Flooded reservoir subdivision layer identification method
By combining curve morphology with AI-powered water-flooded reservoir identification methods, and employing random forest algorithms and the activity function method, the problem of low accuracy and efficiency in subdividing water-flooded reservoirs was solved, achieving efficient and intelligent interpretation.
Patent Information
- Application Number
- CN202411183835.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-27
- Publication Date
- 2026-03-06
AI Technical Summary
Existing methods for subdividing water-flooded reservoirs suffer from low accuracy, high cost, and low efficiency, making it difficult to meet production needs.
By combining curve morphology with AI intelligence, and employing the random forest algorithm and the activity function method, the model is trained using core well data to identify the boundaries of subdivided water-flooded reservoirs, thereby achieving intelligent subdivision of layers.
It improves the accuracy and efficiency of water-flooded layer subdivision interpretation, is highly adaptable, can adapt to changes in complex geological conditions, and reduces the cost of manual annotation.
Smart Images
Figure CN121614992A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of well logging evaluation technology, and in particular to a method for identifying subdivided layers of water-flooded reservoirs. Background Technology
[0002] Traditional subdivision methods can be broadly categorized into two types: manual subdivision, where the reservoir is subdivided according to subdivision principles; and theoretical subdivision methods, such as the activity method, the derivative method, and cluster analysis. These theoretical methods use defined functions to subdivide the reservoir by curve morphology. Manual subdivision offers high accuracy and better meets subdivision requirements, but its drawbacks include high cost and low efficiency. Theoretical subdivision methods are overly reliant on curves, have low accuracy, and cannot incorporate expert experience. Therefore, it is necessary to research a new subdivision method that is intelligent, significantly improving the accuracy and efficiency of water-flooded reservoir subdivision interpretation—a new trend in intelligent interpretation of water-flooded reservoirs. Summary of the Invention
[0003] This invention addresses the shortcomings of existing methods in the background art, such as insufficient accuracy in meeting production needs and low efficiency, by providing a method for identifying subdivided layers of water-flooded reservoirs. This method, based on a combination of curve morphology and AI intelligence, achieves intelligent subdivision identification of water-flooded reservoirs. It is more adaptable and can be applied to the interpretation of infill wells, laying the foundation for the subdivision interpretation of water-flooded layers in thick oil reservoirs.
[0004] The present invention solves its problem through the following technical solution: the method for subdividing and identifying water-flooded reservoirs includes the following steps:
[0005] S1: According to the principle of subdivision, the effective thickness of the coring well is manually subdivided into layers; the results of the subdivision data of the coring well are selected as sample data.
[0006] S2: Based on the subdivided data from the core wells, construct a sample label dataset according to the training sample points;
[0007] S3: The random forest algorithm is used to train the model for labeled well logging curve data;
[0008] S4: Based on the training model of the subdivision layer obtained in step S3, predict the new data, obtain preliminary prediction results, and find the preliminary subdivision layer boundary.
[0009] S5: Subdivide the layers based on the curve shape and find the boundaries of the remaining subdivision layers;
[0010] S6: Summarize the subdivision boundaries found in steps S4 and 5 as the final stratification result, and output the final stratification result.
[0011] Furthermore, the principle for subdividing the effective thickness of the coring well in step S1 is as follows:
[0012] For reservoirs with the same water washing level but significantly different physical properties, further subdivisions should be made;
[0013] For reservoirs with similar physical properties but different water washing levels, further subdivisions are made;
[0014] For reservoirs with interlayers, further subdivision and interpretation are performed.
[0015] Furthermore, step S2 constructs a sample label dataset based on the subdivided data from the cored wells, according to the training sample points; the specific method is as follows:
[0016] Based on the manually interpreted subdivided core logging data, for each depth sampling point within the effective thickness, it is determined whether the current depth is the subdivided layer boundary, and the sample point is labeled as boundary 1 / non-boundary 0; the labeled logging curve data is classified to construct a sample label dataset.
[0017] Furthermore, step S3 uses the random forest algorithm to train the model on the labeled well logging curve data. The specific methods include:
[0018] First, determine the number of decision tree classifiers in the random forest;
[0019] Then, random sampling with replacement is taken from the original sample dataset in step S1 to construct a subset dataset, and the subset dataset is used to build a sub-decision tree; these labeled logging curve data are put into each sub-decision tree, and each sub-decision tree outputs a result; during the training process of the sub-decision tree, the training subset dataset needs to be split into two subset datasets again and again until the number of samples in each subset dataset is less than the set minimum number of samples, and then the splitting stops.
[0020] The random forest algorithm is used to train a classification model for the two classes of samples in the sample label dataset constructed in step S2, resulting in a training model for the subdivided layers.
[0021] Furthermore, the number of decision tree classifiers in a random forest can be adjusted, either by increasing or decreasing it, until it reaches a suitable value that avoids overfitting or underfitting.
[0022] Furthermore, the number of decision tree classifiers in the random forest is 100; when building sub-decision trees, the maximum number of features selected is the total number of curve features.
[0023] Furthermore, step S4 involves predicting the new data to obtain preliminary prediction results and identifying preliminary subdivision boundaries. The specific method is as follows:
[0024] When there is new data at a certain depth point that needs to be used to obtain a predicted classification result through the random forest training model determined in step S3, the feature values of the new data need to be input into the training model of the subdivided layer established in step S3; each sub-decision tree votes on whether the judgment result is 1 or 0, and the result with the most votes is taken as the output result. The probability threshold is 50% by default. When the probability of voting 1 is greater than 50%, the output result is 1; otherwise, the output result is 0.
[0025] Using the subdivision model trained in step S3, the above method is used to predict each depth point within the effective thickness of the test well, obtain preliminary prediction results, and find the preliminary subdivision boundary.
[0026] Furthermore, the method for subdividing layers based on curve shape in S5 includes:
[0027] Based on the curve morphology subdivision method, the reservoir subdivision is achieved by using the activity function method;
[0028] Since well logging data is recorded in a discrete manner, the activity function can be rewritten as the square of the difference between the curve amplitude value at a certain depth and the average curve amplitude of the depth segment at that point, and then summed cumulatively within the current depth segment;
[0029] As can be seen from the definition of activity, the local maximum of the activity curve, i.e. the peak value, is at the layer interface. The boundary of the subdivided layer can be found by the local maximum of the activity curve.
[0030] Furthermore, the activity function is defined as:
[0031]
[0032] In the formula: y(x) represents the curve amplitude at depth x; E(d) represents the activity value at depth d; w represents the filter window length; y(d) represents the average value of y(x) within the well section (dw / 2, d+w / 2).
[0033] Furthermore, the rewritten expression for the activity function is:
[0034]
[0035] Where: N = 2n + 1.
[0036] Compared with the above-mentioned background technology, the present invention has the following beneficial effects:
[0037] (1) The water-flooded reservoir subdivision identification method based on the combination of curve morphology and AI intelligence not only overcomes the situation that relying solely on curve morphology cannot identify the subdivision of high water flooding and ultra-high water flooding due to the insignificant curve changes, but also overcomes the situation that relying solely on AI methods to predict results may not meet business constraints due to a lack of domain knowledge.
[0038] (2) The interpretation method is more scalable and can be retrained for subdivided layer data of different logging series. The sample library and model can be updated in real time, are autonomous and controllable, and are more adaptable to constantly changing complex geological conditions.
[0039] (3) The intelligent subdivision technology has greatly improved the accuracy and efficiency of the subdivision interpretation of the flooded layer, and is a new trend in the intelligent interpretation of the flooded layer.
[0040] (4) The method of water-flooded reservoir subdivision identification based on the combination of curve morphology and AI intelligence is the basis of well logging intelligent interpretation. Its methods and ideas have great guiding significance for reservoir thickness division, parameter interpretation and intelligent interpretation of water-flooded layers. Attached Figure Description
[0041] Figure 1 This is a flowchart of a method for identifying subdivided layers of a flooded reservoir according to the present invention;
[0042] Figure 2 This is a diagram showing the subdivided layer interpretation results of well B1-21-J654 in Embodiment 1 of the present invention;
[0043] Figure 3 This is a diagram showing the results of the detailed layer interpretation of well B1-330-J49 in Embodiment 2 of the present invention. Detailed Implementation
[0044] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.
[0045] Formations are deposited layer by layer, and well logging data reflecting their properties also show different characteristics layer by layer. The changing shapes of the curves contain unique information about the formations, corresponding to different sedimentary microfacies. Therefore, well logging interpretation can be performed layer by layer based on the characteristics of the well logging curves. Extracting the characteristic extreme values of the curves using appropriate rules and highlighting the characteristic locations of the curves are important steps in identifying subdivided layers based on the curve morphology.
[0046] The curve morphology-based subdivision method mainly uses the activity function method to subdivide reservoirs. This method can automatically pick out the logging response characteristic values of different physical properties and oil-bearing segments, and can also automatically extract values for the upper and lower segments of interlayers.
[0047] From an AI perspective, the subdivision layer recognition method can be viewed as a classification problem, specifically identifying the boundaries of subdivision layers. The manually segmented subdivision boundaries are grouped into one category, while non-boundary depth points are grouped into another. Machine learning is then used to predict the boundaries of points with these characteristics.
[0048] Random forest algorithms are easier to implement, can handle high-dimensional features, and can better balance errors in imbalanced data, resulting in high accuracy in classification problems. Therefore, based on the characteristics of the sample data and the prediction results, the supervised clustering random forest big data algorithm is preferred. Entities, relationships, and attributes in the well logging field are expressed in vector form, and then a subdivided layer model is established.
[0049] The detailed interpretation of flooded reservoirs is extremely complex. Due to a lack of domain knowledge, AI model predictions based solely on input data may not meet operational constraints. Furthermore, relying solely on curve morphology for stratification can lead to the failure to stratify sections where curve inflection points are indistinct but flooding levels have changed. Therefore, for detailed stratification, incorporating curve morphology to correct AI identification results is an effective approach to address this problem.
[0050] like Figure 1 As shown, a method for subdividing and identifying water-flooded reservoirs includes the following steps:
[0051] S1: According to the principle of subdivision, the effective thickness of the coring well is manually subdivided into layers; the results of the subdivision data of the coring well are selected as sample data.
[0052] The principle for subdividing the effective thickness of the coring well, and the specific methods include:
[0053] For reservoirs with the same water washing level but significantly different physical properties, they must be subdivided into layers.
[0054] For reservoirs with similar physical properties but different water washing levels, they must be subdivided into layers.
[0055] Reservoirs with interlayers must be subdivided for interpretation.
[0056] Based on the distribution pattern of residual oil, the area near the interlayer is often not thoroughly washed and contains a lot of residual oil. Therefore, the interlayer needs to be subdivided to study the flooding situation.
[0057] Based on the above subdivision principle, the effective thickness of the core well is manually subdivided into layers.
[0058] S2: Based on the sub-layered data from cored wells, construct a sample label dataset according to the training sample points; the specific method is as follows:
[0059] Based on the manually interpreted subdivided core logging data, for each depth sampling point within the effective thickness, it is determined whether the current depth is the subdivided layer boundary, and the sample point is labeled as boundary 1 / non-boundary 0; the labeled logging curve data is classified to construct a sample label dataset.
[0060] S3: The random forest algorithm is used to train the model for labeled well logging data. Specific methods include:
[0061] First, determine that the number of decision tree classifiers in the random forest is 100. This number can be adjusted, either large or small, until it is adjusted to a suitable value that avoids overfitting or underfitting. When building the sub-decision trees, select the maximum number of features equal to the total number of curve features.
[0062] Then, random sampling with replacement is taken from the original sample dataset in step S1 to construct a subset dataset, and the subset dataset is used to build a sub-decision tree; these labeled logging curve data are put into each sub-decision tree, and each sub-decision tree outputs a result; during the training process of the sub-decision tree, the training subset dataset needs to be split into two subset datasets again and again until the number of samples in each subset dataset is less than the set minimum number of samples, and then the splitting stops.
[0063] The random forest algorithm is used to train a classification model for the two classes of samples in the sample label dataset constructed in step S2, resulting in a training model for the subdivided layers.
[0064] S4: Based on the training model of the subdivision layer obtained in step S3, predict the new data to obtain preliminary prediction results and find the preliminary subdivision layer boundaries. The specific method is as follows:
[0065] When there is new data at a certain depth point that needs to be used to obtain a predicted classification result through the random forest training model determined in step S3, the feature values of the new data need to be input into the training model of the subdivided layer established in step S3; each sub-decision tree votes on whether the judgment result is 1 or 0, and the result with the most votes is taken as the output result. The probability threshold is 50% by default. When the probability of voting 1 is greater than 50%, the output result is 1; otherwise, the output result is 0.
[0066] Using the subdivision model trained in step S3, the above method is used to predict each depth point within the effective thickness of the test well, obtain preliminary prediction results, and find the preliminary subdivision boundary.
[0067] S5: The fifth step is to subdivide the layers based on the curve shape and find the boundaries of the remaining subdivision layers.
[0068] The method for subdividing reservoir layers based on curve morphology mainly employs the activity function method to achieve reservoir layer subdivision. The activity function is defined as:
[0069]
[0070] In the formula: y(x) represents the curve amplitude at depth x; E(d) represents the activity value at depth d; w represents the filter window length; y(d) represents the average value of y(x) within the well section (dw / 2, d+w / 2);
[0071] Since well logging data is recorded discretely, the above formula can be rewritten as:
[0072]
[0073] Where: N = 2n + 1
[0074] As can be seen from the definition of activity, the local maximum of the activity curve, i.e. the peak value, is at the layer interface. The boundary of the subdivided layer can be found by the local maximum of the activity curve.
[0075] The present invention will now be described based on embodiments, but it is worth noting that the present invention is not limited to these embodiments. In the following detailed description of the invention, certain specific details are described in detail. However, those skilled in the art will fully understand the invention for the parts not described in detail.
[0076] Furthermore, those skilled in the art should understand that the accompanying drawings are provided only to illustrate the purpose, features, and advantages of the present invention, and are not actually drawn to scale.
[0077] Furthermore, unless the context explicitly requires it, the words "comprising," "including," and similar terms throughout the specification and claims should be interpreted as including rather than exclusive or exhaustive; that is, meaning "including but not limited to."
[0078] Example 1
[0079] The method of the present invention will be illustrated using well B1-21-J654 as an example. Figure 2 This is a detailed subdivision result chart for well B1-21-J654. From left to right, the first line shows the deep lateral resistivity (RLLD), shallow lateral resistivity (RLLS), and microsphere resistivity curve (RXO); the second line shows the microgradient (RMG) and micropotential curve (RMN); the third line shows the wellbore diameter (CAL), spontaneous potential (SP), and natural gamma ray (GR); the fourth line shows the lithological density (DEN) and sonic logging curve (HAC); the fifth line shows the depth; the sixth line shows the oil-bearing formation; the seventh line shows the effective thickness; the eighth line shows the subdivision result manually calculated according to the subdivision principle; and the ninth line shows the model-predicted subdivision result. All logging curve values used in the model prediction can be directly read from the chart. The specific calculation steps are as follows:
[0080] The first step is to select sample data.
[0081] Using the coring well analysis and testing data, the effective thickness of the multiple coring wells participating in the training was subdivided into layers according to the principle of subdivision layering (as shown in Table 1; due to the large amount of data, Table 1 only shows the data of some wells and some subdivision layers).
[0082] The second step is to construct a labeled dataset based on the training sample points.
[0083] Based on the manually interpreted subdivided layer data from cored wells, for each depth sampling point, it is determined whether the current depth is a layer boundary. The sample point is then labeled with a boundary 1 or a non-boundary 0 for classification, thus constructing a sample label dataset. (See Table 2 for the specific format. Due to the large amount of data, Table 2 only illustrates the construction of a small segment of the subdivided layer label dataset from Well Bei 1-6-Jian 26. The same labeling method is used to construct label data for the subdivided layers of other wells.)
[0084] The third step is model training.
[0085] The training set is constructed using artificially stratified data from the core wells obtained in the second step. Learning samples are extracted from the constructed logging curve feature value file (i.e., the logging curve value corresponding to each depth point). A random forest algorithm is used to train classification models for these two types of samples, resulting in training models for the sub-layers. Here, the number of decision tree classifiers in the random forest is 100. This number can be adjusted, either large or small, to achieve a suitable balance that avoids overfitting or underfitting. The maximum number of features selected when building the sub-decision trees is the total number of curve features. Taking the Daqing Oilfield DLS logging series as an example, the logging curves include lateral RLLD, shallow lateral RLLS, microsphere resistivity curve RXO, microgradient RMG, micropotential curve RMN, wellbore caliper CAL, spontaneous potential SP, natural gamma curve GR, lithological density DEN, and sonic logging curve HAC, totaling ten logging curves. The maximum number of features selected is 10.
[0086] The fourth step is model prediction.
[0087] The effective thickness data of well B1-21-J654 and the feature values of the logging curve (see Table 3) are input into the established decision tree. The subdivided training model obtained in the third step is used to predict each depth point within the effective thickness of well B1-21-J654, and preliminary prediction results are obtained.
[0088] The fifth step involves using the activity function method described earlier to calculate the activity function value at each depth point. The local maximum or peak value of the activity curve is located at the layer interface. The boundary of the subdivision layer is found by using the local maximum value of the activity curve.
[0089] As a marker for further layering, the layers are further subdivided at the new marker depth, and the results of the fourth and fifth steps are output as the final layering result.
[0090] The interpretation results of the effective thickness subdivision prediction of well B1-21-J654 are shown in Table 4.
[0091] To verify the feasibility of the method in this embodiment, the newly predicted subdivision layers of well B1-21-J654 were compared with the results of artificial stratification. There were a total of 61 subdivision layers within the effective thickness, with 50 layers matching the prediction, resulting in a matching rate of 82.0% (the accuracy is shown in Table 4).
[0092] Table 1
[0093] hashtag Top depth bottom depth Subdivision thickness / m Taipei 1-6-Prosecutor's Office 26 952.8 953.3 0.5 Taipei 1-6-Prosecutor's Office 26 953.3 953.8 0.57 Taipei 1-6-Prosecutor's Office 26 953.8 954.5 0.63 Taipei 1-6-Prosecutor's Office 26 955.8 956 0.2 Taipei 1-6-Prosecutor's Office 26 956 956.3 0.3 Taipei 1-6-Prosecutor's Office 26 958.8 959.2 0.4 Taipei 1-6-Prosecutor's Office 26 959.2 960.0 0.82 Taipei 1-6-Prosecutor's Office 26 960.0 960.9 0.88 Taipei 1-6-Prosecutor's Office 26 960.9 961.2 0.3 Taipei 1-6-Prosecutor's Office 26 962.2 962.6 0.4 Taipei 1-6-Prosecutor's Office 26 963.3 963.6 0.3 Taipei 1-6-Prosecutor's Office 26 963.6 964.0 0.45 Taipei 1-6-Prosecutor's Office 26 964.0 964.7 0.65 Taipei 1-6-Prosecutor's Office 26 965.1 965.5 0.4 Taipei 1-6-Prosecutor's Office 26 967.7 967.9 0.2 Taipei 1-6-Prosecutor's Office 26 968.1 968.5 0.4 Taipei 1-6-Prosecutor's Office 26 968.9 969.2 0.35 Taipei 1-6-Prosecutor's Office 26 969.2 969.7 0.51 Taipei 1-6-Prosecutor's Office 26 969.7 970.4 0.64 Taipei 1-6-Prosecutor's Office 26 970.4 970.7 0.35 Taipei 1-6-Prosecutor's Office 26 970.7 971.0 0.34 Taipei 1-6-Prosecutor's Office 26 971.0 971.5 0.44 Taipei 1-6-Prosecutor's Office 26 983.5 983.7 0.2 Taipei 1-6-Prosecutor's Office 26 983.7 984.6 0.9
[0094] Table 2
[0095]
[0096] Table 3
[0097] depth DEN RLLD RXO RLLS RMN RMG CALS SP GR HAC Thickness label 913.6 2.34 6.1 7.9 5.2 6.0 5.1 22.5 -24.0 95.0 313.5 0 913.65 2.33 6.1 7.7 5.2 5.8 4.9 22.3 -24.9 95.6 316.8 0 913.7 2.33 6.2 7.5 5.3 5.7 4.9 22.2 -26.8 96.3 317.8 0 913.75 2.31 6.4 7.6 5.4 5.7 4.8 22.1 -28.1 97.6 315.7 0 913.8 2.30 6.8 7.6 5.6 5.8 4.8 21.9 -29.2 99.0 317.4 0 913.85 2.28 7.2 7.7 5.8 5.9 4.9 21.8 -30.3 100.5 319.2 0 913.9 2.27 7.8 8.3 6.3 6.5 5.4 21.7 -31.5 101.6 324.5 0 913.95 2.26 8.3 9.0 6.8 7.2 6.0 21.7 -33.0 102.9 325.7 0 914 2.26 8.8 9.0 7.2 7.6 6.1 21.6 -34.1 103.6 325.3 0 914.05 2.25 9.2 9.5 7.5 7.8 6.1 21.5 -35.2 103.9 325.6 0 914.1 2.24 9.7 10.1 7.9 7.9 6.1 21.5 -36.3 104.1 325.9 0 914.15 2.24 10.0 11.4 8.1 8.1 6.6 21.4 -37.4 104.3 324.7 0 914.2 2.23 10.5 11.6 8.1 8.1 6.6 21.4 -38.3 104.1 324.6 0 914.25 2.23 11.5 10.9 8.6 8.2 6.3 21.4 -39.1 104.6 328.1 0 914.3 2.22 12.6 12.6 9.6 8.8 6.6 21.4 -39.8 105.1 329.9 0 914.35 2.21 13.8 14.2 10.6 9.3 7.0 21.3 -40.5 105.6 331.7 0 914.4 2.20 14.5 16.3 11.2 9.6 7.5 21.3 -40.9 106.0 327.7 0 914.45 2.19 15.2 15.9 11.4 9.4 7.1 21.3 -41.4 106.8 326.1 0 914.5 2.19 16.6 15.4 11.8 9.1 6.6 21.3 -41.8 107.2 328.1 0 914.55 2.18 18.1 19.2 13.3 9.7 7.0 21.3 -42.1 107.5 331.9 1 914.6 2.17 19.5 23.1 14.8 10.2 7.4 21.3 -42.3 107.9 335.7 1 914.65 2.17 20.0 25.0 15.2 11.1 8.2 21.3 -42.6 108.1 339.8 1 914.7 2.17 20.1 26.5 15.8 11.5 8.7 21.3 -42.9 106.9 342.5 1 914.75 2.16 20.0 23.8 15.3 11.6 8.4 21.3 -43.2 105.7 341.6 1 914.8 2.16 20.7 22.6 15.7 11.7 8.5 21.3 -43.5 104.0 340.7 1 914.85 2.16 21.5 21.4 16.2 11.8 8.7 21.3 -43.7 102.3 339.8 1
[0098] Table 4
[0099]
[0100]
[0101]
[0102] Example 2
[0103] Taking well B1-330-J49 as an example, the interpretation results of this well are shown in the figure below. Figure 3 Using the same calculation steps as in Example 1, the interpretation results of the subdivided layers within the effective thickness are shown in Table 5.
[0104] To verify the feasibility of the method in this embodiment, the newly predicted subdivision layers of well B1-330-J49 were compared with the results of artificial stratification. There were a total of 180 subdivision layers within the effective thickness, with 150 layers matching the prediction, resulting in a matching rate of 83.3% (the accuracy is shown in Table 5).
[0105] Table 5
[0106]
[0107]
[0108]
[0109]
[0110]
[0111]
[0112]
[0113]
[0114] The embodiments described above are merely illustrative of implementation methods of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications, equivalent substitutions, and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.
Claims
1. A method for identifying subzones of a watered-out reservoir, the method comprising: The method comprises the following steps: S1: according to the principle of subdividing layers, the effective thickness of the coring well is manually subdivided into layers; the subdivided layer data of the coring well is selected as sample data; S2: based on the subdivided layer data of the coring well, a sample label data set is constructed according to training sample points; S3: the random forest algorithm is used to train the model for the logging curve data with labels; S4: based on the training model of the subdivided layer obtained in step S3, new data is predicted to obtain a preliminary prediction result and find a preliminary subdivided layer boundary; S5: the subdivided layer is subdivided based on the curve shape to find the remaining subdivided layer boundary; S6: the subdivided layer boundaries found in steps S4 and S5 are summarized as the final layering result, and the final layering result is output.
2. The method according to claim 1, characterized in that: The principle of subdividing layers of the effective thickness of the coring well in step S1 is as follows: For reservoirs with the same water washing level but large differences in physical property parameters, the reservoirs are subdivided into layers; For reservoirs with similar physical property parameters but different water washing levels, the reservoirs are subdivided into layers; For reservoirs with interbeds, the reservoirs are subdivided into layers for interpretation.
3. The method according to claim 1, characterized in that: In step S2, the sample label data set is constructed according to training sample points based on the subdivided layer data of the coring well; specifically, according to the subdivided layer data of the coring well interpreted manually, whether the current depth is a subdivided layer boundary is judged for each depth sampling point in the effective thickness, and the sample point is labeled with a boundary 1 / non-boundary 0 label; the logging curve data with labels is classified to construct the sample label data set. In step S3, the random forest algorithm is used to train the model for the logging curve data with labels; specifically, the number of decision tree classifiers in the random forest is determined first; 4. The method according to claim 1, characterized in that: Then, random sampling with replacement is performed on the original sample data set in step S1 to construct a sub-data set, and the sub-data set is used to construct a sub-decision tree; the logging curve data with labels is put into each sub-decision tree, and each sub-decision tree outputs a result; in the training process of the sub-decision tree, the training sub-data set needs to be split into two sub-data sets each time until the number of samples in each sub-data set is less than the set minimum sample number, and the splitting is stopped; The random forest algorithm is used to train the classification model for the two types of samples in the sample label data set constructed in step S2 to obtain the training model of the subdivided layer. The number of decision tree classifiers in the random forest can be adjusted, and can be adjusted to be larger or smaller, and finally adjusted to a suitable value that is not prone to overfitting or underfitting.
6. The watered-out reservoir subdivided layer identification method according to claim 4, wherein:
5. The method according to claim 4, characterized in that: The number of decision tree classifiers in the random forest is 100; when the sub-decision tree is established, the maximum number of selected features is the number of all curve features. In step S4, new data is predicted to obtain a preliminary prediction result and find a preliminary subdivided layer boundary, and the specific method is as follows: 7. The method according to claim 1, characterized in that: When a new certain depth point data needs to be predicted by the random forest training model determined by S3, the feature value of the new data needs to be input into the training model of the sublayer established in step S3; each sub-decision tree votes for the judgment result of 1 or 0, and the most voted result is taken as the output result, and the probability threshold is 50% by default. When the probability of voting 1 is greater than 50%, the output result is 1; Otherwise, the output result is 0. The sublayer model trained in step S3 is applied, and the above method is used to predict each depth point in the effective thickness of the test well to obtain a preliminary prediction result and find a preliminary sublayer boundary.
8. The method according to claim 1, characterized in that: The method for sublayering based on curve morphology in S5 comprises: The method for sublayering based on curve morphology uses an activity function method to realize reservoir sublayering. Since the logging data is recorded in a discrete manner, the activity function can be rewritten as the square of the difference between the curve amplitude value at a certain depth and the average value of the curve amplitude of the depth section where the point is located, and the sum is accumulated within the current depth section. According to the definition of activity, the local maximum value of the activity curve, i.e. the peak value, is the layer interface, and the sublayer boundary is found by the local maximum value of the activity curve.
9. The method according to claim 8, characterized in that: The activity function is defined as: In the formula, y(x) represents the curve amplitude at depth x; E(d) represents the activity value at depth d; w represents the filter window length; and y(d) represents the average value of y(x) in the well section (d-w / 2, d+w / 2).
10. The method according to claim 8, wherein the expression of the activity function after rewriting is: wherein N = 2n + 1.