method, production template curve fitting method, formula method, analogy of historical data method, etc. These methods require researchers to have sufficient background knowledge and well selection experience on reservoir properties, various parameters of oil layers and oil wells, and technical operation parameters related to fracturing technology, and a large amount of corresponding logging interpretation and historical data as a reference. To further reduce the prediction error, some oil well fracturing productivity prediction methods based on various mathematical calculations have recently emerged, including using numerical simulation technology and well selection methods based on fuzzy mathematics to evaluate the effect of fracturing wells. Both of these calculation methods make a quantitative judgment on whether a fracturing measure can be taken for an oil well by considering various factors affecting the fracturing effect. Due to the numerous factors affecting the fracturing effect, the calculation workload of the two methods is relatively large. In addition to using traditional methods to predict the productivity of oil wells, data mining technology is also widely used in the oilfield exploitation process, including using various neural network technologies to extract relevant features from a large amount of oilfield historical data, predicting the oil production or liquid production of fracturing wells after actual oil well fracturing measures, and then identifying candidate wells with fracturing potential in the reservoir.

[0004] One of the prior arts: The patent with the application number CN201710790910.8 discloses a method for selecting wells for refracturing of multi-stage fractured horizontal wells in low-permeability heterogeneous reservoirs, which successively includes the following steps: 1) Statistically analyze the key factors affecting the refracturing effect of each fracturing stage of the candidate wells, including reservoir physical property parameters, primary fracturing construction parameters, production dynamic parameters and their distribution ranges; 2) Determine the parameter values of the ideal fracturing stage with the greatest potential for refracturing according to the parameter distribution ranges of each fracturing stage of the candidate wells; 3) Calculate the Euclidean distance between each fracturing stage of the candidate wells and the ideal fracturing stage by using the method of fuzzy system clustering; 4) Calculate the average Euclidean distance between each fracturing stage of the candidate wells and the ideal fracturing stage, and optimize the refracturing wells according to the size of the average Euclidean distance.

[0005] Another prior art: The patent with the application number CN201710873415.3 discloses a method for selecting wells and fracturing stages for refracturing of shale reservoirs. Specifically, it relates to a method for selecting wells and fracturing stages for refracturing of shale reservoirs, which includes: obtaining basic well data, including gas-bearing property data of the reservoir, reservoir minerals and rock mechanics data, well data, production data after primary fracturing, and primary fracturing construction and monitoring data; determining the gas-bearing property of the candidate wells, including remaining abundance and initial abundance ratio, and determining effective candidate wells and non-candidate wells according to the ratio of the remaining abundance to the initial abundance; determining the volume ratio of production increase and fracturing stage brittleness index of the candidate well sections after primary fracturing; and determining the candidate well sections according to the primary transformation ratio and brittleness index of the candidate well sections.

[0006] In practical applications, it is found that the dynamic and static parameters affecting the fracturing effect of oil wells come from different sources. The current well selection methods simply combine these data information and do not fully utilize some potential relevant and complementary information in the attribute characteristics, ignoring their multi-view characteristics. And over time, there are many cases of missing data, and these technical solutions do not take into account the situation of missing data. They often choose to discard the missing data, resulting in low accuracy of measure well selection. Summary of the Invention

[0007] (1) Technical problems to be solved

[0008] The present invention provides a multi-view clustering oil well fracturing well selection method based on multi-order neighbor cyclic graph constraints to overcome the defects in the prior art that the situation of missing data is not comprehensively considered, and the missing data is directly selected and discarded for processing, resulting in incomplete data and low accuracy of measure well selection.

[0009] (2) Technical solutions

[0010] To solve the above problems, the present invention provides a multi-view clustering oil well fracturing well selection method based on multi-order neighbor cyclic graph constraints, including:

[0011] Step S1: Obtain the dynamic and static parameters affecting the fracturing effect of the oil well, and perform Min-Max normalization on the dynamic and static parameters;

[0012] Step S2: Based on the data normalized in Step S1, divide the multi-view dataset according to the sources of the dynamic and static parameters;

[0013] Step S3: Add a low-level feature extractor to each view divided in Step S2;

[0014] Step S4: On the basis of Step S3, use the self-attention mechanism to extract the high-level semantic features of each view, and calculate the embedded features with information interaction by means of weighted average;

[0015] Step S5: Use a decoder symmetric to the network structure in Step S4 to reconstruct all data, including the missing views;

[0016] Step S6: Supplement the missing views recovered in Step S5 to the original dataset to construct an approximately complete multi-view dataset, and construct a multi-order fusion neighbor graph according to the approximately complete multi-view data;

[0017] Step S7: Calculate the loss function by using the views reconstructed in Step S5 and the multi-order fusion neighbor graph constructed in Step S6, and update the network parameters by using the stochastic gradient descent algorithm;

[0018] Step S8: Perform K-means clustering by using the embedded features finally obtained by the model to complete the well selection for the oil well fracturing measures.

[0019] Preferably, in Step S1, the formula for performing Min-Max normalization on the dynamic and static parameters includes:

[0020]

[0021] Among them, Y represents the original dynamic and static parameters, and X represents the characteristic dynamic and static parameters after Min-Max normalization.

[0022] Preferably, Step S2 includes:

[0023] Based on the normalized data, according to the sources of the dynamic and static parameters affecting the fracturing effect of the oil well, it is divided into a fracturing parameter view X (1) , a production parameter view X (2) , a formation parameter view X (3) , and an engineering parameter view X (4) , where represents that there are m views in total, and d v represents the original feature dimension of the v-th view.

[0024] Preferably, step S3 includes: using a fully connected layer as a low-level feature extractor Wherein, represents the low-level features obtained after passing through the low-level feature extractor, and d e represents the dimension of the original features after passing through the low-level feature extractor, and θ v is the parameter of the fully connected layer Φ v .

[0025] Preferably, step S4 specifically includes: given the number of heads h of the multi-head self-attention mechanism and the low-level semantic features of the i-th sample using h groups of linear transformation parameter matrices to obtain the corresponding mappings of queries, keys, and values

[0026] dividing the entire embedded feature into h parts, each part with dimension d h = d e / h; for the t-th head of sample i, calculate its score matrix according to the following formula:

[0027]

[0028] The above formula calculates the score matrix for the complete view; for the missing view data, use the following formula to calculate the score matrix:

[0029]

[0030] where zerofill() represents filling zero values with -1e 9 and ⊙ is the Hadamard product;

[0031] To ignore the missing views in the cross-view attention score calculation, it is summarized as follows:

[0032]

[0033] where is the high-level semantic feature obtained by the t-th head after completing information interaction; therefore, the high-level semantic features of m views are Then, the high-level semantic features A of each sample are sequentially input into a linear layer, a norm layer, and a multi-layer perceptron to obtain the final embedded features of each sample For all samples That is, the multi-view information fusion encoder Adopt the method of multi-view weighted fusion to obtain the common representation shared by all views:

[0034]

[0035] where, represents the fusion matrix The i-th sample; Equation (5) is designed for the missing views and can be simply calculated using for the complete views. .

[0036] Preferably, step S5 specifically includes:

[0037] The encoder and uses the available original data to impose a partial reconstruction constraint, adding the weighted reconstruction loss L re :

[0038]

[0039] where is the reconstructed data for the i-th sample; in the second stage, the reconstruction loss function becomes:

[0040]

[0041] Preferably, step S6 specifically includes: supplementing the missing views recovered in step S5 to the original dataset to construct an approximately complete multi-view dataset.

[0042]

[0043] where W represents an indicator matrix, W i,j = 1 indicates the j-th view of the i-th sample, otherwise, W i,j = 0 indicates that the j-th view of the i-th sample is missing; constructing a multi-order fusion nearest neighbor graph based on the approximately complete multi-view data:

[0044]

[0045] where

[0046]

[0047]

[0048] where is the multi-order nearest neighbor matrix between the generated sample data, is the o-th order nearest neighbor matrix of the v-th view.

[0049] Preferably, step S7 is specifically implemented as: in the stage of recovering the missing views, the overall loss function is:

[0050] L = L re + βL rg (5)

[0051] β is the penalty parameter for balancing these two losses; after completing the missing views, the loss function is:

[0052] L' = L' re + βL rg (6)

[0053] Where:

[0054]

[0055] Where b represents the sample size of each mini - batch, is the output of the encoder corresponding to the data of each mini - batch, | k represents the k - th training round;

[0056] In the multi - order neighbor graph loss calculation, all the embedding features Z obtained in the previous training round are retained to make full use of the global structural information, and the corresponding parts are updated after the current batch processing; after completing the missing views, the graph formed after the first - stage training is fixed, i.e., G (v) | k-1 = G (v) | k .

[0057] Preferably, step S8 specifically includes: using the finally obtained fused embedding features as the clustering index matrix; performing K - means on Z, setting K to 2, and dividing the samples into two categories, one is the wells with high measure benefits and the other is the wells with low measure benefits, to obtain the final well selection result for fracturing measures.

[0058] (III) Beneficial effects

[0059] The multi - view clustering oil well fracturing well selection method based on multi - order neighbor cyclic graph constraint provided by the present invention makes full use of some relevant and complementary information latent in the dynamic and static parameters of oil well fracturing, applies deep multi - view clustering to the field of oil well fracturing measure well selection, and provides an important reference for oil well fracturing measure well selection. Brief description of the drawings

[0060] Figure 1 It is a schematic flow chart of the multi - view clustering oil well fracturing well selection method based on multi - order neighbor cyclic graph constraint in an embodiment of the present invention. Detailed implementation manners

[0061] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0062] As Figure 1 shown, the present invention provides a well selection method for oil well fracturing measures based on multi-order nearest neighbor cyclic graph constraints, which includes the following steps:

[0063] Step S1: Obtain the dynamic and static parameters affecting the fracturing effect of the oil well, and perform Min-Max normalization on the dynamic and static parameters; the original dynamic and static parameters affecting the fracturing effect of the oil well are normalized using the Min-Max method, and the specific definition is:

[0064]

[0065] where Y represents the original dynamic and static parameters, and X represents the characteristic dynamic and static parameters after Min-Max normalization.

[0066] Step S2: Based on the data normalized in Step S1, divide the multi-view data set according to the sources of the dynamic and static parameters.

[0067] Specifically, it includes: based on the normalized data, according to the sources of the dynamic and static parameters affecting the fracturing effect of the oil well, it is divided into a fracturing parameter view X (1) , a production parameter view X (2) , a formation parameter view X (3) , and an engineering parameter view X (4) , etc. Among them, represents a total of m views, and d v represents the original feature dimension of the v-th view.

[0068] Step S3: Add a low-level feature extractor to each view divided in Step S2.

[0069] Specifically, it includes: using a fully connected layer as the low-level feature extractor where represents the low-level features obtained after passing through the low-level feature extractor, d e represents the dimension of the original features after passing through the low-level feature extractor, and θ v is the parameter of the fully connected layer Φ v .

[0070] Step S4: Based on step S3, use the self-attention mechanism to extract the high-level semantic features of each view, and calculate the embedded features with information interaction by weighted average.

[0071] Specifically, it includes: given the number of heads h of the multi-head self-attention mechanism and the low-level semantic features of the i-th sample Use h groups of linear transformation parameter matrices Obtain the corresponding query, key, and value mappings Divide the entire embedded feature into h parts, each part with dimension d h = d e / h. For the t-th head of sample i, calculate its score matrix according to the following formula:

[0072]

[0073] The above formula calculates the score matrix for the complete view. For the missing view data, use the following formula to calculate the score matrix:

[0074]

[0075] Where zerofill() means filling zero values with -1e9, and ⊙ is the Hadamard product. This is to ignore the missing views in the cross-view attention score calculation. Finally, summarize as follows:

[0076]

[0077] Where is the high-level semantic feature obtained by the t-th head after information interaction. Therefore, the high-level semantic features of m views are Then, input the high-level semantic feature A of each sample into the linear layer, norm layer, and multi-layer perceptron in sequence to obtain the final embedded feature of each sample (for all samples )), that is, the multi-view information fusion encoder Obtain the common representation shared by all views by multi-view weighted fusion:

[0078]

[0079] Where, represents the i-th sample of the fusion matrix . Formula (5) is designed for missing views. For the complete view can simply use to calculate.

[0080] Step S5: Use a decoder symmetric to the network structure in step S4 to reconstruct all data, including the missing views.

[0081] Specifically, it includes: reconstructing all data, including the missing views, using a decoder symmetric to the network structure in step S4, and the encoder and using the available original data to impose partial reconstruction constraints, adding the weighted reconstruction loss L re :

[0082]

[0083] where is the reconstructed data for the i-th sample. In the second stage, the reconstruction loss function becomes:

[0084]

[0085] Step S6: Supplement the missing views recovered in step S5 to the original dataset to construct an approximately complete multi-view dataset, and construct a multi-order fusion neighbor graph based on the approximately complete multi-view data.

[0086] Specifically, it includes: supplementing the missing views recovered in step S5 to the original dataset to construct an approximately complete multi-view dataset, where W represents an indicator matrix, and W i,j = 1 indicates the j-th view of the i-th sample, otherwise, W i,j = 0 indicates that the j-th view of the i-th sample is missing. Construct a multi-order fusion neighbor graph based on the approximately complete multi-view data:

[0087]

[0088] where

[0089]

[0090]

[0091] where is the multi-order neighbor matrix between the generated sample data, is the o-th order neighbor matrix of the v-th view.

[0092] Step S7: Calculate the loss function using the views reconstructed in step S5 and the multi-order fusion neighbor graph constructed in step S6, and update the network parameters using the stochastic gradient descent algorithm. The specific implementation is as follows: In the stage of recovering the missing views, the overall loss function is:

[0093] L = L re + βL rg (12)

[0094] β is a penalty parameter that balances these two losses. After filling in the missing views, the loss function is:

[0095] L' = L' re +βL rg (13)

[0096] Wherein:

[0097]

[0098] Wherein, b represents the sample size of each small batch, is the output of the encoder corresponding to the data of each small batch, | k represents the k-th training round. In the calculation of the multi-order neighbor graph loss, in order to make the most of the global structure information, all the embedding features Z obtained in the previous training round are retained in this paper, and the corresponding parts are updated after the current batch processing. Because the frequent change of the input makes it difficult for the model to converge, after completing the missing view filling, the graph constructed by the first-stage training is fixed instead of being updated by the final output, that is, G (v) | k-1 = G (v) | k .

[0099] Step S7: Calculate the loss function by using the view reconstructed in step S5 and the multi-order fusion neighbor graph constructed in step S6, and update the network parameters by using the stochastic gradient descent algorithm;

[0100] Step S8: Perform K-means clustering by using the embedding features finally obtained by the model to complete the well selection for oil well fracturing measures.

[0101] Specifically, it includes: Step S8 uses the finally obtained fused embedding features as the clustering index matrix. For simplicity, this paper performs K-means on and sets K to 2, dividing the samples into two categories, one is the wells with high measure benefits and the other is the wells with low measure benefits, so as to obtain the final well selection result for fracturing measures.

[0102] The embodiment of the present invention cyclically promotes the recovery of incomplete views and the preservation of the geometric structure within the views, helps to obtain more discriminative semantic fusion information. This enables the model to adapt to the complex problems of multi-scenarios in oil reservoirs, with good robustness, outstanding generalization ability, strong flexibility, can improve the accuracy of well selection for measures, and provides a new method for well selection for oil well fracturing measures.

[0103] The present invention fully considers the multi-view characteristics and data missing situations of the dynamic and static parameters of oil well fracturing, and proposes a multi-order nearest neighbor cycle graph constraint, which can provide multi-level structural description information for data and can provide more explicit clustering guidance for clustering algorithms. On the other hand, more obvious discriminative semantic features can also promote the recovery of missing views, further contributing to constructing a more realistic graph.

[0104] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Those of ordinary skill in the relevant technical fields can also make various changes and modifications without departing from the spirit and scope of the present invention. Therefore, all equivalent technical solutions also belong to the scope of the present invention, and the patent protection scope of the present invention shall be defined by the claims.

Claims

1. A well selection method for oil well fracturing based on multi - order neighbor cyclic graph constraint in multi - view clustering, characterized in that, it includes: Step S1: Obtain the dynamic and static parameters affecting the oil well fracturing effect, and perform Min - Max normalization on the dynamic and static parameters; Step S2: Based on the data normalized in Step S1, divide the multi - view data set according to the sources of the dynamic and static parameters; Step S3: Add low - level feature extractors to each view divided in Step S2; Step S4: On the basis of Step S3, use the self - attention mechanism to extract the high - level semantic features of each view, and calculate the embedded features with information interaction by means of weighted average; Step S5: Use a decoder symmetric to the network structure in Step S4 to reconstruct all data, including the missing views, specifically including: Encoder and utilize the available raw data to impose partial reconstruction constraints and weighted reconstruction loss L re : Among them is the reconstructed data the i-th sample in; in the second stage, the reconstruction loss function becomes: Step S6: Supplement the missing views recovered in Step S5 to the original data set to construct an approximately complete multi - view data set, and construct a multi - order fusion neighbor graph according to the approximately complete multi - view data, specifically including: Supplement the missing views recovered in Step S5 to the original data set to construct an approximately complete multi - view data set, where \(W\) represents an indicator matrix, \(W\) i,j = 1 indicates the \(j\)-th view of the \(i\)-th sample, otherwise, \(W\) i,j = 0 indicates that the \(j\)-th view of the \(i\)-th sample is missing; construct a multi-order fusion nearest neighbor graph based on the approximately complete multi-view data: wherein, Among them, is the multi-order neighbor matrix between the generated sample data, is the o-th order neighbor matrix of the v-th view; Step S7: Calculate the loss function using the views reconstructed in Step S5 and the multi - order fusion neighbor graph constructed in Step S6, and update the network parameters using the stochastic gradient descent algorithm, specifically including: The overall loss function in the stage of recovering the missing views is: L = L re + βL rg (12) β is the penalty parameter that balances these two losses; the loss function after completing the missing views is: L' = L' re + βL rg (13) where: where b represents the sample size of each mini-batch, is the output of the encoder corresponding to the data of each mini-batch, | k represents the k-th training epoch; in the multi-order neighbor graph loss calculation, the global structure information is maximally utilized, all the embedding features Z obtained in the previous training epoch are retained, and the corresponding parts are updated after the current batch processing; after completing the missing views, the graph formed by the first-stage training is fixed, i.e., G (v) | k-1 = G (v) | k ; Step S8: Perform K - means clustering using the embedded features finally obtained by the model to complete the well selection for oil well fracturing measures.

2. The well selection method for oil well fracturing based on multi - order neighbor cyclic graph constraint in multi - view clustering according to claim 1, characterized in that, in the said Step S1, the formula for performing Min - Max normalization on the dynamic and static parameters includes: where, Y represents the original dynamic and static parameters, and X represents the feature dynamic and static parameters after Min - Max normalization.

3. The well selection method for oil well fracturing based on multi - order neighbor cyclic graph constraint in multi - view clustering according to claim 1, characterized in that, the said Step S2 includes: Based on the normalized data, according to the sources of dynamic and static parameters affecting the fracturing effect of oil wells, it is divided into the fracturing parameter view X (1) , the production parameter view X (2) , the formation parameter view X (3) and the engineering parameter view X (4) , where indicates that there are a total of m views, and d v represents the original feature dimension of the v-th view.

4. The well selection method for oil well fracturing based on multi - order neighbor cyclic graph constraint in multi - view clustering according to claim 3, characterized in that, The step S3 includes: using a fully connected layer as a low-level feature extractor wherein represents the low-level features obtained after passing through the low-level feature extractor, and d e represents the dimension of the original features after passing through the low-level feature extractor, and θ v is the parameter of the fully connected layer Φ v .

5. The well selection method for oil well fracturing based on multi - order neighbor cyclic graph constraint in multi - view clustering according to claim 1, characterized in that, Step S4 specifically includes: specifying the number of heads h of the multi-head self-attention mechanism and the low-level semantic features of the i-th sample using h groups of linear transformation parameter matrices to obtain the corresponding mappings of queries, keys, and values Divide the entire embedded feature into h parts, each part with dimension d h = d e / h; For the t-th head of sample i, calculate its score matrix according to the following formula: The above formula calculates the score matrix for the complete view; for the missing view data, the following formula is used to calculate the score matrix: Among them, zerofill() means filling with -1e 9 to fill zero values, and ⊙ is the Hadamard product; To ignore the missing views in the calculation of the cross - view attention scores, it is summarized as follows: Among them is the high-level semantic feature obtained by the t-th head after information interaction; therefore, the high-level semantic features of m views are Then, the high-level semantic feature A of each sample is input into a linear layer, a norm layer, and a multi-layer perceptron in sequence to obtain the final embedding feature of each sample For all samples That is, the multi-view information fusion encoder Adopt the method of multi-view weighted fusion to obtain the common representation shared by all views: Among them, represents the i-th sample of the fusion matrix ; formula (5) is designed for the missing views, and for the complete views it can be simply calculated using .

6. The well selection method for oil well fracturing based on multi - order neighbor cyclic graph constraint in multi - view clustering according to claim 1, characterized in that, The specific steps of step S8 include: using the finally obtained fused embedding features as the clustering index matrix; performing K-means on and setting K to 2 to divide the samples into two categories, one is the wells with high measure benefits and the other is the wells with low measure benefits, so as to obtain the final well selection result for fracturing measures.

Images