Shale oil reservoir dessert intelligent identification method and system

Abstract

The invention provides a shale oil reservoir dessert intelligent identification method and system, relates to the field of shale oil reservoirs, and aims to obtain original logging data and extract curve deep space-time coding features by fusing a hybrid model of a convolutional neural network and a recurrent neural network. Then, multi-source logging features are dynamically fused through a cross-curve feature fusion algorithm, and spatial compression is carried out through feature principal component analysis; thirdly, implicit features strongly related to the geological signs of the sweet spots are screened through semantic level selection based on feature principal components, sweet spot probability prediction is achieved based on a lightweight classification model, an intelligent recognition closed loop without manual intervention is formed, and the dependence of a traditional method on expert experience is broken through; in this way, non-linear correlation in the logging curve is automatically mined through deep learning, and the interpretability and generalization ability of complex reservoir dessert recognition are improved.

Description

Technical Field

[0001] The present application relates to the field of shale oil reservoirs, and more particularly, in an embodiment of the present application, to a method and system for intelligently identifying sweet spots in shale oil reservoirs. Background Art

[0002] As global energy demand continues to grow and conventional oil and gas development becomes increasingly challenging, the strategic importance of unconventional oil and gas resources, such as shale oil, is becoming increasingly prominent. Shale oil reservoirs are typically characterized by low porosity and permeability. Accurately identifying these "sweet spots" (i.e., high-quality reservoir sections with high oil content, high mobility, and high fracturing susceptibility) has become a key technical challenge in determining development profitability.

[0003] Traditional sweet spot identification relies primarily on geologists' empirical interpretation of well logging curves, using manual extraction of curve morphological features (such as gamma-ray anomaly returns and resistivity box-shaped structures) combined with rock physical parameter calculations for comprehensive judgment. However, the strong heterogeneity of shale oil reservoirs leads to complex spatiotemporal coupling in logging responses, and conventional methods face three major technical bottlenecks: First, raw logging data, as high-dimensional time series signals, contain multiple interference factors such as instrument noise and wellbore interference, which can easily lead to model overfitting when directly modeled; second, manual feature design struggles to capture deep nonlinear correlations, such as the indicative role of spatiotemporal synergistic variation patterns between different curves in pore-fracture systems, which can be crucial for sweet spot identification; and third, the traditional well-by-well interpretation model is inefficient and cannot meet the needs of large-scale three-dimensional sweet spot prediction.

[0004] Therefore, an intelligent identification solution for shale oil reservoir sweet spots is desired. Summary of the Invention

[0005] In order to solve the above-mentioned technical problems, the present application is proposed. An embodiment of the present application provides a method and system for intelligent identification of sweet spots in shale oil reservoirs, which obtains original logging data and extracts deep spatiotemporal coding features of the curves through a hybrid model that integrates convolutional neural networks and recurrent neural networks. Subsequently, multi-source logging features are dynamically fused through a cross-curve feature fusion algorithm, and feature principal component analysis is used for spatial compression. Next, implicit features that are strongly correlated with sweet spot geological markers are screened through semantic-level selection based on feature principal components, and sweet spot probability prediction is achieved based on a lightweight classification model, forming an intelligent identification closed loop that does not require human intervention, so as to break through the traditional method's reliance on expert experience. In this way, the nonlinear correlations in logging curves are automatically mined through deep learning, thereby improving the interpretability and generalization ability of complex reservoir sweet spot identification.

[0006] According to one aspect of the present application, a method for intelligently identifying sweet spots in shale oil reservoirs is provided, comprising:

[0007] Acquiring original logging data, wherein the original logging data includes a natural gamma ray curve, a resistivity curve, a sonic time difference curve, a density curve, and a neutron porosity curve;

[0008] Performing spatiotemporal coding fusion analysis on each curve in the original well logging data to obtain a multi-dimensional spatiotemporal fusion coding feature vector of the original well logging data;

[0009] Performing principal component analysis on the original well logging multi-dimensional spatiotemporal fusion coding feature vector to obtain a set of original well logging multi-dimensional spatiotemporal principal component coding feature vectors;

[0010] Performing feature principal component semantic level selection on the set of original well logging multidimensional spatiotemporal principal component encoding feature vectors to obtain a screening set of original well logging multidimensional spatiotemporal principal component encoding feature vectors;

[0011] Based on the screened set of the original well logging multi-dimensional spatiotemporal principal component encoding feature vectors, it is determined whether a sweet spot exists in the target rock formation.

[0012] According to another aspect of the present application, a shale oil reservoir sweet spot intelligent identification system is provided, comprising:

[0013] The original logging data acquisition module is used to acquire the original logging data, wherein the original logging data includes a natural gamma ray curve, a resistivity curve, an acoustic time difference curve, a density curve, and a neutron porosity curve;

[0014] The original well logging spatiotemporal coding fusion analysis module is used to perform spatiotemporal coding fusion analysis on each curve in the original well logging data to obtain a multi-dimensional spatiotemporal fusion coding feature vector of the original well logging;

[0015] An original well logging feature principal component analysis module is used to perform feature principal component analysis on the original well logging multi-dimensional spatiotemporal fusion coding feature vector to obtain a set of original well logging multi-dimensional spatiotemporal principal component coding feature vectors;

[0016] An original well logging feature principal component semantic level selection module is used to perform feature principal component semantic level selection on the set of original well logging multidimensional spatiotemporal principal component encoding feature vectors to obtain a screening set of original well logging multidimensional spatiotemporal principal component encoding feature vectors;

[0017] The target rock formation sweet spot determination module is used to determine whether the target rock formation has a sweet spot based on the screening set of the original well logging multi-dimensional spatiotemporal principal component encoding feature vectors.

[0018] Compared with the existing technology, the present application provides a method and system for intelligent identification of shale oil reservoir sweet spots, which obtains original logging data and extracts deep spatiotemporal coding features of the curves through a hybrid model that integrates convolutional neural networks and recurrent neural networks. Subsequently, multi-source logging features are dynamically fused through a cross-curve feature fusion algorithm, and feature principal component analysis is used for spatial compression. Next, implicit features that are strongly correlated with sweet spot geological signs are screened through semantic-level selection based on feature principal components, and sweet spot probability prediction is achieved based on a lightweight classification model, forming an intelligent identification closed loop that does not require human intervention, so as to break through the traditional method's reliance on expert experience. In this way, the nonlinear correlations in logging curves are automatically mined through deep learning, thereby improving the interpretability and generalization ability of complex reservoir sweet spot identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0020] Figure 1 Flowchart of a method for intelligently identifying sweet spots in shale oil reservoirs according to an embodiment of the present application.

[0021] Figure 2 Schematic diagram of data flow for the intelligent identification method of shale oil reservoir sweet spots according to an embodiment of the present application.

[0022] Figure 3 This is a flowchart of performing semantic-level feature principal component selection on the set of original well logging multi-dimensional spatiotemporal principal component encoding feature vectors in the intelligent identification method for shale oil reservoir sweet spots according to an embodiment of the present application to obtain a screening set of original well logging multi-dimensional spatiotemporal principal component encoding feature vectors.

[0023] Figure 4 This is a flowchart of a method for intelligently identifying sweet spots in shale oil reservoirs according to an embodiment of the present application, in which a coding vector is aggregated based on the multi-dimensional spatiotemporal principal component prior information of the original well logging, and a feature distillation process based on query prompt strength analysis is performed on the multi-dimensional spatiotemporal principal component coding feature vector of the original well logging to be filtered to determine whether to filter the multi-dimensional spatiotemporal principal component coding feature vector of the original well logging to be filtered.

[0024] Figure 5 This is a system block diagram of a shale oil reservoir sweet spot intelligent identification system according to an embodiment of the present application. DETAILED DESCRIPTION

[0025] Various exemplary embodiments, features, and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise indicated.

[0026] The word “exemplary” is used exclusively herein to mean “serving as an example, example, or illustration.” Any embodiment described herein as “exemplary” is not necessarily to be construed as preferred or advantageous over other embodiments.

[0027] In addition, numerous specific details are provided in the following detailed description to better illustrate the present application. Those skilled in the art will appreciate that the present application can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art are not described in detail in order to highlight the main purpose of the present application.

[0028] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. Throughout the description of this application, "plurality" means two or more, unless otherwise specifically defined.

[0029] With the growth of global energy demand and the increasing difficulty of developing conventional oil and gas resources, the importance of unconventional resources such as shale oil is increasing. The low porosity and low permeability characteristics of shale oil reservoirs make the accurate identification of "sweet spots" (i.e., high-quality reservoir sections with high oil content, high mobility, and high fracturing ability) the key to improving development efficiency. Traditional methods rely on geologists' empirical interpretation of well logging curves. However, due to the strong heterogeneity of shale oil reservoirs, the logging response is complex, and there are technical bottlenecks such as high-dimensional data processing prone to overfitting, artificial feature design having difficulty capturing deep nonlinear correlations, and low efficiency of well-by-well interpretation that cannot meet the needs of large-scale three-dimensional prediction. Therefore, improving sweet spot identification technology is crucial to improving the efficiency of shale oil development.

[0030] In response to the above technical problems, in the technical solution of this application, a method for intelligent identification of shale oil reservoir sweet spots is proposed, which realizes end-to-end decision-making from raw data to sweet spot identification by constructing an automated analysis link of multimodal logging data. This method takes key logging curves such as natural gamma and resistivity as input. First, each curve is jointly modeled in time and space by fusing a hybrid model of convolutional neural network (CNN) and recurrent neural network (RNN): the CNN layer captures subtle changes in the local morphology of the curve (such as abnormal fluctuations in the gamma curve and box boundaries in the resistivity curve), while the RNN layer analyzes the evolution of the logging response along the depth of the formation (such as the gradual trend of the acoustic time difference). The two layers work together to extract deep spatiotemporal coding features of the curve with spatiotemporal correlation. Subsequently, the multi-source logging features are dynamically fused through a cross-curve feature fusion algorithm, which enhances the synergistic effect of the combination of parameters such as porosity-resistivity. In order to eliminate noise interference and reduce dimensionality, feature principal component analysis is used to spatially compress the fused features and retain the key components that are sensitive to the sweet spots. Furthermore, semantic-level selection based on principal component features is used to screen implicit features strongly correlated with sweet spot geological markers, forming a low-dimensional, highly discriminative feature subset. Finally, a lightweight classification model is used to predict sweet spot probabilities, creating an intelligent identification closed loop that requires no human intervention. This technical solution transcends the traditional reliance on expert experience. By using deep learning to automatically mine nonlinear correlations in well logging curves, it provides a basis for intelligent identification of sweet spots in shale oil reservoirs, improving the interpretability and generalization of sweet spot identification in complex reservoirs.

[0031] This application proposes an intelligent identification method for shale oil reservoir sweet spots. Figure 1 Flowchart of a method for intelligently identifying sweet spots in shale oil reservoirs according to an embodiment of the present application. Figure 2 Schematic diagram of data flow of the method for intelligently identifying sweet spots in shale oil reservoirs according to an embodiment of the present application. Figure 1 and Figure 2 As shown, the intelligent identification method of shale oil reservoir sweet spots according to an embodiment of the present application includes: S110, obtaining original logging data, wherein the original logging data includes a natural gamma curve, a resistivity curve, an acoustic time difference curve, a density curve, and a neutron porosity curve; S120, performing spatiotemporal coding fusion analysis on each curve in the original logging data to obtain a multi-dimensional spatiotemporal fusion coding feature vector of the original logging; S130, performing characteristic principal component analysis on the multi-dimensional spatiotemporal fusion coding feature vector of the original logging to obtain a set of multi-dimensional spatiotemporal principal component coding feature vectors of the original logging; S140, performing characteristic principal component semantic level selection on the set of multi-dimensional spatiotemporal principal component coding feature vectors of the original logging to obtain a screened set of multi-dimensional spatiotemporal principal component coding feature vectors of the original logging; S150, determining whether there is a sweet spot in the target rock formation based on the screened set of multi-dimensional spatiotemporal principal component coding feature vectors of the original logging.

[0032] In the aforementioned intelligent identification method for shale oil reservoir sweet spots, step S110 acquires raw logging data, which includes natural gamma ray curves, resistivity curves, acoustic transit time curves, density curves, and neutron porosity curves. It should be understood that each curve in the raw logging data reflects different physical properties of the formation, providing a rich information base for subsequent analysis. The natural gamma ray curve records the content of radioactive elements in the formation and can help distinguish between formation types such as sandstone and mudstone. The resistivity curve is used to assess the water and oil content of the formation, inferring the properties of the fluids within it by measuring the formation's resistance to electric current. The acoustic transit time curve provides information about rock porosity and can be used to infer the pore structure of the formation based on the differences in the propagation speed of sound waves in different media. The density curve reveals the density distribution of the formation and is important for understanding the rock skeleton and its internal fillings. The neutron porosity curve indirectly reflects the formation porosity by detecting the number of hydrogen atoms in the formation, making it particularly suitable for identifying formations containing light hydrocarbons. This data not only covers the basic physical properties of the formation but also includes a variety of parameters closely related to oil and gas occurrence, such as oil saturation, pore structure, and fracture development. Through in-depth analysis of this raw logging data, key information reflecting the sweet spots in shale oil reservoirs can be effectively extracted. The acquisition and utilization of this raw logging data lays a solid foundation for the subsequent use of advanced deep learning techniques to automatically discover the deep nonlinear correlations hidden in the massive amount of data.

[0033] To obtain well logging data, field drilling is first required. During this process, drilling equipment penetrates the surface and travels deep underground until it reaches the target formation. Simultaneously, logging-while-drilling (LWD) or wireline logging techniques are employed to collect the necessary data. LWD uses sensors mounted on the drill string to acquire formation parameters in real time while drilling. This method allows for data acquisition without disrupting drilling progress and, because it operates close to the newly exposed formation surface, provides more accurate information. In contrast, wireline logging involves lowering a cable equipped with multiple sensors into the wellbore after drilling is complete. Each logging type has its own unique measurement principles and technical requirements. For example, gamma ray logging relies on detecting gamma rays emitted by radioactive elements such as uranium, thorium, and potassium in the formation. By recording the time and intensity of these rays reaching the detector, a gamma ray curve can be constructed, which can then be used to identify organic-rich shale layers. Resistivity logging uses electrodes to send an electric current into the formation and measures the strength of the return signal to assess the conductivity of the formation, which is crucial for determining the presence of oil reservoirs. Acoustic time-of-day logging typically uses a sound source and receiver to measure the time required for sound waves to travel through the formation. This time is proportional to the porosity of the formation and can therefore effectively reflect the pore structure of the formation. Density logging relies on the degree of scattering or absorption of gamma rays after interacting with formation materials to determine the density of the formation, which is critical for understanding the mineral composition and mechanical properties of the formation. Finally, neutron porosity logging estimates the porosity of the formation by emitting high-energy neutrons into the formation and then slowing down the neutrons after colliding with hydrogen atoms in the formation. This technology is particularly suitable for formations rich in light hydrocarbons.

[0034] Ensuring data quality and accuracy is crucial throughout the data acquisition process. This requires not only the use of advanced logging techniques and equipment, but also a series of measures to minimize external interference. For example, when performing resistivity logging, the impact of borehole conditions on measurement results must be considered, as irregularities in the borehole wall can alter the current path, affecting the final resistivity reading. Similarly, when performing transit-time logging, care must be taken to avoid changes in the acoustic wave propagation path caused by the presence of gas or other anomalies within the borehole. Furthermore, to ensure data consistency and comparability, all logging operations should adhere to standardized standards and specifications, including selecting the appropriate measurement frequency, adjusting gain settings, and calibrating various sensors. Once the raw logging data is acquired, the next task is preliminary data processing and validation. The goal of this stage is to ensure that the acquired data is comprehensive and reliable, meeting the requirements of subsequent analysis. This typically involves steps such as data cleaning, format conversion, and basic quality control checks. For example, data points that are clearly noisy or do not conform to expected patterns may be removed. Consistency checks are also performed across the entire data set to identify logical inconsistencies between parameters.

[0035] In the above-mentioned shale oil reservoir sweet spot intelligent identification method, the step S120, performing time-space coding fusion analysis on each curve in the original logging data to obtain the original logging multi-dimensional time-space fusion coding feature vector, includes: S121, performing time-space mixing feature analysis on each curve in the original logging data to obtain the natural gamma time-space coding feature vector, the resistivity time-space coding feature vector, the acoustic wave time difference time-space coding feature vector, the density time-space coding feature vector and the neutron porosity time-space coding feature vector; S122, performing feature fusion on the natural gamma time-space coding feature vector, the resistivity time-space coding feature vector, the acoustic wave time difference time-space coding feature vector, the density time-space coding feature vector and the neutron porosity time-space coding feature vector to obtain the original logging multi-dimensional time-space fusion coding feature vector.

[0036] Specifically, in step S121, a time-space mixing feature analysis is performed on each curve in the original logging data to obtain a natural gamma-ray time-space coding feature vector, a resistivity time-space coding feature vector, an acoustic time difference time-space coding feature vector, a density time-space coding feature vector, and a neutron porosity time-space coding feature vector. In an embodiment of the present application, the time-space mixing feature analysis is a time-space mixing modeling based on a CNN-RNN hybrid model. It should be understood that, considering that the logging response of shale oil reservoirs has significant time-space coupling characteristics: the logging curve is both a time series signal of the change of formation physical properties with depth (such as the vertical gradient of resistivity reflecting the change of oil saturation), and contains local morphological mutation characteristics (such as the abnormal peak of natural gamma-ray indicating the lithologic interface). Traditional single network architectures (such as using only CNN or RNN) have difficulty balancing the dual characteristics of time and space. That is, while CNN is good at capturing local spatial patterns (such as the boundary characteristics of resistivity box structures), it cannot model the longitudinal temporal dependencies of logging curves. While RNN can process sequence information (such as the gradual trend of acoustic wave transit time), it lacks sensitivity to local morphology. Therefore, in the technical solution of this application, each curve in the original logging data is subjected to spatiotemporal hybrid modeling based on a CNN-RNN hybrid model to obtain natural gamma ray spatiotemporal coding feature vectors, resistivity spatiotemporal coding feature vectors, acoustic wave transit time spatiotemporal coding feature vectors, density spatiotemporal coding feature vectors, and neutron porosity spatiotemporal coding feature vectors. A CNN-RNN hybrid model is used to perform independent spatiotemporal modeling on each well logging curve. The goal is to extract local sensitive features of the curve (such as pore structure changes corresponding to abrupt changes in the density curve) through sliding convolution kernels of the CNN, while simultaneously utilizing the gating mechanism of the RNN to analyze the evolution of logging parameters along the formation depth (such as the periodic fluctuations in the neutron porosity curve reflecting sedimentary cycles). The two work together to uncover deep temporal and spatial correlations implicit in the logging signals. More specifically, by performing spatiotemporal hybrid modeling on each curve based on the CNN-RNN hybrid model, noise interference in the original logging data (such as gamma curve glitches caused by wellbore collapse) is eliminated through spatiotemporal hybrid modeling. The local feature abstraction capability of the CNN is leveraged to filter high-frequency noise, while the temporal smoothing properties of the RNN are used to suppress the influence of outliers. Furthermore, this method overcomes the limitations of manual feature design, automatically capturing nonlinear synergistic patterns between different curves (such as the role of resistivity and acoustic transit time in indicating fracture development), and extracting deep spatiotemporal features that are difficult to quantify using traditional methods (such as the implicit correlation between density curve gradient changes and oil content). In addition, it can also provide a unified representation space for subsequent cross-curve feature fusion. By independently encoding the spatiotemporal characteristics of each curve (such as the lithologic sensitivity characteristics of the natural gamma ray curve and the fluid response characteristics of the neutron porosity curve), it can retain their physical meaning differences while establishing interactive semantic associations.In this way, the spatiotemporal encoding feature vectors of different curves generated in this step not only significantly reduce the dimensionality of the original data, but also enhance the expression of sweet spot sensitive information through deep nonlinear transformation (such as encoding the box-shaped structure of the resistivity curve into a separable pattern in the high-dimensional feature space), so that the weak sweet spot signals that were originally submerged by noise (such as acoustic wave time difference anomalies corresponding to microfractures) can be highlighted, laying a high-information-density feature foundation for subsequent feature fusion and selection, and ultimately improving the generalization ability of the sweet spot recognition model for complex heterogeneous reservoirs.

[0037] Specifically, step S122 involves fusing the gamma ray spatiotemporal encoding feature vector, the resistivity spatiotemporal encoding feature vector, the acoustic transit time spatiotemporal encoding feature vector, the density spatiotemporal encoding feature vector, and the neutron porosity spatiotemporal encoding feature vector to obtain the original well logging multidimensional spatiotemporal fusion encoding feature vector. It should be understood that sweet spot identification in shale oil reservoirs relies on a multi-parameter coordinated response mechanism: a single logging curve only reflects local information about reservoir physical properties (e.g., a gamma ray curve indicates lithology, and a resistivity curve represents oil content), while the formation of sweet spots is controlled by the coupling of multiple factors such as pore structure, fracture development, and fluid saturation. Therefore, it is necessary to capture implicit coordinated patterns through cross-curve feature interactions (e.g., the difference between the density curve and neutron porosity reflects oil saturation, and the temporal offset between acoustic transit time and resistivity indicates fracture connectivity). Traditional artificial feature combinations have difficulty quantifying these nonlinear cross-curve correlations, especially when the logging response is affected by noise (e.g., when wellbore expansion causes distortion of the density curve), and single curve features are prone to misjudgment. Therefore, the natural gamma ray space-time encoding feature vector, the resistivity space-time encoding feature vector, the acoustic transit time space-time encoding feature vector, the density space-time encoding feature vector, and the neutron porosity space-time encoding feature vector are further fused to obtain the original well logging multidimensional space-time fusion encoding feature vector. This feature fusion method maps the space-time encoding vectors of each curve into a unified representation space, aiming to construct a comprehensive feature set with geophysical significance. This not only preserves the physical property information reflected independently by each curve (such as the lithologic sensitivity characteristics of the gamma curve and the pore pressure response of the acoustic transit time), but also strengthens the indicative role of key combinations (such as the directionality of resistivity-density coordinated changes to oil-bearing sweet spots) through dynamic weight allocation between parameters (such as the attention mechanism). In this way, the complementary information of multi-source logging data can be integrated to make up for the characterization limitations of single curve features. For example, when the natural gamma curve fails due to interference from the wellbore environment, the sweet spot can still be effectively identified by integrating the fluid response characteristics of the neutron porosity curve. It also mines the implicit correlation between the curves and encodes complex collaborative patterns that are difficult to describe with human experience (such as the abnormal attenuation of acoustic time differences when resistivity box structures appear) into quantifiable high-dimensional features, breaking through the linear assumptions of traditional rock physics equations. In addition, it can also construct a unified feature space across curves, providing an intermediate expression compatible with geological semantics for subsequent dimensionality reduction and feature selection, avoiding information loss caused by heterogeneity of feature spaces.In this way, the multi-dimensional spatiotemporal fusion encoding feature vector of the original logging effectively amplifies the discriminant signal of the sweet spot through nonlinear interaction (such as forming a high-discrimination pattern by associating the weak box-shaped characteristics of the resistivity curve with the gradient change of the density curve), while suppressing the interference of single curve noise (such as the gamma anomaly caused by wellbore collapse is compensated by the stable characteristics of other curves). The fused feature set not only contains deep spatiotemporal correlation information (such as the multi-curve time series offset caused by fracture development), but also has cross-well consistency, providing a highly robust feature basis for large-scale three-dimensional sweet spot prediction, and ultimately improving the generalization ability and interpretation credibility of the model in complex heterogeneous reservoirs.

[0038] In the above-mentioned shale oil reservoir sweet spot intelligent identification method, step S130 performs a principal component analysis on the original well logging multi-dimensional spatiotemporal fusion encoding feature vector to obtain a set of original well logging multi-dimensional spatiotemporal principal component encoding feature vectors. It should be understood that although cross-curve feature fusion captures multi-parameter collaborative patterns (such as the joint indication of resistivity and acoustic wave time difference on fractures) through nonlinear interaction, the fused high-dimensional feature set (for example, thousands of dimensions) still contains a large number of collinear components (such as repeated representations of the same geological attribute by different curves) and noise interference (such as the residual effect of wellbore collapse in multiple curves). These redundant features not only increase the computational burden, but are more likely to drown out weak but critical sweet spot discrimination signals (such as weak time series offsets of multiple curves caused by microfractures). Therefore, in the technical solution of the present application, a principal component analysis is further performed on the original well logging multi-dimensional spatiotemporal fusion encoding feature vector to obtain a set of original well logging multi-dimensional spatiotemporal principal component encoding feature vectors. Principal component analysis (PCA) is used to perform an orthogonal transformation on the fused features, aiming to discover the projection direction that maximizes the data variance and map the high-dimensional feature space into a low-dimensional orthogonal subspace, thereby eliminating linear correlations between features and highlighting global variation patterns that are sensitive to sweet spots. Feature principal component analysis compresses feature dimensions through linear dimensionality reduction, solving the "curse of dimensionality" problem and providing efficient input for subsequent lightweight classification models. Furthermore, the orthogonal properties of the principal components are utilized to remove noise interference (such as the overall curve offset caused by instrument drift) while retaining core features that can explain the essential differences in reservoir physical properties (such as the principal component direction reflecting sudden changes in oil saturation). Furthermore, a feature base with geostatistical significance is constructed, establishing potential correlations between the principal component vectors and key parameters of the sweet spot (such as fracture density and organic matter content). For example, the first principal component may correspond to a comprehensive indicator of "oil content-fracturability," while the second principal component reflects "pore structure heterogeneity." In terms of performance, the principal component encoding feature set, which selects a low-dimensional feature subset based on the variance maximization principle, not only retains the effective information in the fused features (such as the coordinated variation pattern between the resistivity gradient and the acoustic time derivative), but also significantly improves the robustness of the features. Specifically, the sweet spot discriminant signals originally scattered in multidimensional space (such as oil saturation reflected by the difference in neutron-density curves) are concentrated into a few principal components (for example, the high loading of the third principal component corresponds to the oil sweet spot feature), enabling the subsequent feature distillation network to more accurately select discriminant features with clear geological semantics. At the same time, the orthogonal nature of the principal components effectively decouples the coupling effects of different geological factors (such as separating lithologic changes and fluid responses into different principal components), enhancing the model's adaptability to complex heterogeneous reservoirs and providing high signal-to-noise ratio input features for large-scale three-dimensional sweet spot prediction.

[0039] Figure 3This is a flowchart of performing semantic level feature principal component selection on the set of original well logging multi-dimensional spatiotemporal principal component encoding feature vectors in the shale oil reservoir sweet spot intelligent identification method according to an embodiment of the present application to obtain a screening set of original well logging multi-dimensional spatiotemporal principal component encoding feature vectors. Figure 3 As shown, in an embodiment of the present application, the step S140, performing feature principal component semantic level selection on the set of the original well logging multidimensional spatiotemporal principal component encoding feature vectors to obtain a screening set of the original well logging multidimensional spatiotemporal principal component encoding feature vectors, includes: S141, performing prior information aggregation analysis on the set of the original well logging multidimensional spatiotemporal principal component encoding feature vectors to obtain the original well logging multidimensional spatiotemporal principal component prior information aggregation encoding vector; S142, extracting the kth original well logging multidimensional spatiotemporal principal component encoding feature vector from the set of the original well logging multidimensional spatiotemporal principal component encoding feature vectors as the original well logging multidimensional spatiotemporal principal component encoding feature vector to be filtered; S143, based on the original well logging multidimensional spatiotemporal principal component prior information aggregation encoding vector, performing feature distillation processing based on query prompt strength analysis on the original well logging multidimensional spatiotemporal principal component encoding feature vector to be filtered to determine whether to filter the original well logging multidimensional spatiotemporal principal component encoding feature vector to be filtered. It should be understood that semantic ambiguity and redundancy still exist in the feature space after principal component analysis: although principal component analysis extracts the characteristic direction with the largest variance (such as the third principal component reflecting the oil-bearing mutation) through orthogonal transformation, the geophysical meaning of each principal component is not explicitly associated with the sweet spot discrimination criteria (such as fracture density, organic matter abundance, etc.), resulting in some principal components corresponding to non-critical geological factors (such as residual wellbore environmental noise) or redundant parameter combinations (such as the linear superposition of resistivity-acoustic time difference). At the same time, traditional principal component selection methods rely on statistical indicators such as variance contribution rate and cannot dynamically screen key features based on the geological semantics of the sweet spot (such as the principal component sensitive to fracturability may have a low variance share but strong discriminative power). Therefore, the set of original well logging multidimensional spatiotemporal principal component encoding feature vectors is further subjected to feature principal component semantic level selection to obtain a screened set of original well logging multidimensional spatiotemporal principal component encoding feature vectors. The semantic-level selection method of feature principal components is adopted, aiming to construct a global prior semantic framework (such as fracture development pattern and oil saturation threshold) through self-supervised learning, and dynamically mine the implicit association between principal component features and sweet spot geological signs (such as the nonlinear mapping relationship between a principal component vector and microfracture density) based on the query prompt mechanism to achieve semantic-driven optimization of feature selection.

[0040] Specifically, through semantic-level selection of principal components, self-supervised learning can be used to extract geological prior knowledge from unlabeled well logging data (e.g., low-value natural gamma ray intervals typically correspond to organic-rich layers). This prior information aggregation encoding vector is then generated as a global semantic template (e.g., encoding a "high resistivity + low density" pattern as an indication of oil-bearing sweet spots), providing geophysical constraints for feature distillation and semantic selection. Furthermore, a query prompt mechanism is used to establish an interactive channel between the multidimensional spatiotemporal principal component features of the original well logging data and the sweet spot semantics. For example, a principal component vector (e.g., a principal component reflecting pore structure complexity) is used as a query signal to model the association with the "high fracturability" pattern in the prior template. The nonlinear mapping between the two is captured through the implicit encoding matrix of the feature distillation query prompt (e.g., when the principal component is above a certain threshold, the probability of fracture development increases). Furthermore, a dynamic decision-making mechanism enables adaptive feature selection. For example, when the feature distillation strength descriptor indicates that a principal component has low semantic relevance to the sweet spot (e.g., a descriptor value below a threshold), it is removed to reduce interference, while principal components that are highly synergistic with the global prior are retained (e.g., principal components with high descriptor values correspond to oil saturation-sensitive features). The selection of the selected set achieves dual optimization through semantic-level selection: First, global prior injection based on self-supervised learning (e.g., fracture development patterns in the prior information aggregation encoding vector) enhances the geological interpretability of features, enabling the selected principal components to be directly associated with key sweet spot parameters (e.g., the first selected principal component corresponds to the combined "oil content-fracture density" metric). Second, dynamic distillation under query prompt constraints (e.g., optimizing local-global feature dependencies using a neighborhood empathy gain coefficient matrix) effectively removes statistically significant but semantically irrelevant principal components (e.g., principal components with high variance contribution but only reflecting lithologic variation), while strengthening weak but discriminative implicit features (e.g., principal components with low variance contribution but strong correlation with movable fluid saturation). It is worth mentioning that each screening feature after semantic-level feature selection has a clear geological semantic orientation (for example, the third screening principal component is specifically used to identify high-brittleness layers), enabling subsequent classification models to achieve accurate discrimination based on high-purity feature sets, reducing the misjudgment rate of complex heterogeneous reservoirs, and providing a traceable geophysical basis for three-dimensional sweet spot prediction.

[0041] Specifically, in step S141, a priori information aggregation analysis is performed on the set of the original well logging multidimensional spatiotemporal principal component encoding feature vectors to obtain the original well logging multidimensional spatiotemporal principal component priori information aggregation encoding vector, which is expressed as the original well logging priori information aggregation analysis formula:

[0042] X={x1,x2,...,x i ,...,x n}

[0043]

[0044] Among them, X is the set of original logging multi-dimensional spatiotemporal principal component encoding feature vectors, x1, x2, x i and x n are the first, second, i-th and n-th original well logging multi-dimensional space-time principal component encoding feature vectors in the set of original well logging multi-dimensional space-time principal component encoding feature vectors, f prior (X) is the self-supervisory processing of the intrinsic mode signal prior information of X, max(x i ) and min(x i ) are x i The maximum and minimum values in ε are adjusted to prevent the denominator from being zero. i is the dynamic range coefficient of the multidimensional spatiotemporal principal component characteristic of the original well logging, and a i is the dynamic range weight coefficient of the multidimensional spatiotemporal principal component characteristic of the original well logging, n is the number of vectors in X, V prior is the aggregated prior information encoding vector of the multidimensional spatiotemporal principal components of the original well log. It should be understood that the self-supervised learning framework can autonomously mine geophysical laws from unlabeled well log data without relying on manual labeling. This self-supervised mechanism forces the network to learn the inherent structural patterns of the data through predefined tasks (such as calculating feature dynamic range weights). Performing prior information aggregation analysis can generate a globally context-aware aggregated prior information encoding vector of the multidimensional spatiotemporal principal components of the original well log. Essentially, this involves a structured semantic reconstruction of the multidimensional principal component features, making implicit geological correlations (such as the nonlinear coupling between fracture development patterns and oil saturation) that were previously scattered across independent principal components explicit through the self-supervised task into interpretable global feature templates. By integrating the global statistical properties of multi-curve spatiotemporal encodings (such as the dynamic range weights after maximum-minimum normalization), the aggregated prior information encoding vector of the original well log multidimensional spatiotemporal principal components constructs a prior knowledge representation of key reservoir geological markers (such as the combination of high oil content and high fracturability). This provides strong semantic constraints for subsequent feature distillation and effectively improves the model's ability to capture weak sweet spot signals. In this way, self-supervised learning can be used to compress the redundant information in the original high-dimensional features and generate low-dimensional and high-purity geological semantic carriers. Moreover, by making implicit association patterns explicit, quantifiable geophysical association criteria can be established for the feature selection stage, thereby reducing the complexity of the model while enhancing the reliability of sweet spot identification.

[0045] Specifically, in step S142, the kth original well logging multidimensional space-time principal component coding feature vector is extracted from the set of the original well logging multidimensional space-time principal component coding feature vectors as the original well logging multidimensional space-time principal component coding feature vector to be filtered, and the multidimensional space-time principal component coding feature extraction formula of the original well logging to be filtered is expressed as:

[0046] x target =x k

[0047] Among them, x k is the kth original well logging multidimensional space-time principal component encoding feature vector in the set of original well logging multidimensional space-time principal component encoding feature vectors, x target The kth original well log multidimensional spatiotemporal principal component encoding feature vector is used as the original well log multidimensional spatiotemporal principal component encoding feature vector to be filtered. It should be understood that although principal component analysis can compress dimensions, different principal components may carry redundant geophysical information (such as repeated representations of the same reservoir parameter by multiple curves) or noise interference (such as the residual effect of the wellbore environment in multiple curves). Moreover, the orthogonal transformation process of the principal components leads to the abstraction of their geological semantics, making it difficult to directly associate them with the sweet spot discrimination criteria. Therefore, it is necessary to traverse all original well log multidimensional spatiotemporal principal component encoding feature vectors (represented by the kth one) in the set of original well log multidimensional spatiotemporal principal component encoding feature vectors and dynamically screen potential redundant or low-discriminative features one by one. This establishes an entry point for feature selection. By treating each original well log multidimensional spatiotemporal principal component encoding feature vector as an independent object to be evaluated, it gives it equal screening opportunities and ensures that the subsequent feature distillation process covers all possible redundant or irrelevant features. By achieving refined filtering of the feature space, it is possible to specifically analyze the correlation strength between each original well logging multi-dimensional spatiotemporal principal component encoding feature vector and the geological markers of the sweet spot, dynamically strip off invalid features, and retain feature subsets that are sensitive to key parameters of the sweet spot, such as fracture development and oil saturation. In this way, through the feature-by-feature review mechanism, it is possible to eliminate the redundant noise that still exists after the principal component analysis, improve the purity and discrimination efficiency of the set of original well logging multi-dimensional spatiotemporal principal component encoding feature vectors, and enable adaptive identification and strengthening of implicit features that are strongly related to the semantics of the sweet spot, thereby reducing computational complexity while enhancing the accuracy and interpretability of sweet spot discrimination. Here, the kth original well logging multi-dimensional spatiotemporal principal component encoding feature vector represents any original well logging multi-dimensional spatiotemporal principal component encoding feature vector in the set of original well logging multi-dimensional spatiotemporal principal component encoding feature vectors, and it is regarded as the original well logging multi-dimensional spatiotemporal principal component encoding feature vector to be filtered.

[0048] Figure 4 This is a flowchart of the method for intelligently identifying sweet spots in shale oil reservoirs according to an embodiment of the present application, which aggregates the encoding vector based on the multi-dimensional spatiotemporal principal component prior information of the original well logging, and performs feature distillation processing based on query prompt strength analysis on the multi-dimensional spatiotemporal principal component encoding feature vector of the original well logging to be filtered to determine whether to filter the multi-dimensional spatiotemporal principal component encoding feature vector of the original well logging to be filtered. Figure 4As shown, in an embodiment of the present application, the step S143, based on the original well logging multi-dimensional spatiotemporal principal component prior information aggregate coding vector, performs feature distillation processing based on query prompt strength analysis on the multi-dimensional spatiotemporal principal component coding feature vector of the original well logging to be filtered to determine whether to filter the multi-dimensional spatiotemporal principal component coding feature vector of the original well logging to be filtered, including: S1431, implicitly encoding the multi-dimensional spatiotemporal principal component coding feature vector of the original well logging to be filtered and the multi-dimensional spatiotemporal principal component prior information aggregate coding vector of the original well logging based on feature filtering query to obtain the original well logging multi-dimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix; S1432, based on the original well logging multi-dimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix, determines the original well logging multi-dimensional spatiotemporal principal component feature filtering strength descriptor; S1433, based on the original well logging multi-dimensional spatiotemporal principal component feature filtering strength descriptor, determines whether to filter the original well logging multi-dimensional spatiotemporal principal component coding feature vector to be filtered.

[0049] Specifically, in step S1431, the multi-dimensional spatiotemporal principal component encoding feature vector of the original well logging to be filtered and the multi-dimensional spatiotemporal principal component prior information aggregation encoding vector of the original well logging are implicitly encoded based on feature filtering query to obtain the original well logging multi-dimensional spatiotemporal principal component feature filtering query hint implicit encoding matrix, which is expressed as the implicit encoding formula based on feature filtering query:

[0050]

[0051] Among them, g prompt (x target ,V prior ) is the pair x target and V prior Perform feature filtering query prompt processing, is the matrix multiplication, V prior T It is V prior The transposed vector of V prior Length, M k is x kThe corresponding implicit encoding matrix of the original well log multidimensional spatiotemporal principal component feature filtering query prompts. It should be understood that the implicit encoding based on feature filtering queries introduces query prompts as supervisory signals. This step uses the multidimensional spatiotemporal principal component encoding feature vector of the original well log to be filtered as the query signal, deeply interacting with the aggregated encoding vector of the original well log multidimensional spatiotemporal principal component prior information, which contains global geological prior knowledge, to construct the implicit encoding matrix of the original well log multidimensional spatiotemporal principal component feature filtering query prompts to explicitly demonstrate the correlation strength between the two. This process autonomously explores the implicit mapping relationship between features and sweet spot criteria (such as fracture development pattern and oil saturation threshold) through predefined self-supervised learning tasks (such as feature dynamic range weight calculation), thus giving the originally abstract orthogonal principal component features interpretable geophysical meaning. This establishes a dynamic feature evaluation mechanism that provides an adaptive screening basis for subsequent feature distillation by quantifying the semantic consistency of principal component features with global priors. Among them, through the focusing effect of query prompts, principal component noise irrelevant to the semantics of sweet spots (such as the residual resistivity abnormal fluctuations caused by wellbore collapse) can be filtered out. In addition, the generation of the implicit coding matrix of the original logging multi-dimensional spatiotemporal principal component feature filtering query prompts makes the correlation strength between the principal component features and the geological targets explicit as a computable weight parameter, enabling the model to suppress redundant information while retaining key discriminant features, ultimately improving the generalization ability and decision reliability of sweet spot identification.

[0052] In an embodiment of the present application, the step S1432, based on the implicit coding matrix of the original well logging multi-dimensional spatiotemporal principal component feature filtering query prompt, determines the original well logging multi-dimensional spatiotemporal principal component feature filtering strength descriptor, including: S1432-1, performing fusion optimization based on local-global correlation dependency on the implicit coding matrix of the original well logging multi-dimensional spatiotemporal principal component feature filtering query prompt to obtain the optimized original well logging multi-dimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix; S1432-2, performing matrix-based trace measurement on the optimized original well logging multi-dimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix to obtain the original well logging multi-dimensional spatiotemporal principal component feature filtering strength descriptor.

[0053] Specifically, step S1432-1 performs fusion optimization based on local-global correlation dependency on the original well logging multidimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix to obtain an optimized original well logging multidimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix, which is expressed as a fusion optimization formula based on local-global correlation dependency:

[0054] α k+1,k-1 =||concat(M k+1 ,M k-1 )|| F

[0055]

[0056] Among them, M k+1 and M k-1 They are x k+1 and x k-1 The corresponding original well logging multidimensional spatiotemporal principal component feature filtering query prompts the implicit coding matrix, concat(·,·) is a cascade operation, ||·|| F To calculate the Frobenius norm, α k+1,k-1 is x k The corresponding original logging multidimensional spatiotemporal principal component neighborhood autocorrelation weight coefficient, β k+1,k-1 is x k The corresponding original logging multidimensional space-time principal component neighborhood cosine cross-weight coefficient, ⊙ is the point multiplication of the position, E k+1,k-1 is x k The corresponding original logging multidimensional spatiotemporal principal component sympathetic gain coefficient matrix, M' k It's M k The optimized original well logging multidimensional spatiotemporal principal component feature filtering query suggests an implicit coding matrix. It should be understood that due to the strong heterogeneity and complex spatiotemporal coupling characteristics of shale oil reservoirs, a single principal component feature may only reflect the transient changes of a local well logging curve (such as abnormal resistivity fluctuations), while the formation of sweet spots is often controlled by regional geological processes across wellbores and horizons (such as the global extension of the fracture network). Therefore, it is necessary to deeply integrate the implicit coding features suggested by the original well logging multidimensional spatiotemporal principal component feature filtering query with the spatiotemporal correlation patterns within its neighborhood by introducing the autocorrelation weight coefficient of the original well logging multidimensional spatiotemporal principal component neighborhood (to control local static consistency) and the cosine cross-weight coefficient of the original well logging multidimensional spatiotemporal principal component neighborhood (to control dynamic feature flow). This step constructs a sympathetic gain coefficient matrix for the original well log multidimensional spatiotemporal principal component, explicitly characterizing the distributional sensitivity association between local features and global priors. This allows the implicit encoding matrix suggested by the original well log multidimensional spatiotemporal principal component feature filtering query to retain the complexity of the original high-dimensional features while enhancing the hierarchical expression of geological semantics. This approach overcomes the limitations of traditional single-well interpretation through local-global information complementation, enabling the establishment of a multiscale feature association network and providing an optimized feature base with both local sensitivity and global consistency for subsequent feature distillation. Neighborhood feature coupling significantly enhances the model's ability to capture hidden geological features such as microfractures, and the sympathetic gain mechanism optimizes feature distribution sensitivity. This allows the optimized implicit encoding matrix suggested by the original well log multidimensional spatiotemporal principal component feature filtering query to both reflect the dynamic response characteristics of individual wells and conform to the regional structural-lithological-fluid coupling law, thereby achieving more accurate sweet spot identification in complex reservoir scenarios.

[0057] Specifically, step S1432-2 performs a matrix-based trace measurement on the optimized original well logging multi-dimensional spatiotemporal principal component feature filter query hint implicit coding matrix to obtain the original well logging multi-dimensional spatiotemporal principal component feature filter intensity descriptor, which is expressed as a matrix-based trace measurement formula:

[0058] s k =Tr(M′ k )

[0059] Where Tr(·) is the trace value of the matrix, s k is x k The corresponding raw well log multidimensional spatiotemporal principal component feature filter strength descriptor. It should be understood that although the optimized raw well log multidimensional spatiotemporal principal component feature filter query hint implicit coding matrix, after optimization through local-global correlation dependency fusion, enhances the correlation between features and global priors, its internal structure still contains complex nonlinear interactions. Directly using it for feature screening results in high computational complexity and lacks a unified evaluation standard. Therefore, it is necessary to reduce the high-dimensional optimized raw well log multidimensional spatiotemporal principal component feature filter query hint implicit coding matrix through matrix-based trace metric to extract core metrics that characterize the strength of feature filtering. This step aims to condense the complex feature correlations within the optimized raw well log multidimensional spatiotemporal principal component feature filter query hint implicit coding matrix into a single scalar value. By quantifying the semantic consistency and optimization potential of features with global geological priors, it provides a computable basis for subsequent feature distillation decisions. This approach establishes an objective evaluation system for feature screening, using trace metric to capture the discriminative power of principal component features in the global context, thereby adaptively distinguishing key features from redundant noise. The matrix-based trace measurement can significantly reduce the computational load of subsequent processing and improve the model operation efficiency.

[0060] Specifically, in step S1433, based on the original well logging multi-dimensional spatiotemporal principal component feature filtering strength descriptor, it is determined whether to filter the original well logging multi-dimensional spatiotemporal principal component coding feature vector to be filtered, and the multi-dimensional spatiotemporal principal component coding feature filtering determination formula of the original well logging to be filtered is expressed as:

[0061]

[0062] Among them, Select(x target) determines whether to filter the original well log multidimensional spatiotemporal principal component encoded feature vector to be filtered, where θ is a trainable preset threshold. It should be understood that due to the strong heterogeneity of shale oil reservoirs, the discriminative value of a single feature varies significantly under different geological conditions. Using a static threshold for filtering may result in the inadvertent deletion of key features or the retention of redundant features. Therefore, this step dynamically evaluates the semantic consistency and optimization potential of the current feature with the global geological prior by analyzing the quantized value of the feature filter strength descriptor in real time, thereby establishing an adaptive decision-making mechanism. This step aims to achieve refined control of feature screening, dynamically balancing the discriminative power and redundancy of features through quantitative indicators to ensure that key geological signals are retained while avoiding over-filtering. Specifically, when the original well log multidimensional spatiotemporal principal component feature filter strength descriptor indicates that the current feature already has high expressive power, filtering can be skipped to save computational resources. However, when the original well log multidimensional spatiotemporal principal component feature filter strength descriptor indicates that the feature still has significant room for optimization, filtering is initiated to further improve its representation quality. By dynamically adjusting the filtering strategy, computational efficiency can be optimized while maintaining sweet spot identification accuracy, thereby improving the model's adaptability to complex reservoir environments. This dynamic thresholding of descriptors avoids unnecessary filtering of highly discriminative features, preserving the integrity of key geological signals such as microfracture development. Furthermore, the adaptive strategy reduces redundant computations, enabling efficient operation even with large-scale 3D logging data while maintaining high feature purity and strong interpretability.

[0063] In the above-mentioned shale oil reservoir sweet spot intelligent identification method, the step S150 determines whether there is a sweet spot in the target rock formation based on the filtered set of the original well logging multidimensional spatiotemporal principal component encoding feature vectors, including: S151, feature aggregation of each original well logging multidimensional spatiotemporal principal component encoding feature vector in the filtered set of the original well logging multidimensional spatiotemporal principal component encoding feature vectors to obtain an original well logging multidimensional spatiotemporal principal component fusion representation vector; S152, passing the original well logging multidimensional spatiotemporal principal component fusion representation vector through a classifier-based sweet spot identifier to obtain an identification result, and the identification result is used to indicate whether there is a sweet spot in the target rock formation.

[0064] Specifically, step S151 performs feature aggregation on each of the original well logging multidimensional spatiotemporal principal component encoding feature vectors in the screened set to obtain an original well logging multidimensional spatiotemporal principal component fusion representation vector. It should be understood that feature aggregation is not simply a simple numerical superposition or averaging operation, but rather involves intelligent integration based on the geological significance of each original well logging multidimensional spatiotemporal principal component encoding feature vector and their relationships with each other. Although each original well logging multidimensional spatiotemporal principal component encoding feature vector has been rigorously screened and independently reflects certain specific geological attributes or phenomena, when these features are combined, they can often reveal more complex and subtle formation characteristics. For example, in shale oil reservoirs, the presence and development of fractures have a decisive impact on oil and gas recovery. However, fracture information may be dispersed across multiple different well logging curves, such as abnormal attenuation on acoustic wave transit time curves, gradient changes on resistivity curves, and local fluctuations on density curves. Feature aggregation effectively combines these seemingly unrelated but actually closely connected features, resulting in a more accurate depiction of the overall fracture landscape and enhancing the potential synergy between the multidimensional spatiotemporal principal component encoding feature vectors of different raw well logs. Furthermore, even after principal component analysis, a certain number of raw well log multidimensional spatiotemporal principal component encoding feature vectors may still remain in the filtered set. Directly inputting these features into the classifier for sweet spot prediction not only increases the computational burden but also may lead to overfitting of the model. Feature aggregation reduces the input dimensionality and avoids the risk of overfitting caused by too many features. Furthermore, because the aggregated raw well log multidimensional spatiotemporal principal component fusion representation vectors contain richer geological information, they also help improve the generalization ability of the classification model, enabling it to maintain high accuracy even when faced with unseen data. Specifically, in a specific embodiment of the present application, by using the weighted average method, a corresponding weight coefficient is assigned according to the importance of each original well logging multidimensional space-time principal component encoding feature vector, and then all the original well logging multidimensional space-time principal component encoding feature vectors are added according to the weights to obtain the final original well logging multidimensional space-time principal component fusion representation vector.

[0065] Specifically, in step S152, the multi-dimensional spatiotemporal principal component fusion representation vector of the original well logging is passed through a classifier-based sweet spot identifier to obtain an identification result, and the identification result is used to indicate whether there is a sweet spot in the target rock formation. It should be understood that in this way, the screening features with clear geological semantics (such as the "fracture development index") can be converted into engineering-usable sweet spot discrimination labels, resolving the contradiction in traditional methods where "features are interpretable but decisions are not traceable." In addition, the classification model is based on a high-discriminative feature set after semantic screening to more accurately judge the sweet spots of the target rock formation. In particular, in complex heterogeneous reservoirs (such as interlayer development sections), it can effectively identify "hidden sweet spots" (such as the acoustic time difference-resistivity synergistic anomaly caused by microfracture development) that are easily missed by traditional empirical rules, providing reliable technical support for the efficient development of shale oil. Specifically, the technical solution of the present application automatically determines whether the stratum represented by the input original well logging multi-dimensional spatiotemporal principal component fusion representation vector is a sweet spot area by inputting the trained classification model (such as a support vector machine, random forest or neural network), and using the shale oil reservoir sweet spot related patterns and features learned within the classifier-based sweet spot identifier. This process relies on the learning of a large number of known samples, so that the classifier-based sweet spot identifier can accurately capture the deep nonlinear correlation related to the sweet spot geological signs, and output a probability value or directly give a binary judgment (sweet spot / non-sweet spot) based on this, thereby realizing an intelligent recognition closed loop without human intervention, and improving the interpretability and generalization ability of complex reservoir sweet spot identification. The final recognition result is used to indicate whether there is a sweet spot in the target rock formation, providing a decision-making basis for the effective development of shale oil resources.

[0066] In summary, a method for intelligent identification of shale oil reservoir sweet spots based on an embodiment of the present application is illustrated, which obtains original logging data and extracts deep spatiotemporal coding features of the curves through a hybrid model that integrates convolutional neural networks and recurrent neural networks. Subsequently, multi-source logging features are dynamically fused through a cross-curve feature fusion algorithm, and feature principal component analysis is used for spatial compression. Next, implicit features that are strongly correlated with sweet spot geological markers are screened through semantic-level selection based on feature principal components, and sweet spot probability prediction is achieved based on a lightweight classification model, forming an intelligent identification closed loop that does not require human intervention, so as to break through the traditional method's reliance on expert experience. In this way, the nonlinear correlations in logging curves are automatically mined through deep learning, thereby improving the interpretability and generalization ability of complex reservoir sweet spot identification.

[0067] Figure 5 FIG. 1 is a system block diagram of a shale oil reservoir sweet spot intelligent identification system according to an embodiment of the present application. Figure 5As shown, the shale oil reservoir sweet spot intelligent identification system 100 according to the embodiment of the present application includes: an original logging data acquisition module 110, which is used to acquire original logging data, wherein the original logging data includes a natural gamma curve, a resistivity curve, a sonic time difference curve, a density curve, and a neutron porosity curve; an original logging spatiotemporal coding fusion analysis module 120, which is used to perform spatiotemporal coding fusion analysis on each curve in the original logging data to obtain a multi-dimensional spatiotemporal fusion coding feature vector of the original logging; and an original logging feature principal component analysis module 130, which is used to analyze the original logging features. The original well logging multi-dimensional spatiotemporal fusion coding feature vectors are subjected to characteristic principal component analysis to obtain a set of original well logging multi-dimensional spatiotemporal principal component coding feature vectors; an original well logging characteristic principal component semantic level selection module 140 is used to perform characteristic principal component semantic level selection on the set of original well logging multi-dimensional spatiotemporal principal component coding feature vectors to obtain a screening set of original well logging multi-dimensional spatiotemporal principal component coding feature vectors; a target rock formation sweet spot determination module 150 is used to determine whether a target rock formation has a sweet spot based on the screening set of original well logging multi-dimensional spatiotemporal principal component coding feature vectors.

[0068] Here, those skilled in the art will appreciate that the specific operations of each step in the above-mentioned shale oil reservoir sweet spot intelligent identification system have been described in the above reference. Figures 1 to 4 The description of the intelligent identification method of shale oil reservoir sweet spots has been introduced in detail, and therefore, its repeated description will be omitted.

[0069] As described above, the shale oil reservoir sweet spot intelligent identification system 100 according to an embodiment of the present application can be implemented in various terminal devices. In one example, the shale oil reservoir sweet spot intelligent identification system 100 can be integrated into the terminal device as a software module and / or hardware module. For example, the shale oil reservoir sweet spot intelligent identification system 100 can be a software module in the terminal device's operating system, or can be an application developed for the terminal device. Of course, the shale oil reservoir sweet spot intelligent identification system 100 can also be one of the terminal device's many hardware modules.

[0070] Alternatively, in another example, the shale oil reservoir sweet spot intelligent identification system 100 and the terminal device may also be separate devices, and the shale oil reservoir sweet spot intelligent identification system 100 may be connected to the terminal device via a wired and / or wireless network and transmit interactive information in accordance with an agreed data format.

[0071] In summary, an intelligent identification system for shale oil reservoir sweet spots based on an embodiment of the present application is illustrated, which obtains original logging data and extracts deep spatiotemporal coding features of the curves through a hybrid model that integrates convolutional neural networks and recurrent neural networks. Subsequently, multi-source logging features are dynamically fused through a cross-curve feature fusion algorithm, and feature principal component analysis is used for spatial compression. Next, implicit features that are strongly correlated with sweet spot geological markers are screened through semantic-level selection based on feature principal components, and sweet spot probability prediction is achieved based on a lightweight classification model, forming an intelligent identification closed loop that does not require human intervention, so as to break through the traditional method's reliance on expert experience. In this way, the nonlinear correlations in logging curves are automatically mined through deep learning, thereby improving the interpretability and generalization ability of complex reservoir sweet spot identification.

Claims

1. A method for intelligently identifying sweet spots in shale oil reservoirs, characterized in that: include: Acquiring original logging data, wherein the original logging data includes a natural gamma ray curve, a resistivity curve, a sonic time difference curve, a density curve, and a neutron porosity curve; Performing spatiotemporal coding fusion analysis on each curve in the original well logging data to obtain a multi-dimensional spatiotemporal fusion coding feature vector of the original well logging data; Performing principal component analysis on the original well logging multi-dimensional spatiotemporal fusion coding feature vector to obtain a set of original well logging multi-dimensional spatiotemporal principal component coding feature vectors; Performing feature principal component semantic level selection on the set of original well logging multidimensional spatiotemporal principal component encoding feature vectors to obtain a screening set of original well logging multidimensional spatiotemporal principal component encoding feature vectors; Based on the screened set of the original well logging multi-dimensional spatiotemporal principal component encoding feature vectors, it is determined whether a sweet spot exists in the target rock formation.

2. The method for intelligently identifying sweet spots in shale oil reservoirs according to claim 1, characterized in that: Performing spatiotemporal coding fusion analysis on each curve in the original well logging data to obtain a multi-dimensional spatiotemporal fusion coding feature vector of the original well logging data includes: Performing a time-space mixed characteristic analysis on each curve in the original logging data to obtain a natural gamma ray time-space coding characteristic vector, a resistivity time-space coding characteristic vector, an acoustic time difference time-space coding characteristic vector, a density time-space coding characteristic vector, and a neutron porosity time-space coding characteristic vector; The natural gamma ray space-time coding feature vector, the resistivity space-time coding feature vector, the acoustic time difference space-time coding feature vector, the density space-time coding feature vector and the neutron porosity space-time coding feature vector are feature fused to obtain the original well logging multi-dimensional space-time fusion coding feature vector.

3. The method for intelligently identifying sweet spots in shale oil reservoirs according to claim 2, characterized in that: The spatiotemporal mixing feature analysis is spatiotemporal mixing modeling based on the CNN-RNN hybrid model.

4. The method for intelligently identifying sweet spots in shale oil reservoirs according to claim 3, characterized in that: Performing feature principal component semantic level selection on the set of original well logging multidimensional spatiotemporal principal component encoding feature vectors to obtain a screening set of original well logging multidimensional spatiotemporal principal component encoding feature vectors, including: Performing a priori information aggregation analysis on the set of original well logging multi-dimensional spatiotemporal principal component encoding feature vectors to obtain an original well logging multi-dimensional spatiotemporal principal component priori information aggregation encoding vector; Extracting the kth original well logging multidimensional space-time principal component encoding feature vector from the set of original well logging multidimensional space-time principal component encoding feature vectors as the original well logging multidimensional space-time principal component encoding feature vector to be filtered; Based on the multi-dimensional spatiotemporal principal component prior information aggregation coding vector of the original well logging, feature distillation processing based on query prompt strength analysis is performed on the multi-dimensional spatiotemporal principal component coding feature vector of the original well logging to be filtered to determine whether to filter the multi-dimensional spatiotemporal principal component coding feature vector of the original well logging to be filtered.

5. The method for intelligently identifying sweet spots in shale oil reservoirs according to claim 4, characterized in that: Based on the prior information aggregation coding vector of the original well logging multidimensional spatiotemporal principal component, a feature distillation process based on query prompt strength analysis is performed on the multidimensional spatiotemporal principal component coding feature vector of the original well logging to be filtered to determine whether to filter the multidimensional spatiotemporal principal component coding feature vector of the original well logging to be filtered, including: Performing implicit coding based on feature filtering query on the multi-dimensional spatiotemporal principal component coding feature vector of the original well logging to be filtered and the multi-dimensional spatiotemporal principal component prior information aggregation coding vector of the original well logging to obtain a multi-dimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix of the original well logging; Determine the original well logging multi-dimensional spatiotemporal principal component feature filtering intensity descriptor based on the original well logging multi-dimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix; Based on the original well logging multi-dimensional spatiotemporal principal component feature filtering strength descriptor, it is determined whether to filter the original well logging multi-dimensional spatiotemporal principal component encoding feature vector to be filtered.

6. The method for intelligently identifying sweet spots in shale oil reservoirs according to claim 5, characterized in that: Determining the original well logging multidimensional spatiotemporal principal component feature filtering intensity descriptor based on the original well logging multidimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix includes: Performing fusion optimization based on local-global correlation dependency on the original well logging multi-dimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix to obtain an optimized original well logging multi-dimensional spatiotemporal principal component feature filtering query prompt implicit coding matrix; The optimized original well logging multi-dimensional spatiotemporal principal component feature filter query prompt implicit coding matrix is subjected to matrix-based trace measurement to obtain the original well logging multi-dimensional spatiotemporal principal component feature filter intensity descriptor.

7. The method for intelligently identifying sweet spots in shale oil reservoirs according to claim 6, characterized in that: Determining whether a target rock formation has a sweet spot based on a screening set of multi-dimensional spatiotemporal principal component encoding feature vectors of the original well logging includes: Performing feature aggregation on each original well logging multidimensional space-time principal component encoding feature vector in the screening set of the original well logging multidimensional space-time principal component encoding feature vectors to obtain an original well logging multidimensional space-time principal component fusion representation vector; The multi-dimensional spatiotemporal principal component fusion representation vector of the original well logging is passed through a classifier-based sweet spot identifier to obtain a recognition result, and the recognition result is used to indicate whether a sweet spot exists in the target rock formation.

8. An intelligent identification system for shale oil reservoir sweet spots, characterized in that: include: The original logging data acquisition module is used to acquire the original logging data, wherein the original logging data includes a natural gamma ray curve, a resistivity curve, an acoustic time difference curve, a density curve, and a neutron porosity curve; The original well logging spatiotemporal coding fusion analysis module is used to perform spatiotemporal coding fusion analysis on each curve in the original well logging data to obtain a multi-dimensional spatiotemporal fusion coding feature vector of the original well logging; An original well logging feature principal component analysis module is used to perform feature principal component analysis on the original well logging multi-dimensional spatiotemporal fusion coding feature vector to obtain a set of original well logging multi-dimensional spatiotemporal principal component coding feature vectors; An original well logging feature principal component semantic level selection module is used to perform feature principal component semantic level selection on the set of original well logging multidimensional spatiotemporal principal component encoding feature vectors to obtain a screening set of original well logging multidimensional spatiotemporal principal component encoding feature vectors; The target rock formation sweet spot determination module is used to determine whether the target rock formation has a sweet spot based on the screening set of the original well logging multi-dimensional spatiotemporal principal component encoding feature vectors.

9. The shale oil reservoir sweet spot intelligent identification system according to claim 8, characterized in that: The original well logging spatiotemporal coding fusion analysis module is used to: Performing a time-space mixed characteristic analysis on each curve in the original logging data to obtain a natural gamma ray time-space coding characteristic vector, a resistivity time-space coding characteristic vector, an acoustic time difference time-space coding characteristic vector, a density time-space coding characteristic vector, and a neutron porosity time-space coding characteristic vector; The natural gamma ray space-time coding feature vector, the resistivity space-time coding feature vector, the acoustic time difference space-time coding feature vector, the density space-time coding feature vector and the neutron porosity space-time coding feature vector are feature fused to obtain the original well logging multi-dimensional space-time fusion coding feature vector.

Images