A porosity interval extension prediction method

Abstract

The application discloses a porosity interval extension prediction method, and belongs to the technical field of intelligent interpretation of oil and gas exploration geophysical logging. The application extracts general logging response characteristics by using a frozen pre-training logging base model, calculates curve fusion weights according to potential characteristics corresponding to each logging curve, and constructs a deep index potential sequence based on the curve fusion weights; after local shape extraction and response state memory modeling, the potential logging base characteristics are updated in a residual manner, and then curve sensitive reading is performed on the adapted potential logging characteristics to obtain a basic porosity prediction result; the limited offset correction parameters and the limited affine correction parameters are calculated by using the real labels of the explained interval, and the gated selection or weighted fusion is performed according to the unlabeled response state statistical characteristics; in the training stage, higher loss weight is given to the high porosity sample. The application can improve the stability and generalization ability of porosity prediction under the condition of label scarcity, and improve the problem that the response of the high porosity reservoir is compressed.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation technology of geophysical logging in oil and gas exploration, specifically to a porosity interval extension prediction method. It is applicable to reservoir evaluation scenarios where the interpreted depth interval of the target well has porosity labels and the uninterpreted depth interval to be predicted only has multi-curve logging data. It is particularly applicable to target well porosity extension prediction scenarios where the logging curve response state changes along the well depth direction and the response of high-porosity reservoirs needs to be maintained. Background Technology

[0002] Porosity is one of the key parameters in oil and gas reservoir evaluation. It characterizes the proportion of pore volume to the total volume of rock and serves as an important basis for evaluating reservoir storage capacity, identifying effective reservoirs, estimating reserves, and formulating development plans. Continuous and reliable porosity profiles can reflect the variation of reservoir properties with well depth, which is of great significance for oil and gas exploration and development.

[0003] In practical well logging interpretation, high-reliability porosity data typically originates from core experimental analysis, professional well logging interpretation, or manual comprehensive evaluation results. While these methods offer high reliability, they suffer from drawbacks such as high acquisition costs, long interpretation cycles, limited spatial coverage, and reliance on expert experience. Therefore, within the same target well, only a portion of the interpreted depth range typically possesses reliable porosity (POR) labels, while a large number of uninterpreted depth ranges only have conventional logging curves such as sonic logging, natural gamma logging, and resistivity logging, without forming continuous porosity interpretation results. How to achieve extended prediction of target well porosity ranges using porosity labels from interpreted depth ranges and logging curve data from uninterpreted depth ranges without using true porosity (POR) labels for uninterpreted depth ranges is a crucial technical challenge in intelligent well logging interpretation.

[0004] Conventional logging curves can reflect formation physical property responses from different perspectives. Among them, sonic logging curves are related to rock compaction, pore structure, and porosity changes; natural gamma logging curves are often used to reflect clay content and lithological changes; and resistivity logging curves are related to formation fluid properties, water content, and electrical conductivity. Different logging curves have different measurement physical mechanisms, numerical scales, noise characteristics, and geological response meanings. Their relationship with porosity is not a simple linear mapping but is influenced by multiple factors such as lithology, sedimentary environment, compaction degree, fluid state, and interpretation background.

[0005] Existing porosity prediction methods based on machine learning or deep learning typically use multiple well logging curves as input and porosity as output to establish a nonlinear prediction model between the well logging curves and porosity. These methods can, to some extent, utilize the combined response relationship between multiple curves to improve porosity prediction accuracy. However, existing methods still have the following shortcomings.

[0006] First, existing methods often use random sample partitioning to evaluate model performance. Since samples at adjacent depths within the same well exhibit strong continuity and similarity, random partitioning may result in training and testing samples coming from similar formation response states. The model essentially performs local interpolation, rather than a true extension prediction from the interpreted depth range to the uninterpreted depth range to be predicted. Actual well logging interpretation processes typically do not involve randomly selecting depth points for prediction; instead, they utilize the porosity (POR) label of the interpreted depth range of the target well to extend porosity to subsequent or other uninterpreted depth ranges. Therefore, random sample evaluation results cannot accurately reflect the model's reliability in actual porosity range extension tasks.

[0007] Secondly, there may be differences in response states between different depth ranges of the target well. The same well may experience changes in lithology, compaction degree, sedimentary environment, or fluid state along the depth direction, causing the learned logging curve-porosity relationship in the interpreted depth range to shift in the uninterpreted depth range. Existing models typically lack a controlled migration correction mechanism for changes in the target well's response state, easily leading to insufficient migration or overcorrection problems caused by unconstrained trend fitting.

[0008] Third, while high porosity ranges are significant in reservoir evaluation, they are often few in number in actual sample distributions. Models that primarily optimize global average error are easily dominated by low-to-medium porosity samples, predicting high porosity ranges as medium porosity results close to the average level. This leads to the compression and underestimation of favorable reservoir responses, affecting reservoir selection and development decisions.

[0009] Fourth, general time series models or conventional deep learning models typically treat different logging curves as ordinary multi-channel sequences for unified modeling, making it difficult to fully preserve the differences in physical mechanisms and curve-specific response information among different logging curves. For logging data, different curves are not equivalent channels, but rather multi-curve response systems with clear geophysical meanings. Without curve-sensitive feature reading and fusion mechanisms, the model's ability to identify key logging responses is easily weakened.

[0010] In recent years, pre-trained basic models have provided a new technical approach for intelligent interpretation of well logging. By pre-training on large-scale unlabeled or weakly labeled well logging sequences, basic logging models can learn general well logging response representations, providing prior knowledge for downstream tasks such as porosity prediction. However, pre-trained basic logging representations are not directly equivalent to the porosity prediction capability of the target well. If only frozen representations are used for simple regression, problems such as mismatch in local response states of the target well, insufficient depth direction migration, and compression of predictions in high-porosity regions may still exist. At the same time, if only ordinary same-well calibration or unconstrained trend fitting is performed, overcorrection may be introduced, causing the porosity prediction results in the uninterpreted depth range to deviate from the actual reservoir response.

[0011] Therefore, there is an urgent need for a porosity range extension prediction method that can utilize pre-trained well logging fundamental characterization and achieve the following objectives without using the True Porosity (POR) label for the uninterpreted depth range to be predicted: First, to determine the effective evidence and contribution weight of each curve for porosity prediction based on the differences in physical response of different well logging curves; Second, to transfer calibration information in the interpreted depth range to the uninterpreted depth range to be predicted in a restricted manner, considering the changes in the response state of the target well along the depth direction; Third, to improve the model's ability to preserve the high porosity response of favorable reservoirs, addressing the problems of scarce high porosity samples and prediction compression. Summary of the Invention

[0012] This invention addresses the technical problems of existing porosity prediction methods, such as the scarcity of target well porosity labels, the inability of random sample evaluation to reflect the true range extension capability, insufficient response state transition between different depth ranges, the easy compression of high-porosity reservoir responses, and the difficulty of general time-series models in distinguishing the differences in physical responses of different logging curves. It provides a porosity range extension prediction method based on a logging fundamental model, constrained response state transition correction, and high-porosity tail adaptation. This method takes multi-curve logging data of the target well and porosity labels for the interpreted depth range of the target well as input. It uses a frozen pre-trained logging fundamental model to extract a general logging response characterization. Based on this characterization, it combines curve-sensitive readout, potential time-series state adaptation, constrained response state transition correction, and high-porosity tail adaptation mechanisms to achieve continuous porosity prediction for the uninterpreted depth range of the target well. Among them, the porosity labels of the target well's explained depth range are used for model supervised training, target well response state calibration, correction parameter calculation, and model selection; the unexplained depth range to be predicted only provides multi-curve logging data, curve availability information, potential logging basic characterization, and basic porosity prediction statistics. The true porosity labels of the unexplained depth range to be predicted do not participate in model training, parameter updates, response state calibration, gating selection, high porosity threshold determination, or model selection.

[0013] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A porosity interval extension prediction method, the construction and training process of the porosity interval extension prediction model includes: After preprocessing the acquired multi-curve logging data of the target well and the true porosity labels of the interpreted depth range, a multi-curve logging input token sequence is constructed. Input the multi-curve logging input token sequence into the frozen pre-trained logging base model to extract the potential logging base representation for each depth location and each logging curve; The confidence scores of curves are calculated based on the basic characteristics of each potential well logging curve, and the curve confidence scores are normalized by combining the curve availability identifier to obtain the curve fusion weight. Based on the curve fusion weight, the potential basic characteristics of different well logging curves at the same depth location are weighted and fused to construct a potential sequence of depth index. Local morphological extraction and response state memory modeling are performed on the potential sequence of the depth index. The basic characterization of the potential logging is updated in a residual manner to obtain the adapted potential logging characterization. The adapted potential logging characterization is input into the curve sensitive readout mechanism to generate the porosity candidate prediction value and the latest curve confidence value corresponding to each logging curve. The latest curve confidence value is re-normalized in combination with the curve availability identifier to obtain the adapted curve fusion weight. The adapted curve fusion weight is then used to perform weighted fusion of the porosity candidate prediction values ​​to obtain the basic porosity prediction result. Based on the deviation between the true porosity label in the explained depth range and the basic porosity prediction result, the restricted offset correction result and the restricted affine correction result are calculated; the compatibility of the response state is judged according to the statistical characteristics of the unlabeled response state between the explained depth range and the unexplained depth range to be predicted, and the gated selection or weighted fusion between the restricted offset correction and the restricted affine correction is performed according to the response state compatibility to obtain the final porosity prediction result. During the training phase of the porosity range extension prediction model, samples with porosity higher than the set value are assigned higher loss weights.

[0014] In one embodiment, the step of preprocessing the acquired target well multi-curve logging data and the true porosity labels of the interpreted depth range, and then constructing a multi-curve logging input token sequence, specifically includes: Depth window slicing is performed on the normalized multi-curve logging data of the target well to obtain logging segments corresponding to each depth location and each logging curve. ; through well logging fragment embedding function Generate well logging content embedding Unavailable or missing logging segments are embedded using a shared missing mask. Representation; construct a depth location embedding for each depth location. , The index of the t-th depth location; construct a curve type embedding for each logging curve. , The index of the c-th logging curve; combined into an input token. ,in, The curve availability identifier consists of a multi-curve logging input token sequence comprised of input tokens for all depth locations and all logging curves. , Indicates the number of depth positions; This indicates the number of logging curves.

[0015] In one embodiment, the step of calculating curve confidence based on the basic characterization of each potential well logging, and normalizing the curve confidence by combining the curve availability identifier to obtain the curve fusion weight, specifically includes: For the The depth position Potential logging-based characterization of a single logging curve Through the curve confidence function And the confidence curve is generated by Sigmoid activation. Combined with curve availability indicators Normalize the confidence scores of the curves to obtain the curve fusion weights. , Indicates the first The effective confidence level of all available logging curves at each depth location and , Indicates the first At the depth position of the first Availability indicators for each logging curve; Indicates the first At the depth position of the first The confidence level of each logging curve; This indicates the number of logging curves.

[0016] In one embodiment, the weighted fusion of potential well logging baseline representations of different well logging curves at the same depth location based on the curve fusion weights to construct a potential depth index sequence specifically includes: Based on curve fusion weights Potential logging-based characterization of different logging curves at the same depth location Weighted fusion is performed to obtain the latent representation of the deep index. The target well depth index potential sequence is composed of potential representations of depth indexes at continuous depth locations. , Indicates the number of depth positions.

[0017] In one embodiment, the local morphological extraction and response state memory modeling of the depth index potential sequence are performed to update the potential logging base characterization in a residual manner, resulting in an adapted potential logging characterization. The adapted potential logging characterization is then input into a curve-sensitive readout mechanism to generate porosity candidate prediction values ​​and the latest curve confidence scores for each logging curve. The latest curve confidence scores are then re-normalized using a curve availability identifier to obtain adapted curve fusion weights. These adapted curve fusion weights are then used to weight and fuse the porosity candidate prediction values ​​to obtain the base porosity prediction result. Specifically, this includes: For deep indexed latent sequences After layer normalization, the data is input into a one-dimensional depth convolution module to obtain local depth morphological features. , Indicates the first Local depth morphological features at a depth location Indicates the number of depth locations; indexes the latent sequence by depth. Input the bidirectional response state memory module to obtain the shallow-to-deep directional memory features. Memory features from deep to shallow After fusion, the response state memory features are obtained. ; local depth morphological features With response state memory features The residual fitting terms are obtained by low-rank projection after splicing. ; Indicates a low-rank projection module; This represents the trainable parameters of the low-rank projection module; through the gating coefficients. Controlling the injection intensity of residual adaptation terms to update the underlying logging characterization The potential logging characterization after adaptation was obtained. , This represents a learnable or preset residual scaling factor. This represents the Sigmoid activation function; The adapted potential logging characterization set The input curve-sensitive readout mechanism generates porosity candidate prediction values ​​and curve confidence scores for each adapted potential logging characterization. The curve confidence scores are then re-normalized using a curve availability identifier to obtain adapted curve fusion weights. These weights are then used to weight and fuse the porosity candidate prediction values ​​to obtain the basic porosity prediction result. ,in For curve-sensitive porosity readout function, This represents the trainable parameters in the curve-sensitive porosity readout function.

[0018] In one embodiment, the calculation process for the constrained offset correction result is as follows: Within the explained depth range Within this framework, constrained offset correction parameters are calculated based on the median deviation between the actual porosity label and the baseline porosity prediction. And obtain the constrained offset correction results: ; This represents the boundary truncation function. Indicates the deviation scaling factor; This indicates the explained depth range of the target well. Calculate the median within the range; Indicates the target well In depth The actual porosity label at the location; Indicates the target well In depth Predicted results of basic porosity at the location; These represent the boundary values ​​of the offset correction parameters. Indicates the target well In depth The porosity prediction results after constrained offset correction.

[0019] In one embodiment, the calculation process for the restricted affine correction result is as follows: Within the explained depth range Within, the restricted affine correction result is: ; Target well In depth Porosity prediction results after constrained affine correction; Indicates the target well Proportional correction parameters; Indicates the target well Intercept correction parameters; Proportional correction parameters and intercept correction parameters satisfy: ; ; in, This represents the loss function for fitting the correction parameters; Indicates the explained depth range The number of valid samples in the sample; and These represent the lower and upper boundaries of the proportional correction parameter, respectively. This represents the absolute boundary value of the intercept correction parameter.

[0020] In one embodiment, determining the compatibility of response states based on the statistical characteristics of unlabeled response states between the explained depth range and the unexplained depth range to be predicted specifically includes: Calculate the difference between the base forecast and the mean: ; Indicates the explained depth range With the unexplained depth range to be predicted The difference between the underlying predicted mean values; This represents the mean operation. This represents the mean operation. Indicates the target well In depth Predicted results of basic porosity at the location; Calculate the distance between potential representation centers: ; in, Indicates the explained depth range Unexplained depth range to be predicted Distance between potential representation centers To adapt to the depth index representation obtained from the subsequent potential logging characterization, Represents the L2 norm; Calculate the percentage of unlabeled tails supported: ; Indicates the unexplained depth range to be predicted. The unlabeled tail support ratio is obtained based on the basic porosity prediction results. Indicates an indicator function; This refers to the basic prediction tail judgment threshold determined based on training samples, explained depth ranges, or preset reservoir evaluation standards. When both conditions are met , , The system determines whether the explained and unexplained depth ranges have compatible response states; otherwise, it determines that there are differences in response states. This indicates the threshold for the difference between the baseline predicted mean and the target mean. This represents the threshold for the distance from the potential representation center; This indicates the threshold for the proportion of unlabeled tails that are supported.

[0021] In one embodiment, the step of gating or weighted fusion between restricted offset correction and restricted affine correction based on the response state compatibility to obtain the final porosity prediction result specifically includes: Set the correction method selection coefficient. When it is determined that there is response state compatibility The constrained offset correction result is used; when a difference in response state is determined and the constrained affine correction does not trigger the backoff rule, The results of restricted affine correction are used; when a difference in response state is determined but the restricted affine correction triggers a backoff rule, The prediction was reversed to a restricted offset correction; the final prediction result is: ; in, Indicates the target well In depth Final porosity prediction results at the location; This represents the porosity prediction result after constrained affine correction. This represents the porosity prediction result after constrained offset correction.

[0022] In one embodiment, assigning a higher loss weight to samples with porosity higher than a set value specifically includes: Set a high porosity threshold For the true porosity labels in the training samples Define a high porosity indicator variable ; This indicates an indicator function that takes the value when the condition within the parentheses is true. Otherwise, the value is ; Set the weights for the tail sample loss: ; Indicates the first Loss weights for each training sample; This indicates the loss weight for high-porosity samples; Indicates the first High porosity indicator variable for each training sample; Based on training batches High porosity tail adaptation loss ; This represents the smoothed L1 loss function; Indicates the first Basic porosity prediction results for each training sample; Indicates the first The true porosity label of each training sample; The core parameters of the pre-trained well logging basic model Keep it frozen during the training phase.

[0023] Compared with the prior art, the beneficial technical effects of the present invention are: 1. This invention utilizes a frozen pre-trained logging baseline model to extract general logging response characteristics, so that porosity range extension prediction no longer completely depends on training the model from scratch with limited porosity labels of the target well. This can improve the stability and generalization ability of porosity prediction under label scarcity conditions.

[0024] 2. The present invention adopts a curve-sensitive readout mechanism to generate porosity candidate prediction values ​​and curve contribution weights for different logging curves. This can fully distinguish the differences in physical response of different curves such as sonic logging, natural gamma logging, and resistivity logging, and avoid confusion of response information caused by simply treating different logging curves as equivalent channels.

[0025] 3. This invention introduces a potential temporal state adaptation mechanism in the potential well logging basic characterization space. Through local depth morphology extraction and response state memory modeling, it can characterize the continuous changes in formation and the transfer of response state along the well depth direction of the target well, so that the porosity prediction results are more consistent with the geological response change law in the well depth direction.

[0026] 4. This invention performs constrained response state transfer correction based on the explained depth range of the target well. It can use the real porosity labels in the explained depth range to calibrate the local response deviation of the target well and transfer the correction relationship to the unexplained depth range to be predicted in a controlled manner. This avoids overcorrection caused by unconstrained fitting and prevents the real porosity labels in the unexplained depth range to be predicted from participating in training, correction or gating selection.

[0027] 5. This invention introduces a high porosity tail adaptation mechanism, which assigns higher loss weights to high porosity samples in the explained depth range or training samples during the training phase. This can alleviate the compression problem of high porosity reservoir response by conventional models under the condition of uneven sample distribution, and is conducive to maintaining the porosity prediction response of favorable reservoir ranges.

[0028] 6. This invention organically combines pre-trained logging basic characterization, curve-sensitive readout, potential time-series state adaptation, constrained response state migration correction, and high porosity tail adaptation to form a "calibrated interval - uninterpreted interval prediction" technical route that conforms to the actual logging interpretation process, which can improve the reliability of target well porosity interval extension prediction.

[0029] 7. In one specific embodiment, the method of the present invention is used to predict the depth-direction porosity interval extension of multiple wells in the target area. Compared with the calibrated sequence model comparison method, the method of the present invention has a higher coefficient of determination in the depth-direction interval extension evaluation. Increase to In the evaluation of high-porosity sample retention, the method of this invention achieves... The prediction results show that the present invention can improve the problem of compression in high-porosity reservoirs while maintaining the overall porosity prediction accuracy. Attached Figure Description

[0030] Figure 1 This is a flowchart of the porosity range extension prediction method of the present invention.

[0031] Figure 2 This is a schematic diagram of the overall framework of the present invention.

[0032] Figure 3 This is a schematic diagram illustrating the construction of the depth window input token according to the present invention.

[0033] Figure 4 This is a schematic diagram of the constrained response state migration correction of the present invention.

[0034] Figure 5 This is a schematic diagram of the high porosity tail adapter of the present invention. Detailed Implementation

[0035] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0036] like Figure 1 As shown, the porosity range extension prediction method of the present invention includes the following steps: S1 performs unified preprocessing on the multi-curve logging data of the target well to obtain a multi-curve logging sequence after depth alignment, outlier cleaning, missing value identification, and scale normalization.

[0037] S2. Construct a depth window input sequence for the pre-trained logging basic model, divide the continuous logging curves of the target well into depth window logging segments, and construct a multi-curve logging input token sequence based on logging segment embedding, curve availability mask, curve type embedding, and depth location embedding.

[0038] S3, input the multi-curve logging input token sequence into the frozen pre-trained logging base model, and extract the potential logging base representation corresponding to each depth position and each logging curve; wherein, the backbone parameters of the pre-trained logging base model remain frozen during the downstream porosity range extension prediction process.

[0039] S4. Calculate the curve confidence based on the potential logging baseline characterization of each logging curve, and normalize the curve confidence by combining the curve availability mask to obtain the curve fusion weight.

[0040] S5, Enhance the expression of depth-direction response state based on the potential temporal state adaptation mechanism. Specifically, the potential logging base representations of different logging curves at the same depth location are weighted and fused according to the curve fusion weights to construct a potential depth index sequence; the potential depth index sequence is modeled through a local depth morphology extraction module and a response state memory module, and the potential logging base representations are updated through residual adaptation to obtain adapted potential logging representations; then the adapted potential logging representations are input into a curve-sensitive readout mechanism, the adapted curve fusion weights are recalculated, and the porosity candidate prediction values ​​corresponding to each logging curve are weighted and fused to obtain the basic porosity prediction results.

[0041] S6. Constrained response state migration correction is performed based on the interpreted depth range of the target well. Specifically, within the interpreted depth range of the target well, constrained offset correction parameters are calculated based on the deviation between the true porosity label and the basic porosity prediction result. Simultaneously, constrained affine correction parameters are calculated based on the basic prediction result and the true porosity label within the interpreted depth range. Then, based on the logging response statistical characteristics, potential characterization statistical characteristics, or basic prediction statistical characteristics of the interpreted depth range and the uninterpreted depth range to be predicted, the response state compatibility between the two is determined. Gating selection or weighted fusion is performed between constrained offset correction and constrained affine correction to obtain the final response state migration correction result. The above response state compatibility determination does not use the true porosity label of the uninterpreted depth range to be predicted.

[0042] S7 enhances the ability to retain the response of high-porosity reservoirs based on a high-porosity tail adaptation mechanism. Specifically, a high-porosity threshold is pre-set. During the training phase, high-porosity samples are identified only based on the true porosity labels in the explained depth range or training samples, and high-porosity samples are given higher loss weights. This allows the model to improve its ability to retain the response of high-porosity reservoirs while optimizing the overall porosity prediction error. During the prediction phase of the unexplained depth range, the true porosity labels of the range are not used to identify high-porosity samples.

[0043] S8, based on the trained porosity interval extension prediction model, inputs multi-curve logging data of the uninterpreted depth interval of the target well into the model. After pre-trained logging basic characterization extraction, curve fusion weight calculation, potential time series state adaptation, curve sensitive readout and restricted response state transfer correction, it outputs the continuous porosity prediction results of the uninterpreted depth interval of the target well.

[0044] Figure 2 This invention demonstrates the overall framework of the invention, which is based on a pre-trained well logging basic model, curve-sensitive readout, potential time-series state adaptation, constrained response state migration correction, and high porosity tail adaptation.

[0045] This invention takes multi-curve logging data of the target well and porosity labels of the interpreted depth range of the target well as inputs, extracts general logging response characteristics using a frozen pre-trained logging basic model, and realizes continuous porosity prediction of the uninterpreted depth range of the target well through curve sensitive readout, potential time-series state adaptation, restricted response state migration correction and high porosity tail adaptation.

[0046] In this invention, the interpreted depth range of the target well is denoted as... The unexplained depth range to be predicted is denoted as .in, The true porosity labels in the data can be used for supervised model training, target well response state calibration, correction parameter calculation, and model selection. The data used in this study only includes multi-curve logging data, curve availability information, potential logging fundamental characterization, and basic porosity prediction statistics. The true porosity label in the model is not used in model training, parameter updates, response state correction, gating selection, high porosity threshold determination, or model selection.

[0047] Step 1: Preprocessing of multi-curve logging data from the target well.

[0048] (1) Data acquisition: Acquire multi-curve logging data from the target well along its depth. In one specific implementation, the multi-curve logging data for the target well includes sonic logging curves (AC), natural gamma logging curves (GR), and resistivity logging curves (RT). These curves, together with the porosity labels in the interpreted depth range of the target well, constitute the main data basis for porosity range extension prediction. If a logging curve in the target well is missing, invalid, or unavailable at certain depth locations, that depth location is not deleted; instead, the missing or invalid curve is identified using a curve availability mask during subsequent input sequence construction.

[0049] Simultaneously, porosity labels corresponding to the interpreted depth range of the target well are obtained. These porosity labels are continuous numerical reservoir property parameters used to supervise model training, target well response state calibration, correction parameter calculation, and prediction error constraints. The true porosity labels for the uninterpreted depth range to be predicted are not used as model input, nor are they used for model training, parameter updates, response state correction, gating selection, high porosity threshold determination, or model selection.

[0050] Let the target well be the first... The set of effective depth sampling points for a well is represented as: ; in, Indicates the target well The effective depth sampling point set; Indicates the first One depth sampling point; Indicates the target well The number of effective depth sampling points.

[0051] Let the set of logging curves used for the target well be represented as: ; in, This represents the set of logging curves used for the target well; Indicates the well logging curve number; This indicates the number of logging curves.

[0052] Target well In depth The multi-curve logging input at the location is represented as follows: ; in, Indicates the target well In depth Multi-curve logging input vector at the location; Indicates the target well In depth First The logging values ​​of the logging curves; This indicates transpose.

[0053] Target well In depth The porosity label at a given location is: ; in, Indicates the target well In depth The actual porosity label at the location; This indicates the porosity value at the corresponding depth.

[0054] It should be noted that the definition in this step This is the multi-curve logging input vector at a single depth location. In subsequent depth window slicing and input sequence construction, it is represented by multiple consecutive depth sampling points. Based on this, we further constructed depth window logging segments and multi-curve logging input token sequences.

[0055] (2) Depth interval division: Target well The effective depth range is divided into the explained depth interval and the unexplained depth interval to be predicted. The explained depth interval is denoted as... The unexplained depth range to be predicted is denoted as And satisfy: ; in, Indicates the target well The data contains explained depth ranges with available porosity labels for training, calibration, or model selection; Indicates the target well The unexplained depth range for which porosity extension prediction is required; This represents the empty set.

[0056] for The depth sampling points, and their multi-curve logging inputs and true porosity label It can be used for supervised model training, target well response state calibration, calibration parameter calculation, and model selection.

[0057] for The depth sampling points only use their multi-curve logging input. The curve availability information, as well as the potential well logging basic characterization, basic porosity prediction results or unlabeled response state statistics generated by the subsequent model; its true porosity label is not used in model training, parameter update, response state correction, gating selection, high porosity threshold determination or model selection.

[0058] (3) Depth alignment: Using the target well depth sampling point as a reference, depth alignment is performed on different types of logging curves and porosity labels so that the same depth location corresponds to a set of multi-curve logging response values. For logging curves with inconsistent sampling intervals, linear interpolation, nearest neighbor interpolation, spline interpolation, or resampling methods can be used to unify the vertical resolution; for missing depth locations, their missing state is retained and identified in the subsequent input construction stage using a curve availability mask.

[0059] The aligned target well logging data is represented as follows: ; in, Indicates the target well Aligned multi-curve logging data set; Indicates depth The multi-curve logging input vector at the location.

[0060] (4) Data cleaning and curve availability labeling: The logging curves and porosity labels undergo quality screening, eliminating or marking null values, non-finite values, invalid instrument codes, obvious outliers, non-monotonic depth records, and data samples that do not meet the preset physical value range. For logging curves with large orders of magnitude ranges, such as resistivity RT, logarithmic transformation can be performed to reduce the impact of extreme values ​​on model training.

[0061] For target well In depth First Set up a logging curve availability indicator: ; in, Indicates the target well In depth First Availability indicators for each logging curve; when When, it indicates that the curve is available at the corresponding depth position; when When the value is 0, it indicates that the curve is missing or invalid at the corresponding depth position.

[0062] The availability mask is composed of availability identifiers corresponding to each depth location and each logging curve: ; in, Indicates the target well A set of curve availability masks is used for subsequent input token construction, curve-sensitive readout, and curve weight normalization.

[0063] (5) Scale normalization: The cleaned logging curves are then normalized. Normalization methods can include Z-score standardization, percentile scaling, maximum / minimum normalization, or combinations thereof. For different logging curves, the normalization method can be set according to the curve's physical meaning and numerical distribution.

[0064] In one implementation, for the first The logging curves are normalized as follows: ; in, Represents the normalized i-th Log logging curve values; Indicates that for the first The preprocessing transformation function set for each logging curve, when the curve does not require special transformation. It is an identity mapping; Indicates the first The statistical measure of the mean of the logging curves; Indicates the first The standard deviation statistic of a well logging curve; This represents a numerically stable term, used to prevent the denominator from being zero.

[0065] To avoid information leakage caused by the uninterpreted porosity labels in the undefined depth range, the normalized statistics... and The preferred method is based on training well data and the interpreted depth range of the target well. The logging curve data, or the normalized statistics pre-set in the pre-trained logging basic model, are used to calculate and are fixed for the uninterpreted depth range to be predicted in the target well. Well logging curve processing.

[0066] When it is needed in practical applications When performing label-free scaling correction on well logging curve distribution information, only the following methods are permitted: The logging curve values ​​or curve availability information are used to calculate unlabeled statistics, without using... The true porosity label.

[0067] After the above processing, normalized multi-curve logging data, curve availability mask, and interpreted depth range of the target well are obtained. Porosity labels in the data, and unexplained depth ranges to be predicted. The results above serve as the basis for constructing the depth window slicing and multi-curve logging input token sequence in the second step. The true porosity labels in the model can be used for model training, calibration parameter calculation, and model selection. Only multi-curve logging data and its unlabeled statistics are used; no other data is used. The true porosity label in the data.

[0068] The second step is to construct the depth window slice and the input sequence.

[0069] like Figure 3 As shown, this step uses the target well normalized multi-curve logging data, curve availability mask, and interpreted depth range obtained in the first step. and unexplained depth range to be predicted The input is a multi-curve logging input token sequence used to construct a suitable input pre-trained logging baseline model along the well depth direction. The input token sequence consists of logging segment content information, curve availability information, curve type information, and depth location information, and serves as the input for the third step of freezing the pre-trained logging baseline model to extract potential logging baseline representations. Figure 3 This illustrates the process by which multi-curve logging data from a target well is processed through depth window slicing, fragment content embedding, missing mask embedding, depth location embedding, and curve type embedding to form a multi-curve logging input token sequence.

[0070] (1) Depth window slicing: According to the preset depth window length Sliding slices are used to extract multiple depth window logging segments from the effective logging sequence of the target well. Each depth window contains continuous... Each depth sampling point corresponds to a point in the target well. Several logging curves. Among them, Indicates the length of the logging segment, used to distinguish it from the uninterpreted depth range to be predicted. .

[0071] For the The depth position and the first The logging curve is represented as follows: ; in, Indicates the first Centered on the depth position, the first The logging segment corresponding to each logging curve; Indicates the target well In depth First Normalized logging values ​​for each logging curve; Indicates the length of the logging segment; Indicates the window radius; This indicates transpose.

[0072] When the depth window is located at the boundary of the effective depth range of the target well, it can be processed by boundary completion, mirror completion, zero value completion, nearby effective value completion or window shortening, and the completion or missing position can be identified by curve availability mask.

[0073] For the The depth position and the first For each logging curve, set an availability indicator for the curve segment: ; in, Indicates the first At the depth position of the first Availability identifier for each logging curve segment; when When, it indicates that the curve segment is available; when When this occurs, it indicates that the curve segment is missing or invalid. The curve segment availability identifier is determined by the curve availability mask from the first step. Derived from.

[0074] (2) Embedding of curve content: For available logging segments Perform fragment embedding encoding to obtain the well logging content embedding: ; in, Indicates the first At the depth position of the first Embedding the content of each logging curve; This represents the well logging segment embedding function; This represents the parameters of the well logging segment embedding function.

[0075] The logging segment embedding function can be implemented using a one-dimensional convolutional network, a linear projection layer, a multilayer perceptron, a lightweight sequence encoder, or a combination thereof. In a preferred embodiment, different logging curves share the same logging segment embedding function to reduce the number of model parameters and enhance the comparability of different curve segments in a unified latent space.

[0076] (3) Missing mask embedding For unavailable or missing logging segments, a shared missing mask embedding representation is used: ; in, This indicates the mask embedding corresponding to the missing well logging curve segment; This indicates the embedding dimension of the input token.

[0077] By embedding a missing mask, unavailable or missing logging curve segments are still retained in the multi-curve logging input token sequence, thereby maintaining the consistency of depth location, curve type, and token arrangement structure, and avoiding depth-curve structure disorder caused by curve missingness.

[0078] (4) Depth position embedding: For each depth position Constructing deep location embeddings: ; in, Indicates the first Embedding of depth positions corresponding to depth positions; This represents the depth-position embedding function.

[0079] The depth position embedding is used to characterize the relative or absolute position information of the logging sequence along the well depth direction, and can employ learnable position coding, sine and cosine position coding, relative depth position coding, or continuous position coding based on actual well depth values.

[0080] (5) Curve type embedding: For each logging curve, construct a curve type embedding: ; in, Indicates the first Embed the curve type corresponding to each logging curve; This indicates a curve-type embedded function.

[0081] The curve type embedding is used to identify the curve identity and physical response source of different logging curves, enabling the pre-trained logging baseline model to distinguish the geophysical meaning of different curves such as acoustic logging curves (AC), natural gamma logging curves (GR), and resistivity logging curves (RT). By introducing curve type embedding, even if different curves share the logging segment embedding function, the model can still identify the curve category corresponding to each input token.

[0082] In an alternative implementation, when the multi-curve logging input token sequence adopts a fixed curve arrangement order and the curve identity has been explicitly encoded in the model structure, the curve type embedding can also be replaced by fixed curve position encoding or curve identifier encoding.

[0083] (6) Input token construction: Combining well logging content embedding, missing mask embedding, depth location embedding, and curve type embedding yields the first... The depth position, the first Input tokens corresponding to each logging curve: ; in, This represents the logging input token used to input the pre-trained logging baseline model; Indicates the first At the depth position of the first Availability indicators for each logging curve segment; This indicates that well logging content is embedded; Indicates missing mask embedding; Indicates depth position embedding; This indicates that the curve type is embedded.

[0084] The multi-curve logging input token sequence consists of input tokens corresponding to all depth locations and logging curves: ; in, Indicates the target well Input the multi-curve logging input token sequence into the pre-trained logging baseline model; This indicates the number of depth locations in the input sequence; This indicates the number of logging curves.

[0085] After the above processing, a multi-curve logging input token sequence is obtained, retaining curve content, curve availability, curve type, and depth location information. This is used to extract potential logging basic characteristics from the subsequent frozen pre-trained logging basic model.

[0086] The third step is to freeze the pre-trained logging basic model and extract potential logging basic characteristics.

[0087] This step uses the multi-curve logging input token sequence constructed in the second step as input, and inputs it into the frozen pre-trained logging basic model to extract the potential logging basic representations corresponding to each depth location and each logging curve. This provides a general logging response prior for the curve fusion weight calculation in the fourth step, the potential temporal state adaptation in the fifth step, and the constrained response state transition correction in the sixth step.

[0088] (1) Source of pre-trained well logging basic model: The pre-trained logging baseline model is a logging sequence encoding model pre-trained on multi-curve logging sequences, used to output a general logging response representation; this invention does not limit its specific pre-training task and network structure. It should be noted that the pre-trained logging baseline model is used to obtain general logging response priors; during the prediction of the target well porosity interval extension, the true porosity label of the uninterpreted depth interval to be predicted is not used in the pre-trained model update, downstream model training, response state correction, or model selection.

[0089] (2) Extraction of potential well logging basic characterization: The target well multi-curve logging input token sequence constructed in the second step. Input the pre-trained well logging basic model to obtain the potential well logging basic representation set: ; in, Indicates the target well The potential set of basic logging characterizations; This represents the pre-trained well logging basic model; This represents the core parameters of the pre-trained well logging basic model; This represents the multi-curve logging input token sequence obtained in the second step.

[0090] For the The depth position, the first The potential logging baseline representation output by the pre-trained logging baseline model is represented by the following logging curve: ; in, Indicates the first At the depth position of the first Potential logging fundamental characterization corresponding to each logging curve; This indicates that the output of the pre-trained well logging basic model is related to the first... The depth position and the first The characterization vector corresponding to each logging curve.

[0091] It consists of the underlying logging baseline characterization corresponding to all depth locations and all logging curves: ; in, This indicates the number of depth locations in the input sequence; This indicates the number of logging curves.

[0092] (3) Boundary between trunk freezing and parameter update: In a preferred embodiment of the present invention, the core parameters of the pre-trained well logging basic model... During the downstream porosity range extension prediction process, the parameter update boundary remains frozen and does not participate in backpropagation updates. Its parameter update boundary is expressed as: ; in, This represents the training loss for downstream porosity prediction; This represents the core parameters of the pre-trained well logging basic model.

[0093] In other words, during the downstream training phase, the pre-trained logging baseline model is used only as a general logging response characterization extractor; model training mainly updates the parameters of the curve-sensitive readout module, the potential time-series state adaptation module, and the porosity prediction task output head. In this way, on the one hand, the general logging response priors learned in the pre-training phase are preserved, and on the other hand, the risk of overfitting under the condition of limited samples in the explained depth range of the target well is reduced.

[0094] After the above processing, a potential logging baseline characterization set for the target well is obtained. This characterization set serves as the basis for the fourth step of calculating curve fusion weights and the fifth step of constructing the depth index potential sequence and performing potential time-series state adaptation.

[0095] Step 4: Calculate the curve fusion weights.

[0096] This step uses the potential well logging-based characterization set obtained in step three. The curve availability mask obtained in the first and second steps is used as input to calculate the curve confidence of each logging curve at each depth location, and is normalized in combination with the curve availability identifier to obtain the curve fusion weight.

[0097] Different logging curves have different physical meanings, numerical scales, noise characteristics, and porosity response relationships. Therefore, this invention does not simply splice, average, or treat different logging curves as equivalent channels. Instead, it calculates the curve confidence based on the potential logging fundamental characterization corresponding to each logging curve and obtains the curve fusion weight by combining the curve availability identifier. The curve fusion weight is used to characterize the effective evidence of each logging curve at the current depth position and its contribution to the fusion of potential characterizations, and serves as the basis for constructing the potential sequence of depth index in the fifth step.

[0098] In this way, the model can dynamically determine the contribution of each curve to the potential characterization fusion based on the response state of different logging curves such as AC, GR, and RT at the same depth location, thereby avoiding the confusion of curve physical response information caused by fixed averaging or simple splicing.

[0099] (1) Calculation of curve confidence: For the The depth position, the first Potential logging fundamental characterization corresponding to each logging curve The curve confidence score of the logging curve at the current depth is generated using the curve confidence function. ; in, Indicates the first At the depth position of the first The confidence level of each logging curve; This represents the Sigmoid activation function; Represents the confidence function for the curve; This represents the trainable parameters of the curve confidence function.

[0100] (2) Curve weight normalization: Combining the curve segment availability identifier obtained in the second step The confidence scores of each logging curve are normalized to obtain the curve fusion weight. First, the effective confidence scores of all available logging curves at the current depth are calculated: ; in, Indicates the first The effective confidence level of all available logging curves at a given depth location; Indicates the first At the depth position of the first Availability indicators for each logging curve; Indicates the first At the depth position of the first The confidence level of each logging curve; This indicates the number of logging curves.

[0101] when Greater than the preset minimum effective threshold At that time, the confidence level of the curve is normalized to obtain the first... Curve fusion weights for each logging curve: ; in, Indicates the first At the depth position of the first Curve fusion weights for each logging curve; This represents a numerical stability term, used to improve the numerical stability of normalized calculations. This represents the minimum effective threshold for determining whether the current depth location has a valid logging curve input.

[0102] When the When only one logging curve is available at a given depth location, the curve fusion weight of the available logging curve is 1, and the curve fusion weight of the remaining unavailable logging curves is 0.

[0103] when When this occurs, it indicates that all logging curves at the current depth location are missing, invalid, or have insufficient confidence. In this case, at least one of the following anomaly handling methods can be used: skip this depth location or depth window and exclude it from the training loss calculation; supplement the current depth location using a nearby effective depth window; output a low-confidence flag and use the porosity prediction results of a nearby effective depth location as a backup prediction; or use the porosity statistics of the explained depth range of the target well, the porosity statistics of the training set, or the model's default output as a backup prediction. The above anomaly handling processes do not use the unexplained depth range to be predicted. The true porosity label.

[0104] The curve weight vector at the current depth position is constructed from the curve fusion weights corresponding to each well logging curve: ; Among them, the curve fusion weight vector is used to characterize the degree of contribution of each logging curve to the potential characterization fusion of the current depth position.

[0105] After the above processing, the curve fusion weights are obtained. The curve fusion weights are used in the fifth step to weight and fuse the potential logging baseline characteristics of different logging curves at the same depth location, and to construct a potential depth index sequence.

[0106] Step 5: Potential temporal state adaptation.

[0107] This step uses the potential well logging-based characterization set obtained in step three. The curve fusion weights obtained in step four are used as input to construct a potential sequence of depth indices arranged along the well depth direction. Through local depth morphology extraction, response state memory modeling, and gated residual update, the adapted potential logging characterization and basic porosity prediction results are obtained.

[0108] Because the same target well may exhibit variations in lithology, compaction degree, sedimentary environment, fluid state, or local logging response morphology along its depth, the logging response at a single depth location is insufficient to fully characterize the local formation state. Therefore, this invention introduces a potential temporal state adaptation mechanism within the pre-trained logging foundation representation space. This mechanism enhances the expression of local formation morphology and response state transitions along the depth direction of the target well without compromising the general response priors of the pre-trained logging foundation model.

[0109] Specifically, the potential time-series state adaptation mechanism includes: constructing a potential depth index sequence based on curve fusion weights; capturing peak and valley changes, gradient changes, thin-layer responses, and local logging morphology within adjacent depth ranges through a local depth morphology extraction module; capturing the response state context in the shallow-to-deep and deep-to-shallow directions through a response state memory module; and then injecting the depth state residual into the potential logging basic representation through low-rank projection and gated residual update methods to obtain the adapted potential logging representation.

[0110] (1) Construction of latent sequences in deep indexing: First, based on the curve fusion weights obtained in step four, the potential logging baseline characteristics of different logging curves at the same depth location are weighted and fused to obtain the... Potential representation of depth index at depth locations: ; in, Indicates the first Potential representation of depth index at depth locations; Indicates the first At the depth position of the first Curve fusion weights for each logging curve; Indicates the first At the depth position of the first Potential logging fundamental characterization corresponding to each logging curve; This indicates the number of logging curves.

[0111] The potential sequence of target well depth indexes is composed of potential representations of depth indexes at continuous depth locations: ; in, Indicates the target well The deep index of the latent sequence; This indicates the number of depth positions in the current input sequence.

[0112] In this way, the curve fusion weights obtained in the fourth step are used to weight and fuse the potential logging basic characteristics of different logging curves at the same depth location, forming a state sequence arranged by depth, so that subsequent modules can model the response state changes along the well depth direction.

[0113] (2) Local depth morphology extraction: For deep indexed latent sequences Layer normalization is performed and then input into a one-dimensional depthwise convolution module to obtain local depth morphological features: ; in, This represents the set of local depth morphological features corresponding to all depth positions within the depth window; Indicates the first Local depth morphological features corresponding to each depth location; Indicates the kernel size as One-dimensional depthwise convolution operation; Presentation layer normalization operation; This represents the potential sequence of the target well depth index.

[0114] The local depth morphology features are used to characterize the local variation patterns of logging responses within adjacent depth ranges, including but not limited to peak value variations, valley value variations, gradient variations, local abrupt changes, thin-layer responses, and morphological continuity between adjacent depth sampling points. Through this local depth morphology extraction module, the model can identify local formation morphological variations that are difficult to express using only single-point potential representations.

[0115] In one implementation, the one-dimensional depth convolution module can be implemented using single-layer one-dimensional convolution, multi-layer one-dimensional convolution, dilated convolution, depth-separable convolution, or a combination thereof. The convolution direction expands along the well depth direction without altering the contribution relationship of different logging curves formed in the fourth step.

[0116] (3) Response state memory modeling: Deep indexing of latent sequences Input the response state memory module to obtain the response state memory characteristics: ; in, This represents the set of response state memory features corresponding to all depth locations within the depth window; Indicates the first Response state memory features corresponding to each depth location; This indicates the response state memory module; This represents the trainable parameters of the response state memory module.

[0117] In a preferred embodiment, the response state memory module is implemented using a bidirectional response state memory network to model potential response state changes in both the shallow-to-deep and deep-to-shallow directions. The process can be represented as follows: ; ; ; in, Indicates the first Response state memory features from shallow to deep at each depth location; Indicates the first Response state memory features from deep to shallow direction at each depth location; This indicates vector concatenation; and This represents the feature fusion parameters.

[0118] Through the bidirectional response state memory modeling described above, the model can simultaneously perceive potential response state information above and below the current depth position, thereby capturing formation state transitions, response state continuations, and local response state abrupt changes along the well depth direction.

[0119] (4) Latent representation gated residual update: Local depth morphological features and response state memory features The data is stitched together and mapped back to the pre-trained well logging basic representation space via a low-rank projection module to obtain the residual fitting terms: ; in, Indicates the first Residual adaptation terms corresponding to each depth location; Indicates a low-rank projection module; Represents the trainable parameters of the low-rank projection module; This indicates that local depth morphological features are spliced ​​with response state memory features.

[0120] Furthermore, to avoid excessive disturbance to the frozen pre-trained logging baseline characterization during the initial stage of downstream training, this invention sets a gating coefficient to control the injection intensity of the residual adaptation term. The gating coefficient is expressed as: ; in, Indicates the first Gating coefficient at each depth position; This represents the Sigmoid activation function; Indicates the gated generation function; These represent the trainable parameters of the gated generation function. The gate coefficients... The range of values ​​is to It can be a scalar gate or a vector gate with the same dimension as the residual fit term.

[0121] The original potential logging baseline characterization is updated using a gated residual method to obtain the adapted potential logging characterization: ; in, Indicates the first At the depth position of the first The potential logging characterization after adaptation corresponding to each logging curve; This represents the original potential basic characterization of well logging; Indicates the residual fit term; Indicates the gating coefficient; This represents element-wise multiplication; This indicates a learnable or preset residual scaling factor.

[0122] In one specific implementation, Initialize to a smaller value, or use the gated generation function. The initial output is set to a low level to avoid excessive disturbance to the pre-trained logging baseline representation during the early stages of model training. The residual adaptation term is shared across different logging curves at the same depth location, representing the common local response state changes at that depth location; different logging curves are still represented by their respective original underlying logging baselines. The curve-sensitive readout mechanism maintains the curve difference.

[0123] In alternative implementations, explicit gating coefficients may not be set. Instead, only the residual scaling factor is used. The residual adaptation is scaled and updated; this method is called scaled residual update.

[0124] (5) Generation of basic porosity prediction results: It consists of the adapted potential logging characterization corresponding to all depth locations and all logging curves: ; in, This represents the set of potential logging characterizations after adaptation to potential time-series states.

[0125] The adapted potential logging characterization The input curve-sensitive porosity readout mechanism generates porosity candidate prediction values ​​and curve confidence scores for each adapted potential logging characterization. The curve confidence scores are then re-normalized using a curve availability identifier to obtain the adapted curve fusion weights. These adapted curve fusion weights are then used to weight and fuse the porosity candidate prediction values ​​to obtain the basic porosity prediction results. ; in, Indicates the first The basic porosity prediction results obtained after potential temporal state adaptation at each depth location; This represents the curve-sensitive porosity readout function that sequentially performs porosity candidate prediction, curve confidence calculation, curve fusion weight normalization after adaptation, and candidate prediction weighted fusion on the adapted potential logging characterization. This represents the trainable parameters in the curve-sensitive porosity readout function.

[0126] The basic porosity prediction results Used for constrained response state migration correction in step six. Within the explained depth range of the target well. middle, It can be used in conjunction with the true porosity label to calculate correction parameters; in the unexplained depth range to be predicted. In this study, only multi-curve logging input, adapted potential logging characterization, and basic porosity prediction results are used; no other methods are employed. The true porosity label.

[0127] After the above processing, the adapted potential logging characterization is obtained. and basic porosity prediction results This is used for subsequent confined response state migration correction. The basic porosity prediction results serve as input for the sixth step of confined response state migration correction; the adapted potential logging characterization can also be used in the sixth step to calculate the interpreted depth range. Unexplained depth range to be predicted Statistical characteristics of unlabeled response states between them.

[0128] Step 6: Constrained response state transition correction.

[0129] This step uses the basic porosity prediction results obtained in step five, the adapted potential logging characterization, and the interpreted depth range defined in step one. and unexplained depth range to be predicted As input, used to utilize The calculation of the true porosity label in the model is constrained by correction parameters, and is based on the absence of... The response state compatibility index of the true porosity label will transfer the correction relationship in a controlled manner to... .

[0130] The pre-trained logging baseline model provides a global logging response prior. Although the baseline porosity prediction results obtained in the fifth step have incorporated curve-sensitive information and potential temporal state information, local response deviations may still exist between different depth ranges of the target well. Therefore, this invention utilizes the true porosity labels in the interpreted depth range of the target well to perform constrained response state transfer correction on the baseline porosity prediction results, and transfers the correction relationship to the uninterpreted depth range to be predicted in a controlled manner.

[0131] The basic porosity prediction result in this step comes from step five, that is, the basic prediction result obtained by the curve-sensitive porosity readout mechanism after potential temporal state adaptation, denoted as . ;in, Indicates the target well In depth The predicted results of the basic porosity at the location.

[0132] for Depth location, known true porosity label It can be used to calculate offset correction parameters, restricted affine correction parameters, and evaluation indicators for correction methods. For The depth location is determined using only multi-curve logging input, curve availability information, basic potential logging characterization, adapted potential logging characterization, basic porosity prediction results, and response state statistical characteristics calculated from the aforementioned unlabeled information; no other data is used. The true porosity label in the data does not use the method described by [the source]. The high porosity identifier obtained from the true porosity label is used in calibration parameter fitting, response state compatibility judgment, gating selection, model selection, or parameter update.

[0133] (1) Calculation of constrained offset correction parameters: Within the interpreted depth range of the target well Within this framework, based on the median deviation between the actual porosity label and the basic porosity prediction results, the target well offset correction parameters are calculated: ; in, Indicates the target well Offset correction parameters; Indicates the boundary truncation function; Indicates the deviation scaling factor; This indicates the explained depth range of the target well. Calculate the median within the range; Indicates the target well In depth The actual porosity label at the location; Indicates the target well In depth Predicted results of basic porosity at the location; This represents the boundary values ​​of the offset correction parameters.

[0134] By setting offset boundaries This can avoid overcorrection caused by local anomalies in the explained depth range of the target well, label noise, or sample distribution skew.

[0135] In one specific implementation, Set as , Set as Porosity percentage points.

[0136] The porosity prediction result obtained based on constrained offset correction is expressed as follows: ; in, Indicates the target well In depth The porosity prediction results after constrained offset correction.

[0137] (2) Calculation of restricted affine correction parameters: When the target well has interpreted depth range Unexplained depth range to be predicted When there are changes in response state, the restricted affine correction method can be used to perform proportional and intercept corrections on the basic porosity prediction results. The results of the restricted affine correction are expressed as follows: ; in, Indicates the target well In depth Porosity prediction results after constrained affine correction; Indicates the target well Proportional correction parameters; Indicates the target well The intercept correction parameters.

[0138] Proportional correction parameters and intercept correction parameters Based solely on the interpreted depth range of the target well The actual porosity label and the predicted basic porosity are fitted together and constrained by a preset range. ; ; in, The loss function for fitting the correction parameters can be represented by mean squared error, mean absolute error, SmoothL1 loss, or Huber loss. Indicates the explained depth range The number of valid samples in the sample; and These represent the lower and upper boundaries of the proportional correction parameter, respectively. This represents the absolute boundary value of the intercept correction parameter.

[0139] In one specific implementation, The range of values ​​is , The range of values ​​is Porosity percentage points.

[0140] To avoid overcorrection of the unexplained depth range to be predicted by restricted affine correction, when the proportional correction parameter or intercept correction parameter reaches a preset boundary, when restricted affine correction does not bring a stable error improvement compared to restricted offset correction, or when the unlabeled response statistical characteristics of the unexplained depth range to be predicted are inconsistent with the correction direction, a fallback to restricted offset correction or a reduction in the weight of restricted affine correction may be implemented. The above fallback decision does not use the true porosity label of the unexplained depth range to be predicted.

[0141] (3) Calculation of response status compatibility index: To determine the interpreted depth range of the target well Unexplained depth range to be predicted To determine whether a response state is compatible, this invention constructs a compatibility index based on unlabeled response state statistical features. These response state statistical features include logging curve statistical features, potential characterization statistical features, basic porosity prediction statistical features, or combinations thereof.

[0142] In one specific implementation, the difference between the baseline forecast means is first calculated: ; in, Indicates the explained depth range Unexplained depth range to be predicted The difference between the underlying predicted mean values; This indicates the operation of averaging.

[0143] Further calculate the distance between potential representation centers. Let... Indicates the target well In depth Given the depth index representation obtained from the adapted potential logging representation, then: ; in, Indicates the explained depth range Unexplained depth range to be predicted The distance between potential representation centers; This represents the L2 norm.

[0144] Further calculation of the unlabeled tail support metric for the unexplained depth range to be predicted: ; in, Indicates the unexplained depth range to be predicted. The unlabeled tail support ratio is obtained based on the basic porosity prediction results. Indicates an indicator function; This represents the basic prediction tail judgment threshold determined based on training samples, explained depth ranges, or preset reservoir evaluation criteria. This metric does not use... The true porosity label.

[0145] Further calculations can be performed on each logging curve. and The differences in mean, variance, median, quantile, curve variation, curve availability, or distribution distance, and their correlation with... , and Both are used together to determine response status compatibility.

[0146] (4) Response status compatibility judgment rules: In one specific implementation, a threshold rule is used for judgment. and Response state compatibility. If both conditions are met: ; ; ; Then determine the interpreted depth range of the target well. Unexplained depth range to be predicted If the response states are compatible, the restricted offset correction method is preferred; otherwise, if there is a difference in response states, the restricted affine correction candidate process is entered.

[0147] in, This represents the threshold for the difference between the baseline predicted mean and the target mean. This represents the threshold for the distance from the potential representation center; This indicates the threshold for the proportion of unlabeled tail support. This threshold is predetermined by the interpreted depth range of training wells, validation wells, target wells, or a preset reservoir evaluation standard, and is not based on... The true porosity label is determined.

[0148] When a difference in response state is determined, restricted affine correction is not necessarily applied directly. The reliability of restricted affine correction must be determined in conjunction with the affine correction backoff rules described in Section 2. If the restricted affine correction triggers the backoff rules, then restricted offset correction will ultimately be used, or the weight of restricted affine correction will be reduced.

[0149] (5) Gating selection of calibration method: Based on the response state compatibility judgment results and the affine correction backoff rules, determine the correction method selection coefficient. .

[0150] In a preferred embodiment, For binary gated variables: ; when When, the constrained offset correction result is used; when When using the results of restricted affine correction, the results are employed.

[0151] Specifically, it can be determined according to the following rules: ; The final porosity prediction result is expressed as follows: ; in, Indicates the target well In depth Final porosity prediction results at the location; This represents the porosity prediction result after constrained affine correction. This represents the porosity prediction result after constrained offset correction; This represents the coefficient for selecting the correction method.

[0152] In an alternative implementation, It can also be a continuous fusion coefficient, and its value range is: ; at this time, It can be calculated from the response state compatibility index, the affine correction reliability index, or a combination thereof. The closer When the final prediction result is more biased towards restricted affine correction; when The closer When this happens, the final prediction result tends to be more biased towards restricted offset correction. The continuous fusion method also does not use... The true porosity label.

[0153] After the above processing, the uninterpreted depth range of the target well is obtained. Final porosity prediction results at various depths This result serves as the primary prediction for the continuous porosity prediction curve output in step eight, and together with the high porosity tail adaptation mechanism in step seven, it ensures overall prediction accuracy and the ability to maintain high porosity response.

[0154] Figure 4 The process of calculating restricted offset correction parameters and restricted affine correction parameters based on the true porosity labels of the interpreted depth range of the target well and the basic porosity prediction results, and performing gating selection or fusion based on response state compatibility is illustrated.

[0155] Step 7: High porosity tail fitting.

[0156] This step uses the basic porosity prediction results obtained in step five and the true porosity labels in the explained depth range of the target well as inputs to assign higher loss weights to high-porosity samples during the training phase. This step belongs to the tail adaptation mechanism in the model training phase, used to assign higher loss weights to high-porosity samples based on training samples or the explained depth range of the target well. The true porosity label enhancement model learns weights for high-porosity samples, and together with the constrained response state transition correction in step 6, it serves the unexplained depth range to be predicted in step 8. Porosity output. This step does not indicate that... Prediction phase The actual porosity label in the data is adjusted online.

[0157] High porosity regions typically represent a small percentage of the sample distribution, but they are crucial for identifying favorable reservoirs, evaluating reservoirs, and selecting optimal development horizons. Conventional porosity prediction models, if optimized solely based on the global average error, are easily dominated by low-to-medium porosity samples, predicting high-porosity regions as having medium porosity close to the overall average, thus compressing the response of favorable reservoirs.

[0158] The constrained response state migration correction in step six is ​​mainly used to address the interpreted depth range of the target well. Unexplained depth range to be predicted The overall prediction offset or response state deviation between the two; the high-porosity tail adaptation in this step is mainly used to solve the problem of high-porosity response being compressed by meanization due to the low proportion of high-porosity samples during the training phase. The two solve different problems, but both serve the prediction of the target well porosity range extension.

[0159] In this step, the high-porosity tail adaptation only uses real porosity labels to identify high-porosity samples in the training samples during the model training phase, and adjusts the training loss weights accordingly; in the unexplained depth range to be predicted... In the prediction phase, do not use The true porosity label is not based on The true porosity label is used to identify high-porosity samples.

[0160] (1) Setting the high porosity threshold: Set a high porosity threshold For the true porosity labels in the training samples ,when When, the training sample is defined as a high-porosity sample; when At that time, the training sample is defined as a non-high porosity sample. Its high porosity indicator variable is expressed as: ; in, Indicates the first High porosity indicator variable for each training sample; This indicates an indicator function that takes the value when the condition within the parentheses is true. Otherwise, the value is ; Indicates the first The true porosity label of each training sample; This indicates the high porosity threshold.

[0161] The high porosity threshold Based on the reservoir evaluation criteria for the study area, and the interpreted depth range of training or target wells. The sample distribution, validation set performance, or business interpretation requirements are predetermined. The high porosity threshold... Not based on the unexplained depth range to be predicted The true porosity label has been determined, and it is no longer being viewed. Adjustments are made after labeling the actual porosity.

[0162] In one specific embodiment, the high porosity threshold Set as This value is a parameter for one embodiment and does not constitute the sole limitation on the scope of protection of this invention.

[0163] (2) Weighting of tail sample loss: To enhance the model's learning strength for high-porosity samples, tail sample loss weights are set for the training samples: ; in, Indicates the first Loss weights for each training sample; This indicates the loss weight for high-porosity samples; Indicates the first The high porosity indicator variable for each training sample.

[0164] when hour, The corresponding sample loss weights are ;when hour, The corresponding sample loss weights are .

[0165] The loss weight of the high porosity sample The weights can be preset based on the proportion of high-porosity samples in the training set, the prediction performance on the validation set, or business evaluation requirements. The loss weights for the high-porosity samples are not based on the unexplained depth range to be predicted. Adjust the actual porosity label or high porosity label.

[0166] In one specific implementation, the high porosity sample loss weight Set as This value is a parameter for one embodiment and does not constitute the sole limitation on the scope of protection of this invention.

[0167] (3) Tail-weighted loss function: Let a training batch or training sample set be represented as: ; in, Indicates a training batch or training sample set; Indicates the first Each training sample corresponds to a multi-curve logging input; Indicates the first The true porosity label of each training sample; This indicates the number of samples in the training batch or training sample set.

[0168] For the Using training samples, the basic porosity prediction results are obtained through steps one through five. ;in, Indicates the first The basic porosity prediction results corresponding to each training sample.

[0169] The high-porosity tail-fit loss function based on training batches or training sample sets is expressed as follows: ; in, This indicates the tail-end adaptation loss in high-porosity applications; Indicates the first Loss weights for each training sample; This represents the smoothed L1 loss function; Indicates the first Basic porosity prediction results for each training sample; Indicates the first The true porosity label of each training sample.

[0170] This loss function is calculated only using the true porosity labels during the model training phase and is not used for the unexplained depth range to be predicted. The online prediction phase utilizes this loss function to optimize the overall porosity prediction error while increasing the contribution of high-porosity samples to the training objective, thus mitigating the problem of high-porosity reservoir responses being compressed into medium-porosity prediction values.

[0171] In alternative implementations, the basic loss function may also be mean squared error, mean absolute error, Huber loss, or a combination thereof; the weighting method for high porosity samples may also be piecewise weighting, continuous weighting, quantile weighting, or a weighting function based on reservoir evaluation level.

[0172] (4) Training parameter update range: During the model training phase, the core parameters of the pre-trained well logging basic model are kept frozen, and the parameters of the curve sensitive readout module, the potential time series state adaptation module, and the porosity prediction task output head are mainly updated.

[0173] Let the main parameters of the pre-trained well logging basic model be... The parameters of the curve-sensitive readout module are: The parameters of the potential time-series state adaptation module are: The porosity prediction task output header parameters are: The parameter update boundary during the training phase is represented as: ; in, This represents the core parameters of the pre-trained well logging basic model; This indicates the parameters of the curve-sensitive readout module; Indicates the parameters of the potential timing state adaptation module; This represents the output header parameters of the porosity prediction task.

[0174] In this way, the present invention retains the general logging response prior of the pre-trained logging basic model while enabling the downstream adaptation module to learn more fully from high porosity samples in the interpreted depth range of the target well.

[0175] (5) Information boundaries during the training, correction, and prediction phases: To avoid the high porosity tail adaptation mechanism being misunderstood as using the true label of the interval to be predicted in the prediction stage, this invention further limits the information boundaries of the training, correction and prediction stages.

[0176] First, during the offline training phase, the depth range can be explained using training wells, validation wells, or target wells. The calculation of high porosity indicator variables using true porosity labels in the data. Tail sample loss weights and high porosity tail adaptation loss Based on this, the parameters of the curve-sensitive readout module, the potential timing state adaptation module, and the porosity prediction task output head are updated.

[0177] Second, during the target well calibration phase, the restricted offset calibration parameters, restricted affine calibration parameters, and calibration method gating parameters in step six are based only on the interpreted depth range of the target well. The true porosity label, and and The calculation of unlabeled logging response statistics is permitted; however, the following methods are prohibited: The true porosity label or the high porosity identifier obtained from the true porosity label.

[0178] Third, in the unexplained depth range to be predicted In the online prediction phase, the parameters of the curve-sensitive readout module, the potential temporal state adaptation module, and the porosity prediction task output head are fixed; the correction parameters and gating rules determined in step six are fixed; and the high porosity threshold is... and tail sample loss weights Fixed. Only input at this point. The system uses multi-curve logging data, curve availability information, potential logging baseline characterization, and baseline porosity prediction results to output the final porosity prediction result. It does not read or use... The true porosity label.

[0179] After the above processing, a porosity range extension prediction model trained with high porosity tail adaptation is obtained. The model's parameters remain fixed in the eighth prediction stage, receiving only... The system uses multi-curve logging data, curve availability information, potential logging basic characterization, and basic porosity prediction results, and outputs the final porosity prediction results through the constrained response state migration correction parameters determined in step six.

[0180] Figure 5 This paper demonstrates a process for identifying high-porosity samples based on real porosity labels in the explained depth range or training samples during the training phase, and enhancing the response retention capability of high-porosity reservoirs through tail-weighted loss.

[0181] Step 8: Model training and porosity range extension prediction.

[0182] This step utilizes the porosity range extension prediction model built and trained in steps one through seven, and the unexplained depth range to be predicted obtained in step one. Multi-curve logging data, along with the constrained response state migration correction parameters determined in step six, are used as inputs to... Outputs continuous porosity prediction results. The prediction stage does not use... The true porosity label in the data is not based on... Adjust model parameters, high porosity threshold, correction parameters, or gating rules based on the actual porosity label.

[0183] It should be noted that the unexplained depth range to be predicted... Not limited to areas within the explained depth range The subsequent depth range can also be located in Previously, located The interval between them, or other depth intervals in the target well that lack reliable porosity labels but have multi-curve logging data. In a preferred application scenario, Located along the well depth direction The subsequent depth range was not explained.

[0184] (1) Training sample construction: Based on the interpreted depth range of the target well Construct supervised training samples. For The depth location, its multi-curve logging input, curve availability mask, potential logging basic characterization, adapted potential logging characterization, and true porosity label can be used for model supervised training, validation, calibration parameter calculation, and model selection.

[0185] For the unexplained depth range of the target well It only uses its multi-curve logging input, curve availability mask, potential logging basic characterization, adapted potential logging characterization, basic porosity prediction results, and unlabeled response state statistics, without using... The true porosity label in the model is used for model training, parameter updates, response state correction, gating selection, high porosity threshold determination, or model selection.

[0186] The training sample set can be represented as: ; in, This represents the training sample set consisting of the explained depth range of the target well; Indicates the target well In depth Multi-curve logging input at the location; Indicates the target well In depth The actual porosity label at the location.

[0187] (2) Model training implementation parameters: Gradient descent-type optimization algorithms are used to train the curve-sensitive readout module, the latent temporal state adaptation module, and the porosity prediction task output head. The training loss function adopts the high-porosity tail adaptation loss described in step 7, or a combination of smoothing L1 loss, mean square error, mean absolute error, and high-porosity tail weighted loss.

[0188] In one specific implementation, the input sequence length is set to 32, the training batch size is set to 48, the number of training epochs is set to 80, and the learning rate is set to... The learning rate warm-up ratio is set to 0.05, the weight decay coefficient is set to 0.008, the gradient clipping threshold is set to 1.0, and the number of early stopping rounds is set to 16. The optimizer can be AdamW, and the loss function can be a combination of SmoothL1 loss and high porosity tail-weighted loss.

[0189] The training parameters described above are for one specific embodiment and do not constitute the sole limitation on the scope of protection of this invention. In other embodiments, the sequence length, batch size, number of training epochs, learning rate, weight decay coefficient, gradient pruning threshold, and early stopping strategy can be adjusted according to the training sample size, the number of logging curves, the number of target well labels, the model size, and computing resources.

[0190] (3) Determination of target well response state correction parameters: After the model training is completed, based on the explained depth range of the target well... Based on the true porosity label and the basic porosity prediction results obtained in step 5, the constrained offset correction parameters are calculated according to the method described in step 6. Constrained affine correction parameters , And determine the correction method and gating coefficient. .

[0191] The correction parameters and gating coefficients are based only on The true porosity label and the allowed unlabeled response state statistics are determined, without using... The true porosity label in the data.

[0192] Once the correction parameters are determined, in the unexplained depth range to be predicted... The prediction phase remains fixed and is no longer based on Adjust the actual porosity label.

[0193] (4) Unexplained depth interval prediction: During the prediction phase, the unexplained depth range to be predicted will be... The multi-curve logging data is input into the trained porosity interval extension prediction model, and the following processes are performed sequentially: First, according to the first step... Multi-curve logging data is processed for depth alignment, outlier cleaning, missing value identification, and scale normalization. Second, construct according to the second step. Depth window logging segments and multi-curve logging input token sequences; Third, following the third step, input the input token sequence into the frozen pre-trained well logging basic model to extract potential well logging basic representations; Fourth, calculate the curve fusion weights based on the potential well logging baseline characterization of each well logging curve in step four; Fifth, following the fifth step, construct a potential depth index sequence using the curve fusion weights and perform potential temporal state adaptation. Then, perform curve-sensitive readout on the adapted potential logging characterization to obtain the basic porosity prediction results. Sixth, following step six, use the determined constrained response state transition correction parameters to... The basic porosity prediction results are corrected to obtain the final porosity prediction results.

[0194] for The final porosity prediction result is expressed as follows: ; in, Indicates the target well In the unpredicted, unexplained depth range Inner depth Porosity prediction results at the location; This represents the final porosity prediction result after constrained response state migration correction.

[0195] Depend on The porosity prediction results at all depth locations constitute a continuous porosity prediction set: ; in, This indicates the unexplained depth range of the target well. The set of continuous porosity prediction results.

[0196] (5) Prediction results output: The output includes the uninterpreted depth range of the target well. The predicted porosity values ​​at various depths are provided. The prediction results can be output in the form of depth-porosity data tables, continuous porosity curves, reservoir evaluation maps, digital logging interpretation results, or model prediction files, to assist in reservoir property evaluation, favorable reservoir identification, reservoir stratification interpretation, and oil and gas development decisions.

[0197] In one implementation, curve fusion weights, curve availability indicators, basic porosity prediction results, constrained response state migration correction results, and low confidence indicators corresponding to each depth location can also be output simultaneously for subsequent manual verification and well logging interpretation quality control.

[0198] After the above processing, this invention realizes a complete technical process from pre-trained logging basic characterization to target well porosity range extension prediction. This process retains the general logging response priors of the pre-trained logging basic model, and achieves continuous porosity prediction of the uninterpreted depth range of the target well through curve-sensitive readout, potential time-series state adaptation, constrained response state migration correction, and high porosity tail adaptation.

[0199] This invention does not directly utilize the pre-trained well logging base model for frozen representation regression. Instead, it constructs a downstream adaptation structure on the frozen pre-trained well logging base representation for the target well porosity range extension task. This allows the pre-trained well logging base model to provide general well logging response priors, and the downstream module to complete the target well response state adaptation.

[0200] This invention does not simply splice, average, or model different logging curves as equivalent channels. Instead, it uses a curve-sensitive readout mechanism to generate porosity candidate prediction values ​​and curve contribution weights for each logging curve, and then performs adaptive fusion based on curve availability.

[0201] This invention does not rely solely on the logging response at a single depth point to predict porosity. Instead, it constructs a potential sequence of depth indexes in the potential logging-based characterization space and expresses the change in the response state of the target well along the depth direction through local depth morphology extraction and response state memory modeling.

[0202] This invention does not perform unconstrained trend fitting on the prediction results of the target well. Instead, it uses the real porosity labels of the interpreted depth range of the target well to calculate constrained offset correction parameters and constrained affine correction parameters, and controls the selection or fusion of correction methods based on the response state compatibility index that does not contain the real porosity labels of the uninterpreted depth range to be predicted.

[0203] This invention does not simply optimize the porosity prediction model with global average error, but introduces a high porosity tail adaptation mechanism during the training phase, assigning higher loss weights to high porosity samples in the training samples, in order to alleviate the problem that the response of high porosity reservoirs is compressed into medium porosity prediction results.

[0204] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the invention. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0205] It should be understood that although the steps in the flowcharts of the accompanying drawings are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some of the steps in the flowcharts of the accompanying drawings may include multiple steps or stages, which are not necessarily completed at the same time, but may be executed at different times, and the execution order of these steps or stages is not necessarily sequential, but may be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0206] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0207] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0208] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for predicting porosity range extension, characterized in that, The construction and training process of the porosity interval extension prediction model includes: After preprocessing the acquired multi-curve logging data of the target well and the true porosity labels of the interpreted depth range, a multi-curve logging input token sequence is constructed. Input the multi-curve logging input token sequence into the frozen pre-trained logging base model to extract the potential logging base representation for each depth location and each logging curve; The confidence scores of curves are calculated based on the basic characteristics of each potential well logging curve, and the curve confidence scores are normalized by combining the curve availability identifier to obtain the curve fusion weight. Based on the curve fusion weight, the potential basic characteristics of different well logging curves at the same depth location are weighted and fused to construct a potential sequence of depth index. Local morphological extraction and response state memory modeling are performed on the potential sequence of the depth index. The basic characterization of the potential logging is updated in a residual manner to obtain the adapted potential logging characterization. The adapted potential logging characterization is input into the curve sensitive readout mechanism to generate the porosity candidate prediction value and the latest curve confidence value corresponding to each logging curve. The latest curve confidence value is re-normalized in combination with the curve availability identifier to obtain the adapted curve fusion weight. The adapted curve fusion weight is then used to perform weighted fusion of the porosity candidate prediction values ​​to obtain the basic porosity prediction result. Based on the deviation between the true porosity label in the explained depth range and the basic porosity prediction result, the restricted offset correction result and the restricted affine correction result are calculated; the compatibility of the response state is judged according to the statistical characteristics of the unlabeled response state between the explained depth range and the unexplained depth range to be predicted, and the gated selection or weighted fusion between the restricted offset correction and the restricted affine correction is performed according to the response state compatibility to obtain the final porosity prediction result. During the training phase of the porosity range extension prediction model, samples with porosity higher than a set value are assigned a higher loss weight.

2. The porosity interval extension prediction method according to claim 1, characterized in that, After preprocessing the acquired multi-curve logging data of the target well and the true porosity labels of the interpreted depth range, a multi-curve logging input token sequence is constructed, specifically including: Depth window slicing is performed on the normalized multi-curve logging data of the target well to obtain logging segments corresponding to each depth location and each logging curve. ; through well logging fragment embedding function Generate well logging content embedding Unavailable or missing logging segments are embedded using a shared missing mask. Representation; construct a depth location embedding for each depth location. , The index of the t-th depth location; construct a curve type embedding for each logging curve. , The index of the c-th logging curve; combined into an input token. ,in, The curve availability identifier consists of a multi-curve logging input token sequence comprised of input tokens for all depth locations and all logging curves. , Indicates the number of depth positions; This indicates the number of logging curves.

3. The porosity interval extension prediction method according to claim 1, characterized in that, The process involves calculating curve confidence based on the fundamental characterization of each potential well log, and then normalizing the curve confidence using curve availability indicators to obtain curve fusion weights. Specifically, this includes: For the first The depth position Potential logging-based characterization of a single logging curve Through the curve confidence function And the confidence curve is generated by Sigmoid activation. Combined with curve availability indicators Normalize the confidence scores of the curves to obtain the curve fusion weights. , Indicates the first The effective confidence level of all available logging curves at each depth location and , Indicates the first At the depth position of the first Availability indicators for each logging curve; Indicates the first At the depth position of the first The confidence level of each logging curve; This indicates the number of logging curves.

4. The porosity interval extension prediction method according to claim 1, characterized in that, The process of weighted fusion of potential well logging fundamental characteristics of different well logging curves at the same depth location based on the curve fusion weights to construct a potential depth index sequence specifically includes: Based on curve fusion weights Potential logging-based characterization of different logging curves at the same depth location Weighted fusion is performed to obtain the latent representation of the deep index. The target well depth index potential sequence is composed of potential representations of depth indexes at continuous depth locations. , Indicates the number of depth positions.

5. The porosity interval extension prediction method according to claim 1, characterized in that, The process involves local morphological extraction and response state memory modeling of the potential depth index sequence, updating the underlying potential logging representation using residuals to obtain an adapted potential logging representation. This adapted representation is then input into a curve-sensitive readout mechanism to generate candidate porosity prediction values ​​and the latest curve confidence scores for each logging curve. The latest curve confidence scores are then re-normalized using a curve availability identifier to obtain adapted curve fusion weights. These weights are then used to weight and fuse the candidate porosity prediction values ​​to obtain the basic porosity prediction result. Specifically, this includes: For deep indexed latent sequences After layer normalization, the data is input into a one-dimensional depth convolution module to obtain local depth morphological features. , Indicates the first Local depth morphological features at a depth location Indicates the number of depth locations; indexes the latent sequence by depth. Input the bidirectional response state memory module to obtain the shallow-to-deep directional memory features. Memory features from deep to shallow After fusion, the response state memory features are obtained. ; local depth morphological features With response state memory features The residual fitting terms are obtained by low-rank projection after splicing. ; Indicates a low-rank projection module; This represents the trainable parameters of the low-rank projection module; through the gating coefficients. Controlling the injection intensity of residual adaptation terms to update the underlying logging fundamental characterization The potential logging characterization after adaptation was obtained. , This represents a learnable or preset residual scaling factor. This represents the Sigmoid activation function; The adapted potential logging characterization set The input curve-sensitive readout mechanism generates porosity candidate prediction values ​​and curve confidence scores for each adapted potential logging characterization. The curve confidence scores are then re-normalized using a curve availability identifier to obtain adapted curve fusion weights. These weights are then used to weight and fuse the porosity candidate prediction values ​​to obtain the basic porosity prediction result. ,in For curve-sensitive porosity readout function, This represents the trainable parameters in the curve-sensitive porosity readout function.

6. The porosity interval extension prediction method according to claim 1, characterized in that, The calculation process for the constrained offset correction result is as follows: Within the explained depth range Within this framework, the constrained offset correction parameter is calculated based on the median deviation between the actual porosity label and the basic porosity prediction result. And obtain the constrained offset correction results: ； This represents the boundary truncation function. Indicates the deviation scaling factor; This indicates the explained depth range of the target well. Calculate the median within the range; Indicates the target well In depth The actual porosity label at the location; Indicates the target well In depth Predicted results of basic porosity at the location; These represent the boundary values ​​of the offset correction parameters. Indicates the target well In depth The porosity prediction results after constrained offset correction.

7. The porosity interval extension prediction method according to claim 6, characterized in that, The calculation process for the restricted affine correction result is as follows: Within the explained depth range Within, the restricted affine correction result is: ； Target well In depth Porosity prediction results after constrained affine correction; Indicates the target well Proportional correction parameters; Indicates the target well The intercept correction parameters; Proportional correction parameters and intercept correction parameters satisfy: ； ； in, This represents the loss function for fitting the correction parameters; Indicates the explained depth range The number of valid samples in the sample; and These represent the lower and upper boundaries of the proportional correction parameter, respectively. This represents the absolute boundary value of the intercept correction parameter.

8. The porosity interval extension prediction method according to claim 1, characterized in that, The step of determining response state compatibility based on the statistical characteristics of unlabeled response states between the explained depth range and the unexplained depth range to be predicted specifically includes: Calculate the difference between the base forecast and the mean: ； Indicates the explained depth range With the unexplained depth range to be predicted The difference between the underlying predicted mean values; This represents the mean operation. This represents the mean operation. Indicates the target well In depth Predicted results of basic porosity at the location; Calculate the distance between potential representation centers: ； in, Indicates the explained depth range Unexplained depth range to be predicted Distance between potential representation centers To adapt to the depth index representation obtained from the subsequent potential logging characterization, Represents the L2 norm; Calculate the percentage of unlabeled tails supported: ； Indicates the unexplained depth range to be predicted. The unlabeled tail support ratio is obtained based on the basic porosity prediction results. Indicates an indicator function; This refers to the basic prediction tail judgment threshold determined based on training samples, explained depth ranges, or preset reservoir evaluation standards. When both conditions are met , , The system determines whether the explained and unexplained depth ranges have compatible response states; otherwise, it determines that there are differences in response states. This represents the threshold for the difference between the baseline predicted mean and the target mean. This represents the threshold for the distance from the potential representation center; This indicates the threshold for the proportion of unlabeled tails that are supported.

9. The porosity interval extension prediction method according to claim 1, characterized in that, The process of gating or weighted fusion between restricted offset correction and restricted affine correction based on the response state compatibility to obtain the final porosity prediction result specifically includes: Set the correction method selection coefficient. When it is determined that there is response state compatibility The constrained offset correction result is used; when a difference in response state is determined and the constrained affine correction does not trigger the backoff rule, The results of restricted affine correction are used; when a difference in response state is determined but the restricted affine correction triggers a backoff rule, The prediction was reversed to a restricted offset correction; the final prediction result is: ； in, Indicates the target well In depth Final porosity prediction results at the location; This represents the porosity prediction result after constrained affine correction. This represents the porosity prediction result after constrained offset correction.

10. The porosity interval extension prediction method according to claim 1, characterized in that, The method of assigning higher loss weights to samples with porosity higher than a set value specifically includes: Set a high porosity threshold For the true porosity labels in the training samples Define high porosity indicator variables ; This indicates an indicator function that takes the value when the condition within the parentheses is true. Otherwise, the value is ; Set the weights for the tail sample loss: ； Indicates the first Loss weights for each training sample; This indicates the loss weight for high-porosity samples; Indicates the first High porosity indicator variable for each training sample; Based on training batches High porosity tail adaptation loss ; This represents the smoothed L1 loss function; Indicates the first Basic porosity prediction results for each training sample; Indicates the first The true porosity label of each training sample; The core parameters of the pre-trained well logging basic model Keep it frozen during the training phase.

Images