Drilling stratum profile prediction method based on multiple algorithm interpolation of well logging curve

Abstract

The present application relates to a method for predicting pre-drilling stratigraphic profile based on multi-algorithm interpolation of well logging curves, comprising: predicting the top / bottom and thickness of each layer of the target well based on stratified data; aligning the normalized depth within the layer; predicting multiple well logging curves by using multi-algorithm (at least IDW and Kriging) interpolation; correcting the proximity and mapping the depth based on the well deviation trajectory; outputting the multi-depth reference predicted curve and the stratigraphic position QC result. Further comprising lithology profile prediction: obtaining the classification probability by using stratified supervised classification, and fusing the spatial prior probability obtained based on the lithology of adjacent wells and distance weight to output the lithology interval of the target well and the profile comparison chart. This method comprehensively utilizes the well logging curves of adjacent drilled wells, stratigraphic position stratified data, well location and well deviation trajectory information, and under the unified depth reference, performs multi-algorithm horizontal interpolation prediction on multiple well logging curves of the target well, and further combines supervised learning and spatial prior to realize lithology profile prediction, and outputs the stratigraphic profile results which can be used for pre-drilling decision and well location deployment.

Description

Technical Field

[0001] This technical solution relates to the field of digital technology in oil and gas exploration, and in particular to a pre-drilling formation profile prediction method for target wells (planned wells, design wells, and undrilled wells). This method comprehensively utilizes logging curves from adjacent drilled wells, stratigraphic data, and well location and deviation trajectory information. Under a unified depth benchmark, it performs multi-algorithm lateral interpolation prediction on multiple logging curves of the target well, and further combines supervised learning and spatial priors to achieve lithological profile prediction, outputting formation profile results that can be used for pre-drilling decision-making and well location deployment. Background Technology

[0002] Well logging curves (such as natural gamma ray (GR), resistivity (RT), sonic transit time (AC / DT), compensation seed (CNL), and density (DEN)) are crucial continuous data foundations for characterizing formation lithology, pore structure, hydrocarbon potential, and fluid properties in oil and gas exploration and development. When deploying new wells or adjusting well locations, engineers and geologists often hope to obtain the target well's "expected logging response" and "lithology / stratum profile" before drilling to support drilling risk assessment, reservoir correlation, sweet spot evaluation, and well trajectory optimization.

[0003] In existing technologies, pre-drilling logging curve prediction often employs the following approaches: simple interpolation between adjacent wells using the wellhead horizontal distance as a weight, or interpolation after a rough depth mapping of the formation. These methods typically have the following shortcomings:

[0004] (1) The lack of detailed stratigraphic division and thickness constraints makes it impossible to effectively align the same layer between different wells, and the predicted curves are prone to morphological distortion near the layer boundary;

[0005] (2) Calculating weights based solely on wellhead distances fails to reflect the “true geological proximity” of adjacent wells in scenarios involving deviated wells, horizontal wells, or well trajectories with significant differences, leading to deviations in adjacent well selection and weight allocation.

[0006] (3) A single interpolation algorithm (such as inverse distance weighted interpolation IDW) is often used, which makes it difficult to take into account different well network densities, geological heterogeneity and data noise conditions.

[0007] (4) The lack of quantitative quality control and parameter optimization mechanisms makes it difficult to assess the reliability of prediction results;

[0008] (5) For lithological profile prediction, if we rely solely on full-section monitoring and classification or rely solely on the proportion of adjacent wells for inference, it is easy to encounter problems such as excessive smoothing of the same lithology or layer in the entire section, which cannot meet the requirements for fine stratigraphic profile prediction.

[0009] Therefore, there is a need for a method for predicting stratigraphic profiles that can unify depth benchmarks, introduce stratigraphic stratification constraints, integrate multiple interpolation algorithms, and provide evaluable and traceable outputs in pre-drilling scenarios. Summary of the Invention

[0010] This technical solution aims to provide a pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves, addressing issues such as "difficulty in aligning layers, inaccurate thickness, distortion of well proximity, poor algorithm adaptability, lack of quality evaluation, and unstable lithological profiles" in pre-drilling prediction. Given the project well location and adjacent well data, this method can automatically complete adjacent well selection, layer thickness prediction, layer alignment, curve prediction, depth correction, lithological profile prediction, and result export and visualization, providing high-precision, interpretable, and evaluable prediction results for pre-drilling oil and gas exploration.

[0011] The technical solution adopted by this invention to achieve the above objectives is: a pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves, comprising the following steps:

[0012] Data access and depth benchmark unification: Read logging curves, well location coordinates, stratification data and lithological annotation data from adjacent wells; unify the depth axes of each well to the selected working axis, and establish a consistent mapping relationship between different depth axes;

[0013] Neighbor selection and proximity correction: Constructing a set of neighboring wells When an inclined shaft trajectory exists, calculate the trajectory correction distance. and distance from the wellhead Form an optional distance metric;

[0014] Layered framework construction and layer prediction: Based on layered data and distance metrics, the top and bottom depths of each well within each layer are determined; the layer information of the target well is predicted by distance weighting, and boundary correction is performed by introducing continuity constraints of adjacent layers;

[0015] Intra-layer alignment and resampling: Normalized intra-layer coordinates are constructed within each layer segment, and a system is established... Mapping; in a uniform sampling grid Above, the curves of adjacent wells are changed from the depth domain. Resampling ;

[0016] Multi-algorithm interpolation prediction and adaptive fusion: in each normalized coordinate At this point, at least two spatial interpolation algorithms are executed, and the fusion weights are adaptively determined. The final prediction curve is obtained. ;

[0017] Depth mapping correction: for each layer segment The prediction curves for the entire well section of the target well are formed by splicing them together in depth order. When the target well is an inclined well, the prediction results are mapped between TVD and MD.

[0018] Lithology prediction: Using the prediction curve as a feature, combined with lithology annotations from adjacent wells and a layered framework, the classifier probability is obtained through at least one strategy. Simultaneously, spatial prior probabilities are constructed based on the lithology and distance weights of adjacent wells. And obtained through posterior fusion Output lithology identification Including confidence and uncertainty indicators, to achieve lithological profile prediction.

[0019] The adjacent well screening and proximity correction include the following steps:

[0020] Candidate set determination and neighbor set generation: based on the wellhead coordinates of the target well O. Candidate well set { w Wellhead coordinates And the set neighbor well selection constraint parameters: nearest neighbor parameters k radius constraint R Constructing a set of adjacent wells ;

[0021] Distance calculation between the target well and adjacent wells: for each candidate well w Calculate the Euclidean distance between wells ;

[0022] When the candidate well is a vertical well, the Euclidean distance between wells is calculated based on the wellhead coordinates. ;

[0023] When the candidate well is an inclined well, the trajectory correction and anisotropic distance calculation are performed as follows:

[0024] (1) When selecting distance correction based on trajectory: select representative vertical depth Select the medium-deep target layer In candidate wells w adjacent sampling points of the trajectory and Interlinear interpolation , to be used as wellhead coordinates calculate ;in,

[0025] ,

[0026] .

[0027] Indicates the first j The corresponding medium and deep points , Indicates the first j The coordinates of the points;

[0028] (2) When selecting anisotropic distance: set the principal axis azimuth angle Compared with anisotropy , , a , b These represent the distances along the primary and secondary axes, respectively; for wellhead coordinates... The displacement vector is rotated and scaled to obtain the anisotropic distance. As ;

[0029] (3) When both trajectory correction and anisotropy are selected: then... Wellhead coordinates calculate And as the corrected wellhead distance ;

[0030] Output updated distance and parameter groups ( k , R , ratio z ).

[0031] The hierarchical skeleton construction and segment prediction include the following steps:

[0032] S3-1, Layered Data Reading and Depth Benchmark Unification

[0033] Read the layer table: containing well name, layer number (layer_id), top depth, and bottom depth;

[0034] Determine the working depth benchmark When only MD is provided and MD-TVD trajectories exist, convert to TVD first and then perform layer prediction and alignment; when only a single depth reference is provided, use that reference as the output and then remap.

[0035] For each adjacent well w Extract the top depth (top) for each layer number (layer_id). w Bottom depth w Validate top w <bottom w ;

[0036] S3-2, Target Well Stratification and Thickness Prediction

[0037] For each layer number, the distance is determined based on the distance obtained from the adjacent well screening and proximity correction steps. As And calculate IDW weights ;

[0038] Weighted predictions are performed on the top and bottom layers respectively to obtain the target well layer top. O with bottom O And calculate the thickness ;in,

[0039] ; ; , Indicates the first i The top and bottom depths of the layer; the thickness can also be obtained by directly interpolating the thickness.

[0040] and order ;

[0041] S3-3, Boundary Continuity Correction and Layer QC Output

[0042] (3.1) Apply continuity constraints to adjacent layer segments that share layer boundaries, so that... :

[0043] For the two adjacent layers after sorting and public boundary And correct the target well number i Bottom depth Target well number i Top floor depth Simultaneously satisfy ;

[0044] (3.2) Handling of abnormal situations: When the thickness of a certain layer is less than the thickness threshold ,according to Perform the trimming, or return to step S3-2 and record the QC flag.

[0045] (3.3) Output layer QC table: top O bottom O , Parameter group ( k , R , (ratio).

[0046] The intra-layer alignment and resampling includes the following steps:

[0047] S4-1. In-layer alignment and sampling mesh construction

[0048] (4.1) For each layer number, in the candidate well w Define normalized intra-layer coordinates within the layer segment ,in For vertical depth, top w bottom w These are the top depth and the bottom depth, respectively. ;

[0049] (4.2) For target well O, according to The predicted depth grid for the reconstructed segment; Indicates the grid depth;

[0050] S4-2, Resampling and Preprocessing within Adjacent Well Curve Layers

[0051] For each adjacent well w and each logging curve f Extracting the top segment w bottom w Internal data;

[0052] Interpolation in one dimension Resampling :

[0053]

[0054] Indicates the resampling curve. This represents a one-dimensional interpolation operator. Represents the depth domain of adjacent well curves. Indicates the normalized coordinates within the layer The mapping relationship with the actual depth z of the adjacent well. Indicates the first j Normalized relative depth coordinates corresponding to sampling points within each layer .

[0055] The multi-algorithm interpolation prediction and adaptive fusion includes the following steps:

[0056] S5-1, Multi-algorithm spatial interpolation prediction and fusion;

[0057] For resampling curves The method can perform single-algorithm interpolation prediction or use multiple interpolation algorithms for fusion prediction; the algorithm is one of IDW interpolation, ordinary kriging, or RBF interpolation.

[0058] The method of using multi-algorithm interpolation for fusion prediction is as follows:

[0059] Based on the performance of each algorithm under LOOCV or Bootstrap s Interpolation algorithm error Calculate the first s Fusion weights of interpolation algorithms And obtain the final prediction. ;in,n This represents the algorithm index in normalized summation. Indicates the first s Algorithm error conduct The result of the power transformation is used to adjust the sensitivity of the error in the weight calculation. This represents a stable term to prevent the denominator from being zero. s Indicates the algorithm index. ,in The number of algorithms participating in the fusion. Indicates the first s Interpolation algorithms sample points within the layer The predicted value at that location;

[0060] S5-2, Layer splicing and output

[0061] Predicting each floor number By splicing the data by depth, a complete prediction curve sequence for the target well O is obtained.

[0062] The depth mapping correction includes the following steps:

[0063] When the target well O is an inclined well and provides discrete points for the MD-TVD trajectory, a piecewise linear mapping is used to convert the TVD grid results into MD grid results; TVD and MD represent vertical depth and measurement depth, respectively.

[0064] TVD→MD mapping, given discrete points in the trajectory table For any TVD=z located at : ; z Indicates vertical depth. j Indicates the discrete point index;

[0065] The MD→TVD inverse mapping, for any MD=m located at... :

[0066] , m Indicates the depth being measured.

[0067] The lithological prediction includes the following steps:

[0068] S6-1 Training Data Construction and Feature Engineering

[0069] Read the lithology interval table and logging curves of adjacent wells; map the lithology intervals to the working depth axis. Discretize it into sample points;

[0070] Basic features representing lithology are extracted from logging curves of adjacent wells to form a feature vector. x ( z );

[0071] By analyzing the feature vector x ( z Normalization and grouping verification by well name are used to construct a training sample set;

[0072] S6-2, Lithology Prediction Strategy and Classifier Training

[0073] Training is performed using a single training method based on the training sample set; the training method is one of the following: whole-segment supervision method, hierarchical supervision method, or adjacent well interpolation method; wherein...

[0074] Full-segment supervised method: Train a global classifier, and output the classifier probability point by point for the entire segment of the target well O. , z Indicates vertical depth. c Indicates lithological category index, , C This represents the number of lithological categories;

[0075] Hierarchical supervision method: A classifier is trained for each layer number, and predictions are made only within the layer segment corresponding to the target well O;

[0076] Neighbor interpolation method: Without training a classifier, spatial prior probabilities are constructed solely based on the lithology and distance weights of neighboring wells. Directly determine lithology;

[0077] Alternatively, training can be performed using a hybrid fusion method based on the training sample set: the data obtained through the aforementioned whole-segment supervision method can be used for training. , and obtained through adjacent well interpolation method Obtained through posterior fusion Subsequent judgment.

[0078] In S6-2, the construction of the spatial prior probability in the lithology prediction strategy and classifier training includes the following steps:

[0079] Given a layer number and a depth point z, determine the set of adjacent wells that are valid at that depth. W ( z For each adjacent well w Search for it at depth z Lithology ;

[0080] The distance obtained based on the neighbor well screening and proximity correction steps Calculate weights Power is p Set parameters Preventing division by zero: intermediate variables , ,in For adjacent wells The weight of the target well O, u For adjacent well index, For adjacent wells;

[0081] We perform weighted statistics and normalization on each lithological category to obtain the spatial prior probability. The details are as follows:

[0082] The weight of the adjacent well w is Its lithological category is ;but: , c Indicates lithological category index;

[0083] The hybrid fusion method calculates the classifier probability for each depth point z. With spatial prior probability And merge to obtain the posterior probability. ;in:

[0084] Classifier probability: the probability of the feature vector x ( z Input a trained classifier to obtain the probability output for each lithology category. Where Pr represents the probability operator, Y Represents a random variable indicating lithological category. c Indicates lithological category index, Indicates the parameters of the classifier model;

[0085] Spatial prior probability: obtained by weighted statistics on the lithology of adjacent wells based on the construction of spatial prior probability. ;

[0086] Posterior fusion: settings α ≥0、 β ≥0 and smoothing terms First, calculate the unnormalized posterior probability:

[0087] Expressed as a linear fusion of the logarithmic field:

[0088] ;

[0089] Normalization: ,make ;

[0090] This represents the lithology category index in the normalized summation. Represents an exponential function;

[0091] Lithology identification and confidence level: final lithology classification And obtain the confidence level. ;in, This represents the lithology category index that maximizes the objective function. c ;

[0092] Uncertainty assessment: using posterior entropy Measure classification uncertainty.

[0093] The pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves also includes visualization, which is achieved through the results output module, as detailed below:

[0094] Intervalization: For point-by-point category sequences Adjacent and similar elements are merged to form initial lithological intervals. and categories c ; , These represent the starting and ending depths of the lithological interval, respectively.

[0095] Minimum thickness constraint: for thickness Less than the threshold The thin layers are merged based on the posterior probability, confidence level, and uncertainty of adjacent intervals, i.e., merged into... higher or H Lower neighbor intervals;

[0096] Stratigraphic constraints: All lithological intervals are limited to the range of the corresponding stratigraphic layer number. Inside, the boundary is cut off and updated at the layer boundary;

[0097] Export Results: Output the lithological interval table of the target well O, including layer number, top depth, bottom depth, lithology, confidence level, and uncertainty, and output a point-by-point probability table containing posterior probability, confidence level, and uncertainty for quality control;

[0098] Chart drawing: Draw lithological profiles of sample wells and target wells side by side, with superimposed layer boundaries and connections across wells; for well O, there is uncertainty in lithological labeling.

[0099] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the pre-drilling formation profile prediction method based on multi-algorithm interpolation of logging curves.

[0100] The present invention has the following beneficial effects and advantages:

[0101] (1) Introduce “top / bottom of layer - thickness” constraint: Before curve prediction, quantitative prediction of the top / bottom and thickness of each layer in well O is performed, and the layer boundary is corrected by continuity constraint, which significantly reduces the risk of mismatch and distortion near the layer boundary.

[0102] (2) Proximity correction of inclined wells: In the screening and weight calculation of adjacent wells, it is supported to obtain the XY coordinates along the inclined trajectory at the representative TVD (e.g., the middle and deep layers), so as to replace the "wellhead distance" with "geological proximity" and improve the prediction accuracy of the inclined well scenario.

[0103] (3) Multi-algorithm interpolation and adaptive fusion: Multiple interpolation algorithms such as IDW and Kriging can be used in parallel within the same layer segment, and adaptive fusion can be performed based on cross-validation index or confidence estimation to improve adaptability under different well network densities and heterogeneous conditions.

[0104] (4) Evaluable parameter optimization mechanism: Provides error evaluation and grid optimization based on leave-one-out method, which can evaluate IDW power and neighbor well number. k Rapid filtering based on parameters such as radius and anisotropy improves the traceability of results.

[0105] (5) The framework of “layered supervision + spatial prior + curve uncertainty” for lithological prediction: avoids distortion problems such as single lithology throughout the whole section, and takes into account the lateral geological trend while ensuring the continuity of the layer, thereby improving the stability and interpretability of stratigraphic profile prediction. Attached Figure Description

[0106] Figure 1 Overall flowchart of the method;

[0107] Figure 2 Schematic diagram of inclined shaft proximity correction;

[0108] Figure 3 Schematic diagram of IDW prediction and continuity constraints for the top / bottom and thickness of the segment;

[0109] Figure 4 Schematic diagram of intra-layer normalized depth alignment and multi-algorithm interpolation;

[0110] Figure 5 Schematic diagram of well logging curve prediction;

[0111] Figure 6 Schematic diagram of lithological profile prediction. Detailed Implementation

[0112] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.

[0113] This invention, based on the principles of "layered skeleton constraint + intra-layer normalization alignment + multi-algorithm interpolation fusion + depth benchmark mapping + probabilistic fusion lithology prediction + quality control output," achieves pre-drilling formation profile and logging curve, as well as lithology distribution prediction for target well O. Figure 1 The overall process includes the following steps:

[0114] (A) Data access and depth benchmark unification: Read the logging curves of adjacent wells, well location coordinates, layer / segment data, well inclination trajectory and wellhead benchmark (KB) information; unify the depth axis of each well to the selected working axis (MD / TVD / TVDSS), and establish a consistent mapping relationship between different depth axes to provide a unified benchmark for subsequent interpolation, splicing and output.

[0115] (B) Adjacent well screening and proximity correction: based on k Nearest neighbor / radius rule constructs neighbor well set When an inclined shaft trajectory exists, the method of following the trajectory is adopted. Calculate the trajectory correction distance by taking XY values. and distance from the wellhead An optional distance metric is formed for the unified calculation of interpolation weights and spatial prior weights. Figure 2 ).

[0116] (C) Layered framework construction and top, bottom, and thickness prediction: Based on the layering table (or activity function to identify layer boundaries), determine the top / bottom depth of each well within each layer_id; for the target well O... , and Distance-weighted prediction is performed, and continuity constraints between adjacent layers are introduced (e.g., Boundary correction) to ensure consistency of stratigraphic boundaries and geological rationality ( Figure 3 ).

[0117] (D) Intra-layer alignment and resampling: Constructing normalized intra-layer coordinates within each layer segment. and establish Mapping; in a uniform sampling grid Above, the curves of adjacent wells are changed from the depth domain. Resampling It also fills in missing parts and records the missing rate and resampling metadata.

[0118] (E) Multi-algorithm interpolation prediction and adaptive fusion: in each At least two spatial interpolation algorithms (such as IDW, ordinary kriging OK, RBF, etc.) are executed in parallel to obtain the desired result. Based on leave-one-out method and Bootstrap error Adaptive determination of fusion weights Output the final prediction And can provide uncertainty characterization indicators ( Figure 4 ).

[0119] (F) Depth mapping correction and curve stitching output: Outputting the depth of each segment... The prediction curves for the entire well section of well O are spliced ​​together in depth order. When the target well is an inclined well, the results are mapped between TVD and MD, and derived depths such as TVDSS are calculated. The multi-depth benchmark curve file and the layer QC table that can be used before drilling are output.

[0120] (G): Using the predicted curve as a feature, combined with adjacent well lithology annotations and a layered framework, the classifier probability is obtained under strategies such as "whole-section supervision, layered supervision, adjacent well interpolation, and hybrid fusion". Simultaneously, spatial prior probabilities are constructed based on the lithology and distance weights of adjacent wells. And obtained through posterior fusion Output lithology identification Including uncertainty indicators such as confidence level and entropy, to achieve interpretable lithological profile prediction results.

[0121] The specific steps of this invention are as follows:

[0122] (a) Data input and preprocessing

[0123] The input data should include at least:

[0124] (1) Adjacent well logging curve data: Each drilled well must contain at least a depth column (measured depth MD or depth / vertical depth TVD / vertical depth below sea level TVDSS) and several curve columns (natural gamma GR, resistivity RT, sonic transit time AC / DT, compensation seed CNL, density DEN, etc.).

[0125] (2) Well location coordinates: including the plane coordinates (X, Y) of the target well O and the adjacent wells, used for distance calculation and weight allocation.

[0126] (3) Layer / Layer data: includes well name, layer number, top depth, bottom depth (top / bottom can correspond to one of MD / TVD / TVDSS).

[0127] (4) Well inclination trajectory data (optional but recommended): includes well name, vertical depth tvd, x, y (or measured depth md, well inclination angle inc, azimuth angle azm can be converted), used for well proximity correction and O well TVD-MD mapping.

[0128] (5) Lithological labeling data (used for lithological profile prediction): includes well name, top depth, bottom depth, lithology (and optional Chinese name column).

[0129] Preprocessing key points:

[0130] a) Standardize the format of well names (remove BOM, remove spaces, and adhere to capitalization rules);

[0131] b) Convert depth columns to numerical values, sort, remove duplicates, and eliminate missing segments;

[0132] c) Curve column name alignment: When there are different names for the same physical quantity, use synonym mapping to select the optimal column (e.g., RT can be degenerated into the column that is "most like resistivity").

[0133] (II) Selection of adjacent wells and calculation of distance

[0134] Step S1: Determine the candidate set of neighboring wells and generate the set of neighboring wells.

[0135] (1) Input the coordinates of the target well O wellhead Candidate well set { w Wellhead coordinates And optional well deviation trajectory data (for well proximity correction).

[0136] (2) Set the neighbor selection constraint parameters: k nearest neighbor parameter k radius constraint R If set at the same time k and R Then it is preferable to press first. R Filter, then sort by distance and select the top ones. k A well.

[0137] (3) Data integrity filtering: remove wells that lack necessary curve / layer / coordinate information to prevent incomplete data from entering the interpolation and training process.

[0138] (4) Output adjacent well set It and its distance-sorted list serve as a unified input for subsequent weight calculations, top / bottom prediction of layers, curve interpolation, and spatial prior construction.

[0139] Step S2: Calculate the basic distance.

[0140] (1) For each candidate well w Calculate the Euclidean distance between wells based on the wellhead coordinates. (See Formula 1).

[0141] (2) If trajectory correction and anisotropy are not enabled, then... This serves as the distance between the subsequent spatial weights and the neighboring well ranking.

[0142] Step S3: Trajectory correction and anisotropic distance calculation.

[0143] (1) If “correct distance by trajectory” is selected, select a representative vertical depth. Preferred selection of the target layer is the medium-deep layer. Or specify depth, medium to deep is the representative vertical depth of the target interval to obtain the point coordinates on the trajectory; at adjacent sampling points on the trajectory of well w and linear interpolation is obtained between (see Equation 2).

[0144] (2) If "anisotropic distance" is selected, set the major axis azimuth angle and the anisotropy ratio , perform rotation and scale transformation on the displacement vector according to Equation 3 to obtain the anisotropic distance .

[0145] (3) Distance selection rule: If both trajectory correction and anisotropy are selected, then use as the coordinates to calculate ; otherwise, use the wellhead distance or the Euclidean distance based on trajectory correction.

[0146] (4) Output the distance used for weight calculation, and record the parameters( p , k , R , , ratio, z ) for reproducibility and quality control (QC), where p represents the distance power exponent of IDW.

[0147] [Equation (1) Wellhead distance] .

[0148] [Equation (2) Trajectory correction coordinates] Given the representative vertical depth , at adjacent sampling points on the trajectory and linear interpolation:[[]] , . Use to replace the wellhead coordinates to calculate .

[0149] [Equation (3) Anisotropic distance] Let the displacement vector , the major axis azimuth angle . The rotation matrix , the scale matrix (a is the scale related to the major axis, b is the scale related to the minor axis, 0 < b ≤ a). Let , then:[[]] . If the anisotropy ratio is expressed as ratio = a / b, it can also be written as .

[0150] Note: The distance quantity It is used in key processes such as neighbor well sorting and screening (k-nearest neighbor / radius / automatic selection), IDW weight calculation (formula 4), top / bottom and thickness prediction of the layer (formulas 5-6), curve interpolation (formula 10), and spatial prior construction (formula 19).

[0151] (III) Prediction of the top / bottom and thickness of the O-section of the target well

[0152] Step S4: Read layered data and unify depth benchmark.

[0153] (1) Read the layer table, which contains at least the fields of well name, layer number, top depth, and bottom depth; the depth type of top / bottom can be MD, TVD or TVDSS.

[0154] (2) Determine the working depth benchmark TVD is preferred; when only MD is provided and MD-TVD trajectory exists, it can be converted to TVD first before performing segment prediction and alignment; when only a single depth reference is provided, this reference is used and then remapped at the output.

[0155] (3) For each well w Extract the top value for each layer_id. w bottom w Validate top w <bottom w ; Calculate the depth This refers to the mid-depth of the current adjacent well section. The mid-depth varies in different adjacent well sections and is used for deviated well proximity correction or parameter diagnosis.

[0156] (4) Handling missing layers: When a well is missing a layer number or the layer has no effective depth range, the well is marked as invalid in the layer and will not participate in the weighting and interpolation calculation of that layer.

[0157] Step S5: Target well level (top) O bottom O ) and thickness prediction.

[0158] (1) For each layer number, based on the distance obtained in steps S1-S3 Calculate the IDW weights according to Formula 4. .

[0159] (2) Perform weighted predictions on top and bottom according to Formula 5 to obtain the top. O with bottom O And calculate the thickness .

[0160] (3) Optional thickness through-hole: First, determine the thickness Perform interpolation, then use top O Rebuild bottom O This reduces the correlation between top / bottom errors and enhances the stability of segment thickness prediction.

[0161] Step S6: Layer boundary continuity correction and layer QC output.

[0162] (1) For adjacent segments sharing the same boundary, apply continuity constraints according to Formula 6, so that... To avoid gaps or overlaps between layers.

[0163] (2) Handling abnormal situations: If negative thickness or insufficient thickness occurs, the minimum thickness threshold can be applied. Perform trimming or revert to the weighted average thickness of adjacent wells, and record the QC flag.

[0164] (3) Output layer QC table: layer_id, top_O, bottom_O, thickness_O, used_wells involved in modeling, parameter group ( p , k , R , The ratio and missing rate are used for LOOCV, plot annotation, and result verification.

[0165] [Formula 4 IDW Weights] for adjacent wells i =1.. N The distance is The power is p Set a very small amount ε >0 prevents division by zero: , , j This represents the index used to iterate through all neighboring wells when performing a summation.

[0166] [Formula 5 Layer / Thickness Prediction] ; ;thickness Alternatively, thickness interpolation can be performed directly: and order .

[0167] [Formula 6 Adjacent Layer Continuity Constraint] For the sorted adjacent layers... and public boundary and correct , Simultaneously satisfy .

[0168] (iv) Intra-layer alignment, sampling strategy and multi-algorithm interpolation prediction

[0169] Step S7: Intra-layer alignment and sampling mesh construction.

[0170] (1) For each layer number, in the well w Define normalized intra-layer coordinates within the layer segment (See Formula 7), where .

[0171] (2) For target well O, according to Reconstruct the predicted depth grid for this layer (see Equation 8).

[0172] (3) Construct a unified r Sampling grid ( M The number of normalized sampling points within a given layer segment (representing the number of normalized sampling points) can be specified by the user or adaptively determined by the layer thickness and the desired sampling interval; different layer numbers can use different [samples / intervals]. M To maintain consistent vertical resolution.

[0173] Step S8: Resampling and preprocessing within the adjacent well curve layer.

[0174] (1) For each adjacent well w and each curve f (Well logging curves from adjacent wells, such as SP, GR, etc. for each well), extract this section [top] w bottom w For valid samples within the range, perform optional outlier handling and denoising (e.g., IQR removal, median filtering).

[0175] (2) Interpolate by one dimension Resampling (See Formula 9); When local missing values ​​cause interpolation failure, the most recent valid value, the mean within the layer, or linear extrapolation can be used to fill in the missing values, and the missing rate should be recorded.

[0176] (3) For curves with positive resistivity that vary across orders of magnitude, logarithmic domain processing can be used during interpolation and error evaluation: ,exist g The domain is used to perform the prediction and then the inverse transformation is performed to improve numerical stability.

[0177] Step S9: Multi-algorithm spatial interpolation prediction and fusion.

[0178] (1) IDW interpolation: in each Place, according to calculate (See Formula 10).

[0179] (2) Ordinary Kriging (optional): Under the assumption of second-order stationarity, the weights are solved based on the semi-variogram γ(h). The Kriging curve is obtained. and satisfy , Indicates the first i Sampling points of adjacent wells within the layer The resampling curve value at the location (see Formula 11).

[0180] (3) RBF interpolation (optional): RBF curve , where kernel function Gaussian kernel can be selected wait. The coefficients (weighting coefficients) of the RBF interpolation model are obtained by solving for the interpolation conditions satisfied by the sample points of adjacent wells. Indicates the first i Spatial location vectors of neighboring well sample points This represents the spatial location vector of the point to be predicted (the sampling point of the target well within the corresponding layer); , These represent the spatial distance and the scale parameter of the Gaussian kernel, respectively.

[0181] (4) Fusion (optional): If multi-algorithm fusion is enabled, then based on the first iteration of each algorithm under LOOCV or Bootstrap... s Interpolation algorithm error Calculate the first s Fusion weights of interpolation algorithms (See Formula 12) and obtain the final prediction. ;in, n This represents the algorithm index in the normalized summation (summing over all algorithms involved in the fusion). Indicates the first s Algorithm error conduct The result of the power transformation is used to adjust the sensitivity of the error in the weight calculation. This represents a stable term (a very small positive number) to prevent the denominator from being zero. s Indicates the algorithm index. ,in The number of algorithms participating in the fusion. Indicates the first s Interpolation algorithms sample points within the layer The predicted value at the location; (see Formula 13).

[0182] Step S10: Segment splicing, curve post-processing and output organization.

[0183] (1) Predict the numbers of each floor By splicing the data by depth, a complete prediction curve sequence for the target well O is obtained.

[0184] (2) Perform curve post-processing and physical constraints: for example, RT>0, DEN takes a reasonable range, GR is non-negative; optional peak suppression and smoothing processing can be performed, and the processing parameters are recorded for reproduction.

[0185] (3) Output the prediction curve table based on TVD (including depth and each curve column), and simultaneously output the number of sampling points for each layer. M Metadata such as missing rate and fusion weight / uncertainty are used for subsequent lithology prediction and quality control.

[0186] [Formula 7 Normalized In-Layer Coordinates] .

[0187] [Formula 8 Depth Reconstruction of Target Well Section O] .

[0188] [Formula 9: Resampling within adjacent well curve layers] .

[0189] Indicates the resampling curve. This represents a one-dimensional interpolation operator (such as linear interpolation, spline interpolation, or nearest neighbor interpolation). Represents the depth domain of adjacent well curves. Indicates the normalized coordinates within the layer The mapping relationship with the actual depth z of the adjacent well. This represents the normalized relative depth coordinates corresponding to the sampling point within the j-th layer. ;

[0190] [Formula 10 IDW Interpolation Prediction] .

[0191] [Formula 11 Ordinary Kriging Estimation] N represents the number of sampling points within the layer. The number of valid adjacent well sample points participating in ordinary kriging estimation;

[0192] [Formula 12: Multi-algorithm fusion weights] .

[0193] [Formula 13 Fusion Prediction] .

[0194] (v) Uncertainty estimation and quality control

[0195] Step S11: Uncertainty estimation.

[0196] (1) Bootstrap resampling of neighboring well sets: While keeping the number of sample wells unchanged, the neighboring well index is sampled with replacement to form the first set of samples. b Subsample set.

[0197] (2) Repeat steps S5-S10 for each resampled sample set to obtain the first... b Secondary prediction curve .

[0198] (3) Calculate the predicted mean μ(z) and standard deviation according to formula 14. σ ( z It can further construct prediction intervals to indicate prediction uncertainty, identify unreliable depth segments, and assist in setting weights for multi-algorithm fusion.

[0199] Step S12: Leave-one-out evaluation (LOOCV).

[0200] (1) For each well in the stratification table w For each layer number, temporarily designate the well as a "pseudo-target well" and use the remaining wells to perform steps S5-S6 to predict its layer top / bottom and thickness.

[0201] (2) Calculate and summarize the layer error indices (e.g., MAE (top), MAE (bottom), MAE (thickness)), and if necessary, evaluate the curve prediction error (RMSE, MAE, MAPE, etc., see Formula 15).

[0202] (3) Output LOOCV detail table and summary table to determine the rationality of adjacent well selection, distance parameters and interpolation parameters, and provide objective function for parameter optimization.

[0203] Step S13: Parameter mesh optimization.

[0204] (1) Construct a set of candidate parameters, such as IDW exponentiation. p Number of adjacent wells k ,radius R Anisotropy parameters With ratio, and fusion weight parameters q wait.

[0205] (2) Perform LOOCV on each set of parameters and calculate the objective function. J (For example, a weighted combination of thickness MAE or layer MAE can also be used to add curve error terms).

[0206] (3) Select the one that makes J The smallest set of parameters is taken as the optimal parameters and automatically backfilled into the prediction settings. At the same time, the search space, the optimal solution and the suboptimal solution are recorded to form a traceable parameter optimization report.

[0207] [Formula 14 Bootstrap Uncertainty] Repeat B Secondary resampling Mean Standard deviation .

[0208] [Formula 15 Error Index] mean is the arithmetic mean of all samples (depth points or layers) participating in the evaluation, pred represents the predicted value, and true represents the true value. , These represent the predicted and actual values ​​of a curve or indicator at the corresponding sample points, respectively.

[0209]

[0210]

[0211]

[0212]

[0213]

[0214]

[0215] (vi) Deep mapping and output results

[0216] Step S14: Depth mapping.

[0217] (1) When the target well O is an inclined well and provides discrete points of MD-TVD trajectory, the TVD grid result is converted into the MD grid result by piecewise linear mapping. The TVD→MD mapping is shown in Formula 16, and the MD→TVD inverse mapping is shown in Formula 17.

[0218] (2) When the wellhead reference KB is valid, calculate TVDSS according to formula 18, and simultaneously provide three sets of depth columns: MD, TVD and TVDSS in the output to meet different interpretation scenarios.

[0219] (3) When the target well is a vertical well or has no trajectory data, MD can be set to TVD according to the project agreement, or only a single depth benchmark can be output and the depth type can be marked.

[0220] Step S15: Output and visualization of results.

[0221] (1) Prediction curve output: Export the TVD baseline prediction curve file (e.g., Proposed_O_Multi.csv) and the MD-corrected prediction curve file (e.g., Proposed_O_Multi_MD.csv), and retain the necessary metadata (parameter group, adjacent well list, depth type, etc.).

[0222] (2) Layer QC output: Export the target well layer prediction table (layer_id, top_O, bottom_O, thickness_O, QC flag, etc.) and LOOCV details table for easy review by interpreters.

[0223] (3) Visual preview: Multiple curves share the same depth axis for plotting, and resistivity curves can use logarithmic coordinates; the target well boundary dashed line and layer label are superimposed, and the curves can be compared with any adjacent well curves.

[0224] [Formula 16 TVD→MD Mapping] Given discrete points in the trajectory table For any TVD= z lie in :

[0225] .

[0226] [Formula 17 MD→TVD Inverse Mapping] For any MD=m located at : .

[0227] [Formula 18 TVDSS] When the wellhead reference KB is known: .

[0228] (vii) Lithological profile prediction methods

[0229] Lithological profile prediction is used to infer the lithological columnar distribution of the target well O based on the prediction curve before drilling, and to compare and display it with the profiles of adjacent wells.

[0230] Step S16: Training data construction and feature engineering.

[0231] (1) Read the lithology interval table of adjacent wells (field examples: well, top, bottom, litho) and the logging curves of adjacent wells; map the lithology intervals to the working depth axis. The data is then discretized into sample points or sample windows. The lithological interval table can be obtained by selecting specific depth intervals from the imported lithological data.

[0232] (2) Feature extraction: Extract basic features (such as GR, RT, AC / DT, DEN, CNL, etc.) and optional derived features (gradient, sliding window statistics, intra-layer normalized coordinates) from the logging curves.r (etc.), forming feature vectors x ( z ).

[0233] (3) Standardization and missing value handling: Standardize / normalize the feature columns; impute missing values ​​and record the missing value rate; prune outliers when necessary to improve the robustness of the model.

[0234] (4) Sample division and leakage control: cross-validation or training-validation division is carried out by well name to avoid overfitting evaluation bias caused by leakage of samples from the same well.

[0235] Step S17: Lithology prediction strategy and classifier training.

[0236] (1) Full-segment supervision (Method 1): A global classifier is trained based on all training samples, and the classifier probability is output point by point for the entire segment of the target well O. . z Indicates vertical depth. c Indicates lithological category index, , C This represents the number of lithology categories (e.g., mudstone, sandstone, dolomite, limestone, etc. can be defined according to a lithology dictionary).

[0237] (2) Hierarchical supervision (method 2): A classifier is trained for each layer number and prediction is made only within the layer segment corresponding to the target well to enhance the adaptability under the difference in response between different layers.

[0238] (3) Neighbor interpolation (method 3): Without training a classifier, spatial prior probabilities are constructed solely based on the lithology and distance weights of neighboring wells. This allows for direct identification of lithology.

[0239] (4) Hybrid fusion (method 4): Simultaneous computation and Obtained through posterior fusion The subsequent judgment takes into account both well logging evidence and spatial continuity.

[0240] (5) Classifier implementation options: Random Forest, Gradient Boosting Tree, Support Vector Machine, Naive Bayes, Neural Network, etc.; Optional probability calibration can be performed to improve performance. Explainability.

[0241] Step S18: Spatial Prior Probability Construction (IDW).

[0242] (1) For a given layer number and depth point z Determine the set of adjacent wells that are effective at that depth. W ( z For each adjacent well w Search for it at depth zLithology .

[0243] (2) The distance obtained from steps S1-S3 Calculate weights (The normalized weight form of Formula 4 can be reused). As The power is p ,parameter Preventing division by zero: intermediate variables , ,in For adjacent wells The weight of the target well O, where u is the neighbor well index. For adjacent wells.

[0244] (3) Perform weighted statistics and normalization on each lithological category according to Formula 19 to obtain the spatial prior probability. ; Optional Slight smoothing is applied along the depth to suppress discrete noise.

[0245] Step S19: Posterior probability fusion, lithology discrimination and uncertainty assessment.

[0246] For each depth point z (or normalized coordinates within the layer) r Calculate the classifier probability for each corresponding depth point. With spatial prior probability And merge to obtain the posterior probability. :

[0247] (1) Classifier probability: Input the feature vector x(z) into the trained classifier (hierarchical supervision or whole-segment supervision) to obtain the probability output of each lithology category. (See Formula 20). Where Pr represents the probability operator. Y Represents a random variable indicating lithological category. c Indicates lithological category index, This represents the classifier model parameters (such as weight parameters, thresholds, or hyperparameter set).

[0248] (2) Spatial prior: obtained by weighted statistical analysis of lithology of adjacent wells according to step S18. (See Formula 19).

[0249] (3) Posterior fusion: setting α ≥0、 β ≥0 and smoothing terms First, calculate the unnormalized posterior probability:

[0250] Equivalently, this can be written as a linear fusion of the logarithmic field: (See Formula 21).

[0251] (4) Normalization: Let ,make .

[0252] This represents the lithology category index (dummy variable) in the normalized summation. Represents an exponential function;

[0253] (5) Lithology identification and confidence level: final lithology category And the confidence level can be defined. Used for post-processing and quality control.

[0254] (6) Uncertainty assessment: using posterior entropy The classification uncertainty is measured (see Formula 22) and used to identify low-confidence intervals, drive thin-layer merging / smoothing strategies, or prompt manual review.

[0255] (7) Parameter settings (optional): α , β It can be set manually, or the lithological error can be minimized through cross-validation / LOOCV, or the uncertainty of curve prediction can be combined. σ ( z Adaptive adjustment (e.g.) σ ( z Increase when it increases β (To enhance spatial constraints).

[0256] Step S20: Interval output, in-layer / layer position constraints and drawing of the drawing board.

[0257] The point-by-point lithology identification results are converted into engineering-usable lithology intervals, and comparison charts are generated:

[0258] (1) Intervalization: For point-by-point category sequences Perform run-length encoding to merge adjacent elements of the same type to form the initial lithological interval. and categories c .

[0259] (2) Minimum thickness constraint: for thickness Less than the threshold Thin layers are selected based on the posterior probability / confidence and uncertainty of adjacent intervals (e.g., preferential merging into...). higher or H Merge lower neighboring intervals; if necessary, perform sliding window majority voting / median filtering within the layer before intervalization.

[0260] (3) Stratigraphic constraints: All lithological intervals should be limited to the range of the corresponding strata number. Inside, at the layer boundaries, the sections are cut off and the boundaries are updated to ensure that the profile is consistent with the layer lines.

[0261] (4) Export Results: Output the lithological interval table of target well O (suggested fields: layer_id, top, bottom, litho, conf_mean, H_mean, etc.), and optionally output the point-by-point probability table (P_post, conf, ...). H (This is used for quality control.)

[0262] (5) Drawing: Draw the lithological profile of the sample well and the target well side by side (default TVD or TVDSS depth axis), overlay the layer boundaries and connect across wells; uncertainties can be marked with color / shading next to the lithological trace of well O. H (or 1-conf), used for pre-drilling interpretation and production applications.

[0263] [Formula 19 Spatial Prior Probability] Let the set of neighboring wells be... W ( z (in depth) z Effective adjacent wells), adjacent wells w The weight is (From Formula 4 and distance) (Calculation), its lithological category is .but: .

[0264] [Formula 20 Classifier Probability] For the input feature vector x ( z (From the predicted curve in) z The values ​​(composed of the given values) are standardized and then input into the classifier. .

[0265] [Formula 21 Posterior Fusion] and normalization Final lithological determination .

[0266] [Formula 22 Uncertainty Measure] Possible posterior entropy: .

[0267] IX. Examples

[0268] Example 1: Predicting the curve of well O and generating a lithological profile based on interpolation of adjacent wells using multi-curve interpolation

[0269] (1) Data preparation: Collect logging curves of several adjacent wells (including at least MD or TVD columns and GR / RT / AC curves, etc.) and organize the well coordinates; prepare the layers_table.csv for each well; if there are deviated wells, prepare the deviation table.csv (which can be in well-tvd-xy or well-md-inc-azm format).

[0270] (2) Adjacent well selection: setting k =6, radius R =1500 m (optional), and set anisotropy parameters (such as azim=45°, ratio=2.0) to match the dominant geological strike; the system automatically calculates the distance and selects neighboring wells.

[0271] (3) Layer prediction and QC: Perform IDW prediction on each layer_id to obtain top_O / bottom_O / thickness_O of well O and output the QC table; perform LOOCV to obtain MAE index, and perform grid optimization to update parameters when necessary.

[0272] (4) Curve prediction: Select curves such as GR, RT, and AC, and set dz_out=0.1 m or adaptive sampling; the system performs intra-layer normalization alignment for each layer_id, and calculates the IDW and Kriging prediction values ​​in parallel, and fuses the output when necessary; obtain the prediction curve file of the TVD depth benchmark of well O ( Figure 5 ).

[0273] (5) Depth mapping: If well O is an inclined well, the trajectory relationship is used to map the TVD prediction results to MD, and the TVDSS is calculated and the MD baseline prediction curve file is exported.

[0274] (6) Lithological profile prediction: Read the lithological interval table of adjacent wells, select the Hybrid strategy; enable spatial prior and set α , β (Or automatically); predict lithology point by point using the predicted curve as a feature and divide it into intervals, draw stratigraphic profiles and export the results. Figure 6 ).

[0275] Optional solutions and extensions

[0276] (1) Automated layering: In addition to directly reading the layering table, methods such as activity function can be used to automatically identify the layer boundaries from key logging curves and compare them with artificial layers to determine the final layering.

[0277] (2) Trend and zoning modeling: When the well spacing is large or there are obvious changes in sedimentary facies zones, trend surface or zoning model can be introduced to train and interpolate different regions respectively.

[0278] (3) Multi-source data fusion: Core, logging, seismic attributes or geological constraints can be added to the features to improve the accuracy of lithology prediction.

[0279] (4) Visualization of outcome uncertainty: or The entropy is output as an uncertainty curve and used for pre-drilling risk warning.

Claims

1. A pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves, characterized in that, Includes the following steps: Data access and depth benchmark unification: Read logging curves from adjacent wells, well location coordinates, stratigraphic data, and lithological annotation data; Unify the depth axes of each well to the selected working axis and establish a consistent mapping relationship between different depth axes; Neighbor selection and proximity correction: Constructing a set of neighboring wells ; When an inclined shaft trajectory exists, calculate the trajectory correction distance. and distance from the wellhead Form an optional distance metric; Layered framework construction and layer prediction: Based on layered data and distance metrics, the top and bottom depths of each well within each layer are determined; the layer information of the target well is predicted by distance weighting, and boundary correction is performed by introducing continuity constraints of adjacent layers; Intra-layer alignment and resampling: Normalized intra-layer coordinates are constructed within each layer segment, and a system is established... Mapping; in a uniform sampling grid Above, the curves of adjacent wells are changed from the depth domain. Resampling ; Multi-algorithm interpolation prediction and adaptive fusion: in each normalized coordinate At this point, at least two spatial interpolation algorithms are executed, and the fusion weights are adaptively determined. The final prediction curve is obtained. ; Depth mapping correction: for each layer segment The prediction curves for the entire well section of the target well are formed by splicing them together in depth order. When the target well is an inclined well, the prediction results are mapped between TVD and MD. Lithology prediction: Using the prediction curve as a feature, combined with lithology annotations from adjacent wells and a layered framework, the classifier probability is obtained through at least one strategy. Simultaneously, spatial prior probabilities are constructed based on the lithology and distance weights of adjacent wells. And obtained through posterior fusion Output lithology identification Including confidence and uncertainty indicators, to achieve lithological profile prediction.

2. The pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves according to claim 1, characterized in that, The adjacent well screening and proximity correction include the following steps: Candidate set determination and neighbor set generation: based on the wellhead coordinates of the target well O. Candidate well set { w Wellhead coordinates And the set neighbor well selection constraint parameters: nearest neighbor parameters k radius constraint R Constructing a set of adjacent wells ; Distance calculation between the target well and adjacent wells: for each candidate well w Calculate the Euclidean distance between wells ; When the candidate well is a vertical well, the Euclidean distance between wells is calculated based on the wellhead coordinates. ; When the candidate well is an inclined well, the trajectory correction and anisotropic distance calculation are performed as follows: (1) When selecting distance correction based on trajectory: select representative vertical depth Select the medium-deep target layer In candidate wells w adjacent sampling points of the trajectory and Interlinear interpolation , to be used as wellhead coordinates calculate ;in, ， ， Indicates the first j The corresponding medium and deep points , Indicates the first j The coordinates of the points; (2) When selecting anisotropic distance: set the principal axis azimuth angle Compared with anisotropy , , a , b These represent the distances along the primary and secondary axes, respectively; for wellhead coordinates... The displacement vector is rotated and scaled to obtain the anisotropic distance. As ; (3) When both trajectory correction and anisotropy are selected: then... Wellhead coordinates calculate And as the corrected wellhead distance ; Output updated distance and parameter groups ( k , R , ratio z ).

3. The pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves according to claim 1, characterized in that, The hierarchical skeleton construction and segment prediction include the following steps: S3-1, Layered Data Reading and Depth Benchmark Unification Read the layer table: containing well name, layer number (layer_id), top depth, and bottom depth; Determine the working depth benchmark When only MD is provided and MD-TVD trajectories exist, convert to TVD first and then perform layer prediction and alignment; when only a single depth reference is provided, use that reference as the output and then remap. For each adjacent well w Extract the top depth (top) for each layer number (layer_id). w Bottom depth w Validate top w bottom w ; S3-2, Target Well Stratification and Thickness Prediction For each layer number, the distance is determined based on the distance obtained from the adjacent well screening and proximity correction steps. As And calculate IDW weights ; Weighted predictions are performed on the top and bottom layers respectively to obtain the target well layer top. O with bottom O And calculate the thickness ;in, ; ; , Indicates the first i The top and bottom depths of the layer; the thickness can also be obtained by directly interpolating the thickness. and order ; S3-3, Boundary Continuity Correction and Layer QC Output (3.1) Apply continuity constraints to adjacent layer segments that share layer boundaries, so that... : For the two adjacent layers after sorting and public boundary And correct the target well number i Bottom depth Target well number i Top floor depth Simultaneously satisfy ; (3.2) Handling of abnormal situations: When the thickness of a certain layer is less than the thickness threshold ,according to Perform the trimming, or return to step S3-2 and record the QC flag. (3.3) Output layer QC table: top O bottom O , Parameter group ( k , R , (ratio).

4. The pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves according to claim 1, characterized in that, The intra-layer alignment and resampling includes the following steps: S4-1. In-layer alignment and sampling mesh construction (4.1) For each layer number, in the candidate well w Define normalized intra-layer coordinates within the layer segment ,in For vertical depth, top w bottom w These are the top depth and the bottom depth, respectively. ; (4.2) For target well O, according to The predicted depth grid for the reconstructed segment; Indicates the grid depth; S4-2, Resampling and Preprocessing within Adjacent Well Curve Layers For each adjacent well w and each logging curve f Extracting the top segment w bottom w Internal data; According to one-dimensional interpolation Resampling : ； Indicates the resampling curve. This represents a one-dimensional interpolation operator. Represents the depth domain of adjacent well curves. Indicates the normalized coordinates within the layer The mapping relationship with the actual depth z of the adjacent well. Indicates the first j Normalized relative depth coordinates corresponding to sampling points within each layer .

5. The pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves according to claim 1, characterized in that, The multi-algorithm interpolation prediction and adaptive fusion includes the following steps: S5-1, Multi-algorithm spatial interpolation prediction and fusion; For resampling curves The method can perform single-algorithm interpolation prediction or use multiple interpolation algorithms for fusion prediction; the algorithm is one of IDW interpolation, ordinary kriging, or RBF interpolation. The method of using multi-algorithm interpolation for fusion prediction is as follows: Based on the performance of each algorithm under LOOCV or Bootstrap s Interpolation algorithm error Calculate the first s Fusion weights of interpolation algorithms And obtain the final prediction. ;in, n This represents the algorithm index in normalized summation. Indicates the first s Algorithm error conduct The result of the power transformation is used to adjust the sensitivity of the error in the weight calculation. This represents a stable term to prevent the denominator from being zero. s Indicates the algorithm index. ,in The number of algorithms participating in the fusion. Indicates the first s Interpolation algorithms sample points within the layer The predicted value at that location; S5-2, Layer splicing and output Predicting each floor number By splicing the data by depth, a complete prediction curve sequence for the target well O is obtained.

6. The pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves according to claim 1, characterized in that, The depth mapping correction includes the following steps: When the target well O is an inclined well and provides discrete points for the MD-TVD trajectory, a piecewise linear mapping is used to convert the TVD grid results into MD grid results; TVD and MD represent vertical depth and measurement depth, respectively. TVD→MD mapping, given discrete points in the trajectory table For any TVD=z located at : ; z Indicates vertical depth. j Indicates the discrete point index; The MD→TVD inverse mapping, for any MD=m located at... : , m Indicates the depth being measured.

7. The pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves according to claim 1, characterized in that, The lithological prediction includes the following steps: S6-1 Training Data Construction and Feature Engineering Read the lithology interval table and logging curves of adjacent wells; map the lithology intervals to the working depth axis. Discretize it into sample points; Basic features representing lithology are extracted from logging curves of adjacent wells to form a feature vector. x ( z ); By analyzing the feature vector x ( z Normalization and grouping verification by well name are used to construct a training sample set; S6-2, Lithology Prediction Strategy and Classifier Training Training is performed using a single training method based on the training sample set; the training method is one of the following: whole-segment supervision method, hierarchical supervision method, or adjacent well interpolation method; wherein... Full-segment supervised method: Train a global classifier, and output the classifier probability point by point for the entire segment of the target well O. , z Indicates vertical depth. c Indicates lithological category index, , C This represents the number of lithological categories; Hierarchical supervision method: A classifier is trained for each layer number, and predictions are made only within the layer segment corresponding to the target well O; Neighbor interpolation method: Without training a classifier, spatial prior probabilities are constructed solely based on the lithology and distance weights of neighboring wells. Directly determine lithology; Alternatively, training can be performed using a hybrid fusion method based on the training sample set: the data obtained through the aforementioned whole-segment supervision method can be used for training. , and obtained through adjacent well interpolation method Obtained through posterior fusion Subsequent judgment.

8. The pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves according to claim 7, characterized in that, In S6-2, the construction of the spatial prior probability in the lithology prediction strategy and classifier training includes the following steps: Given a layer number and a depth point z, determine the set of adjacent wells that are valid at that depth. W ( z For each adjacent well w Search for it at depth z Lithology ; The distance obtained based on the neighbor well screening and proximity correction steps Calculate weights Power is p Set parameters Preventing division by zero: intermediate variables , ,in For adjacent wells The weight of the target well O, u For adjacent well index, For adjacent wells; We perform weighted statistics and normalization on each lithological category to obtain the spatial prior probability. The details are as follows: The weight of the adjacent well w is Its lithological category is ;but: , c Indicates lithological category index; The hybrid fusion method calculates the classifier probability for each depth point z. With spatial prior probability And merge to obtain the posterior probability. ;in: Classifier probability: the probability of the feature vector x ( z Input a trained classifier to obtain the probability output for each lithology category. Where Pr represents the probability operator, Y Represents a random variable indicating lithological category. c Indicates lithological category index, Indicates the parameters of the classifier model; Spatial prior probability: obtained by weighted statistics on the lithology of adjacent wells based on the construction of spatial prior probability. ; Posterior fusion: settings α ≥0、 β ≥0 and smoothing terms First, calculate the unnormalized posterior probability: Expressed as a linear fusion of the logarithmic field: ； Normalization: ,make ; This represents the lithological category index in the normalized summation. Represents an exponential function; Lithology identification and confidence level: final lithology classification And obtain the confidence level. ;in, This represents the lithology category index that maximizes the objective function. c ; Uncertainty assessment: using posterior entropy Measure classification uncertainty.

9. A pre-drilling formation profile prediction method based on multi-algorithm interpolation of well logging curves, characterized in that, include: The data reading module is used to read logging curves, well location coordinates, stratigraphic data, and lithological annotation data from adjacent wells. Unify the depth axes of each well to the selected working axis and establish a consistent mapping relationship between different depth axes; The adjacent well selection module is used to construct adjacent well sets. ; When an inclined shaft trajectory exists, calculate the trajectory correction distance. and distance from the wellhead Form an optional distance metric; The layer prediction module is used to determine the top and bottom depths of each well within each layer based on layered data and distance metrics; it performs distance-weighted prediction of the layer information of the target well and introduces the continuity constraint of adjacent layers for boundary correction; The intra-layer alignment and resampling module is used to construct normalized intra-layer coordinates within each layer segment and establish... Mapping; in a uniform sampling grid Above, the curves of adjacent wells are changed from the depth domain. Resampling ; Curve prediction module, used for each normalized coordinate At this point, at least two spatial interpolation algorithms are executed, and the fusion weights are adaptively determined. The final prediction curve is obtained. ; The depth mapping module is used for mapping each layer segment. The prediction curves for the entire well section of the target well are formed by splicing them together in depth order. When the target well is an inclined well, the prediction results are mapped between TVD and MD. The lithology prediction module is used to obtain classifier probabilities by using the prediction curve as a feature, combined with adjacent well lithology annotations and a layered framework, through at least one strategy. Simultaneously, spatial prior probabilities are constructed based on the lithology and distance weights of adjacent wells. And obtained through posterior fusion Output lithology identification Including confidence and uncertainty indicators, to achieve lithological profile prediction.

10. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the pre-drilling formation profile prediction method based on multi-algorithm interpolation of logging curves as described in any one of claims 1-9.

Images