A multi-point geostatistical sedimentary microfacies prediction method based on deep learning

Abstract

The present invention provides a multi-point geostatistical sedimentary microfacies prediction method based on deep learning, and the specific method includes: converting geological constraints such as sedimentary patterns and phase sequence combination rules into regularized tensors that can be embedded in neural networks; using a hybrid architecture of convolutional neural networks and graph neural networks to simultaneously extract local microscopic features and regional macroscopic distribution patterns of sedimentary microfacies; fine-tuning the target work area based on a pre-trained model, and combining adversarial sample generation technology to enhance the model generalization ability; generating a three-dimensional sedimentary microfacies distribution model with confidence assessment. This technical solution innovatively realizes: coupled optimization of geostatistical spatial correlation constraints and deep learning feature extraction; nonlinear mapping and collaborative characterization of multi-source data in implicit feature space; and visualization of the dynamic evolution of geological processes of sedimentary microfacies prediction results.

Description

Technical Field

[0001] The present invention belongs to the technical field of oil and gas reservoir geological modeling and reservoir description, and specifically relates to an intelligent sedimentary microfacies prediction method and system integrating deep learning. In particular, based on an improved convolutional neural network architecture, the present invention constructs a sedimentary microfacies spatial prediction model integrating multi-point geostatistical feature constraints, and develops a matching automated prediction device, realizing the intelligent quantitative characterization of the sedimentary microfacies spatial distribution pattern under complex geological conditions. Background Art

[0002] In the field of reservoir characterization, the reconstruction of the spatial distribution pattern of sedimentary microfacies is the core link in building a high-precision geological model, which directly determines the reliability of numerical simulation of oil and gas reservoirs and the economic feasibility of development plans. The traditional method system has three main technical bottlenecks: (1) Methodological level: Conventional mathematical statistics methods (such as two-point geostatistics) are difficult to effectively characterize the spatial structural characteristics of complex sedimentary systems. Although multi-point geostatistics can extract high-order statistical features through training images, it still has problems such as difficulty in constructing variograms and poor model adaptability when dealing with multi-scale and non-stationary sedimentary patterns. (2) Technical implementation level: The current mainstream workflow relies on geological experts to manually draw sedimentary microfacies plane maps, which face three major technical challenges: ① There are significant subjective differences in the manual interpretation process. Studies have shown that the error in the judgment of microfacies boundaries by different interpreters in the same work area can reach 35%-60% of the well spacing; ② In the face of dense well network work areas with more than 200 wells / km², traditional manual mapping is inefficient. Statistics show that it takes 80-120 man-hours per square kilometer to make detailed maps; ③ Mechanical repetitive labor accounts for more than 60%, which is easy to cause interpretation standard drift and systematic error accumulation. (3) Intelligent level: Existing machine learning methods mostly adopt a pure data-driven model, which has two major defects: ① Large-scale labeled samples are required to support model training, while actual exploration blocks often only have data from dozens to hundreds of wells; ② The prediction results often show spatial configuration relationships that violate the basic principles of sedimentology, resulting in reduced geological credibility. Summary of the invention

[0003] In response to the above technical bottlenecks, the present invention provides a multi-point geostatistical sedimentary microfacies prediction method based on deep learning. This application innovatively constructs a machine learning framework guided by geological knowledge. By developing a multi-scale feature fusion network architecture, integrating image recognition, feature learning and statistical inference modules, intelligent prediction of sedimentary microfacies under small sample conditions is achieved. The technical breakthroughs are reflected in: ① Establishing a loss function system constrained by geological laws to ensure that the prediction results conform to the principles of sedimentary dynamics; ② Developing an adaptive feature migration algorithm to achieve reliable expansion of local knowledge to regional models; ③ Constructing a three-dimensional space verification mechanism to control the plane prediction error within 12% of the well spacing. Field verification has shown that this method can increase the efficiency of sedimentary microfacies drawing by 5-7 times, while improving the rationality of spatial configuration by more than 37%.

[0004] The technical solution adopted in this application is: a multi-point geostatistical sedimentary microfacies prediction method based on deep learning, comprising the following steps:

[0005] Step 1: Data preprocessing: Multi-source data fusion, integrating well logging curves, core descriptions, seismic data and dynamic production data to build the underlying data set;

[0006] (1) Collect and organize logging data, including natural gamma, acoustic time difference, resistivity, density, and shale content logging curves;

[0007] (2) Preprocess the logging curves, including unit correction, depth correction, and removal of outliers;

[0008] (3) Collect and organize core descriptions and lithology data, and assign digital values ​​to different lithologies;

[0009] (4) Processing of seismic data, including attribute extraction and seismic phase division;

[0010] (5) Use logging data, seismic data and core data to divide the sedimentary microfacies of all wells;

[0011] (6) Digitally assign values ​​to accurately divided sedimentary microfacies;

[0012] (7) Prepare a sedimentary microfacies division report, recording the division process, results and conclusions in detail;

[0013] Step 2, geological rule coding module: convert geological laws such as sedimentary phase sequence law and sand body superposition pattern into digital constraint matrix and embed them into feature extraction process; select corresponding geological laws according to geological characteristics of target blocks; use random forest algorithm to process missing values ​​in the conversion of geological laws into digital constraint matrix, and calculate multivariate linear correlation with Pearson correlation coefficient; geological rule tensor coding specifically includes: constructing sedimentary phase contact relationship matrix, defining the probability of river channel-natural levee-breach fan phase sequence transition; quantifying the source direction into 8-azimuth rose diagram tensor; using constraint matrix coding to represent the classification characteristics of basin evolution stage;

[0014] Step 3: Build a deep learning model architecture: including a multi-scale feature fusion network; use an improved ResNet-50 architecture as the feature extraction network for the backbone network; implement a transfer learning strategy and initialize parameters with a pre-trained model; set a dynamic learning rate adjustment mechanism, and set the initial learning rate and the decay coefficient per cycle according to the example;

[0015] Step 4: Sample training mechanism: By using multi-point geostatistics methods, data templates are used to scan training images to obtain data events to reflect the corresponding geological patterns; the frequency of occurrence of different data events is approximately the joint distribution probability of multiple points in space; the idea of ​​multi-point geostatistics is to use limited geological data to establish training images through sedimentological analysis, and under the constraints of conditional data, find the optimal data events from the training images as the basis for simulation sampling of the estimated points; it is believed that the training image is one of the key factors that determine the simulation effect; specifically, the following steps are included:

[0016] (1) Select an appropriate range, use the accurate sedimentary microfacies generated in the data preprocessing and feature engineering module, and combine human-computer interaction to select an appropriate range that best meets geological rationality.

[0017] (2) Digitally scan the small-scale sedimentary microfacies plane map that has been obtained and generate a rasterized result with a geographic coordinate system;

[0018] (3) Using the digitally scanned sedimentary microfacies plane map as a sample, a high-order compatibility algorithm is used to obtain the most matching training image;

[0019] (4) Using the acquired training images as geological models and using the data event repetition probability statistical algorithm, a sedimentary microfacies plane map matching the actual area is obtained;

[0020] Step 5. Deep learning: Combined with relevant algorithms, using training images as samples, obtain the geological model that best meets the geological conditions of the study area, and obtain the final result through human-computer interaction: Using the deep learning results as constraints and combining them with the basic geological data of the study area, interactive corrections are made in the geological expert system to obtain the final sedimentary microfacies plane map of the test area.

[0021] The beneficial effects of this application are: by integrating multi-source heterogeneous data such as various basic geological parameters, a deep learning framework with geological interpretability is constructed. The specific implementation process includes: converting geological constraints such as sedimentary patterns and phase sequence combination rules into regularized tensors that can be embedded in neural networks; using a hybrid architecture of convolutional neural networks and graph neural networks to simultaneously extract local microscopic features and regional macroscopic distribution patterns of sedimentary microfacies; fine-tuning the target work area based on the pre-trained model, and combining adversarial sample generation technology to enhance the model generalization ability; generating a three-dimensional sedimentary microfacies distribution model with confidence assessment. This technical solution innovatively realizes: the coupled optimization of geostatistical spatial correlation constraints and deep learning feature extraction; nonlinear mapping and collaborative characterization of multi-source data in implicit feature space; and visualization of the dynamic evolution of geological processes of sedimentary microfacies prediction results. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 A schematic diagram of the process of this application is presented;

[0023] Figure 2 A schematic diagram of the target block of an embodiment of the present application is presented;

[0024] Figure 3 The mud content and sedimentary microfacies division results of the standard wells in the target block of the embodiment of the present application are presented;

[0025] Figure 4 A plan view of sedimentary microfacies in the test area within the target block of the embodiment of the present application is presented;

[0026] Figure 5 The digital scanning results of the sedimentary microfacies plane map of the test area in the target block of the embodiment of the present application are presented;

[0027] Figure 6 The training image obtained by using the digital scanning result of sedimentary microfacies in the test area in the target block in the embodiment of the present application is shown;

[0028] Figure 7 The sedimentary microfacies plane map of the entire target area obtained by deep learning the training image in the embodiment of the present application is shown;

[0029] Figure 8 The sedimentary microfacies plane map that best conforms to the geological laws of the target block obtained after human-computer interaction in the embodiment of the present application is presented;

[0030] Figure 9 A schematic diagram of the application module is presented. DETAILED DESCRIPTION

[0031] The following is an introduction to the multi-point geostatistical sedimentary microfacies prediction method based on deep learning provided by the present application with reference to examples and drawings.

[0032] like Figure 1 As shown, the present application provides a multi-point geostatistical sedimentary microfacies prediction method based on deep learning, the method comprising:

[0033] Step 1: Data preprocessing: Multi-source data fusion, integrating well logging curves, core descriptions, seismic data and dynamic production data, and constructing the underlying data set. The results are as follows: Figure 2 As shown. Specifically including the following steps:

[0034] 1. There are 4032 wells in the study area (only 20 wells are listed in this application). The test area is located in the northwest corner of the study area, covering an area of ​​about one-fourth of the total area of ​​the study area. Collect and organize well logging data, including natural gamma, acoustic wave time difference, resistivity, density, mud content and other well logging curves. Input the basic geological data that has been organized, including well location data, boundary data, logging data, etc.

[0035] 2. Preprocess the logging curves, including unit correction, depth correction, removal of outliers, etc.

[0036] Based on the well logging curve, relevant geological parameters can be obtained. The results of mud content are as follows: Figure 3 As shown:

[0037] Natural gamma method to solve the mud content: As the mud content of the formation increases, the radioactivity adsorbed by the mud increases, and the readings of natural gamma logging increase. Therefore, natural gamma logging can be used to determine the mud content (Vsh) of the formation. The formula is as follows:

[0038] (1)

[0039] (2)

[0040] In formula (1) and (2): V sh —— formation mud content; GR min ——Minimum value of natural gamma in the logging curve, in API units; GCUR --coefficient; GR max ——The maximum value of natural gamma in the logging curve, in API; GR——The natural gamma logging value of the formation, in API; I sh ——Mud content of target layer.

[0041] The main purposes of using the natural gamma method to solve the mud content are as follows: ① Determine the lithology of the formation: Natural gamma logging can be used to divide the geological profile of the borehole and determine the mud content of the sandstone in the sandstone mudstone profile. In sedimentary rocks, the more mud it contains, the stronger its radioactivity. The natural gamma logging results can be used to distinguish rock formations with different mud contents. ② Evaluate reservoir characteristics: Mud content is an important basic parameter for the calculation and evaluation of mud sandstone formation parameters. It is closely related to the effective porosity, water saturation, bound water saturation, permeability and other parameters of the reservoir. Natural gamma logging can determine the mud content of the formation, thereby providing a basis for the calculation of reservoir parameters and then evaluating the oil and gas storage capacity of the reservoir. ③ Study the sedimentary environment: Natural gamma spectral logging can determine what kind of radioactive elements are contained in the formation based on the measured gamma-ray characteristic peak, and then study the sedimentary environment and distinguish clay minerals.

[0042] Natural potential method to solve the mud content: Natural potential logging gives a baseline value at the mudstone, but an anomaly appears in the permeability. The size of the anomaly is affected by the mud content. The higher the mud content, the smaller the natural potential anomaly. Therefore, the natural potential logging curve can be used to determine the mud content of the mud sandstone formation. The method is as follows:

[0043] (3)

[0044] (4)

[0045] In formulas (3) and (4): SP , SP sd , SP sh ——The natural potential value of rock, sandstone and mudstone based on a certain baseline, in mV; PSP ——Pseudo static natural potential value ,PSP=SP-SP sh Unit: mV; SSP ——static natural potential value, SSP=SP sd -SP sh , unit: mV; GCUR --coefficient; V sh —— formation mud content; I sh ——Mud content of target layer.

[0046] The main purposes of using the natural potential method to solve the mud content are as follows: ① Division of permeable strata: Natural potential logging is an effective tool for dividing permeable layers. Obvious natural potential anomalies are significant features of permeable layers, and the interface of permeable layers can be determined by the natural potential curve. The level of mud content will affect the size of the natural potential anomaly value. The higher the mud content, the smaller the natural potential anomaly value. Therefore, the natural potential method can be used to determine the mud content of mud sandstone strata, and then divide the permeable strata. ② Analysis of lithology changes: The natural potential curve is closely related to lithology, and the curve changes can reflect the changes in lithology. When the lithology of the stratum becomes finer and the mud content increases, the natural potential amplitude will decrease. Therefore, the natural potential method can be used to analyze the changes in stratum lithology, especially the changes in mud content. ③ Stratigraphic correlation: Stratigraphic correlation is the basis of all geological work, and the natural potential curve is an important data for stratigraphic correlation. Since the natural potential curve is closely related to lithology and the curve changes are more vivid, the changing characteristics of the natural potential curve can be combined to carry out stratigraphic correlation and determine the corresponding relationship between different stratigraphic units.

[0047] 3. Collect and organize core descriptions and lithology data, and digitally assign values ​​to different lithologies.

[0048] 4. Process seismic data, including attribute extraction and seismic phase division.

[0049] The ICA algorithm is used for the deep domain seismic attribute fusion method. Independent component analysis (ICA) is based on principal component analysis (PCA) and is a rapidly developing and growing data and signal analysis method. It mainly seeks and obtains non-singular transformations of multivariate data and tries to make the data components after non-singular transformations independent of each other.

[0050] Independent component analysis (ICA) can be defined as follows: Assume that there are n initial variables, defined as x 1 ,..,x n , where x i It can be expressed as:

[0051] x i =a i1 s 1 +a i2 s 2 +…+a in s n ,i=1,…,n (5)

[0052] In formula (5): X is the initial variable; a is the matrix position; s is the mixed signal set.

[0053] In the model, set the initial variable s i, are statistically independent of each other. This model is the most basic ICA model, namely the transient model. Define x as a random vector including the mixed signal x 1 ,..,x n The set of is s. Using vector and matrix representation, Equation 5 can also be written as:

[0054] x=As (6)

[0055] In formula (6): x is the initial variable; A is the matrix; s is the mixed signal set.

[0056] The model represented by Formula 6 is the ICA model. Independent component analysis starts by assuming that the initial variables are statistically independent of each other. Secondly, it is assumed that the mixing matrix is ​​an unknown square matrix. After the matrix A is estimated, its inverse matrix is ​​calculated and recorded as W. Then the independent components are:

[0057] s=Wx(7)

[0058] In formula (7): s is the mixed signal set; W is the inverse matrix; x is the initial variable.

[0059] ICA can be said to be one of the most commonly used methods for blind source separation, so it has great relevance. The nonlinear form, delay effect, and convolution effect of ICA should exist in practical applications, and in some cases there is interference noise, that is, the noisy independent component analysis model.

[0060] The notable characteristics of ICA are uncertainty and fuzziness. The waveform amplitude and signal arrangement order are as follows:

[0061] x=As=∑ j (1 / a j *h j )(s j a j )(8)

[0062] In formula (8): x is the initial variable; A is the matrix; s is the mixed signal set; a is the matrix position; h is the eigenvalue; i, j are the eigenvectors.

[0063] From formula 8, we can see that x remains unchanged and is not affected by a. j After unified variance processing, the real amplitude and complex amplitude have uncertain characteristics of sign and phase respectively. Four non-Gaussian independent signals are introduced as experimental data to implement the independent component analysis algorithm, which verifies the feasibility of the algorithm.

[0064] The SOM method is applied in the seismic phase analysis of waveform classification. SOM is a widely used unsupervised deep learning algorithm. The model consists of and only consists of an input layer and an output layer. The SOM training process is different from the reverse transfer based on the loss function of general neural networks. It uses a competitive learning strategy and relies on the competition between neurons to select the best matching neurons (Equation 9) to gradually optimize the network, and maintain the topological structure of the input space according to the proximity relationship function (Equation 10).

[0065] y i =argmindiff(x,w i )(9)

[0066] In formula (9), x is the input data; w is i ——weight coefficient; y i ——Output node.

[0067] dist(d)=(1-d 2 / θ 2 )exp(-d 2 / 2θ 2 ) (10)

[0068] In formula (10), θ is the scale factor; d is the topological distance between the weight w and the input attribute vector x.

[0069] By using the principle of SOM method, on the one hand, unsupervised division of seismic phases is performed to provide a seismic phase analysis result as a reference; on the other hand, the seismic phase analysis result is used as the basis for selecting labels and training data sets. Thus, a fast and automatic labeling method can be obtained, creating conditions for the next step of supervised deep learning algorithm.

[0070] 5. Use logging data, seismic data and core data to divide the sedimentary microfacies of all wells. Obtain the sedimentary microfacies of the test area that conforms to the basic laws of mechanistic theory. The results are as follows: Figure 4 The specific steps are as follows:

[0071] The logging curve is used to obtain relevant parameters such as porosity and permeability, and to divide the dominant phases. It is generally believed that areas with high porosity, high permeability and high oil and gas saturation are more suitable for oil and gas exploration.

[0072] Neutron logging is used to obtain the total porosity of the formation:

[0073] (11)

[0074] In formula (11): l Nma 、l f——neutron logging values ​​of rock skeleton and mudstone respectively; l N ——neutron logging value of target layer; f t ——total porosity of the formation; f N Neutron logging porosity.

[0075] The total porosity of the formation is obtained using density logging:

[0076] (12)

[0077] In formula (12): r ma ,r f ——density values ​​of rock skeleton and formation fluid respectively; r b ——Density logging value of target layer; f t ——total porosity of the formation; f D Density logging porosity.

[0078] The total porosity of the formation is obtained by using acoustic logging:

[0079] (13)

[0080] In formula (13): Δt ma , Δt f ——the acoustic time difference of rock skeleton and formation fluid respectively; Δt—— Acoustic time difference logging value of target layer; f t ——total porosity of the formation; f S Sonic logging porosity.

[0081] Calculate the total porosity of the formation using the neutron-density geometric mean:

[0082] (14)

[0083] In formula (14): f t ——total porosity of the formation; f D , f N ——are density and neutron porosity, respectively.

[0084] Calculate the effective porosity of the formation using neutron logging:

[0085] (15)

[0086] In formula (15): f t ——Total porosity of the formation; f N Neutron log porosity; l f ——Neutron log value of shale; V sh ——Shale content of the formation.

[0087] Calculate the effective porosity of the formation using density logging:

[0088] (16)

[0089] In formula (16): f ——Effective porosity of the formation; f D Density log porosity; V sh ——Shale content of the formation; r f ——Neutron log value of shale.

[0090] Calculate the effective porosity of the formation using acoustic logging:

[0091] (17)

[0092] In formula (17): f ——Effective porosity of the formation; Δt—— Acoustic travel time log value of the target layer; V sh ——Shale content of the formation; Δt ma 、 Δt f ——Acoustic travel times of the rock matrix and formation fluid respectively; CP——Acoustic compaction correction coefficient.

[0093] According to the capillary seepage law and Darcy's law, the relationship between permeability and porosity can be obtained as follows:

[0094] (18)

[0095] In formula (18): K ——Relationship between permeability and porosity; f ——Effective porosity of the formation; r——Capillary inner diameter of the rock.

[0096] For the calculation of formation water saturation, Archie's formula is mainly used. Archie's formula includes two formulas:

[0097] The Archie formula for the formation factor is:

[0098] (19)

[0099] In formula (19): F - Archie formula for formation factors; R 0 ——Resistivity of the formation 100% saturated with formation water, unit: Ω·m; R W ——Resistivity of formation water, unit: Ω·m; f ——porosity; m——rock cementation index; a——proportional coefficient related to rock.

[0100] The resistance increase factor is:

[0101] (20)

[0102] In formula (20): I ——resistance increase coefficient; R t ——Resistivity of oil and gas formation, unit: Ω·m; n——saturation index; b——constant related to rock; S W ——Water saturation.

[0103] 6. Digitally assign values ​​to accurately divided sedimentary microfacies;

[0104] 7. Prepare a sedimentary microfacies division report, recording the division process, results and conclusions in detail.

[0105] Sedimentary subfacies 1 (underwater distributary channel subfacies): ① Logging curve characteristics: The natural gamma curve usually shows a bell-shaped, funnel-shaped or box-shaped shape with medium to high amplitude; the natural potential curve usually shows a negative anomaly, and the curve shape is diverse, such as bell-shaped, funnel-shaped or box-shaped; the acoustic time difference curve in the underwater distributary channel subfacies section has a large acoustic time difference value due to the coarse sand particles and high porosity of the channel sand; the density logging curve shows a lower density value; the neutron logging curve shows a higher neutron porosity value. ② Core characteristics: It is mainly composed of sandstone, mostly light gray or gray in color, with coarse particles, good sorting, and obvious bedding structures, such as parallel bedding and cross-bedding; it often has scour-fill structures, parallel bedding, cross-bedding, etc., reflecting the high-energy sedimentary environment and the rapid changes in water flow; there are relatively few biological fossils, because the high-energy water flow environment is not conducive to the survival of organisms and the preservation of fossils. ③Seismic characteristics: usually manifested as a group of reflection waves with high amplitude and good continuity. This is because the sandstone section of the river channel phase has good reflection characteristics and can produce stronger reflection waves; the reflection waves may show tongue-shaped, belt-shaped and other characteristics on the seismic profile, reflecting the extension direction and sedimentary range of the river channel; the reflection waves usually have short-axis, strong amplitude reflection characteristics.

[0106] Sedimentary subfacies 2 (intra-channel front sheet sand subfacies): ① Logging curve characteristics: The natural gamma curve usually shows a medium-amplitude box or bell shape, reflecting the alternating distribution of sandstone and mudstone sections. The sandstone section has a lower natural gamma value, while the mudstone section has a higher natural gamma value. The natural potential curve shows a negative anomaly in the sandstone section, and the curve shape may be bell-shaped, funnel-shaped or box-shaped, reflecting that the sandstone section of this subfacies has a lower mud content and a more stable grain size characteristic. The sonic time difference curve has a larger sonic time difference value in the sandstone section due to coarser particles and higher porosity, while the opposite is true in the mudstone section. The density logging curve usually shows a lower density value in the sandstone section, while the density value is higher in the mudstone section, which is consistent with the changing trend of the natural gamma and natural potential curves. The neutron logging curve usually shows a higher neutron porosity value in the sandstone section, while it is lower in the mudstone section, reflecting that the sandstone section has a higher porosity. ② Core characteristics: mainly composed of light gray siltstone and muddy siltstone, with thin layers of gray mudstone in some parts, and the sand is relatively pure; wavy bedding and parallel bedding are common, which reflect the relatively stable sedimentary environment and water flow conditions; due to the increase of diagenesis, the sheet-like sand bodies of the source may present a lens-shaped sand-mud composite rhythm after superposition. ③ Seismic characteristics: usually manifested as a reflection wave group with medium amplitude and good continuity; the reflection wave may show tongue-shaped, belt-shaped and other characteristics on the seismic profile, reflecting the extension direction and sedimentary range of the front edge sheet-like sand in the river channel, and the reflection wave has a short axis and medium amplitude reflection characteristics.

[0107] Sedimentary subfacies 3 (sheet sand subfacies at the front of the river): ① Logging curve characteristics: The natural gamma curve usually shows a low-amplitude finger shape; the natural potential curve also shows a low-amplitude finger shape in the sandstone section. ② Core characteristics: It is mainly composed of light gray siltstone and muddy siltstone, with a relatively simple lithology and pure sand quality. Due to the increase in diagenesis, the sheet sand bodies of the source are superimposed to present a lens-shaped sand-mud composite rhythm. ③ Seismic characteristics: The sheet sand subfacies at the front of the river and the sheet sand at the front of the river are similar in terms of reflection wave group characteristics, reflection wave morphology, and phase and frequency of reflection waves, but the reflection wave group of the sheet sand at the front of the river is more continuous and stable because it is less affected by the action of the river.

[0108] Sedimentary subfacies 4 (subfacies between underwater distributary channels): ① Logging curve characteristics: The natural gamma curve shows the characteristics of "two lows and one high", that is, there is a mudstone section with a high natural gamma value between two sandstone sections with low natural gamma values; the natural potential curve shows a low-amplitude finger or box shape; the acoustic time difference curve shows a smaller value; the density logging curve shows a higher density value; the neutron logging curve shows a lower neutron porosity value. ② Core characteristics: It is mainly composed of light gray siltstone, argillaceous siltstone and mudstone, with a relatively simple lithology and pure sand, and thin layers of gray mudstone may be sandwiched in some places. ③ Seismic characteristics: The reflection wave group shows a medium-weak amplitude continuous, parallel-subparallel reflection seismic phase; the reflection wave has medium and low frequency reflection characteristics.

[0109] Step 2: Geological rule encoding module: convert geological laws such as sedimentary phase sequence and sand body superposition pattern into digital constraint matrix and embed it into feature extraction process.

[0110] The law of sedimentary phase sequence means that only those phases and phase regions that can be observed to be adjacent to each other can overlap natively. This is the famous Walter phase law, also known as the phase contrast principle. The sequential changes of adjacent sedimentary phases in the longitudinal direction are consistent with the sequential changes in the lateral direction, that is, the changes of adjacent sedimentary phases in the lateral direction or longitudinal direction can be predicted based on the changes of adjacent sedimentary phases in the longitudinal direction or lateral direction.

[0111] The random forest model is used as a digital constraint matrix to constrain geological laws such as sedimentary phase sequence and sand body superposition pattern.

[0112] In a random forest, each tree in the ensemble is built from samples drawn with replacement (i.e., bootstrap samples) from the training set. In addition, when splitting each node in the process of building the tree, the best split is found from all input features or a random subset. The purpose of these two randomnesses is to reduce the variance of the forest estimator. Because, individual decision trees usually exhibit high variance and tend to overfit. The randomness injected into the forest produces decision trees with somewhat decoupled prediction errors. By averaging these predictions, some of the errors can be offset. Random forests reduce variance by combining different trees, sometimes at the cost of a slight increase in bias. In the process of predicting recovery, the variance reduction is usually significant, thus producing an overall better model.

[0113] Suppose there is a random vector tree I , so that the prediction tree h(X,Θ) Use numerical values. Assume that the training set is independent of the random vector X,Y , then the predicted value h(X) The mean square generalization error is:

[0114] (twenty one)

[0115] In formula (21), E X,Y —— X and Y The mathematical expectation of Y ——output vector; h(X) ——Predicted value.

[0116] By taking the average value of k trees, we can form { h(X,Θ k )}, as the number of trees in the forest approaches infinity:

[0117] (twenty two)

[0118] In formula (22), E X,Y —— X and Y The mathematical expectation of Y ——output vector; E Θ --vector I expectations; h(X, I k ) --average value.

[0119] Formula (22) can be expressed as PE*(forest) , as the generalization error of the forest. Therefore, the average generalization error of the tree is defined as:

[0120] (twenty three)

[0121] In the formula, PE* ——the generalization error of the tree; E Θ --vector I expectations; E X,Y —— X and Y The mathematical expectation of Y ——output vector; h(X,Θ k ) --average value.

[0122] Assume that for all vector trees I , E Y =E x h(X,Θ k ) have

[0123] (twenty four)

[0124] In formula (24): PE*(forest) ——Generalization error of the forest; r —weighted correlation between residual and generalization error; h(X,Θ k ) --average value; E Y ——Y’s expectations; E X ——X's expectations

[0125] Random forest by I The factor reduces the average error of the tree. In the specific process of predicting the recovery rate, a custom module is coded in python, and the sklearn.ensemble.RandomForestRegressor interface is called to combine classifiers by averaging their probability predictions, instead of letting each classifier vote for a single class. This part is different from the original concept.

[0126] Correlation analysis:

[0127] Use the Pearson correlation coefficient to calculate multivariate linear correlation:

[0128] (25)

[0129] In formula (25): r x,y ——correlation coefficient; n——sample size;x i , y i ——No. x variables and y variables. Pearson correlation coefficient 0~0.3 is weak correlation, 0.3~0.6 is moderate correlation, and 0.6~1 is strong correlation. Use python to encode custom modules and call seaborn interface to assist in implementation;

[0130] Normalization or standardization:

[0131] (26)

[0132] (27)

[0133] In formulas (26) and (27): (X max ,X min ) — characteristic range; X std ——Discrete standardization; X ——Input characteristic value; X scaled ——Relative scaling of features; m ——sample mean; S - standard deviation; Z --standardization.

[0134] Through the standardization calculation, the feature values ​​are centered and scaled, and the mean and standard deviation are stored. Use python to encode a custom module, call sklearn.preprocessing.StandardScaler, and use the PolynomialFeatures interface to assist in the implementation.

[0135] Step 3: Build a deep learning model architecture: including a multi-scale feature fusion network.

[0136] Multi-scale feature fusion network: The backbone network adopts an improved ResNet-50 architecture and adds a hole convolution layer to capture multi-scale spatial features;

[0137] The residual neural network (ResNet) is a deep neural network whose main feature is the use of a residual block structure. This structure solves the gradient vanishing and gradient exploding problems in deep neural networks by adding cross-layer connections. This structure enables ResNet to effectively train deep neural networks.

[0138] This paper proposes a class imbalance image classification algorithm based on improved ResNet-50. The overall framework of the algorithm is divided into a regional significant information suppression and feature fusion module (RSIS-FFM) and an algorithm module for alleviating class imbalance (ACI). Based on the residual network ResNet, the feature output of the middle layer of the network is used as the module input in the RSIS-FFM module. Average pooling is applied in the channel direction to obtain the regional attention map. The salient regions are extracted with a certain window size and the features of the most salient regions are suppressed, so that the next layer of the network pays attention to more discriminative regions, and the feature fusion of the most salient region feature map is performed. In the ACI module, the network is trained using decoupled representation learning and classifier methods. In the first stage, the feature extraction network and the classifier are combined. In the second stage, the classifier is fine-tuned and trained in combination with class balanced sampling. Improvements are made in the class balance loss to improve the classification accuracy of low- and medium-frequency category images.

[0139] By applying average pooling in the channel direction to obtain the attention feature map, most areas in the feature map will not contain highly significant features, but some areas have significant features, and the features of these areas can be extracted for feature fusion to further improve the network's representation ability. Suppressing the feature information of the extracted feature-significant areas will help the network learn the distinguishable features of other more significant areas in the next stage, increasing the distribution of distinguishable areas of the network.

[0140] First, given the input feature map X∈R C×H×W , where C、H、W are the number of channels, height, and width of the input feature map respectively. X The position information of the salient features in the channel is obtained by applying average pooling in the channel direction and using the activation function to transform the input feature map X∈R C×H×W Convert to attention map X att ∈R C×H×W , and get the attention score. In the attention score, each value represents the significance of the feature information at each pixel position.

[0141] (28)

[0142] In formula (28), X att ——Feature attention map; f(x)——channel-wise average pooling function; Sigmoid ( ) —— Sigmoid Activation function.

[0143] Fine-grained image classification tasks generally have small differences between classes and large differences within classes. The network often needs to extract more subtle features to make judgments. Therefore, we can further learn feature maps at different levels to further improve the expressiveness of features. Set the window size to 7×7 and add the feature attention map X att It is divided into multiple regions and implemented using a convolution layer with a convolution kernel size of 7×7. This method generates a set of attention strengths after processing. P∈R 1×(H / 7)×(W / 7) , the set contains (H / 7)×(W / 7) The attention intensity of regions of equal size.

[0144] (29)

[0145] In formula (29): P ——Feature attention strength; Sigmoid( ) —— Sigmoid Activation function; conv7_7 ( ) ——Convolutional layer operation with a convolution kernel size of 7×7; X att ——Feature Attention Map.

[0146] ResNet-50 is used as the feature extraction network for convolution The output feature map of convolution is visualized and displayed in the form of a heat map. The output feature map size is (1024, 28, 28). After partitioning, a 4×4 grid feature map is generated, which generates a set of attention strengths containing 16 feature regions. P ={ P i |i=1, 2, , 16}, region set P The size is (1, 4, 4). In this case, each region has an attention strength, which effectively approximates the spatial distribution of the most distinctive regions.

[0147] Regional attention strength set P The position with the largest attention intensity is selected (m, n), and the corresponding position mapped to the output is:

[0148] IA={X(x,y)|7m≤x≤7(m+1), 7n≤y≤7(n+1), x∈N, y∈N} (30)

[0149] In formula (30): IA∈ RC×7×7 ——Input feature map corresponding to the position (m, n) with the maximum attention intensity Regional feature map in ; X ——Input feature map; x,y ——Eigenvector.

[0150] The attention mechanism module is introduced to dynamically weight the contribution of different deposition units; channel attention is designed to model the correlation between different channels and feature maps. In essence, it automatically obtains the importance of each feature channel through network learning, and finally assigns different weight coefficients to each channel to strengthen the attention to important features. The use of one-dimensional convolution can effectively reduce the amount of calculation of the attention module.

[0151] In addition to the correlation between channels, the spatial distribution of features in a single channel feature map is also very important. By modeling the spatial information of the feature map to obtain the correlation of each pixel in the space, the feature map can focus more on the target information and find the target location more easily.

[0152] Step 4, sample training mechanism: By using multi-point geostatistics methods, data templates are used to scan training images to obtain data events to reflect the corresponding geological patterns. The frequency of occurrence of different data events is approximately the joint distribution probability of multiple points in space. The idea of ​​multi-point geostatistics is to use limited geological data, establish training images through sedimentological analysis, and find the optimal data events from the training images under the constraints of conditional data as the basis for simulation sampling of the points to be estimated. It can be considered that the training image is one of the key factors that determine the effectiveness of the simulation. Specifically, the following steps are included:

[0153] 1. Select the appropriate range, use the accurate sedimentary microfacies generated in the data preprocessing and feature engineering module, and combine human-computer interaction to select the appropriate range that best meets geological rationality.

[0154] 2. Digitally scan the small-scale sedimentary microfacies plane map that has been obtained. Generate a rasterized result with a geographic coordinate system. The result is as follows: Figure 5 shown.

[0155] 3. Use the digitally scanned sedimentary microfacies plane map as a sample and use a high-order compatibility algorithm to obtain the most matching training image. The details are as follows:

[0156] By scanning the training images with data samples, the number of repetitions of data events is calculated. R i,j , and then determine the relative frequency of data events in different training images F i,j Through normalization, the method calculates the relative compatibility of each training imageC j , in order to evaluate the matching degree between the training images and the actual geological data. However, this method mainly focuses on the number of conditional data points, while ignoring the differences in the spatial distribution of data points, which may lead to deviations in the compatibility assessment between the training images and the actual data events.

[0157] Among them, the relative frequency F i,j is a key parameter that quantifies how often a particular data event occurs in a training image. Specifically, the relative frequency F i,j Describes the i Data events in j The number of times the data event appears repeatedly in the training images is related to the number of times the data event appears in all t The ratio of the total number of repetitions in the training image. This ratio is an important indicator to measure the matching degree between the training image and the actual geological model. The calculation formula is as follows:

[0158] (31)

[0159] In formula (31), R i,j ——No. Data events in The number of repetitions in the training images; Denominator - the sum of the number of repetitions of the data event in all training images; i,j --constant.

[0160] This calculation enables us to assess how representative each training image is for a specific data event, providing more accurate input for predictions across the region.

[0161] Relative compatibility C j It is an important indicator to measure the matching degree between training images and data events. j The ratio of the sum of the relative frequencies of all data events in a training image to the sum of the relative frequencies of the corresponding data events in all training images. This metric helps to evaluate the representativeness and applicability of training images. Specifically, relative compatibility C j The calculation formula is as follows:

[0162] (32)

[0163] In formula (32), C j ——Relative compatibility; F i,j ——No. i Data events inj relative frequency in the training images; n — the total number of data events; t ——The total number of training images; i,j --constant.

[0164] By calculation C j , we can quantify the consistency and reliability of each training image in describing the characteristics of the data event distribution.

[0165] Absolute compatibility M j It is an intuitive measure that directly reflects the j The metric is evaluated by a simple binary system, that is, if the first i Data events in training images, then the corresponding indicator variable Y i,j is assigned a value of 1; if it does not appear, Y i,j is 0. The absolute compatibility of the training image is then determined by calculating the ratio of the total number of times all data events appear in the training image to the total number of data events. This ratio provides a quantitative perspective to assess the completeness and accuracy of the training image in reproducing geological phenomena. Absolute compatibility M j The calculation formula is as follows:

[0166] (33)

[0167] In formula (33), M j ——Absolute compatibility; Y i,j - indicator variables; — the total number of data events; i,j --constant.

[0168] Absolute compatibility M j The higher the value of , the better the match between the training image and the data event, thus providing a more reliable reference for geostatistical modeling. The high-order compatibility method is a technique used in geostatistics to evaluate the match between training images and actual geological data. The core idea of ​​this method is that the higher the compatibility of the training image, the better its consistency with the data event.

[0169] By scanning the training images with data templates and calculating the number of repetitions of data events, the relative frequencies of data events in different training images are determined. Through normalization, this method calculates the relative compatibility of each training image to evaluate the matching degree between the training image and the actual geological data. However, this method mainly focuses on the number of conditional data points and ignores the differences in the spatial distribution of data points, which may lead to deviations in the compatibility assessment between the training images and the actual data events. Among them, relative frequency is a key parameter that quantifies the frequency of occurrence of a specific data event in the training image.

[0170] 4. Using the acquired training images as geological models and using the data event repetition probability statistical algorithm, a sedimentary microfacies plane map matching the actual area is obtained. The results are as follows Figure 6 As shown, the details are as follows:

[0171] This method aims to reflect the distribution characteristics of specific data events in the training image by quantifying their occurrence frequency. In the specific operation, the conditional data is first scanned using the specified template to identify n A collection of data events DE . Then, in each training image, count the Data events DE i The number of occurrences of is recorded as R i,j Based on these data, we further calculate the distribution characteristics of data events in different training images, including the variance of data event repetition probability. s j No match rate with data events UNF j These indicators, namely the variance of the repetition probability and the no-match rate, provide a more comprehensive and accurate evaluation criterion for the optimization of training images.

[0172] In geostatistics, the number of repetitions of a single data event. R i,j The ratio between the sum of the number of repetitions of all data events in the training image is defined as the repetition probability of the data event. FT i,j This metric is a key parameter for measuring the frequency of occurrence of specific data events in training images, and it helps to evaluate the reliability of training images in simulating geological phenomena. Specifically, the probability of data event repetition FT i,j The calculation formula is as follows:

[0173] (34)

[0174] In formula (34), R i,j——No. i Data events in j The denominator is the sum of the number of repetitions of all data events in the training images. i,j --constant.

[0175] This formula allows us to quantify the frequency of occurrence of each data event in the training images, thereby providing more accurate input for geostatistical modeling.

[0176] Repeat probability variance s j is an important statistic that measures the degree of dispersion of the probability of repetition of data events in the training image. This indicator reflects the uniformity of the distribution of different data events in the training image. Specifically, the variance of the repetition probability s j The calculation involves the probability of repetition of each data event FT i,j Its average repetition probability in training images` FT j The calculation formula is as follows:

[0177] (35)

[0178] In formula (35), FT i,j ——No. i Data events in j The probability of repetition in training images; FT j ——The average value of the repetition probability of all data events in the training image; n ——The total number of data events.

[0179] Repeat probability variance s j The smaller the value of , the more uniform the distribution of data events in the training image, and vice versa, the more discrete the distribution.

[0180] In geostatistics, the selection of training images depends not only on the repetition probability of data events, but also on the matching degree of data events. D i,j , used to mark whether a matching data event is found in the training image. i Data events in j training images are successfully matched. D i,j If the data event is not matched, it is recorded as 1; otherwise, it is recorded as 0. Based on these indicator values, the proportion of events with no matching data can be calculated, that is, the data event no matching rate UNFj , and its calculation formula is as follows:

[0181] (36)

[0182] In formula (36), UNF j ——Event no match rate; D i,j - indicated value; n — the total number of data events; i, j --constant.

[0183] No match rate UNF j The lower the value of, the more data events in the training image match the actual geological pattern, which indicates that the geological pattern of the training image is richer. The smaller the value of , the more stable the distribution of data events in the training image, which further indicates that the training image is more consistent with the actual geological model.

[0184] In geostatistics, the ratio between the number of repetitions of a single data event and the sum of the number of repetitions of all data events in the training image is defined as the repetition probability of the data event. This indicator is a key parameter for measuring the frequency of occurrence of specific data events in the training image, which helps to evaluate the reliability of the training image in simulating geological phenomena. Therefore, an ideal training image should have a lower no-match rate and a smaller variance of the repetition probability. Such a training image can more accurately reflect geological phenomena and provide a more reliable basis for geostatistical modeling.

[0185] Step 5: Deep learning: Combined with relevant algorithms, using training images as samples, we can obtain the geological model that best matches the geological conditions of the study area. Figure 7 In the deep learning process of this example, the workload of manual interpretation was reduced by 62%, the plane prediction error rate was ≤10%, the hardware and improved architecture increased the prediction efficiency by 3 times, the time consumption of deep learning was reduced by 50%, the prediction accuracy was improved by 40% compared with the baseline method, and the model iteration cycle was shortened to 1 / 4 of the conventional process.

[0186] Human-computer interaction to obtain the final results: Using the deep learning results as constraints, combined with the basic geological data of the study area, the final sedimentary microfacies plane map of the test area was obtained through interactive correction in the geological expert system. The results are as follows: Figure 8 As shown in the figure, the human-computer interaction mechanism improves the rationality of the spatial configuration of the deep learning prediction model and the sedimentation law by 37% through the sedimentation pattern knowledge graph constraint.

[0187] This application breaks through the bottleneck of traditional methods for characterizing complex spatial patterns by establishing an intelligent prediction model with geological constraints. The following is a systematic summary from the two levels of technical implementation and innovative advantages:

[0188] ① Technical implementation: Construct a deep learning model architecture, including a multi-scale feature fusion network, with the backbone network adopting an improved ResNet-50 architecture as the feature extraction network to simultaneously extract the local microscopic characteristics and regional macroscopic distribution patterns of sedimentary microfacies; introduce an attention mechanism to achieve feature enhancement of key geological interfaces.

[0189] Table 1 Advantages of this application compared with the prior art

[0190]

[0191] ②Innovative advantages: Break through the limitations of traditional end-member methods, realize dynamic matching of geological model library and data characteristics through probabilistic embedding layer; improve multi-point geostatistical algorithms to make the prediction results conform to the basic laws of geology; support geological experts to correct the prediction results in real time.

[0192] This method realizes the quantitative characterization of all elements of sedimentary microfacies "morphology-structure-probability" for the first time through the essential integration of deep learning and geostatistics, providing a new generation of technical paradigm for the intelligent study of complex sedimentary systems. Compared with similar international technologies (such as Schlumberger's Petrel geological modeling software), it has certain advantages in key indicators such as prediction accuracy and geological rationality.

[0193] This application includes the following modules: Figure 9 As shown, 1. Data preprocessing and feature engineering module, used to fuse multi-source data, integrate logging curves, core descriptions, seismic data and dynamic production data, and construct underlying data sets. 2. Geological rule encoding module, used to convert geological laws such as sedimentary phase sequence and sand body superposition pattern into digital constraint matrices and embed them into feature extraction processes. 3. Deep learning module, used to combine high-order compatibility algorithms and data repetition probability statistics with multi-point geostatistics to establish a prediction model. 4. Human-computer interaction correction module, combines the traditional geological experience of drawing sedimentary microfacies with deep learning results to obtain the results that best conform to the basic laws of geology. 5. Output module, used to convert the final results into content that can be needed in the actual exploration process.

[0194] The working principle of the multi-point geostatistical sedimentary microfacies prediction device based on deep learning provided in the present application is to organize the required basic geological data through the data preprocessing and feature engineering modules; activate the special geological laws required for the target block through the geological rule coding module; establish training images through the deep learning module to obtain the sedimentary pattern that conforms to the geological laws of the target block; obtain the final sedimentary microfacies plane map of the target block through the human-computer interaction correction module combined with the traditional sedimentary microfacies drawing experience; and select the dominant phase through the output module, and use the obtained results as the goal of oil and gas exploration and development.

Claims

1. A multi-point geostatistical sedimentary microfacies prediction method based on deep learning, comprising the following steps: Step 1: Data preprocessing: Multi-source data fusion, integrating well logging curves, core descriptions, seismic data and dynamic production data to build the underlying data set; (1) Collect and organize logging data, including natural gamma, acoustic time difference, resistivity, density, and shale content logging curves; (2) Preprocess the logging curves, including unit correction, depth correction, and removal of outliers; (3) Collect and organize core descriptions and lithology data, and assign digital values ​​to different lithologies; (4) Processing of seismic data, including attribute extraction and seismic phase division; (5) Sedimentary microfacies classification of all wells using well logging data, seismic data, and core data; (6) Digitally assign values ​​to accurately divided sedimentary microfacies; (7) Prepare a sedimentary microfacies division report, recording the division process, results and conclusions in detail; Step 2, geological law encoding module: geological laws are converted into digital constraint matrices and embedded in the feature extraction process; Select the corresponding geological laws according to the geological characteristics of the target block; use the random forest algorithm to handle missing values ​​when converting geological laws into digital constraint matrices, and use the Pearson correlation coefficient to calculate the multivariate linear correlation coefficient; The geological law coding specifically includes: constructing a sedimentary phase contact relationship matrix to define the probability of phase transition from river channel to natural levee to breach fan; quantifying the source direction into an 8-azimuth rose diagram tensor; using constraint matrix coding to represent the classification characteristics of basin evolution stages; Step 3: Build a deep learning model architecture: including a multi-scale feature fusion network; use the improved ResNet-50 architecture as the feature extraction network in the backbone network, and add a hole convolution layer to capture multi-scale spatial features; implement a transfer learning strategy and initialize parameters with a pre-trained model; set a dynamic learning rate adjustment mechanism, and set the initial learning rate and the decay coefficient per cycle according to the example; Step 4: Sample training mechanism: by using multi-point geostatistics method, using data template to scan training images to obtain data events to reflect the corresponding geological model; the frequency of occurrence of different data events is approximately the joint distribution probability of multiple points in space; specifically including the following steps: (1) Select an appropriate range, using the accurate sedimentary microfacies of the small blocks generated in the data preprocessing and feature engineering module, combined with human-computer interaction to select an appropriate range that best meets geological rationality; (2) Digitally scan the small-scale sedimentary microfacies plane map that has been obtained to generate a rasterized result with a geographic coordinate system; (3) Using the digitally scanned sedimentary microfacies plane map as a sample, a high-order compatibility algorithm is used to obtain the most matching training image; (4) Using the acquired training images as geological models and using the data event repetition probability statistical algorithm, a sedimentary microfacies plane map matching the actual area is obtained; Step 5. Deep learning: Using training images as samples, obtain the geological model that best meets the geological conditions of the study area, and obtain the final result through human-computer interaction: Using the deep learning results as constraints, combined with the basic geological data of the study area, interactive corrections are made in the geological expert system to obtain the final sedimentary microfacies plane map of the test area.

2. The multi-point geostatistical sedimentary microfacies prediction method based on deep learning according to claim 1, characterized in that: In step 3, an improved ResNet-50-based image classification algorithm for class imbalance is proposed. The overall framework of the algorithm is divided into a regional significant information suppression and feature fusion module (RSIS-FFM) and an algorithm module for alleviating class imbalance (ACI). The residual network ResNet is used as the basic model. In the RSIS-FFM module, the feature output of the middle layer of the network is used as the module input. Average pooling is applied in the channel direction to obtain the regional attention map. The significant area is extracted with a certain window size and the most significant regional features are suppressed. In the ACI module, the decoupled representation learning and classifier methods are used to train the network. In the first stage, the feature extraction network and the classifier are combined. In the second stage, the classifier is fine-tuned with class-balanced sampling. An attention mechanism module is introduced to dynamically weight the contribution of different deposition units.

3. The multi-point geostatistical sedimentary microfacies prediction method based on deep learning according to claim 1, characterized in that: In step 4, a high-order compatibility algorithm and data repetition probability statistics method are used to predict sedimentary microfacies using training images as search samples using multi-point geostatistics; microscopic features of sedimentary structures are extracted using a random field algorithm; The graph attention network is used to capture the macroscopic distribution of regional sedimentary patterns; ultimately a training image set that conforms to geological statistical laws is generated.

4. The multi-point geostatistical sedimentary microfacies prediction method based on deep learning according to claim 1, characterized in that: In step 4 (3), the number of repetitions of the data event R is calculated by scanning the training image as a data sample. i,j , and then determine the relative frequency F of data events in different training images i,j ; Through normalization, this method calculates the relative compatibility C of each training image j , in order to evaluate the matching degree between the training images and the actual geological data; however, this method mainly focuses on the number of conditional data points and ignores the differences in the spatial distribution of data points, which may lead to deviations in the compatibility assessment between the training images and the actual data events; Among them, the relative frequency F i,j is a key parameter that quantifies how often a particular data event occurs in a training image; specifically, the relative frequency F i,j Describes the ratio of the number of repetitions of the i-th data event in the j-th training image to the total number of repetitions of the data event in all t training images; this ratio is an important indicator to measure the matching degree between the training image and the actual geological model, and its calculation formula is as follows: In formula (31), R i,j ——The number of repetitions of the i-th data event in the j-th training image; Denominator——The sum of the number of repetitions of the data event in all training images; i, j——Constants; Relative compatibility C j It is an important indicator to measure the matching degree between training images and data events; it reflects the ratio of the sum of relative frequencies of all data events in the jth training image to the sum of relative frequencies of corresponding data events in all training images; specifically, relative compatibility C j The calculation formula is as follows: In formula (32), C j ——Relative compatibility; F i,j ——relative frequency of the ith data event in the jth training image; n——the total number of data events; t——the total number of training images; i, j——constants; Absolute compatibility j is an intuitive measure that directly reflects the occurrence of data events in the jth training image; this indicator is evaluated by a simple binary system, that is, if the i-th data event appears in the j-th training image, the corresponding indicator variable Y i,j is assigned a value of 1; if it does not appear, Y i,j is 0; then, the absolute compatibility of the training image is determined by calculating the ratio of the total number of times all data events appear in the training image to the total number of data events; the absolute compatibility M j The calculation formula is as follows: In formula (33), M j ——Absolute compatibility; Y i,j ——indicator variable; n——the total number of data events; i, j——constants; Absolute compatibility j The higher the value of, the better the match between the training image and the data event, the higher the compatibility of the training image, and the better its consistency with the data event; By scanning the training images for data templates and calculating the number of repetitions of data events, the relative frequencies of data events in different training images are determined. Through normalization, the method calculates the relative compatibility of each training image to evaluate the matching degree between the training images and the actual geological data. Among them, relative frequency is a key parameter that quantifies the frequency of occurrence of specific data events in training images.

5. The multi-point geostatistical sedimentary microfacies prediction method based on deep learning according to claim 1, characterized in that: In step 4 (4), this method aims to reflect the distribution characteristics of specific data events in the training image by quantifying their occurrence frequency; In the specific operation, the conditional data is first scanned using the specified template to identify a set DE of n data events; Then, in each training image, the i-th data event DE is counted i The number of occurrences of i,j ; Based on these data, we further calculate the distribution characteristics of data events in different training images, including the variance of data event repetition probability σ j No match rate UNF for data events j ; In geostatistics, the number of repetitions of a single data event, R i,j The ratio between the sum of the number of repetitions of all data events in the training image is defined as the repetition probability FT of the data event. i,j ; This indicator is a key parameter to measure the frequency of occurrence of specific data events in training images, which helps to evaluate the reliability of training images in simulating geological phenomena; specifically, the data event repetition probability FT i,j The calculation formula is as follows: In formula (34), R i,j ——The number of repetitions of the i-th data event in the j-th training image; Denominator——The sum of the number of repetitions of all data events in the training image; i, j——Constants; Repeat probability variance σ j is an important statistic that measures the degree of discreteness of the probability of repetition of data events in the training image; this indicator reflects the uniformity of the distribution of different data events in the training image; specifically, the variance of the repetition probability σ j The calculation involves the repetition probability FT of each data event i,j The average probability of its repetition in the training images The difference between; the calculation formula is as follows: In formula (35), FT i,j ——The probability of repetition of the i-th data event in the j-th training image; FT j ——the average repetition probability of all data events in the training image; n——the total number of data events; Repeat probability variance σ j The smaller the value of, the more uniform the distribution of data events in the training image, and vice versa, the more discrete the distribution; Introducing indicator value D i,j , used to mark whether a matching data event is found in the training image; if the i-th data event is successfully matched in the j-th training image, D i,j is recorded as 1; otherwise, it is recorded as 0. Based on these indicator values, the proportion of unmatched data events can be calculated, that is, the data event unmatched rate UNF j , and its calculation formula is as follows: In formula (36), UNF j ——Event no match rate; D i,j ——indicative value; n——the total number of data events; i, j——constants; No match rate UNF j The lower the value of , the more data events in the training image match the actual geological pattern, which indicates that the geological pattern of the training image is richer; at the same time, the repetition probability variance σ j The smaller the value of , the more stable the distribution of data events in the training image.

Images