A machine learning-based multi-dimensional index coupling driving lost circulation prediction method

Abstract

The application discloses a kind of multi-dimensional index coupling driving lost circulation prediction method based on machine learning, belong to geological engineering integration technical field.The application carries out pre-processing to seismic, well logging, well logging multi-source data, obtains standardization data body and fracture distribution structure body, and builds work area fine geological structure model;Statistical lost circulation spatial distribution and loss amount, after normalizing using Sigmoid function, generate risk trend model CLRI1 by spatial interpolation;According to four-level logic, the index system of nine static indexes is constructed;Adopt symbolic regression machine learning training to obtain prediction master model CLRI2;After three-dimensional connectivity analysis and weight adjustment to CLRI2, weighted fusion with CLRI1, generate comprehensive loss risk model CLRI.The application realizes the accurate quantification of multi-source data fusion and nonlinear coupling relationship, can output risk distribution and dominant factor, effectively guide drilling design and risk prevention and control.

Description

Technical Field

[0001] This invention relates to the field of integrated geological engineering research technology, and in particular to a multi-dimensional index coupling-driven well leakage prediction method based on machine learning. Background Technology

[0002] Accurate prediction of well leakage is a key prerequisite for efficient well leakage control, and clarifying the inducing factors of well leakage is the foundation for accurate prediction. Well leakage refers to the loss of a significant portion or all of the drilling fluid from the wellbore into a highly permeable formation. Well leakage is the result of the synergistic effect of geological conditions, engineering parameters, and fluid characteristics, requiring multi-factor coupled analysis for accurate prediction and control. First, in terms of material basis, there needs to be an effective leakage channel that can seep through; the formation needs to have natural or potential leakage channels with a size larger than the particle size of the drilling fluid solids. Second, there must be mechanical conditions that satisfy the opening or expansion of the channel, i.e., parameters such as formation rock strength and tectonic stress state can promote the transformation of potential channels into effective seepage channels. Finally, the external triggering condition is that the positive pressure difference formed by the engineering mud parameters reaches the critical leakage threshold, driving the drilling fluid to seep into the formation, ultimately causing well leakage.

[0003] However, existing well leakage prediction methods still have significant shortcomings. First, the data types are limited, with well logging data, seismic data, and other data not being combined. Second, the prediction indicators are singular and isolated, resulting in an incomplete static indicator system. Third, a dedicated static indicator coupling modeling process has not been developed for well leakage prediction, making it impossible to deeply integrate multi-dimensional static indicators to form an accurate risk characterization field. Furthermore, existing prediction models do not fully explore the correlation patterns between static indicators, making it difficult to simultaneously predict leakage risk and its dominant factors based on static data. Summary of the Invention

[0004] To address the aforementioned problems, this invention provides a multi-dimensional index coupling-driven well leakage prediction method based on machine learning, which can solve the problems of incomplete static index system, inability to predict leakage type, and insufficient prediction pertinence and accuracy in existing technologies.

[0005] The technical solution of this invention is:

[0006] A machine learning-based multidimensional index coupling-driven well leakage prediction method includes:

[0007] S1. Preprocess multi-source heterogeneous data from seismic, well logging, and well logging sources to obtain standardized multi-source data volumes and fracture distribution structures;

[0008] S2. Based on the standardized multi-source data volume and fracture distribution structure, a fine geological structure model of the work area is built using a well-seismic combined method;

[0009] S3. Based on the detailed geological structure model of the work area, a multi-dimensional static index system structure containing nine static indicators is constructed according to the four-level logic of macro-geological control, micro-characterization of leakage channels, stratigraphic mechanical property constraints, and engineering geological coupling threshold determination.

[0010] S4. Statistically analyze the spatial distribution and leakage of well leakage in the work area, use the Sigmoid function to perform risk normalization assignment, and generate the well leakage risk trend structure model CLRI1 through spatial interpolation.

[0011] S5. Using the aforementioned multi-dimensional static index system architecture as input and the measured well leakage risk value as the objective, a symbolic regression machine learning algorithm is used for training to derive an analytical formula containing static indicators, thus obtaining the well leakage prediction master model structure CLRI2.

[0012] S6. Perform three-dimensional spatial connectivity analysis on the well leakage prediction master model structure CLRI2, increase the risk weight for areas with strong connectivity and decrease the weight for isolated small areas to obtain the optimized well leakage prediction master model structure CLRI2'.

[0013] S7. The well leakage risk trend structure model CLRI1 and the optimized well leakage prediction master model structure CLRI2' are fused through a weighted fusion mechanism to generate a comprehensive leakage risk index structure model CLRI.

[0014] Furthermore, the specific steps of S1 are as follows:

[0015] The logging curves are trimmed and denoised. Based on the geological survey results and GR curves, lithology is classified. A standardized multi-source data volume is obtained through logging curve calculation. The standardized multi-source data volume includes information on porosity, permeability, saturation and rock cementation.

[0016] The seismic body is subjected to structural smoothing and amplitude contrast processing to reduce the signal-to-noise ratio. Variance and chaotic bodies are generated for boundary detection. A passive-then-active tracking method is adopted, supplemented by low-dip attitude control to extract the crack distribution structure.

[0017] Furthermore, S2 specifically includes:

[0018] Based on the standardized multi-source data obtained from S1, a detailed geological model of the work area was constructed. First, seismic interpretation fault data was imported to clarify the fault structure pattern, and a three-dimensional mesh framework was generated according to the work area boundary. Then, seismic interpretation layer data and well logging interpretation layer data were imported. Seismic layers were used to provide regional macro trends, and well logging layer data were used for high-precision constraints. Through a well-seismic joint collaborative modeling method, a three-dimensional geological model under dual constraints was constructed.

[0019] Furthermore, the four-level logic described in S3—macro-geological control, micro-characterization of leakage channels, stratigraphic mechanical property constraints, and engineering geological coupling threshold determination—specifically consists of:

[0020] Level 1: Macro-geological control, including sedimentary facies index X1 and lithological index X2, used to characterize the macro-control effect of regional sedimentary background and rock type on well leakage risk;

[0021] Level 2: Microscopic characterization of leakage channels, including porosity-cementation index X3, crack index X4 and weak surface index X5, used to characterize the development degree, connectivity and opening conditions of leakage channels.

[0022] Level 3: Formation mechanical property constraints, including the brittleness index X6 and the pore pressure anomaly index X7, are used to characterize the formation mechanical properties and the sensitivity of pressure state to leakage.

[0023] Level 4: Engineering-geological coupling threshold determination, including leakage pressure X8 and safe mud density window X9, is used to characterize the critical conditions that trigger well leakage under the coupling effect of drilling engineering parameters and geological conditions.

[0024] Furthermore, in S4, the well leakage risk trend structure model CLRI1 is generated using the Kriging space interpolation method.

[0025] Furthermore, the lithological index X2 is assigned a value based on a preset lithological risk assignment table, wherein mudstone and shale are assigned a value of 0.0 to 0.1, and limestone and dolomite are assigned a value of 0.8 to 1.0; the fracture index X4 is assigned a value by combining fracture development intensity, activation status, connectivity, and distance from fault; the weak surface index X5 is superimposed with a coefficient of 1.1 to 1.2 on unconformities or volcanic rock bodies, and a high threshold is directly assigned to the identified karst cave development areas.

[0026] Furthermore, the analytical formula containing static indicators described in S5 is as follows:

[0027] CLRI2=0.132+0.591*X4-0.147*X3*X4+0.105*exp(X3)+0.774*pow(X6,0.5),

[0028] Among them, X3 is the porosity-cementation index; X4 is the crack index; and X6 is the brittleness index.

[0029] Furthermore, the specific method for performing three-dimensional spatial connectivity analysis on the well leakage prediction master model structure CLRI2 described in S6 is as follows: calculate the volume or spatial continuity range of the connected volume to which each grid point belongs in the three-dimensional attribute volume of CLRI2, determine the continuous distribution area with a connected volume greater than a preset threshold as a region with strong connectivity, and determine the isolated distribution area with a connected volume less than a preset threshold as an isolated small region.

[0030] Furthermore, the specific rules for increasing the risk weight of highly connected regions and decreasing the weight of isolated small regions as described in S6 are as follows: for continuous regions with a connected volume greater than the first threshold, the original risk value is multiplied by a weighting coefficient greater than 1; for isolated regions with a connected volume less than the second threshold, the original risk value is multiplied by a weighting coefficient less than 1; the original risk value of intermediate regions remains unchanged, so as to avoid the homogenization of risk values ​​between small isolated risk areas and large continuous risk areas.

[0031] Furthermore, the specific formula for the weighted fusion described in S7 is as follows:

[0032] CLRI=α×CLRI1+(1-α)×CLRI2',

[0033] Wherein, CLRI1 is the well leakage risk trend structure model; CLRI2' is the optimized well leakage prediction master model structure; and α is the preset weight coefficient.

[0034] The beneficial effects of this invention are:

[0035] 1. In S5 of this invention, the analytical formula is automatically derived from the training data by using the symbolic regression algorithm, and in S7, the well leakage risk trend model and the machine learning master model are weighted and fused together, taking into account geological experience and data-driven approaches, which significantly improves the prediction accuracy of well leakage risk and solves the problems of insufficient prediction targeting and low accuracy of existing technologies.

[0036] 2. According to the four-level logic of macro-geological control, micro-characterization of leakage channels, stratigraphic mechanical property constraints, and engineering geological coupling threshold determination, this invention S3 constructs a multi-dimensional static index system structure containing nine static indicators. It deeply integrates multiple source indicators such as sedimentary facies, lithology, porosity-permeability cementation, fractures, weak surfaces, brittleness index, pore pressure anomaly, leakage pressure, and safe mud density window, overcoming the defects of existing technical index systems that are incomplete and lack correlation between indicators.

[0037] 3. The present invention S6 performs connectivity analysis on the three-dimensional risk body, increases the weight of continuous risk areas and decreases the weight of isolated small areas, thus avoiding risk homogenization. At the same time, based on the contribution of static indicators in the analytical formula, the dominant factors of leakage in each area can be identified simultaneously, guiding drilling design, mud parameter adjustment and well location optimization, thus solving the problem that the existing technology cannot predict the dominant factors of leakage.

[0038] 4. In this invention S2, a fault framework is constructed using seismic interpretation fault data, and regional macroscopic trends are provided using seismic interpretation layer data. At the same time, well logging layer data is used as a high-precision hard constraint, and well-seismic joint collaborative modeling technology is adopted to achieve a dual fusion of macroscopic structural trends and precise well depths.

[0039] 5. In this invention, S4 utilizes the Sigmoid function to accurately reflect the evolution of well leakage risk, namely, the slow growth of low leakage and the rapid rise of medium and high leakage. Compared with linear normalization, it is more in line with the actual leakage characteristics. In this step, Kriging space interpolation is based on the analysis of the spatial autocorrelation of sample points using the variogram function to accurately estimate unknown grid points, providing an auxiliary model that reflects the macroscopic leakage trend for subsequent dual-model fusion.

[0040] 6. This invention establishes a complete, reliable, and interpretable well leakage prediction technology solution, which can simultaneously output the spatial distribution of risk and the judgment results of dominant factors. It provides an intuitive and quantitative decision-making basis for drilling trajectory design, mud density window optimization, and well location deployment, and has significant engineering practical value and promotion prospects. Attached Figure Description

[0041] Figure 1 This is a flowchart of the method of the present invention;

[0042] Figure 2 This is a schematic diagram of seismic data preprocessing.

[0043] Figure 3 A schematic diagram of the well leakage prediction trend model CLRI1;

[0044] Figure 4 This is a schematic diagram of the X3 attribute model;

[0045] Figure 5 This is a schematic diagram of a crack model;

[0046] Figure 6 This is a schematic diagram of the X4 property model;

[0047] Figure 7 This is a schematic diagram of the X6 attribute model;

[0048] Figure 8 A schematic diagram of CLRI2, a multi-index coupled well leakage prediction model based on machine learning;

[0049] Figure 9 A schematic diagram of the CLRI dual-model weighted fusion well leakage prediction model, "CLRI1+CLRI2";

[0050] Figure 10 This is a chart showing the main indicators for judging well leakage risk. Detailed Implementation

[0051] The following will combine Figures 1-10The technical solutions in the embodiments of the present invention will be clearly and completely described herein. The described embodiments are merely some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments in this application without creative effort are within the scope of protection of this application.

[0052] Example 1:

[0053] This invention provides a machine learning-based multi-dimensional index coupling-driven well leakage prediction method, comprising:

[0054] S1. Preprocess multi-source heterogeneous data from seismic, well logging, and well logging sources to obtain standardized multi-source data volumes and fracture distribution structures;

[0055] Further, the specific steps of S1 are as follows: the well logging curves are trimmed and denoised; lithology is classified based on geological survey results and GR curves; a standardized multi-source data volume is obtained through well logging curve calculations, the standardized multi-source data volume including porosity, permeability, saturation and rock cementation information; the seismic volume is structurally smoothed and amplitude contrast is processed to reduce the signal-to-noise ratio; a variance volume and a chaotic volume are generated for boundary detection; a passive-then-active tracking method is adopted and supplemented by low dip angle attitude control to extract the fracture distribution structure.

[0056] Specifically, the well logging curves are first trimmed and denoised. In this embodiment, wavelet thresholding is used to filter out high-frequency noise. Then, lithology is classified based on natural gamma ray (GR) curves, with shale having the highest GR value, followed by sandstone, and carbonate and salt rocks having the lowest. This basic lithological classification is completed in conjunction with geological survey results. Next, a standardized multi-source data volume is obtained through well logging curve calculations. This data volume includes information on porosity, permeability, saturation, and rock cementation. Porosity is calculated using the sonic transit time curve based on the Wyllie time-averaged formula. Permeability is obtained by establishing a porosity-permeability regression model based on core measurement data. Saturation is derived from formation resistivity using the Archie formula to invert water saturation. The rock cementation is evaluated qualitatively and semi-quantitatively by comprehensively considering the response characteristics of sonic transit time, density, and resistivity curves. To reduce the signal-to-noise ratio, the seismic body undergoes structural smoothing and amplitude contrast processing. Structural smoothing employs median filtering, while amplitude contrast is used to equalize energy and enhance discontinuous boundary signals. Then, variance and chaotic volumes are generated for boundary detection. The variance volume highlights discontinuous regions by calculating local amplitude variance, while the chaotic volume quantifies the degree of data disorder based on structural tensor analysis. Following this, three-dimensional boundary enhancement processing is performed, and ant tracking technology is used to extract the fracture distribution structure: virtual ants move along paths with high pheromone concentrations. First, passive tracking is performed to identify macroscopic fractures with a high threshold, then active tracking is performed to capture secondary fractures with a low threshold. Simultaneously, attitude control is used to filter low-dip horizon information, ultimately yielding the ant-like structure as the fracture distribution structure. Figure 2 As shown. Finally, the above standardized multi-source data volume and fracture distribution structure are unified into the spatial coordinate system of the geological model of the work area through spatial resampling and coordinate mapping, for use in the subsequent construction of a fine geological structure model.

[0057] S2. Based on standardized multi-source data volumes, a fine geological structure model of the work area is built using a well-seismic combined method. Further, S2 specifically includes: importing seismic interpretation fault data, clarifying the fault structure pattern, and generating a three-dimensional mesh framework based on the work area boundary; then importing seismic interpretation layer data and well logging interpretation layer data, using seismic layers to provide regional macro trends, using well logging layer data for high-precision constraints, and using a well-seismic joint collaborative modeling method to realize the construction of a three-dimensional geological model under dual constraints.

[0058] Specifically, the process begins by importing seismic interpretation fault data, classifying the faults, and removing outliers. A three-dimensional fault frame mesh is then generated using a fault frame modeling method. Next, the mesh step size is set according to the work area boundary and study depth: a planar step size of 25 meters by 25 meters, and a vertical step size ranging from 1 to 5 meters depending on the stratigraphic thickness and accuracy requirements. Corner grid technology is used to generate a three-dimensional mesh frame covering the entire work area, ensuring the mesh is discontinuous along the fault planes. Then, seismic interpretation bedding plane data and well logging interpretation bedding plane data are imported. The seismic interpretation bedding plane data provides the regional macroscopic trend, while the well logging interpretation bedding plane data provides high-precision discrete point depths. Using the seismic bedding plane as the trend surface and the well logging stratification depth as the hard constraint point, a stratigraphic top surface structure map is constructed layer by layer using kriging or discrete smooth interpolation methods. Finally, the bedding planes are stacked in the order of deposition from oldest to youngest. A stratigraphic framework is constructed. Then, the fault framework, stratigraphic bedding planes, and 3D mesh framework are integrated. Well logging data is used to constrain the seismic trend surface at well points, with well depth as the reference point and increasing weight as the distance from the well point increases. This generates a 3D geological structure model that maintains the regional tectonic trend while accurately matching the well point depth. The fracture distribution structure (ant-like volume) extracted from S1 is then resampled and coordinate-transformed before being embedded into the model mesh. Attributes such as fracture development intensity, fracture orientation, and fracture density are stored as fracture attribute volumes associated with the mesh nodes. Finally, the model undergoes quality checks, including mesh distortion verification, stratigraphic contact relationship verification on both sides of the fault, and comparison of slices along arbitrary directions with the original seismic profile and well logging data. After passing the quality check, a usable, detailed geological structure model for the work area is obtained.

[0059] S3. Based on the detailed geological structure model of the work area, a multi-dimensional static index system structure containing nine static indicators is constructed according to the four-level logic of macro-geological control, micro-characterization of leakage channels, stratigraphic mechanical property constraints, and engineering geological coupling threshold determination.

[0060] Furthermore, the four-level logic of S3—macro-geological control, micro-characterization of leakage channels, formation mechanical property constraints, and engineering geological coupling threshold determination—is as follows: Level 1: Macro-geological control, including sedimentary facies index X1 and lithology index X2, used to characterize the macro-control effect of regional sedimentary background and rock type on well leakage risk; Level 2: Micro-characterization of leakage channels, including porosity-cementation index X3, fracture index X4, and weak surface index X5, used to characterize the development degree, connectivity, and opening conditions of leakage channels; Level 3: Formation mechanical property constraints, including brittleness index X6 and pore pressure anomaly index X7, used to characterize the sensitivity of formation mechanical properties and pressure state to leakage occurrence; Level 4: Engineering geological coupling threshold determination, including leakage pressure X8 and safe mud density window X9, used to characterize the critical conditions for triggering well leakage under the coupling effect of drilling engineering parameters and geological conditions. Furthermore, the lithological index X2 is assigned a value based on the preset lithological risk assignment table, with mudstone and shale assigned a value of 0.0 to 0.1, and limestone and dolomite assigned a value of 0.8 to 1.0; the fracture index X4 is assigned a value by combining fracture development intensity, activation status, connectivity and distance from fault; the weak surface index X5 is superimposed with a coefficient of 1.1 to 1.2 times on unconformities or volcanic rock bodies, and a high threshold is directly assigned to the identified karst cave development areas.

[0061] Specifically, based on the detailed geological structure model of the work area, a multi-dimensional static index system containing nine static indicators is constructed according to a four-level logic: macro-geological control, micro-characterization of leakage channels, stratigraphic mechanical property constraints, and engineering geological coupling threshold determination. The first level of macro-geological control includes sedimentary facies index X1 and lithology index X2, used to characterize the macro-control effect of regional sedimentary background and rock type on well leakage risk. Among them, lithology index X2 is assigned values ​​according to a preset lithology risk assignment table, with mudstone and shale assigned values ​​of 0.0 to 0.1, limestone and dolomite assigned values ​​of 0.8 to 1.0, and other lithologies assigned values ​​in the middle range according to grain size and permeability. The second level of micro-characterization of leakage channels includes porosity-cementation index X3, fracture index X4, and weak surface index X5, used to characterize the development degree, connectivity, and opening conditions of leakage channels. The porosity-cementation index X3 is obtained by linearly weighting the normalized values ​​of three parameters: porosity φ, permeability K, and cementation degree C. The weights of these three parameters reflect an empirical ranking of their contributions to the development of leakage channels. The fracture index X4 integrates factors such as fracture development intensity, activation status, fracture connectivity, and distance from faults to assign a risk value. The weak surface index X5 considers weak surfaces including geological interfaces and fault planes, quantifying the risk value using distance. If an unconformity or volcanic rock mass is present, a coefficient of 1.1 to 1.2 is added to the basic risk value. If karst development is identified, a higher threshold is directly assigned. The basic risk value of the weak surface index X5 is jointly determined by the type of weak surface (unconformity, fault plane, volcanic rock mass boundary, etc.) and the spatial distance from the weak surface. For a given spatial grid point, if its distance to the nearest weak surface is d (in meters), then the basic risk value R is... base =exp(-d / D0), where D0 is the characteristic distance (which can be 50-100m). If the point is located within an unconformity surface or volcanic rock mass, the basic risk value is multiplied by a coefficient of 1.15 (i.e., the median of 1.1-1.2). For karst development areas identified through seismic attributes (such as amplitude anomalies, inverted low-impedance bodies) or drilling logging (such as sudden drops during drilling, venting), X5 is directly assigned a value of 0.95 (high threshold). If there is no weak surface influence, X5 is 0. The third-level formation mechanical property constraints include the brittleness index X6 and the pore pressure anomaly index X7. The higher the brittleness index, the higher the leakage risk. The pore pressure anomaly index characterizes the width of the safe density window by delineating abnormally low-pressure areas or high-pressure areas. The fourth-level engineering geological coupling threshold determination includes leakage pressure (X8) and a safe mud density window (X9). Leakage pressure defines the critical risk of formation leakage, while the safe mud density window quantifies the tolerance space for drilling operations. A low leakage pressure and a narrow window indicate a high risk of well leakage. All nine indicators are established in the form of a three-dimensional attribute model and unified into the spatial grid of the work area's fine geological structure model, forming a multi-dimensional static indicator system as the input feature set for subsequent machine learning modeling. The reasons for assigning lithological risk values ​​are shown in Table 1.

[0062] Table 1 Lithological Risk Assignment Table

[0063] mudstone, shale 0.0～0.1 The rock is dense and has extremely low permeability, with poorly developed leakage channels, resulting in a very low risk of well leakage. siltstone 0.2～0.3 The particles are fine, and the porosity and permeability are weak, so there is only a slight possibility of leakage. Salt rock, gypsum rock 0.1～0.3 The rock is dense and highly plastic, making it less prone to leakage, but it is prone to collapse, posing a slightly higher risk than mudstone. fine sandstone 0.4～0.5 The sandstone has good physical properties and well-developed pores, giving it a certain degree of permeability, with a moderate risk of leakage. medium sandstone 0.5～0.6 The grain size is coarser, the porosity and permeability conditions are better than fine sandstone, the leakage channels are more developed, and the risk is higher. coarse sandstone 0.6～0.7 The particles are large and have strong pore connectivity, which easily forms high permeability channels and poses a high risk of leakage. Conglomerate, sandstone 0.8～0.9 With large intergranular pores and extremely high permeability, it is a typical high-permeability, loss-prone formation. coal seam 0.7～0.9 The area has well-developed stratification and cleavage, which easily form natural seepage channels, resulting in a high risk of leakage. limestone, dolomite 0.8～1.0 The area is characterized by well-developed cracks and dissolution cavities, resulting in complex and highly interconnected leakage channels that make it extremely prone to severe leakage.

[0064] It should be noted that Table 1 is used to convert the lithological index X2 from a discrete categorical variable into a continuous numerical risk value, enabling it to participate in symbolic regression machine learning calculations. Based on the permeability, porosity, fracture and dissolution cavity development characteristics of different rock types, and combined with leakage experience statistics in the petroleum engineering field, the table provides a risk value range of 0 to 1 for each lithology. The denser the lithology and the lower the permeability, the closer the value is to 0; the looser the lithology or the more developed the fractures and cavities, the closer the value is to 1. In practical applications, for any spatial grid point within the work area, the lithological type of the point is first determined based on the lithological results interpreted from well logging (identified by a combination of curves such as natural gamma ray GR, resistivity, and neutron density). Then, referring to the risk value range in Table 1, and combining the specific pore structure characteristics of the point (such as the organic matter content in mudstone and the density of dissolution cavities in limestone), fine-tuning is performed within the given range. If the lithology of a grid point is a transitional type not directly included in the list (such as gravelly fine sandstone), then the assigned value is determined by linear interpolation based on its median grain size and permeability and the risk range of the closest lithology in the table. Through Table 1, the qualitative geological factor of lithology is quantified into a continuous and computable feature variable, providing a unified, standardized, and physically meaningful input for subsequent machine learning modeling.

[0065] It should also be noted that the sedimentary facies index X1 is continuously assigned numerical values ​​based on the contribution of different sedimentary facies types to the leakage risk, specifically using the sedimentary facies-risk correspondence shown in Table 2. In practical applications, it can be fine-tuned based on the statistical distribution of leakage samples from drilled wells within the work area, but the relative risk ranking among various sedimentary facies must remain unchanged.

[0066] Table 2. Assignment Table of Sedimentary Facies Index X1

[0067] Braided river channel phase 0.7～0.9 The sand body is thick, coarse-grained, and has good porosity and permeability, posing the highest risk of leakage. meandering river channel phase 0.6～0.8 The river channel has well-developed sand bodies, but with numerous muddy interlayers, posing a slightly lower risk than braided rivers. Delta front phase 0.5～0.7 Underwater distributary channels have well-developed sand bodies and good connectivity. Beach and dam phase 0.4～0.6 Sandy beach-bar reservoirs have moderate physical properties and moderate risk. Turbid phase 0.3～0.5 Turbidification sand bodies are typically isolated lenses with poor connectivity. Shallow marine shelf facies 0.1～0.3 The sedimentary deposits are mainly muddy, and the leakage channels are not well developed. Semi-deep-deep ocean phase 0.0～0.2 Primarily composed of mudstone and shale, with extremely low risk.

[0068] S4. Statistically analyze the spatial distribution and leakage amount of well leakage in the work area, use the Sigmoid function to perform risk normalization assignment, and generate the well leakage risk trend structure model CLRI1 through spatial interpolation; further, in S3, the Kriging spatial interpolation method is used to generate the well leakage risk trend structure model CLRI1.

[0069] Specifically, firstly, the spatial locations of lost circulation (WS) events and the corresponding drilling fluid losses within the work area are statistically analyzed. Based on the degree of loss, WS is categorized into several levels, including minor, moderate, and severe WS. For each WS level, a Sigmoid function is used for normalization. This function, f(x) = 1 / (1 + e^(-x)), reflects the characteristics of a slow increase in low-level WS and a rapid increase in medium-to-high-level WS, mapping WS risk to a continuous interval from 0 to 1. Then, using the normalized risk values ​​of each well point as samples, a three-dimensional risk attribute volume, i.e., the WS risk trend structure model CLRI1, is generated using the Kriging space interpolation method. Figure 3 As shown, Figure 3 This is a schematic diagram of the well leakage prediction trend model CLRI1. Kriging interpolation analyzes the spatial autocorrelation of sample points based on the variogram function, and performs optimal unbiased estimation of the risk value of unknown grid points, thereby obtaining a continuous risk distribution field covering the entire work area. CLRI1 only reflects the macroscopic spatial distribution trend of well leakage risk and does not have precise quantitative prediction capabilities; it serves as an auxiliary model in the subsequent fusion of two models.

[0070] S5. Using a multi-dimensional static indicator system architecture as input and measured well leakage risk value as the objective, a symbolic regression machine learning algorithm is used for training to derive analytical formulas containing static indicators, resulting in the well leakage prediction master model structure CLRI2; further, the analytical formulas containing static indicators in S5 are as follows:

[0071] CLRI2=0.132+0.591*X4-0.147*X3*X4+0.105*exp(X3)+0.774*pow(X6,0.5),

[0072] Among them, X3 is the porosity-cementation index; X4 is the crack index; and X6 is the brittleness index.

[0073] It should be noted that this step uses a symbolic regression algorithm to model nine static indicators (X1~X9) and measured well leakage risk values. During training, the symbolic regression algorithm automatically performs feature selection and nonlinear combination search. The final analytical formula output only includes the indicators that contribute most significantly to leakage risk prediction, while redundant indicators with weak predictive ability or multicollinearity are removed. In the training data (8 wells) of this embodiment, after multiple iterations and cross-validation of symbolic regression, X3 (porosity-cementation index), X4 (fracture index), and X6 (brittleness index) were determined to have the strongest explanatory power for leakage risk. The other six indicators (X1 sedimentary facies, X2 lithology, X5 weak surface, X7 pore pressure anomaly, X8 leakage pressure, X9 safe mud density window) were automatically removed because they failed to significantly improve the model fit on the training set or posed an overfit risk. This screening result reflects that the core controlling factors of leakage risk in this work area are reservoir properties, fracture development degree, and formation brittleness, which is consistent with geological understanding.

[0074] Specifically, nine indicators from the multi-dimensional static indicator system constructed by S4 are used as input features, and the measured well leakage risk value in the work area is used as the prediction target. The symbolic regression machine learning algorithm is used for training. In this embodiment, a total of 10 wells were drilled in the work area, of which 8 wells were used as training wells for model training, and the remaining 2 wells were used as test wells to verify the accuracy and reliability of the model. The core of the symbolic regression algorithm is to automatically search for the optimal analytical expression containing the input indicators and basic mathematical operations from the training data, without needing to pre-specify the function form. After comparing and optimizing various candidate formulas, the analytical formula with the best fit and no overfitting problem is finally derived: CLRI2=0.132+0.591×X4-0.147×X3×X4+0.105×exp(X3)+0.774×sqrt(X6), where X3 is the porosity-cementation index, X4 is the fracture index, and X6 is the brittleness index. The model's mean relative error (MAPE) is only 11%, with minimal deviation between model predictions and measured values. Furthermore, well testing has verified the absence of overfitting issues, and its generalization ability and prediction accuracy meet the practical application requirements of well logging interpretation in petroleum engineering. This analytical formula is the main model structure CLRI2 for well leakage prediction, used to quantitatively characterize the nonlinear coupling relationship between various static indicators and well leakage risk.

[0075] The above formulas are based on measured logging, well logging, and lost circulation records from multiple drilled wells in the study area, and the optimal parameters were obtained through fitting using the SR method. A nonlinear regression model for well leakage risk was constructed, using leakage labels as supervised samples. A stepwise regression method was used to automatically select the optimal nonlinear function form and fit the corresponding coefficients, ultimately obtaining the above-mentioned optimal parameters. Validation on an independent test set showed that the nonlinear model corresponding to this set of parameters significantly outperformed the linear model in predictive performance, and therefore adopted it. This method automatically examines the improvement effect of different function forms on the model's goodness of fit during the fitting process, ultimately selecting the nonlinear function form with the strongest explanatory power for well leakage risk and its corresponding optimal coefficients. This ensures that the model fully conforms to the nonlinear response laws of the geological and engineering aspects of the study area, avoids the fitting bias of linear models, and guarantees the objectivity and generalization of the model.

[0076] S6. Perform three-dimensional spatial connectivity analysis on the well leakage prediction master model structure CLRI2, increase the risk weight for areas with strong connectivity and decrease the weight for isolated small areas to obtain the optimized well leakage prediction master model structure CLRI2'.

[0077] Furthermore, the specific method for performing three-dimensional spatial connectivity analysis on the CLRI2 main model structure for well leakage prediction in S6 is as follows: The volume or spatial continuity range of the connected volume to which each grid point belongs in the CLRI2 three-dimensional attribute volume is calculated. Continuous distribution areas with a connected volume greater than a preset threshold are identified as areas with strong connectivity, while isolated distribution areas with a connected volume less than the preset threshold are identified as isolated small areas. Furthermore, the specific rules for increasing the risk weight for areas with strong connectivity and decreasing the weight for isolated small areas in S6 are as follows: For continuous areas with a connected volume greater than the first threshold, the original risk value is multiplied by a weighting factor greater than 1; for isolated areas with a connected volume less than the second threshold, the original risk value is multiplied by a weighting factor less than 1; the original risk value remains unchanged in intermediate areas to avoid homogenization of risk values ​​between small-scale isolated risk areas and large-scale continuous risk areas.

[0078] Specifically, a three-dimensional spatial connectivity analysis is performed on the CLRI2 master model structure for well leakage prediction obtained from S5 to distinguish between continuous risk areas and isolated risk areas, avoiding the homogenization of risk values ​​between small-scale isolated low-connectivity risk areas and large-scale continuous risk areas. Specifically, each grid point in the CLRI2 three-dimensional attribute volume is first treated as a node. Based on the connection relationship between adjacent grid points, a connected component labeling algorithm is used to identify each independent connected component, and the number of grid points or spatial volume contained in each connected component is counted. Continuous distribution areas with a connected volume greater than a first threshold are identified as highly connected areas, isolated distribution areas with a connected volume less than a second threshold are identified as small isolated areas, and areas in between are identified as intermediate areas. Then, differentiated weight adjustments are made for different types of areas: for highly connected areas, the original risk value is multiplied by a weighting coefficient greater than 1 to highlight the importance of large-scale continuous risk areas; for small isolated areas, the original risk value is multiplied by a weighting coefficient less than 1 to reduce the misleading impact of accidental isolated high risks; the original risk value remains unchanged for intermediate areas. After the above connectivity analysis and weight adjustment, the optimized well leakage prediction master model structure CLRI2′ is obtained. This model more accurately reflects the spatial continuity characteristics of risk and provides a more reliable master model input for subsequent dual-model fusion.

[0079] It should be noted that the first threshold (connected volume greater than the preset threshold) and the second threshold (connected volume less than the preset threshold) mentioned in S6 are not fixed absolute values, but are dynamically determined based on the overall scale of the three-dimensional geological model of the work area and the spatial statistical characteristics of the drilled leakage channels. A specific and repeatable threshold determination method is given below: First, the total volume of the three-dimensional mesh frame of the work area is statistically analyzed. Then, the spatial distribution of typical leakage channels that have occurred within the work area is analyzed: the actual spatial volume of regional large-scale leakage channels (such as large fault zones, unconformities, and thick high-permeability zones) usually accounts for more than 5% of the total volume of the work area; while the volume of local isolated leakage channels (such as single karst caves and small-scale micro-fracture clusters) is generally less than 0.5% of the total volume of the work area. Based on the above statistical patterns, in this embodiment, the first threshold is set to 5% of the total volume of the work area, that is, connected bodies with a continuous spatial range accounting for 5% or more of the total volume of the work area are judged as "strongly connected areas"; the second threshold is set to 0.5% of the total volume of the work area, that is, isolated connected bodies with a spatial volume less than 0.5% of the total volume of the work area are judged as "isolated small areas". Areas in between are considered intermediate areas.

[0080] S7. The well leakage risk trend structure model CLRI1 and the optimized well leakage prediction master model structure CLRI2' are fused through a weighted fusion mechanism to generate the comprehensive leakage risk index structure model CLRI.

[0081] Furthermore, the specific formula for S7 weighted fusion is as follows:

[0082] CLRI=α×CLRI1+(1-α)×CLRI2',

[0083] Wherein, CLRI1 is the well leakage risk trend structure model; CLRI2' is the optimized well leakage prediction master model structure; and α is the preset weight coefficient.

[0084] Specifically, the well leakage risk trend structure model CLRI1 generated in S3 and the well leakage prediction master model structure CLRI2′ optimized in S6 are fused through a weighted fusion mechanism to generate a comprehensive leakage risk index model CLRI that balances the rationality of geological experience with the accuracy of data-driven analysis. Since CLRI1 is only obtained based on spatial interpolation of well leakage points, it reflects the macroscopic and fuzzy trend of well leakage risk and lacks precise quantitative prediction capabilities. In contrast, CLRI2′ is based on machine learning to extract nonlinear coupling patterns from multi-dimensional static indicators, resulting in higher quantitative prediction accuracy. Therefore, in this embodiment, CLRI1 is given a lower weight, and CLRI2′ is given a higher weight. The specific fusion formula is: CLRI = α × CLRI1 + (1-α) × CLRI2′, where α is a preset weight coefficient with a value of 0.2. Before performing the weighted fusion, CLRI1 and CLRI2′ are unified to the same three-dimensional grid system and resolution through spatial resampling to ensure that the two models are weighted point-by-point within the same spatial coordinate framework. The resulting integrated leakage risk index model, CLRI, retains the constraints of geological experience on macro-level risk trends while fully leveraging machine learning's ability to finely characterize complex nonlinear relationships, providing accurate and reliable support for subsequent 3D visualization and risk early warning. For example... Figure 9 As shown, Figure 9 This is a schematic diagram of the CLRI dual-model weighted fusion well leakage prediction model, which is "CLRI1+CLRI2".

[0085] Finally, the prediction results are transformed into actionable engineering decision-making information. The Comprehensive Leakage Risk Index (CLRI) is visualized in three dimensions using a color gradient mapping method, ranging from blue (low risk) to red (high risk). Multi-view (I, J, K) slice analysis is possible to accurately locate high-risk areas. Based on the model's calculations, attribute models Xmax (Xmax=MAX(X3, X4, X6)), Xsub-max (second largest value of the well leakage index), and Xsub2-max (third largest value of the well leakage index) are generated. The number of dominant factors for well leakage is determined according to the table below. For single-indicator dominant areas, each indicator is assigned a characteristic color according to Xmax=Xn (indicator), achieving a color-coded representation of the dominant indicators. For combined dual-indicator and multi-indicator dominant areas, representative colors are directly assigned. An additional backtracking step is added, projecting the single-dimensional indicator onto the comprehensive risk map, changing the transparency, and then performing overlap analysis to determine which factors are dominant.

[0086] It should be summarized that although the Comprehensive Loss Risk Index (CLRI) model can quantitatively characterize the spatial distribution of well leakage risk, in actual drilling engineering, the main controlling factors of leakage in different areas are often different, and the corresponding prevention and control measures are also quite different. For example, fracture-dominated leakage requires optimization of the plugging slurry formula and fracture sealing process, while brittleness index-dominated leakage requires adjustment of the mud density window to avoid inducing fracture propagation, and porosity-cementation-dominated leakage relies more on drilling-while-drilling sealing of permeable formations. To this end, while outputting the CLRI risk model, this invention further constructs well leakage dominant factor identification rules as shown in Table 3 based on the local differentiation characteristics of nine static indicators (X1~X9). Specifically, for each grid point in the CLRI three-dimensional attribute volume, the risk assignment of the three core indicators X3 (porosity-cementation), X4 (fracture), and X5 (brittleness index) at that point is extracted, and their maximum, second largest, and third largest values ​​are calculated. Based on the differences between the maximum and second-largest values, and between the second-largest and third-largest values, the dominant leakage factors are categorized into three types: single-indicator dominant, dual-indicator combined dominant, and multi-indicator combined dominant, and assigned different visual identifiers. For the remaining six indicators (X1, X2, X5, X7, X8, X9) in Table 3 that are not directly involved in the analytical formula, although they do not appear in the dominant factor determination table, they can serve as auxiliary analytical factors in the case of "multi-indicator combined dominant," and be comprehensively identified in conjunction with the geological background. Through the determination rules in Table 3, engineers can quickly identify the dominant control factors of leakage in the target area, thereby optimizing drilling design in a targeted manner: in areas dominated by a single indicator, control plans can be developed focusing on that indicator; in areas dominated by dual or multi-indicator combined dominant, the synergistic effect of multiple factors needs to be comprehensively considered, and a composite leakage plugging strategy should be adopted. Thus, this invention extends from "risk distribution prediction" to "dominant factor identification," forming a complete well leakage prediction and diagnosis technology chain, providing a more refined and operable basis for on-site engineering decisions.

[0087] Table 3. Determination of Dominant Factors of Well Leakage

[0088] Xmax−Xsub−max>0.1 Single indicator dominance Xmax−Xsub−max≤0.1 and Xsub−max−Xsub2−max>0.1 Dual indicators jointly dominate Xmax−Xsub−max≤0.1 and Xsub−max−Xsub2−max≤0.1 Multiple indicators jointly dominate

[0089] The embodiments described above are merely illustrative of specific implementations of the present invention, and while the descriptions are detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A multi-dimensional index coupling-driven well leakage prediction method based on machine learning, characterized in that, include: S1. Preprocess multi-source heterogeneous data from seismic, well logging, and well logging sources to obtain standardized multi-source data volumes and fracture distribution structures; S2. Based on the standardized multi-source data volume and fracture distribution structure, a fine geological structure model of the work area is built using a well-seismic combined method; S3. Based on the detailed geological structure model of the work area, a multi-dimensional static index system structure containing nine static indicators is constructed according to the four-level logic of macro-geological control, micro-characterization of leakage channels, stratigraphic mechanical property constraints, and engineering geological coupling threshold determination. S4. Statistically analyze the spatial distribution and leakage of well leakage in the work area, use the Sigmoid function to perform risk normalization assignment, and generate the well leakage risk trend structure model CLRI1 through spatial interpolation. S5. Using the aforementioned multi-dimensional static index system architecture as input and the measured well leakage risk value as the objective, a symbolic regression machine learning algorithm is used for training to derive an analytical formula containing static indicators, thus obtaining the well leakage prediction master model structure CLRI2. S6. Perform three-dimensional spatial connectivity analysis on the well leakage prediction master model structure CLRI2, increase the risk weight for areas with strong connectivity and decrease the weight for isolated small areas to obtain the optimized well leakage prediction master model structure CLRI2'. S7. The well leakage risk trend structure model CLRI1 and the optimized well leakage prediction master model structure CLRI2' are fused through a weighted fusion mechanism to generate a comprehensive leakage risk index structure model CLRI.

2. The method according to claim 1, characterized in that, The specific steps of S1 are as follows: The logging curves are trimmed and denoised. Based on the geological survey results and GR curves, lithology is classified. A standardized multi-source data volume is obtained through logging curve calculation. The standardized multi-source data volume includes information on porosity, permeability, saturation and rock cementation. The seismic body is subjected to structural smoothing and amplitude contrast processing to reduce the signal-to-noise ratio. Variance and chaotic bodies are generated for boundary detection. A passive-then-active tracking method is adopted, supplemented by low-dip attitude control to extract the crack distribution structure.

3. The method according to claim 1, characterized in that, S2 specifically includes: Based on the standardized multi-source data volume obtained from S1, seismic interpretation fault data is imported to clarify the fault structure pattern, and a three-dimensional mesh framework is generated according to the work area boundary. Subsequently, seismic interpretation layer data and well logging interpretation layer data are imported. Seismic layers are used to provide regional macro trends, and well logging layer data is used for high-precision constraints. Through the well-seismic joint collaborative modeling method, a fine geological model of the work area is constructed.

4. The method according to claim 1, characterized in that, In S4, the well leakage risk trend structure model CLRI1 is generated using the Kriging space interpolation method.

5. The method according to claim 1, characterized in that, The four-level logic described in S3—macro-geological control, micro-characterization of leakage channels, stratigraphic mechanical property constraints, and engineering geological coupling threshold determination—is as follows: Level 1: Macro-geological control, including sedimentary facies index X1 and lithological index X2, used to characterize the macro-control effect of regional sedimentary background and rock type on well leakage risk; Level 2: Microscopic characterization of leakage channels, including porosity-cementation index X3, crack index X4 and weak surface index X5, used to characterize the development degree, connectivity and opening conditions of leakage channels. Level 3: Formation mechanical property constraints, including the brittleness index X6 and the pore pressure anomaly index X7, are used to characterize the formation mechanical properties and the sensitivity of pressure state to leakage. Level 4: Engineering-geological coupling threshold determination, including leakage pressure X8 and safe mud density window X9, is used to characterize the critical conditions that trigger well leakage under the coupling effect of drilling engineering parameters and geological conditions.

6. The method according to claim 5, characterized in that, The lithological index X2 is assigned a value based on a preset lithological risk assignment table, where mudstone and shale are assigned a value of 0.0 to 0.1, and limestone and dolomite are assigned a value of 0.8 to 1.0; the fracture index X4 is assigned a value by combining fracture development intensity, activation status, connectivity, and distance from fault; the weak surface index X5 is superimposed with a coefficient of 1.1 to 1.2 on unconformities or volcanic rock bodies, and a high threshold is directly assigned to the identified karst cave development areas.

7. The method according to claim 1, characterized in that, The analytical formula containing static indicators described in S5 is as follows: CLRI2=0.132+0.591*X4-0.147*X3*X4+0.105*exp(X3)+0.774*pow(X6,0.5), Among them, X3 is the porosity-cementation index; X4 is the crack index; and X6 is the brittleness index.

8. The method according to claim 1, characterized in that, The specific method for performing three-dimensional spatial connectivity analysis on the well leakage prediction master model structure CLRI2 described in S6 is as follows: calculate the volume or spatial continuity range of the connected body to which each grid point belongs in the CLRI2 three-dimensional attribute volume, determine the continuous distribution area with a connected volume greater than a preset threshold as a region with strong connectivity, and determine the isolated distribution area with a connected volume less than a preset threshold as an isolated small region.

9. The method according to claim 1, characterized in that, The specific rules for increasing the risk weight of highly connected regions and decreasing the weight of isolated small regions, as described in S6, are as follows: for continuous regions with a connected volume greater than the first threshold, the original risk value is multiplied by a weighting coefficient greater than 1; for isolated regions with a connected volume less than the second threshold, the original risk value is multiplied by a weighting coefficient less than 1; the original risk value of intermediate regions remains unchanged, so as to avoid the homogenization of risk values ​​between small isolated risk areas and large continuous risk areas.

10. The method according to claim 1, characterized in that, The specific formula for weighted fusion described in S7 is as follows: CLRI=α×CLRI1+(1-α)×CLRI2', Wherein, CLRI1 is the well leakage risk trend structure model; CLRI2' is the optimized well leakage prediction master model structure; and α is the preset weight coefficient.

Images