A shale oil sweet spot prediction method and system based on multi-attribute fusion

Abstract

This invention relates to a method and system for predicting shale oil sweet spots based on multi-attribute fusion. Under lithofacies constraints, the invention establishes a three-dimensional distribution model of porosity, permeability, oil saturation, total organic carbon content, and brittleness index, constructing a feature set including single attributes and attribute interaction terms. Based on a machine learning model, a nonlinear mapping relationship between multiple attributes and oil production is established, and the contribution value of each attribute and interaction term is calculated using the SHAP method to determine the corresponding weight coefficients. On this basis, a three-dimensional prediction model for the sweet spot coefficient is constructed, and the horizontal well trajectory is dynamically optimized according to the distribution of the sweet spot coefficient. Multi-attribute nonlinear modeling and objective determination of weights improve the accuracy and reliability of shale oil sweet spot prediction, effectively guiding horizontal well deployment and trajectory adjustment, and significantly improving sand body encounter rate and production level.

Description

Technical Field

[0001] This invention relates to the field of oil and gas exploration and development technology, specifically to a method and system for predicting shale oil sweet spots based on multi-attribute fusion. Background Technology

[0002] Shale oil refers to petroleum resources hosted in organic-rich shale formations, characterized by self-generation and self-storage, and integrated source and reservoir formation. The Chang 7 oil-bearing formation in the Ordos Basin is an important target formation for the exploration and development of continental shale oil in my country. Its reservoirs are mainly composed of fine sandstone and siltstone, characterized by tight formations with low porosity and low permeability. The matrix permeability is generally 0.01~1mD, and the porosity is 5%~18%.

[0003] Shale oil sweet spots refer to favorable sections within shale formations that possess relatively high production capacity. Accurate prediction of the spatial distribution of sweet spots is crucial for selecting horizontal well targets, optimizing well trajectories, and formulating development plans. The formation of shale oil sweet spots is controlled by multiple geological factors, primarily including reservoir properties (porosity and permeability), oil-bearing properties (oil saturation), source rock quality (total organic carbon content), and rock mechanical properties (brittleness index), among which complex nonlinear interactions exist.

[0004] Currently, the technical solutions for predicting the "sweet spot" of shale oil have the following main shortcomings:

[0005] First, single-attribute parameter modeling leads to incomplete predictions. Existing techniques typically select only reservoir physical properties (such as porosity or permeability) or source rock parameters (such as... Modeling based on a single attribute in shale oil production data cannot fully reflect the actual situation where the "sweet spot" of shale oil is controlled by multiple factors such as reservoir performance, source rock quality, and engineering modifiability, resulting in a large deviation between the prediction results and actual production data.

[0006] Second, deterministic modeling methods struggle to characterize reservoir spatial heterogeneity. Existing deterministic modeling methods based on simple mathematical interpolation functions do not consider the spatial correlation of reservoir parameters, failing to accurately describe the heterogeneous characteristics of the reservoir and resulting in significant inter-well reservoir prediction errors.

[0007] Third, objective-based stochastic modeling methods are ill-suited to high-density well network conditions. For shale oil reservoirs with diverse lithofacies and strong heterogeneity, objective-based stochastic simulation methods struggle to determine the geometric parameters of each lithofacies. Furthermore, under high-density well network conditions, simulation results often fail to conform to well data, reducing modeling accuracy.

[0008] Fourth, there is a lack of systematic methods for multi-attribute fusion and weight quantification. Existing technologies lack a systematic method for quantitatively evaluating the contribution of each attribute parameter to the shale oil "sweet spot," making it impossible to objectively determine the weight ratio of each attribute parameter in sweet spot prediction. This results in inconsistent comprehensive evaluation standards for sweet spots and unstable prediction accuracy.

[0009] Fifth, there is a lack of a complete technical solution for directly applying 3D prediction models to horizontal well trajectory optimization. Existing technologies typically separate geological modeling from horizontal well design into independent workflows, failing to provide direct guidance from 3D sweet spot prediction models for horizontal well target selection, real-time trajectory verification, and drilling encounter rate optimization, resulting in low drilling encounter rates in sand and oil layers of horizontal wells.

[0010] Furthermore, existing technologies mostly employ deterministic modeling methods or spatial models based on simple mathematical interpolation, failing to fully consider the heterogeneity of reservoir space and the coupling relationships between properties, further increasing prediction errors and uncertainties. Therefore, existing technologies cannot provide a high-precision shale oil sweet spot prediction method applicable to engineering optimization. Summary of the Invention

[0011] The purpose of this invention is to overcome the shortcomings of the prior art and provide a shale oil sweet spot prediction method based on multi-attribute fusion, which achieves accurate prediction of shale oil sweet spots.

[0012] To achieve the above objectives, the present invention adopts the following technical solution: A method for predicting shale oil sweetness based on multi-attribute fusion includes the following steps: Step 1: Based on the drilling data, logging data and geological stratification data of the target block, establish a three-dimensional structural model as a spatial framework; on the basis of the structural model, establish a lithofacies model using the truncated Gaussian simulation method, wherein the lithofacies model contains the three-dimensional spatial distribution of at least two lithofacies types; Step 2: Select attribute parameters that reflect the sweet spot characteristics of shale oil. These attribute parameters include at least porosity, permeability, oil saturation, brittleness index, and total organic carbon content. Based on core analysis data, calibrate the logging interpretation model for each attribute parameter and perform logging interpretation calculations on all wells in the target block. Under the constraints of the lithofacies model, establish a three-dimensional distribution model for each attribute parameter using the sequential Gaussian simulation method. Step 3: Using wells with stable production data that are already in production within the target block as the data source, extract the values ​​of each attribute parameter from the three-dimensional distribution model of each attribute parameter along the perforation depth range of each sample unit. Calculate the weighted average of each attribute parameter within each sample unit as the feature value. Perform Min-Max normalization on the feature value to map each attribute parameter to the [0,1] interval. The calculation formula is: ; in, The original attribute value; These are the normalized attribute values; and The first The preset lower bound and preset upper bound values ​​of each attribute parameter on the training set; In this process, the penetration rate parameter is first subjected to a common logarithmic transformation and then normalized. The normalized parameter is calculated and saved during the training phase. If the original permeability is directly normalized using Min-Max, the vast majority of samples (low-permeability range) will be compressed into an extremely narrow range close to zero due to the numerical values ​​spanning multiple orders of magnitude. For example, in the range of 0.001~1mD, 0.005mD is normalized to only 0.004, and 0.05mD is normalized to only 0.049. The permeability difference of 10 times between the two is almost indistinguishable after normalization. However, in actual development, this magnitude of permeability difference often has a decisive impact on oil production. After taking the logarithm, the differences between different orders of magnitude are transformed into equally spaced linear differences. After normalization, the samples in each interval are evenly distributed in the range of [0,1], effectively preserving the geological information resolution capability of the low-permeability interval. Using the normalized attribute parameters as single-attribute features, and explicitly constructing selected interaction item attribute features, the interaction item attribute features are defined as the product of two normalized attribute values; using the oil production of each sample unit as a label; using the feature set composed of single-attribute features and interaction item attribute features as input and the label as output, a mapping relationship between the feature set and the label is established through a gradient boosting ensemble learning model based on decision trees; Step 4: Based on the ensemble learning model, the TreeSHAP method is used to calculate the SHAP interaction value matrix for each sample in the training set; from the SHAP interaction value matrix, the diagonal elements corresponding to the original attributes are extracted as the main effect contribution values ​​of each single attribute, and the off-diagonal elements of the original attribute pairs corresponding to the selected interaction items are extracted as the interaction effect contribution values; after taking the absolute value of each contribution value, a statistical average is performed on all samples in the training set to obtain the single attribute contribution statistics and the interaction contribution statistics; Single attribute contribution statistics: ; Interaction contribution statistics: ; in, For the first Attribute contribution statistics for each attribute; For the first In the nth sample The contribution value of each attribute; For the first The first attribute and the first The interaction contribution statistics of attribute pairs consisting of 1 attribute; For the first In the nth sample, the nth The first attribute and the first The interaction contribution value of attribute pairs consisting of 1 attribute; The total number of samples in the target block; For the first One sample; The contribution statistics of a single attribute are normalized within that single attribute to obtain the weight coefficient of that single attribute. The interaction contribution statistics are normalized within the interaction items to obtain the weight coefficients of the interaction items. The specific process is as follows: Weight coefficient of a single attribute: ; Weight coefficients of interaction items: ; in, For the first The weight coefficient of each attribute; For the first Contribution statistics of each attribute; The sum of the contribution statistics for all single attributes; For the first The first attribute and the first The weight coefficient of the interaction item of the attribute pair consisting of 1 attribute; For the first The first attribute and the first The interaction contribution statistics of attribute pairs consisting of 1 attribute; The sum of the contribution statistics of all selected interactions. The total number of all single attributes; This application utilizes the inherent structure of the SHAP interaction value matrix to decompose the contribution value into two independent levels: diagonal elements, the main effect contribution value of each attribute, and normalized single attribute weight coefficients; off-diagonal elements, the interaction effect contribution value of attribute pairs, and normalized interaction item weight coefficients. The two sets of weight coefficients are independently normalized. This dual-layer independent normalization weight architecture effectively improves the authenticity of desserts. Step 5: Weighting coefficients based on single attributes and the weight coefficient of the interaction item The original dessert score is obtained by weighting the features. The original dessert score is then normalized to obtain the dessert coefficient value. The dessert coefficient value is calculated for each grid node in the three-dimensional distribution model to construct a three-dimensional prediction model for the dessert coefficient. The dessert's original score The following formula is used to obtain: ; In the formula: The original score for the target node dessert; For the first Normalized values ​​of each attribute parameter; For the first Individual attribute weight coefficients; For the first The first attribute and the first Interactive items consisting of attribute pairs of attributes; This is the adjustment coefficient for the interaction term; For the first The first attribute and the first The weight coefficient of the interaction item of the attribute pair consisting of 1 attribute; The normalization includes: ; in: This is the dessert coefficient; The original score for the target node dessert; The theoretical minimum score for the dessert is 0. The theoretical maximum value of the original score for the dessert is 1 + α; Specifically, the interaction term adjustment coefficient α is determined by cross-validation. The α controls the weighting of the interaction term relative to the single attribute term in the sweet spot coefficient formula. The optimal interaction term adjustment coefficient is applied to the reduction calculation of the interaction term contribution statistic. α only applies to the interaction term, and the weighting coefficient of the single attribute term is not affected by α. The interaction term enters in the form of a product, directly capturing the synergistic effect of the two attributes. When both attributes are at high values, the contribution of the product term increases significantly, reflecting the geological mechanism of "synergistic yield increase from both high and low attributes"; when either attribute is at a low value, the product term naturally decays, reflecting the "weakest link effect". α is determined through cross-validation, adaptively balancing the relative strength of the two types of contributions. Step 6: Based on the three-dimensional prediction model of the sweet spot coefficient, the area where the sweet spot coefficient is greater than the preset threshold is identified as the high sweet spot value area, which is used for the selection of horizontal well targets and the design of well trajectory optimization. Based on the three-dimensional prediction model of the sweet spot coefficient, the area with a sweet spot coefficient greater than a preset threshold in three-dimensional space is identified as the high sweet spot value area. During the drilling process, the well trajectory is dynamically adjusted in combination with logging-while-drilling data. When the sweet spot coefficient of the corresponding position of the well trajectory is lower than the preset threshold, the direction of the well trajectory is adjusted to shift it towards the area with a higher sweet spot coefficient. The preset threshold is determined based on the target block's oil production demand and the statistical distribution of the sweet spot coefficient, with a preferred value range of 0.6 to 0.8.

[0013] Preferably, the lithofacies model is established using a truncated Gaussian stochastic simulation method. This method converts a continuous Gaussian random field into discrete lithofacies types by setting multiple thresholds and controls the spatial continuity of lithofacies through a variogram function to construct a spatial constraint framework for joint modeling of multiple attribute parameters.

[0014] The machine learning model is a supervised learning model that can output feature contribution values. It is used to establish the mapping relationship between multi-attribute features and oil production, and is used for subsequent attribute weight calculation.

[0015] The three-dimensional distribution model is established using the sequential Gaussian stochastic simulation method to construct a spatial constraint framework for joint modeling of multiple attribute parameters.

[0016] Preferably, the method further includes a dynamic optimization method, wherein the dynamic optimization includes: updating the machine learning model based on the newly added production well data, and recalculating the contribution value and weight coefficient of each attribute parameter and interaction item to update the three-dimensional prediction model of the sweet spot coefficient.

[0017] Preferably, in step two, the well logging interpretation models for each attribute parameter are established in the following ways: Porosity, permeability, oil saturation, total organic carbon content, and brittleness index were respectively established based on well logging data and core analysis data to establish corresponding well logging interpretation models; The attribute parameters are used to construct a unified multi-attribute feature set, and further construct inter-attribute interaction term features to characterize the nonlinear coupling effect between reservoir attributes, thereby providing input features for subsequent machine learning-based oil production prediction models.

[0018] Under the constraints of the lithofacies model, the following property model was established using Sequential Gaussian Stochastic Simulation (SGS): porosity penetration rate Oil saturation , brittleness index , The modeling process includes: setting a random path to traverse the grid, calculating the conditional cumulative distribution function based on known data, and randomly sampling node values ​​from the distribution.

[0019] The values ​​used in the following formulas are illustrative parameters and need to be recalibrated based on the core data of the target block: ① Porosity logging interpretation model; Based on the core repositioning, a cross-plot of sonic transit time and porosity was compiled using core analysis data, and a porosity logging interpretation model was established through linear regression. (Formula 1); In the formula: Porosity, expressed as % The measured time difference of sound waves is expressed in μs·m. -1 .

[0020] ②Permeability logging interpretation model; Based on the porosity logging interpretation model, the relationship between porosity and permeability was established using core analysis data. After processing with rock electrical properties and multiple regression, a permeability logging interpretation model was established. (Formula 2); In the formula: Permeability, in mD; Porosity, expressed as a percentage.

[0021] ③ Oil saturation logging interpretation model; Based on resistivity logging data, an interpretation model for oil saturation logging is established using Archie's formula: (Formula 3); In the formula: f represents the oil saturation. R w The resistivity of formation water is given in Ω·m. R t The resistivity of the target layer is given in Ω·m. a and b are lithology coefficients; m is the bonding coefficient; This is the saturation index.

[0022] ④ Total organic carbon content Well logging interpretation model; Using the ΔLogR technique, a method is established by superimposing the acoustic transit time and resistivity logging curves. Well logging interpretation model: ; (Formula 4); The corresponding LOM value is 11; coefficient The value is 0.02; R is the resistivity measured by the logging tool, in Ω·m; The resistivity value, in Ω·m, is the resistivity value when the acoustic transit time curve and resistivity curve of the non-hydrocarbon source rock section coincide. The measured acoustic transit time is expressed in μs·m. -1 ; The acoustic transit time curve and resistivity curve of the non-source rock section coincide, expressed in μs·m. -1 ; The above equation can be rewritten using the normalization of the logging curves as follows: (Formula 5), ​​where c, d, and e are coefficients.

[0023] ⑤ Brittleness index logging interpretation model; Dynamic elastic parameters are calculated based on P-wave and S-wave velocities, and then the brittleness index is calculated: Based on the elastic modulus and Poisson's ratio in rock mechanics, the brittleness index of shale reservoirs is calculated using the following formula: E brit =(E-10) / (80-10)×100; μ brit =(0.4-μ) / (0.4-0.1)×100; Brit = 0.5 × E brit +0.5×μ brit (Formula 7); In the formula, Brit is the brittleness index; E brit Normalized elastic modulus; GPa; μ brit is the normalized Poisson's ratio; E is the elastic modulus, and μ is the Poisson's ratio; Both E and μ are determined by the transverse wave time difference. The rock density ρ is calculated using formulas 8 and 9. (Formula 8); (Formula 9); In the formula The transverse wave time difference is expressed in μs·m. -1 ; The longitudinal sonic transit time in well logging is expressed in μs·m. -1 ; Mechanical parameters such as rock density ρ, longitudinal and transverse wave velocities, dynamic Poisson's ratio, and dynamic elastic modulus were obtained, and the undetermined coefficients a', b', and c' were determined based on Formula 10. (Formula 10); Longitudinal sonic transit time in well logging The density ρ of the rock can be calculated.

[0024] Preferably, the interaction term includes a combination of attributes used to characterize the nonlinear coupling relationship between different reservoir properties. The combination of attributes includes porosity and permeability, oil saturation and total organic carbon content, and oil saturation and brittleness index. The interaction term is used to participate in the training of the machine learning model and the calculation of its contribution value.

[0025] A shale oil sweetness prediction system based on multi-attribute fusion includes: The data acquisition and preprocessing module is used to acquire the well location coordinates, stratification data, single well elevation data, logging data, well logging data, sedimentary facies data and core analysis data of the target block, and to perform quality control, outlier processing and normalization on the data, and output attribute data in a unified format. The construction modeling module, connected to the data acquisition and preprocessing module, is used to establish the top structural surface of each oil layer subgroup based on the layered data and single-well elevation data, and to construct a three-dimensional mesh model based on preset mesh parameters to generate a three-dimensional structural model of the target block. The lithofacies modeling module, connected to the structural modeling module, is used to generate a three-dimensional lithofacies model within the three-dimensional spatial framework provided by the three-dimensional structural model, based on the spatial distribution probability of lithofacies types and the variogram parameters, using the truncated Gaussian simulation method. The multi-attribute parameter modeling module, connected to the lithofacies modeling module, is used to establish porosity, permeability, oil saturation, and total organic carbon content, respectively, using the lithofacies 3D model as a constraint framework. A well logging interpretation model for the brittleness index was developed, and a three-dimensional distribution model of each attribute parameter was generated using the sequential Gaussian simulation method. The weight coefficient calculation and dessert model construction module, connected to the multi-attribute parameter modeling module, is used to normalize each attribute parameter and construct a feature set containing single attributes and interaction terms; based on the trained machine learning model, it calculates the contribution value of each attribute parameter and interaction term to the oil production prediction result; it performs statistical processing on the absolute value of the contribution value to obtain the single attribute contribution statistic and the interaction contribution statistic; it obtains the single attribute weight coefficient and the interaction weight coefficient; and it constructs the dessert coefficient model based on the single attribute weight coefficient and the interaction weight coefficient. The horizontal well trajectory optimization module is connected to the weight coefficient calculation and sweet spot model construction module. Based on the three-dimensional prediction model of the sweet spot coefficient, it determines the area where the sweet spot coefficient is higher than the preset threshold. During the drilling process, it updates the sweet spot coefficient corresponding to the well location in real time by combining the logging-while-drilling data. When the updated sweet spot coefficient is lower than the preset threshold, it adjusts the well trajectory direction so that the well trajectory shifts to the area with a higher sweet spot coefficient.

[0026] Preferably, the weight coefficient calculation and dessert model construction module includes: a processing unit, used to normalize the attribute values ​​in the three-dimensional distribution model of each attribute parameter, and construct a feature set containing single attributes and interaction items; The contribution calculation unit is used to calculate the contribution value of each attribute parameter and interaction term to the oil production prediction result based on the trained machine learning model. The statistics and weighting calculation unit is used to perform statistical processing on the absolute value of the contribution value to obtain the single attribute contribution statistic and the interaction contribution statistic, and to obtain the single attribute weight coefficient and the interaction weight coefficient through normalization processing, respectively. The dessert coefficient calculation and dessert 3D prediction model construction unit is used to calculate the shale oil dessert score based on the single attribute weight coefficient and interaction weight coefficient, normalize it to obtain the dessert coefficient value, calculate the dessert coefficient value for each grid node in the 3D distribution model, and construct the dessert coefficient 3D prediction model.

[0027] Compared with the prior art, the present invention has the following beneficial effects: Compared to traditional methods that calculate shale oil sweet spots based on only a single reservoir attribute, this application introduces a multi-attribute fusion scheme, which considers multiple dimensions such as porosity, permeability, oil saturation, total organic carbon content, and brittleness index, thereby enhancing the ability of sweet spot prediction to describe reservoir complexity and improving the accuracy of multi-dimensional processing.

[0028] Furthermore, based on multiple attributes, interaction features between attributes are introduced to model the nonlinear influence between attributes, enhancing the model's sensitivity and predictive ability to complex reservoir characteristics. By analyzing the contribution of interactive attributes, the excessive influence of a single attribute or interaction term on the prediction results can be avoided. A multi-attribute nonlinear model is established, which differs from the traditional linear weighting method and avoids errors caused by human weighting; interaction terms are introduced to characterize synergistic effects, improving the sweet spot identification capability; and closed-loop engineering control of sweet spot prediction and well trajectory optimization is achieved.

[0029] This application quantifies the importance of single attributes and interaction attributes using feature contribution and designs a statistical method for contribution weight coefficients to ensure that the influence of each attribute is appropriate. To address the issue of excessively large interaction contributions, an interaction term adjustment coefficient α is introduced to reduce the weight of excessively large interaction contributions, thereby mitigating the distortion caused by the interaction weight product and improving the actual performance of dessert prediction.

[0030] In the process of calculating the dessert coefficient, the original dessert scores are averaged to effectively eliminate the impact of calculation errors on the prediction results and improve the accuracy and stability of the dessert coefficient.

[0031] Through the above-mentioned technical design, this application has achieved comprehensive processing of the multi-dimensional, interactive and spatial characteristics of shale oil sweet spots, which has significantly improved the accuracy and reliability of sweet spot prediction and provided technical support for subsequent drilling trajectory optimization and reservoir development.

[0032] This invention employs stochastic modeling under lithofacies constraints to accurately characterize reservoir spatial heterogeneity. It utilizes a truncated Gaussian simulation method to establish a three-dimensional lithofacies model, effectively avoiding phase jumps in uncontrolled areas. Under the constraints of the lithofacies model, a sequential Gaussian simulation method is used to establish a three-dimensional distribution model of each attribute parameter, ensuring that the modeling results faithfully reflect well data and accurately characterize reservoir spatial heterogeneity.

[0033] This invention enables the direct guidance of a 3D sweet spot prediction model for horizontal well design, improving the drilling encounter rate. The invention directly applies a 3D sweet spot coefficient prediction model to horizontal well target selection and real-time trajectory verification, achieving integration of geological modeling and horizontal well design. This effectively improves the drilling encounter rate of sand and oil layers, and advances the integrated process of shale oil exploration and development. Attached Figure Description

[0034] Figure 1 Geographical map of Baimayaoxian area of ​​Dingbian Oilfield (a) and stratigraphic column of Chang 7 oil layer of Triassic Yanchang Formation in Baimayaoxian area (b).

[0035] Figure 2 Three-dimensional structural model of the Chang 7 oil layer of the Triassic Yanchang Formation in Baimayaoxian area; (a) southwest view; (b) south view.

[0036] Figure 3 Figure 1 shows the vertical proportion and variation function analysis of different lithofacies in the Chang 7 oil-bearing subgroups of the Triassic Yanchang Formation in the Baimayaoxian area; (a) Chang 7 1 Distribution percentage of each lithofacies along the vertical grid channel fine unit; (b) Length 7 1 Analysis of principal directional variation function of fine sandstone facies in underwater distributary channels; (c) Length 7 1 Analysis diagram of phase direction variation function of underwater distributary sandstone; (d) Length 7 2 Distribution percentage of each lithofacies along the vertical grid cell; (e) Length 7 2 Analysis of principal directional variation function of fine sandstone facies in underwater distributary channels; (f) Length 7 2 Analysis diagram of the directional variation function of fine sandstone phases in underwater distributary channels.

[0037] Figure 4 Three-dimensional lithofacies model of each oil-bearing subgroup of the Chang 7 oil layer in the Triassic Yanchang Formation of Baimayaoxian District; (a) Chang 7 1 (b) Three-dimensional model of lithofacies; 7 2 Three-dimensional model of lithofacies; (c) Length 7 3 Three-dimensional model of lithofacies.

[0038] Figure 5 Vertical slices of the lithofacies models of the Chang 7 oil-bearing subgroups of the Triassic Yanchang Formation in the Baima Yaoxian area.

[0039] Figure 6 Three-dimensional model diagrams of porosity, permeability, and oil saturation of sandstone reservoirs in the Chang 7 oil-bearing subgroup of the Triassic Yanchang Formation in the Baimayaoxian area; (a) Chang 7 1 (b) Porosity model diagram of sandstone; 7 2 Porosity model diagram of sandstone; (c) Length 7 1 Permeability model diagram of sandstone; (d) Length 7 2 Sandstone permeability distribution map; (e) Length 7 1 Sandstone oil saturation model diagram; (f) Length 7 2 Distribution of oil saturation in sandstone.

[0040] Figure 7 Three-dimensional TOC model of source rocks of the Chang 7 oil-bearing formation in the Triassic Yanchang Formation of Baimayaoxian area; (a) Chang 7 2 (a) TOC 3D model diagram of mudstone and shale; (b) Length 7 3 TOC 3D model diagram of mudstone and shale; (c) Length 7 2 and length 7 3 Longitudinal slice of the TOC model of the source rocks in the oil reservoir subgroup.

[0041] Figure 8 Correlation charts of various attribute parameters and daily oil production in the Chang 7 oil layer of the Triassic Yanchang Formation in the Baima Yaoxian area: (a) Correlation chart of porosity and daily oil production; (b) Correlation chart of permeability and daily oil production; (c) Correlation chart of oil saturation and daily oil production; (d) Correlation chart of brittleness index and daily oil production.

[0042] Figure 9 The 7th Longitudinal Formation of the Triassic System in Baimayaoxian District 1 7 2 Three-dimensional model diagram of sweet spot coefficient of shale oil in oil reservoir subgroup. Detailed Implementation

[0043] Example 1: The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0044] like Figures 1 to 9 As shown, this paper takes the shale oil of the Chang 7 oil layer of the Triassic Yanchang Formation in the Baimayaoxian area of ​​the Dingbian Oilfield in the Ordos Basin as the research object, and describes the implementation process of the shale oil sweet spot prediction method based on multi-attribute fusion described in this invention.

[0045] The study area is located in the semi-deep lacustrine facies hydrocarbon source rock development zone of the Ordos Basin. The Chang 7 oil-bearing formation is mainly a meandering river delta sedimentary system controlled by northeastern provenance. A total of 517 wells were drilled in this block, with a well density of 2.58 wells / km². Data collected included well location coordinates, stratigraphic data, single-well elevation data, conventional logging data, core analysis data, well logging data, sedimentary facies data, and historical production data. The raw data underwent preprocessing, including outlier removal, depth alignment correction, missing value interpolation and completion, and format standardization.

[0046] Figure 1 This section shows the geographical location and stratigraphic sequence of the study area. Among them: Figure 1 (a) indicates the location of the study area within the region; Figure 1 (b) represents the stratigraphic structure of the Chang 7 oil reservoir group, providing regional geological background and stratigraphic basis for subsequent geological modeling.

[0047] Figure 2 It represents a three-dimensional structural model, which is constructed based on well location, layer, and elevation data to build a three-dimensional structural framework.

[0048] Step S1: Constructing the model; Based on the well location coordinates, stratigraphic data, and single-well elevation data of 517 wells in the study area, top structural surfaces of the oil-bearing subgroups of Chang 7¹, Chang 7², Chang 7³, and Chang 8 were established. After establishing the top structural surfaces, local corrections were performed on each top structural surface using single-well elevation data, and consistency checks were conducted between adjacent stratigraphic layers. Based on this, a three-dimensional mesh was established with a planar mesh step size of 50m × 50m and a vertical mesh step size of 2m to create a three-dimensional structural model of the study area.

[0049] Step S2: Lithofacies modeling; Within the three-dimensional spatial framework provided by the structural model, and in conjunction with well logging, well logging, and sedimentary facies data, identification criteria for different lithofacies types in the study area were established. From shallow water to deep water, the study area successively developed fine sandstone facies in the delta front underwater distributary channels, siltstone facies in the interdistributary bays of the delta front, shallow lacustrine mudstone facies, and semi-deep lacustrine to deep lacustrine shale facies.

[0050] To address the diverse lithofacies and high well density in the study area, a truncated Gaussian simulation method was employed to establish lithofacies models. First, the distribution probabilities of different lithofacies along the vertical grid in each oil-bearing subgroup were statistically analyzed. Combined with the results of two-dimensional sedimentary facies plane studies, the variogram parameters were determined. Then, constrained by the spatial distribution of sedimentary facies, a three-dimensional lithofacies model for each oil-bearing subgroup was generated using the truncated Gaussian simulation method. In the implementation, the principal range direction was approximately 30°, the principal range distance was approximately 1500 m, and the secondary range distance was approximately 600 m.

[0051] Lithofacies modeling method selection: Considering the characteristics of the Chang 7 shale oil facies—diverse and heterogeneous—it's difficult to accurately determine the geometric parameters such as length, width, and thickness of each facies, and it's prone to deviating from well data under the high-density well network conditions of the study area. In contrast, pixel-based (grid-based) stochastic modeling methods ensure model fidelity to well data and showcase the spatial distribution characteristics of inter-well lithofacies within the framework of sedimentary geological features. Therefore, the pixel-based truncated Gaussian simulation method was chosen for lithofacies modeling.

[0052] The truncated Gaussian simulation method truncates the three-dimensional continuous Gaussian distribution through a series of threshold values ​​and truncation rules, establishing the three-dimensional spatial distribution of lithofacies type variables. It is suitable for simulating sedimentary facies with ordered facies zones. In the study area, the sedimentary water bodies developed during the Chang 7 period range from shallow to deep delta front underwater distributary channels with fine sandstone facies to semi-deep lacustrine to deep lacustrine shale facies. The lithofacies are arranged in an orderly manner, and the use of truncated Gaussian simulation can effectively avoid facies jumps in areas without well control.

[0053] Variation function analysis: First, quantitative identification and data extraction of lithofacies were performed along the wellbore of 517 single wells in the study area. Statistical analysis was then conducted on the vertical probability distribution and thickness variation distribution of four lithofacies in each oil-bearing subgroup of the Chang 7 reservoir. Using the two-dimensional sedimentary facies plane study results of each sublayer as a framework, the geometric morphology (length, width, principal axis direction) and connectivity properties of the underwater distributary channel sand bodies in the delta front were determined. This provides a basis for determining the key parameters of the variation function (principal direction range, secondary direction range, arch height, nugget value) in the truncated Gaussian simulation. (See [link to Gaussian simulation]). Figure 3 .

[0054] Spatial correlation parameters were set using the variogram function: main range direction: approximately 30°, main range distance: approximately 1500m, secondary range distance: approximately 600m. Under the constraint of the spatial distribution of sedimentary facies, a truncated Gaussian simulation method was used to establish a three-dimensional lithofacies model of each oil-bearing subgroup. (See [reference]). Figure 4 .

[0055] Figure 5 Vertical slices of the lithofacies model are used to show the lithofacies distribution at different depths, verifying whether the model conforms to sedimentary patterns.

[0056] Step S3: Multi-attribute parameter modeling; Using a three-dimensional lithofacies model as a constraint framework, porosity, permeability, oil saturation, and total organic carbon content were established respectively. A well logging interpretation model for the brittleness index was developed, and a three-dimensional distribution model of each attribute parameter was generated using the sequential Gaussian simulation method.

[0057] The sequential Gaussian simulation method includes: traversing grid nodes according to a random path; calculating the conditional cumulative distribution function of the current node based on known sample data and simulated node data; sampling the attribute values ​​of the current node from the conditional cumulative distribution function; and assigning attribute parameters to each node in three-dimensional space. Figure 6 A three-dimensional model of porosity, permeability, and oil saturation. Figure 7 for A three-dimensional model. Reflects the spatial distribution of reservoir properties. Characterizes the spatial distribution of source rock quality. Figure 8 This is a graph showing the relationship between various attribute parameters and daily oil production. It reveals that single-attribute relationships are discrete, and there is non-linearity between attributes, necessitating the use of multi-attribute fusion and interactive modeling.

[0058] (1) Porosity logging interpretation model; Sonic logging is widely used in the study area. Based on core relocation, a cross-plot of sonic transit time and porosity was compiled using core analysis data from 56 sandstone samples from 8 wells in the study area. This allows for the interpretation of porosity using sonic logging. The relationship between porosity and sonic transit time was obtained through linear regression. (Formula 1); In the formula Porosity, % The measured acoustic transit time is expressed in μs·m. -1 .

[0059] (2) Permeability logging interpretation model; Based on the porosity logging interpretation model, core analysis data from 45 sandstone samples from 8 wells in the study area were selected to establish a porosity-permeability relationship diagram. After rock electrical property restoration and multivariate regression processing, the calculation formula for the permeability of the Chang 7 reservoir was obtained: (Formula 2); In the formula Permeability, mD; Porosity, %.

[0060] (3) Oil saturation logging interpretation model; An interpretation model for oil saturation logging was established using Archie's formula. First, rock electrical experimental data from 41 sandstone samples in 5 wells in the study area were plotted to determine the values ​​of a, b, m, and n in Archie's formula as 1.2718, 1.3097, 1.221, and 3.748, respectively. Then, based on the average total salinity of formation water in Chang 7 formation (31099.97 mg / L) and a formation temperature of 69.0℃, the equivalent salinity was calculated to be 30404.3 mg / L through ion conversion. The formation water resistivity was then calculated to be 0.118 Ω·m. Finally, the parameters of Archie's formula were determined, and the oil saturation calculation model was established. ; (Formula 3); In the formula: f represents the water saturation level. Oil saturation; f; The resistivity of formation water is given in Ω·m. ρ is the resistivity of the target layer, Ω·m; a and b are lithological coefficients; The bonding coefficient; This is the saturation index.

[0061] (4) Total organic carbon content Well logging interpretation model; Traditional ΔlogR model explanation ; ; (Formula 4); In the formula, LOM is related to organic matter maturity. Related, research area The main distribution is between 0.7% and 1.3%, corresponding to a LOM value of 11; coefficient The value is 0.02; R is the resistivity measured by the logging tool, in Ω·m; The resistivity value, in Ω·m, is the resistivity value when the acoustic transit time curve and resistivity curve of the non-hydrocarbon source rock section coincide. The sonic transit time in well logging is expressed in μs·m. -1 ; The acoustic transit time curve and resistivity curve of the non-source rock section coincide, expressed in μs·m. -1 Because baseline readings are sometimes subject to significant errors due to subjective factors and are inconvenient to calculate, research has found that for wells with similar or identical sedimentary tectonic backgrounds, the above formula can be rewritten using the normalization of logging curves: (Formula 5); This study used 15 mudstone samples from the Chang 7 oil-bearing formation in the research area. The logging curves from the measured data were calibrated, and a method suitable for the study area was established using multiple linear regression. Explanation of the formula: (Formula 6); (5) Brittleness index logging interpretation model; Based on the elastic modulus and Poisson's ratio in rock mechanics, the brittleness index of shale reservoirs is calculated using the following formula: ; ; (Formula 7); In the formula, Brit is the brittleness index, abbreviated as BI below; E brit For the normalized elastic modulus, Gpa; μ brit The normalized Poisson's ratio; E is the elastic modulus, μ is the Poisson's ratio, and both E and μ can be derived from the transverse wave time difference. The rock density ρ is calculated using formulas 8 and 9. (Formula 8); (Formula 9); In the formula The transverse wave time difference is expressed in μs·m. -1 ; The sonic transit time in well logging is expressed in μs·m. -1 ; and All samples exhibit a linear relationship as per Formula 10. Ten samples from the Chang 7 oil layer in the study area were selected for uniaxial compression mechanical testing. Mechanical parameters such as longitudinal and transverse wave velocities, dynamic Poisson's ratio, and dynamic elastic modulus were obtained. Based on Formula 10, the undetermined coefficients a', b', and c' were determined to be 1.49, 134.03, and -357.47, respectively. (Formula 10); See Formula 11, which uses the longitudinal sonic transit time of well logging. The rock density ρ can be calculated; (Formula 11); Log interpretation calculations were performed on all 517 wells with complete logging data within the target block to obtain continuous curves of five attribute parameters for each layer in each well.

[0062] Step S4: Attribute normalization and feature construction; Prior to this embodiment, a scheme for calculating the dessert coefficient based on single-attribute linear weighting had been established in the study area. This scheme employs Z-score normalization and exponential regression correlation coefficients between each attribute and yield. Normalization determines the weights, and this scheme was validated in 45 wells, showing the correlation coefficient between the sweet spot coefficient and the yield. It is 0.5332.

[0063] This approach has the following limitations: (1) Incomplete attribute parameters: Only four reservoir attributes are included: porosity, permeability, oil saturation and brittleness index. The total organic carbon content, which reflects the quality of the source rock, is not included. For the Chang 7 shale oil reservoir, which is a self-generated and self-storing integrated source-reservoir type, This reflects the source rock's ability to infuse adjacent reservoirs and is an important factor influencing the distribution of sweet spots.

[0064] (2) The weighting method is crude: the exponential regression of each attribute with output is used. As a basis for weighting, it can only capture univariate exponential relationships and cannot reflect nonlinear interaction effects and multicollinearity among multiple variables.

[0065] (3) Lack of interaction effects between attributes: Pure linear weighting cannot characterize the nonlinear synergistic effects between attributes such as porosimetry coupling and source-storage matching.

[0066] Step S5: Machine learning modeling and contribution value calculation; Training sample construction: Using 45 vertical and directional wells with stable production data as data sources, each well had a perforated section of the 7-layer oil formation as a sample unit, resulting in a total of 45 samples. Along the depth range of the perforated section in each sample unit, five attribute parameter values ​​were extracted from the three-dimensional distribution model of each attribute parameter. The effective thickness-weighted average of each attribute parameter within each sample unit was calculated as the feature value. The statistical characteristics of each attribute parameter in the 45 wells are shown in Table 1. Table 1 shows the statistical characteristics of various attribute parameters of 45 wells; ; The permeability ranges from 0.10 to 0.60 mD, spanning approximately 6 times, and the coefficient of variation is 0.551, requiring logarithmic transformation. The fragility index had a coefficient of variation of only 0.013 and a very narrow range of variation (35.56%~38.02%), indicating that the fragility index of this block did not vary much. The porosity ranges from 6.0% to 17.0%, showing a significant range and good distinguishability.

[0067] Normalization process: The weighted average of the five attribute parameters is normalized using Min-Max, mapping each attribute parameter to the [0,1] interval. The permeability K is first transformed using a common logarithm (base 10) before normalization.

[0068] The reason for using Min-Max normalization instead of Z-score normalization is that: Min-Max normalization results in all attribute values ​​being normalized. This makes interactive items The normalized value is always non-negative. When an attribute value is under unfavorable conditions (close to the lower bound), its normalized value approaches 0, and the corresponding interaction item also approaches 0, preventing a false boost to dessert scores due to "negative times negative equals positive". The normalized parameters for each attribute in this embodiment are shown in Table 2: Table 2 shows the normalized parameters for each attribute. ; The above normalization parameters are calculated and saved during the training phase and used for subsequent calculation of the sweet spot coefficient at the grid level.

[0069] Tag definition; The label value for each sample unit is the average daily oil production from day 30 to day 90 after the well is put into production.

[0070] During model training, the oil production is logarithmically transformed by y=ln(q+1) to improve the normality of the label distribution.

[0071] Dataset partitioning; Stratification by daily oil production, down to the quartile: High-yield layer (≥5.35t / d): Nos. 1-6 and 8-9, a total of 8 wells → 7 wells for training and 1 well for testing; Medium to high yield layer (3.05-5.34t / d): Nos. 7, 11-18, 21, 23, 24, 25-32, a total of 20 wells → 16 wells for training, 4 wells for testing; Medium and low yield layer (1.06-3.04t / d): serial numbers 10, 19, 20, 22, 34, 39, a total of 6 ports → 5 ports for training, 1 port for testing; Low-yield layer (<1.06t / d): Nos. 33, 35-38, 40-45, a total of 11 batches → 8 batches for training, 3 batches for testing; The training set consists of 36 wells, and the test set consists of 9 wells.

[0072] The normalized attribute values ​​and daily oil production data of 45 wells are shown in Table 3. Table 3 shows the normalized attribute values ​​and daily oil production of 45 wells; ; First, an initial LightGBM model is trained using five single-attribute normalized features. A large sample size should be used; this application selects a subset of samples to demonstrate the technical process using 45 wells as data points: 36 training samples and 9 validation samples. Given the sample size of 36, a conservative hyperparameter configuration is adopted to prevent overfitting: 200 trees, maximum depth 4, 12 leaf nodes, learning rate 0.05, minimum leaf node sample size 5, L1 regularization coefficient 0.5, L2 regularization coefficient 2.0, row sampling ratio 0.8, and column sampling ratio 0.8. When building a model using a larger sample size or actual production areas, new parameter configurations can be selected to ensure the accuracy of the conclusions; the current selection is based on the existing samples.

[0073] Interactive item filtering; Based on the initial model, the TreeSHAP method was used to calculate the SHAP interaction value matrix for all 36 samples in the training set. The SHAP interaction value matrix is ​​a D×D square matrix (D is the number of features), where the diagonal elements... Represents the main effect contribution value of the i-th feature, off-diagonal elements. This represents the contribution value of the interaction effect of the interaction term feature formed by the i-th attribute and the j-th attribute.

[0074] For all C(5,2)=10 original attributes, calculate the sample mean of the absolute value of the interaction effect for each pairwise combination. Sort in descending order, the results are shown in Table 4: Table 4. Results of interactive item filtering; ; The cumulative contribution rate of the first three items is 73.6%, with porosity × permeability selected as the ratio. × Oil saturation × total organic carbon content × Oil saturation × brittleness index × Three interactive items.

[0075] The geological significance of the three interaction terms is as follows: (1) × This reflects the porosity-permeability coupling effect in the Chang 7 tight reservoir. High porosity accompanied by high permeability represents a favorable reservoir section with well-developed nano-micro pore throats, and the combined contribution of the two to productivity is far greater than the sum of their individual contributions.

[0076] (2) × This reflects the matching relationship between the quality of source rocks and the oil-bearing capacity of reservoirs in the Chang 7 self-generated and self-storage shale oil system. High-quality source rocks (high...) The spatial coupling with high oil saturation represents a sweet spot segment that is both effective in generating hydrocarbons and effectively charged.

[0077] (3) × This reflects the synergistic effect of geological sweet spots and engineering sweet spots. Reservoir sections with high oil saturation and high brittleness possess both material basis (oil content) and modifiability (fractureability), and their economic development value is highest when these two are coupled.

[0078] Formal model training; Three selected interaction terms were added to the feature set as explicit features. Interaction term attribute features were defined as the product of two normalized attribute values. Using eight features (five single-attribute features and three interaction term features) as input, and the logarithmically transformed daily oil production y=ln(q+1) as the label, the LightGBM model was retrained with the same hyperparameter configuration as the initial model. The performance of the formal model was evaluated using Leave-One-Well-Group Cross-Validation (LOOCV), and the results are shown in Table 5. Table 5. Performance evaluation of the formal model; ; The difference between the coefficient of determination r² of the training set and the coefficient of determination r² of leave-one-out cross-validation is 0.13, which is within an acceptable range under the small sample condition of 45 wells, and the model does not have serious overfitting.

[0079] SHAP contribution value calculation and weight determination; Based on the formal model, the TreeSHAP method was used to calculate the SHAP interaction value matrix of 36 samples in the training set. From the interaction value matrix, the diagonal elements corresponding to the five original attributes were extracted as the single-attribute main effect contribution values, and the off-diagonal elements corresponding to the three selected interaction items were extracted as the interaction effect contribution values. The absolute values ​​of each contribution value were taken and statistically averaged over all 36 samples in the training set to obtain the contribution statistics, and further, the weight coefficients of the single attributes were obtained. and the weight coefficient of the interaction item ; The SHAP method was used to calculate the contribution of each feature. SHAP definition: ; F is the set of all features. The total number of features (in this embodiment) (including 5 single-attribute features and 10 interaction features). S is any subset of F that does not contain feature i. The number of features in subset S; This represents the predicted output value of the model when only a feature subset S is used; The predicted output value of the model after adding feature i to the feature subset S; This represents the marginal contribution of feature i to the prediction result after it is added to subset S; The weighting coefficients represent the probability of subset S appearing in all possible permutations of features; SHAP value The meaning is: the contribution of feature i to the prediction result of sample x relative to the baseline prediction value (the mean of all sample prediction values). This indicates that feature i has a positive contribution to the predicted value of the sample (increasing the predicted value). This indicates a negative contribution.

[0080] The SHAP values ​​of all features satisfy additivity: ;

[0081] This refers to the actual predicted output value of the model for sample x, which in this application is the oil production predicted by the LightGBM model for a certain well. in The baseline prediction value (the mean of all training sample predictions) is the sum of the SHAP values ​​of each feature, which is exactly equal to the difference between the model prediction value and the baseline value for that sample.

[0082] SHAP Interaction Values: Furthermore, the TreeSHAP method can compute the SHAP interaction value matrix. The SHAP value of feature i is decomposed into main effects and interaction effects: ;

[0083] in: The diagonal elements of the SHAP interaction value matrix represent the main effect contribution value of feature i, that is, the contribution of feature i to the prediction result independently of other features; (I≠J) represents the off-diagonal elements of the SHAP interaction value matrix, indicating the contribution value of the interaction effect between feature I and feature J, that is, the additional contribution of the synergistic effect of the two features to the prediction result.

[0084] The SHAP interaction value matrix satisfies symmetry: .

[0085] This invention utilizes the aforementioned properties of the SHAP interaction value matrix to extract the main effect contribution value of each single attribute feature from the diagonal elements and the interaction effect contribution value of the interaction term from the off-diagonal elements. These values ​​are used to determine the single attribute weight coefficient and the interaction term weight coefficient, respectively, thereby realizing data-driven objective weight determination.

[0086] Single attribute contribution statistics: ; Interaction contribution statistics: ; Based on the single-attribute contribution statistics and interaction contribution statistics, the weight coefficients of the single attributes are obtained. and the weight coefficient of the interaction item The specific process is as follows: Weight coefficient of a single attribute: Weight coefficients of interaction items: ; Table 6. Contribution statistics and weighting coefficients; ; Oil saturation contributed the highest statistical value (0.251), consistent with the understanding that oil content is the primary indicator in shale oil sweetness evaluation. Brittleness index contributed the lowest value (0.046), consistent with the extremely small variation range of brittleness index in this study area (coefficient of variation only 1.4%). A comparison of the weights with existing schemes is shown in Table 7. Table 7 shows a weight comparison of the existing schemes; ; The following differences can be found by comparing the two weighting schemes: (1) Oil saturation The existing scheme has a weight of 0.43, while this scheme reduces it to 0.325. The reason for the weight reduction is not... The importance of [the concept] has decreased, but this solution introduces [the concept]. and Two interactive items, will Some of the contributions were transferred from single-attribute items to interaction items. If the interaction items... The contribution of the relevant part is converted back Its overall contribution remains the highest.

[0087] (2) Brittleness index The existing scheme has a weight of 0.17, while this scheme significantly reduces it to 0.059. The SHAP method determines the weight based on the model's actual predictive contribution to each sample, accurately identifying [the specific data points] even with minimal changes in the fragility index (only 2.46 percentage points, coefficient of variation 1.4%). The actual contribution to yield differentiation is minimal. Existing schemes' R² methods force a fit through exponential regression, in... It was still assigned a weight of 0.17 even with minimal changes, overestimating the actual contribution of the fragility index.

[0088] (3) Porosity and penetration rate The two schemes ranked the same (2nd and 3rd respectively) and had similar weight values, indicating that the two methods were basically consistent in their evaluation of these two attributes.

[0089] (4) This scheme, which was not included in the existing scheme, assigns a weight of 0.166 (4th place) to fill the gap in the evaluation of source rock quality.

[0090] Determining the interaction term adjustment coefficient α; The interaction term adjustment coefficient α controls the weighting of interaction terms relative to single-attribute terms in the dessert factor formula. α only applies to interaction terms and does not change the weighting coefficients of single-attribute terms. When α=0, it degenerates into a pure single-attribute linear weighted model; as α increases, the contribution of the interaction term is enhanced.

[0091] The optimal α was determined using a one-well-hold cross-validation method. In the training set of 36 wells, 31 candidate α values ​​were set, ranging from 0.00 to 1.50 with a step size of 0.05. For each candidate α value, one well was used as the validation well, and the remaining 35 wells were used as reference wells, using predetermined weighting coefficients. and The current α value is used to calculate the sweet spot coefficient, and the mean square error (MSE) between the sweet spot coefficient and the actual daily oil production is evaluated. The α value corresponding to the minimum MSE is selected as the optimal interaction term adjustment coefficient. The α search results are shown in Table 8. Table 8 shows the search results for α; ; The optimal interaction term adjustment coefficient α = 0.60. When α increases from 0 to 0.60, the MSE decreases from 0.592 to 0.388, a decrease of 34.5%, quantitatively demonstrating the significant contribution of attribute interaction effects to the prediction accuracy of desserts. The MSE rebounds after α exceeds 0.60, indicating that excessively large interaction term weights actually reduce prediction accuracy. Table 8 lists the search results for some representative α candidate values.

[0092] S6: Dessert coefficient calculation and three-dimensional prediction model establishment; After completing the 3D modeling, normalization, and weight coefficient calculation of the attribute parameters, a dessert coefficient model is constructed. This dessert coefficient model is based on single-attribute weight coefficients, interaction weight coefficients, single-attribute features, and interaction item features. The original scoring formula for desserts is: ; In the formula: The original score for the target node dessert; For the first Normalized values ​​of each attribute parameter; For the first The weight coefficient of each attribute; For attribute interaction items; No. Weighting coefficients for interaction items; These are the weight constraint coefficients; Obtain the original score for dessert Subsequently, to eliminate differences in output scale between different blocks and models, interval normalization was performed to obtain the sweet spot coefficient. ; ; Due to all normalized attribute values All scores are within the interval [0,1]. After normalization, the sum of all weight coefficients is 1. This is the original score for the dessert. Theoretical minimum value =0 (when all) (At the time of acquisition), theoretical maximum value =1+ (when all) (obtained at the time), that is .

[0093] The weighting coefficients in this embodiment and Substituting 0.60, we get... .

[0094] Calculation Example; Example 1 (Type I Sweet Well): No. 5, porosity 17.0%, permeability 0.59mD, oil saturation 51.92%, brittleness index 38.02%, TOC 2.924%. Daily oil production is 8.56t / d.

[0095] Normalization: =1.000, =0.986, =0.941, =1.000, =1.000.

[0096] Single attribute item: 0.325×0.941+0.236×1.000+0.214×0.986+0.166×1.000+0.059×1.000=0.978.

[0097] Interaction item: 0.60×(0.395×1.000×0.986+0.331×0.941×1.000+0.274×0.941×1.000)=0.575.

[0098] =0.978+0.575=1.553, P=1.5553 / 1.60=0.971.

[0099] All five attribute parameters of the well are close to the optimal values ​​(normalized values ​​are close to 1), and both the single attribute and interaction terms have obtained high scores. The sweet spot coefficient is close to the theoretical maximum value of 1.0, which is consistent with the high production performance of 8.56 t / d of actual daily oil production.

[0100] Example 2 (Type II Sweet Well): No. 23, porosity 9.4%, permeability 0.193 mD, oil saturation 48.28%, brittleness index 36.17%. The percentage was 2.297%. Daily oil production was 5.08 tons per day.

[0101] Normalization: =0.309, =0.345, =0.786, =0.248, =0.672.

[0102] Single attribute item: 0.325×0.786+0.236×0.309+0.214×0.345+0.166×0.672+0.059×0.248=0.528.

[0103] Interactive items: Item: 0.395×0.309×0.345=0.042.

[0104] × Item: 0.331 × 0.786 × 0.672 = 0.175; S oi ×BI item 0.274×0.786×0.248=0.053.

[0105] Total interaction items: 0.60×(0.042+0.175+0.053)=0.60×0.270=0.162.

[0106] , .

[0107] The well has a high oil saturation. =0.786) but the porosity and permeability conditions are moderate ( =0.309, =0.345, the brittleness index is relatively low ( =0.248). × The interaction term contributed the most (0.175), reflecting the relationship between oil saturation and... Good coupling; and and The interaction term contributed relatively little, reflecting the inadequacy of the porosity and compressibility conditions.

[0108] Example 3 (Non-sweet spot well): No. 42, porosity 6.6%, permeability 0.114 mD, oil saturation 29.79%, brittleness index 35.79%. The percentage is 1.620%. Daily oil production is 0.18 t / d.

[0109] Normalization: =0.055, =0.054, =0.000 (truncated) =0.093, =0.312.

[0110] Single attribute item: 0.325×0+0.236×0.055+0.214×0.054+0.166×0.312+0.059×0.093=0.082.

[0111] Interaction item: Oil saturation normalized value is 0. × and × All interactive items are 0.

[0112] Item: 0.395 × 0.055 × 0.054 = 0.00117.

[0113] Total interaction items: 0.60 × 0.00117 = 0.00070.

[0114] =0.082 + 0.001 = 0.083, =0.083 / 1.60=0.052.

[0115] The oil saturation of this well is close to the lower limit (29.79%, truncated to 0 after normalization). The failure of oil saturation as the primary indicator leads to the largest weight term among the single attribute terms. The contribution of ) is zero, and at the same time with The two related interaction items were also automatically reset to zero. This demonstrates the technical advantage of Min-Max normalization: the normalized value of unfavorable attributes approaches zero, automatically suppressing spurious contributions from related interaction items.

[0116] Calculate the sweet spot coefficient for each of the 45 wells; The results are shown in Table 9, with the sweetness coefficients of 45 wells calculated. ; Dessert grading and production verification; Based on the size of the dessert coefficient, they are divided into four levels, and the production statistics of each level are shown in Table 10. Table 10 shows the dessert grading standards and production statistics; ; Explanation of abnormal samples: Well No. 6 has a daily oil production of 6.92 t / d, corresponding to a sweet spot coefficient. =0.239; Well No. 8 has a daily oil production of 5.35 t / d, corresponding to a sweet spot coefficient. =0.139.

[0117] The above-mentioned well points show a certain deviation between actual production and sweet spot evaluation results. Analysis suggests that well number 6 may have been affected by engineering factors such as fracturing effects, resulting in an actual production higher than its geological condition-corresponding evaluation level. Well number 8, due to its low normalized oil saturation value of only 0.152, has a lower overall geological evaluation. This indicates that the sweet spot coefficient mainly characterizes the favorableness of geological conditions and does not directly reflect the effects of engineering modifications; therefore, deviations at a few well points are reasonable.

[0118] Comparison and verification with existing solutions; The proposed scheme was compared with the existing single-attribute linear weighted scheme using data from the same 45 wells. The results are shown in Table 11. Table 11 shows the comparison results between this scheme and the existing single-attribute linear weighted scheme on the same 45 well data; ; Linear regression prediction based on the dessert coefficient formula of this scheme The value is 0.693, which is better than existing linear weighting schemes. After simplifying the model weights to a dessert coefficient formula, the dessert / non-dessert production ratio was 7.59, indicating good dessert differentiation.

[0119] The improvement mainly comes from three aspects: (1) increase The attribute parameters provide information on the quality of the source rock; (2) Three interaction terms are introduced to capture the synergistic effect between attributes; (3) The SHAP method quantifies the weight of each attribute more accurately, especially by reducing the weight of the brittleness index from 0.20 to 0.059.

[0120] Horizontal well application verification; The three-dimensional prediction model of the sweet spot coefficient calculated in this scheme was applied to the trajectory design verification of two horizontal wells, P28-1 and P28-3, in the study area.

[0121] Well P28-1 has a horizontal section length of 917m, and the average sweet spot coefficient P=0.982 in the horizontal perforated section, classifying it as a Class I sweet spot. The initial average daily oil production after commissioning was... Cumulative oil production .

[0122] Well P28-3 has a horizontal section length of 992m, and the average sweet spot coefficient P=0.855 in the horizontal perforated section, classifying it as a Class I sweet spot. The initial average daily oil production after commissioning of this well was... Cumulative oil production .

[0123] The sweet spot coefficient of well P28-1 (0.982) is higher than that of well P28-3 (0.855), consistent with its higher average daily and cumulative oil production. The validation results of the two horizontal wells demonstrate that the three-dimensional prediction model for the sweet spot coefficient can effectively guide the selection of horizontal well targets and the optimization of well trajectory design.

[0124] Calculation of dessert coefficient at the grid level; For each grid node in the three-dimensional distribution model of attribute parameters established in step S4, extract the original values ​​of five attribute parameters, perform Min-Max normalization using the normalization parameters saved in Table 2, and then substitute them into the sweet spot coefficient formula to calculate the sweet spot coefficient P of that grid node. The sweet spot coefficients of all grid nodes constitute the three-dimensional distribution model of the sweet spot coefficient, which is used to guide the selection of horizontal well targets and the optimization design of well trajectories.

[0125] The rationale for transferring the weighting coefficients obtained from well-level training to grid-level computation is based on the following verification: (1) The weighting coefficient reflects the relative importance of each attribute parameter to the contribution of oil production. It belongs to the geological attributes of the target block and does not change with the spatial scale.

[0126] (2) The attribute parameter values ​​in the training and application phases are all from the same three-dimensional attribute distribution model, and the data sources are consistent.

[0127] (3) The Kolmogorov-Smirnov test was performed on the attribute distribution of the 36 well samples in the training set and the attribute distribution of all grid nodes. The p-values ​​of the five attribute parameters were all greater than 0.05, and the null hypothesis that the two distributions were consistent was not rejected.

[0128] Step S6: Horizontal well trajectory optimization design; Based on the three-dimensional distribution model of the sweet spot coefficient, the following steps are used to optimize the horizontal well trajectory design: (1) Identify the spatial distribution of type I and type II sweet spots in the three-dimensional distribution model of sweet spot coefficient, and select the continuous high value area of ​​sweet spot coefficient as the candidate area of ​​horizontal well target.

[0129] (2) Within the candidate region, the landing point position, azimuth angle and trajectory direction of the horizontal segment are optimized by maximizing the weighted average of the sweet spot coefficient along the horizontal segment trajectory as the objective function.

[0130] (3) Constraints include: the vertical depth variation caused by the dip angle of the strata does not exceed ±15m, and faults and abnormal pressure zones are avoided.

[0131] (4) Output the optimized three-dimensional coordinates of the horizontal well trajectory and the sweet spot coefficient distribution curve along the trajectory to provide a basis for target design for on-site drilling.

[0132] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention.

Claims

1. A method for predicting shale oil sweetness based on multi-attribute fusion, characterized in that, Includes the following steps: Step 1: Construct a three-dimensional geological model in the target area. The three-dimensional geological model includes a structural model and a lithofacies model. Step 2: Using the lithofacies model as a constraint framework, establish parameters for porosity, permeability, oil saturation, and total organic carbon content. A well logging interpretation model for the brittleness index was developed, and a three-dimensional distribution model of each attribute parameter was generated using the sequential Gaussian simulation method. Step 3: Extract the values ​​of each attribute parameter from the three-dimensional distribution model of each attribute parameter, normalize the values ​​of each attribute parameter, and construct a feature set based on the normalized attribute parameters. The feature set includes single attribute features and interaction item attribute features. Based on historical production well oil production data within the target block, a mapping relationship between the feature set and oil production is established using a machine learning model; Step 4: Based on the machine learning model, calculate the contribution value of each attribute feature and interaction item attribute feature to the oil production prediction result, and perform statistical processing on the contribution values ​​to obtain the single attribute contribution statistic. and interaction contribution statistics And based on single-attribute contribution statistics and interaction contribution statistics To obtain the weight coefficient of a single attribute. and the weight coefficient of the interaction item ; Step 5: Weighting coefficients based on single attributes and the weight coefficient of the interaction item The original score of the dessert is obtained by weighting each feature. : ； In the formula: The original score for the target node dessert; For the first Normalized values ​​of each attribute parameter; For the first Individual attribute weight coefficients; For the first The first attribute and the first Interactive items consisting of attribute pairs of attributes; The interaction term adjustment coefficient is determined using cross-validation. For the first The first attribute and the first The weight coefficient of the interaction item of the attribute pair consisting of 1 attribute; The original scores of the desserts are normalized to obtain dessert coefficient values. The dessert coefficient values ​​are calculated for each grid node in the three-dimensional distribution model of the attribute parameters to construct a three-dimensional prediction model of dessert coefficients. Step 6: Based on the three-dimensional prediction model of the sweet spot coefficient, the area where the sweet spot coefficient is greater than the preset threshold is identified as the high sweet spot value area. During the drilling process, the sweet spot coefficient of the corresponding position of the well trajectory is updated in real time in combination with the logging-while-drilling data. When the sweet spot coefficient is lower than the preset threshold, the direction of the well trajectory is adjusted to shift it towards the area with a higher sweet spot coefficient.

2. The shale oil sweet spot prediction method based on multi-attribute fusion according to claim 1, characterized in that, In step four, based on the machine learning model, the contribution values ​​of each attribute parameter and interaction item attribute parameter to the oil production prediction results are calculated, and the contribution values ​​are statistically processed to obtain the single attribute contribution statistics and interaction contribution statistics. The calculation process is as follows: Single attribute contribution statistics: ; Interaction contribution statistics: ; in, For the first Attribute contribution statistics for each attribute; For the first In the nth sample The contribution value of each attribute; For the first The first attribute and the first The interaction contribution statistics of attribute pairs consisting of 1 attribute; For the first In the nth sample, the nth The first attribute and the first The interaction contribution value of attribute pairs consisting of 1 attribute; The total number of samples in the target block; For the first One sample; Based on the single-attribute contribution statistics and interaction contribution statistics, the weight coefficients of the single attributes are obtained. and the weight coefficient of the interaction item The specific process is as follows: Weight coefficient of a single attribute: ; Weight coefficients of interaction items: ; in, For the first The weight coefficient of each attribute; For the first Contribution statistics of each attribute; The sum of the contribution statistics for all single attributes; For the first The first attribute and the first The weight coefficient of the interaction item of the attribute pair consisting of 1 attribute; For the first The first attribute and the first The interaction contribution statistics of attribute pairs consisting of 1 attribute; The sum of the statistics for all selected interactions.

3. The shale oil sweet spot prediction method based on multi-attribute fusion according to claim 1, characterized in that, The cross-validation method involves determining the optimal interaction term adjustment coefficient based on the oil production data of drilled wells, iterating through the values ​​of α within a preset range [0, 1.50], calculating the mean square error between the sweet spot coefficient and the actual oil production using leave-one-out cross-validation for each α value, selecting the α value with the smallest mean square error as the optimal interaction term adjustment coefficient, and applying the optimal interaction term adjustment coefficient to the reduction calculation of the interaction term contribution statistics.

4. The shale oil sweet spot prediction method based on multi-attribute fusion according to claim 1, characterized in that, The normalization in step five includes ； in: This is the dessert coefficient; The original score for the target node dessert; This represents the theoretical minimum of the dessert's original score. This represents the theoretical maximum value of the dessert's original score.

5. The shale oil sweet spot prediction method based on multi-attribute fusion according to claim 1, characterized in that, The machine learning model is an ensemble learning model based on decision trees, and the contribution value is calculated by using the TreeSHAP method to calculate the SHAP interaction value matrix.

6. The shale oil sweet spot prediction method based on multi-attribute fusion according to claim 1, characterized in that, The lithofacies model is established using the truncated Gaussian stochastic simulation method. Specifically, the continuous Gaussian random field is converted into discrete lithofacies types by setting multiple thresholds, and the spatial continuity of lithofacies is controlled by the variogram function to construct a spatial constraint framework for joint modeling of multiple attribute parameters.

7. The shale oil sweet spot prediction method based on multi-attribute fusion according to claim 1, characterized in that, The interaction term attribute features include attribute combinations used to characterize the nonlinear relationship between different reservoir attributes, and the attribute combinations include porosity and permeability, oil saturation and total organic carbon content, and oil saturation and brittleness index.

8. A shale oil sweetness prediction system based on multi-attribute fusion, characterized in that, include: The data acquisition and preprocessing module is used to acquire the well location coordinates, stratification data, single well elevation data, logging data, well logging data, sedimentary facies data and core analysis data of the target block, and to perform quality control, outlier processing and normalization on the data, and output attribute data in a unified format. The construction modeling module, connected to the data acquisition and preprocessing module, is used to establish the top structural surface of each oil layer subgroup based on the layered data and single-well elevation data, and to construct a three-dimensional mesh model based on preset mesh parameters to generate a three-dimensional structural model of the target block. The lithofacies modeling module, connected to the structural modeling module, is used to generate a three-dimensional lithofacies model within the three-dimensional spatial framework provided by the three-dimensional structural model, based on the spatial distribution probability of lithofacies types and the variogram parameters, using the truncated Gaussian simulation method. The multi-attribute parameter modeling module, connected to the lithofacies modeling module, is used to establish porosity, permeability, oil saturation, and total organic carbon content, respectively, using the lithofacies 3D model as a constraint framework. A well logging interpretation model for the brittleness index was developed, and a three-dimensional distribution model of each attribute parameter was generated using the sequential Gaussian simulation method. The weight coefficient calculation and dessert model construction module is connected to the multi-attribute parameter modeling module and is used to normalize each attribute parameter and construct a feature set containing single attributes and interaction items. Based on the trained machine learning model, the contribution values ​​of each attribute parameter and interaction item to the oil production prediction results are calculated; the contribution values ​​are statistically processed to obtain single attribute contribution statistics and interaction contribution statistics; based on the single attribute contribution statistics and interaction contribution statistics, single attribute weight coefficients and interaction weight coefficients are obtained; based on the single attribute weight coefficients and interaction weight coefficients, the original score of the dessert is determined; the original score of the dessert is normalized to obtain the dessert coefficient value; the dessert coefficient is calculated for each grid node in the three-dimensional distribution model to construct a three-dimensional prediction model of the dessert coefficient. The horizontal well trajectory optimization module is connected to the weight coefficient calculation and sweet spot model construction module, and determines the area where the sweet spot coefficient is higher than the preset threshold based on the three-dimensional prediction model of the sweet spot coefficient. During drilling, the sweet spot coefficient corresponding to the well location is updated in real time based on logging-while-drilling data; when the updated sweet spot coefficient is lower than the preset threshold, the well trajectory direction is adjusted so that the well trajectory shifts towards the area with a higher sweet spot coefficient.

9. A shale oil sweet spot prediction system based on multi-attribute fusion according to claim 8, characterized in that, The weighting coefficient calculation and dessert model construction module includes: The processing unit is used to normalize the attribute values ​​in the three-dimensional distribution model of each attribute parameter and construct a feature set containing single attributes and interaction items. The contribution calculation unit is used to calculate the contribution value of each attribute parameter and interaction term to the oil production prediction result based on the trained machine learning model. The statistics and weighting calculation unit is used to perform statistical processing on the absolute value of the contribution value to obtain the single attribute contribution statistic and the interaction contribution statistic, and to obtain the single attribute weight coefficient and the interaction weight coefficient based on the single attribute contribution statistic and the interaction contribution statistic. The unit for calculating the sweetness coefficient and constructing a three-dimensional prediction model for the sweetness coefficient is used to calculate the original score of the sweetness coefficient of shale oil based on the single attribute weight coefficient and the interaction weight coefficient, normalize the original score of the sweetness coefficient to obtain the sweetness coefficient value, calculate the sweetness coefficient for each grid node in the three-dimensional distribution model, and construct a three-dimensional prediction model for the sweetness coefficient.

Images