A method for quantitatively evaluating formation water resistivity heterogeneity by well logging

Abstract

The application discloses a kind of well fusion stratum water resistivity heterogeneity quantitative evaluation method, it is related to petroleum exploration and development technical field.The implementation process is as follows: obtain the stratum signal data of various well logging acquisition to calculate reference stratum water resistivity;Based on the feature data extracted in conventional well logging, a machine learning model is constructed, and the stratum water resistivity is the target variable of the machine learning model.The stratum water resistivity prediction model is obtained by training the machine learning model;Collect conventional well logging data, extract the feature data in conventional well logging data, input the stratum water resistivity prediction model, and output the stratum resistivity prediction value;Stratum resistivity prediction value at multiple depths forms a continuous prediction curve as the stratum water resistivity heterogeneity quantitative evaluation result, solves the accurate evaluation question under the lack of nuclear magnetic resonance logging data.

Description

Technical Field

[0001] This invention relates to the field of petroleum exploration and development technology, and more specifically, to a method for quantitative evaluation of formation water resistivity heterogeneity using well logging integration. Background Technology

[0002] Formation water resistivity is a core parameter for oil and gas exploration and development. Its distribution directly reflects the heterogeneity of the formation and is crucial for key tasks such as reservoir evaluation and reserve calculation. Accurate quantitative characterization of heterogeneous formation water resistivity is a key challenge in well logging interpretation.

[0003] Existing methods for evaluating formation water resistivity have significant shortcomings: (1) Archie formula: Archie is a classic method for calculating formation water saturation. Its core is to establish the relationship between formation resistivity, porosity, and water saturation. It relies on core experiments to obtain rock electrical parameters, which is costly and has poor adaptability, making it difficult to cope with complex and heterogeneous formations; (2) Conventional logging interpretation method: It uses conventional logging curves (such as spontaneous potential, resistivity, sonic waves, density, etc.) to estimate formation water resistivity through empirical formulas or charts. It is easily affected by lithology and mud content. Interferences and limited accuracy; (3) Nuclear magnetic resonance logging and other technologies can directly measure the distribution and properties of pore fluids, providing more accurate key parameters such as bound water saturation and mobile fluids, thus providing high-quality basic data for calculating formation water resistivity. Although it can provide accurate basic data, it is costly and difficult to promote, and most wells to be evaluated lack this data; (4) Although deep learning has been initially applied to well logging interpretation, existing models mostly rely on conventional well logging data and do not integrate the advantages of unconventional well logging, and cannot solve the problem of accurate evaluation of wells lacking nuclear magnetic data. Summary of the Invention

[0004] The purpose of this application is to provide a method for quantitative evaluation of formation water resistivity heterogeneity by integrating well logging, which solves the problem of accurate evaluation when there is a lack of NMR data.

[0005] To solve the above-mentioned technical problems, the solution adopted in this application is as follows:

[0006] A method for quantitatively evaluating the heterogeneity of formation water resistivity using well logging integration is described below:

[0007] Feature data is extracted from formation signal data collected by conventional well logging and input into the formation water resistivity prediction model, and the predicted formation resistivity value is output; the predicted formation resistivity values ​​at multiple depths constitute a continuous prediction curve of formation water resistivity.

[0008] Calculate the gradient of the continuous prediction curve of formation water resistivity in the depth direction, and at the same time calculate the gradient of the input deep resistivity curve. Compare and analyze the two curves to determine whether they maintain physical and logical consistency in the main trend of change. If they are consistent, the geological and physical consistency check is passed.

[0009] The validated continuous prediction curve of formation water resistivity is used as the final quantitative evaluation result of formation water resistivity heterogeneity.

[0010] The input deep resistivity curve is obtained from formation signal data acquired through conventional well logging;

[0011] The process of constructing the formation water resistivity prediction model is as follows:

[0012] The baseline formation water resistivity was calculated based on formation signal data acquired from various well logging sources. Multiple logging methods are available, including conventional logging and nuclear magnetic resonance logging.

[0013] A machine learning model is constructed based on feature data extracted from formation signal data acquired through conventional well logging, with a benchmark formation water resistivity. The target variable of the machine learning model is used to train the machine learning model to obtain a formation water resistivity prediction model; the supervised dataset for training the machine learning model is the feature data extracted from formation signal data acquired by conventional well logging.

[0014] In some feasible implementation schemes, the feature data extracted from conventional well logging includes fine-scale feature data, mesoscale feature data, and coarse-scale feature data that characterize different dimensions of geological information;

[0015] The process of acquiring fine-scale feature data is as follows: read depth sampling points All conventional logging curve values ​​are used as fine-scale feature data;

[0016] The process of acquiring mesoscale feature data is as follows: based on the morphological abrupt change detection of conventional well logging on lithology-sensitive curves, small-layer units are divided, and depth sampling points are calculated. The statistical characteristics of each curve in the corresponding sub-layer are used as mesoscale feature data.

[0017] The process of obtaining coarse-scale feature data is as follows: based on the geological strata pre-divided by humans, depth sampling points are calculated. In some feasible implementation schemes, the global statistical analysis of each logging curve in the geological interval is used as coarse-scale characteristic data.

[0018] In some feasible implementation schemes, the specific process of constructing a machine learning model based on feature data extracted from formation signal data acquired through conventional well logging is as follows:

[0019] For fine-scale, medium-scale, and coarse-scale feature data, three structurally identical but parameter-independent sub-network branches are constructed respectively. Each sub-network branch consists of at least two fully connected layers, with ReLU activation function used for non-linear transformation between layers, and the outputs are fine-scale feature vectors. Mesoscale eigenvectors and coarse-scale eigenvectors ;

[0020] The fine-scale feature vectors of the output of the sub-network branches Mesoscale eigenvectors and coarse-scale eigenvectors The data is concatenated, input into the attention computation unit, and the attention computation unit outputs the fused features. ;

[0021]

[0022] In the formula, Represents fine-scale attention weights, Represents mesoscale attention weights, This represents coarse-scale attention weights; ;

[0023] Fusion features The input is fed into a fully connected output layer, directly mapped to the final predicted formation water resistivity value. .

[0024] In some feasible implementations, the process of obtaining a formation water resistivity prediction model from a machine learning model is as follows:

[0025] Construct a composite loss function to train the constrained model;

[0026] The backpropagation algorithm based on the adaptive optimizer uses a supervised learning dataset to iteratively train the network parameters of the machine learning model in order to minimize the composite loss function. When the composite loss function no longer decreases significantly over multiple consecutive training cycles on the validation set, training is stopped, and the formation water resistivity prediction model is obtained.

[0027] Composite loss function for:

[0028] ;

[0029] In the formula, Mean squared error loss: used to constrain the predicted formation water resistivity. Compared with the reference formation water resistivity The overall deviation between them ensures prediction accuracy; Gradient consistency loss is used to constrain the predicted formation water resistivity. The gradient of change in the vertical direction is made to maintain physical consistency with the gradient of change in the deep resistivity curve in terms of trend; To weigh the importance of the two losses, the hyperparameter is set to 0.1.

[0030] In some feasible implementations, a baseline formation water resistivity is calculated based on formation signal data acquired from various well logging sources. The implementation process is as follows:

[0031] The true resistivity of the formation is obtained from formation signal data acquired through conventional well logging.

[0032] Based on the formation signal data acquired by nuclear magnetic resonance logging, oil-water signal decoupling is performed using the nuclear magnetic multi-relaxation mechanism to calculate the oil saturation of the formation.

[0033] Based on Archie's formula, the baseline formation water resistivity is calculated by substituting the formation true resistivity, porosity, rock electrical parameters, and oil saturation. :

[0034] ;

[0035] In the formula, Porosity, % Oil saturation, % The true resistivity of the formation is given in Ω·m. A coefficient related to lithology; The cementation index; It is the saturation index;

[0036] The true resistivity of the formation is obtained through conventional well logging curves; lithology-related coefficients. Cementation index and saturation index It was obtained through core experiments.

[0037] In some feasible implementations, the supervised dataset is obtained as follows:

[0038] From the set of conventional well logging curves, the subset of curves with the strongest interpretability of formation water resistivity was selected.

[0039] Based on the selected optimal combination of logging curves, sampling points at each depth... Extract fine-scale feature data, meso-scale feature data, and coarse-scale feature data;

[0040] The dataset is linearly normalized using the minimum-maximum scaling method for fine-scale, medium-scale, and coarse-scale feature data.

[0041] Normalized feature data Compared with the corresponding baseline formation resistivity prediction value Pairing is performed to construct the final supervised learning dataset. .

[0042] In some feasible implementation schemes, the process of selecting the subset of curves with the strongest interpretability of formation water resistivity from the set of conventional well logging curves is as follows:

[0043] Different curve combinations are generated based on the curve set acquired by conventional well logging; the curve set includes: gamma ray (GR) curve, deep resistivity (RT) curve, acoustic transit time (AC) curve, density (DEN) curve, and neutron porosity (CNL) curve.

[0044] Based on the same depth measurement, a multiple linear regression model is constructed with the curve response value of each curve combination as the independent variable and the reference formation water resistivity as the dependent variable.

[0045] Calculate the coefficient of determination of the multiple linear regression model corresponding to each curve combination, and select the curve combination with the highest coefficient of determination as the subset of curves with the strongest explanatory power for formation water resistivity.

[0046] In some feasible implementation schemes, the validated continuous prediction curve is used as the final quantitative evaluation result of formation water resistivity heterogeneity, and the Archie formula is used for inversion verification and accuracy assessment based on measured oil saturation data.

[0047] In some feasible implementations, the process of calculating oil saturation is as follows:

[0048] Determining the effective porosity of formations Cutoff value, obtain greater than The cutoff value is a fluid signal, where fluid includes recoverable oil and / or water;

[0049] ;

[0050] In the formula, Indicates temporary Cutoff value, which is from arrive Dynamic changes cause the measured porosity to differ from the effective porosity. The mean square error is minimized at this point. That is, the effective porosity Cutoff value ; express Distribution function;

[0051] Determining oil and water Determine the cutoff value and calculate the oil-bearing porosity;

[0052] ;

[0053] In the formula, Indicates oil-bearing porosity; The lower limit of integration represents the oil-water content. The cutoff value was obtained through experimental calibration. Indicates the maximum value of the points; express Distribution function;

[0054] Calculate the oil saturation of the formation ;

[0055] .

[0056] The technical solution of this application has at least the following advantages and beneficial effects:

[0057] This invention provides a method for quantitatively evaluating the heterogeneity of formation water resistivity using fusion logging data. It employs machine learning with multiple fusion logging datasets, combining the strong fitting and generalization prediction capabilities of deep neural networks to achieve quantitative evaluation of formation water resistivity heterogeneity. First, the oil saturation of a benchmark well is obtained through decoupling inversion of nuclear magnetic resonance logging. Second, a high-information-density, multi-scale training dataset is constructed based on this. Then, a multi-branch attention-based deep network with geological cognitive capabilities is designed for learning. Finally, accurate and high-resolution prediction of the heterogeneous distribution of formation water resistivity in the well to be evaluated is achieved.

[0058] This invention uses the structure output by the established formation water resistivity prediction model to quantitatively evaluate the heterogeneity of formation water resistivity. The input of this model only requires conventional well logging data, thus solving the problem of accurate evaluation in the absence of nuclear magnetic resonance well logging data. Attached Figure Description

[0059] When considered in conjunction with the accompanying drawings, the invention can be more fully and better understood by referring to the following detailed description, and the scientific and practical nature of the invention can be better illustrated. However, the accompanying drawings, which are provided to provide a further understanding of the invention and constitute a part of this invention, are used to explain the invention and do not constitute an undue limitation of the invention.

[0060] Figure 1 The flowchart below illustrates the method for quantitatively evaluating the heterogeneity of formation water resistivity in this embodiment.

[0061] Figure 2 In the examples, the determining coefficients for different curve combinations are shown.

[0062] Figure 3 In this example, the evaluation results of formation water resistivity distribution characteristics and oil saturation were verified. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0064] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0065] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0066] This invention discloses a method for quantitatively evaluating the heterogeneity of formation water resistivity by integrating well logging data. The overall idea is as follows: a supervised learning dataset is constructed by fusing multiple well logging data, and a branch attention network model is constructed based on conventional well logging data. The network model is trained using the supervised learning dataset, and finally a formation water resistivity heterogeneity evaluation model is established to evaluate the formation water resistivity, thereby enabling the evaluation to be obtained by collecting conventional well logging data.

[0067] Example

[0068] Please refer to Figure 1 This invention provides a method for quantitatively evaluating the heterogeneity of formation water resistivity using well logging integration, the implementation process of which is as follows:

[0069] Step S100: Core formation signals of the target block are acquired by nuclear magnetic resonance logging. Oil and water signals are decoupled based on the nuclear magnetic multi-relaxation mechanism, and the oil saturation of the formation is calculated.

[0070] The specific implementation process of this step is as follows:

[0071] Step S101: Determine the T2 cutoff value of the effective porosity of the formation, obtaining a value greater than [value missing]. The cutoff value is a fluid signal, where fluid includes recoverable oil and / or water;

[0072] ;

[0073] In the formula, Indicates temporary Cutoff value, which is from arrive Dynamic changes cause the measured porosity to differ from the effective porosity. The mean square error is minimized at this point. That is, the effective porosity Cutoff value ; express Distribution function.

[0074] Due to nuclear magnetic resonance In the spectrum, the signal comes from both movable and bound fluids. The effective pore usually refers to the pore occupied by the movable fluid, in which case it is necessary to find a... Cutoff value, will The spectrum is divided into two parts: the fluid smaller than the value is bound fluid (water trapped in tiny pores by capillary forces or adsorption forces), and the fluid larger than the value is mobile fluid (oil and / or water that exists in larger pores and can flow freely). The mobile fluid in the formation pores can be distinguished through step S101 for the next step of oil-water separation.

[0075] Step S102: Determine the oil-water ratio Determine the cutoff value and calculate the oil-bearing porosity;

[0076] ;

[0077] In the formula, Indicates oil-bearing porosity; The lower limit of integration represents the oil-water content. The cutoff value was obtained through experimental calibration. Indicates the maximum value of the points; express Distribution function.

[0078] In this step, the oil and water content in the fluid is determined. The cutoff value depends on laboratory core calibration, specifically:

[0079] Core samples from the formation to be evaluated were obtained and saturated with known oil-water ratios in a laboratory to simulate formation conditions. 2D NMR spectroscopy was performed on the samples. Due to the significant difference in diffusion coefficients between oil and water, their signals were clearly separated in the 2D spectrum. Analysis of the 2D spectrum revealed the specific oil-water ratios. The oil and water signals are labeled in the dimension to distinguish them. Cutoff value;

[0080] Step S103: Calculate the oil saturation of the formation. ;

[0081] .

[0082] Step S200: Based on Archie's formula, substitute resistivity, porosity, rock electrical parameters, and oil saturation to calculate the reference formation water resistivity. ;

[0083] ;

[0084] In the formula, Porosity, % Oil saturation, % The true resistivity of the formation is given in Ω·m. A coefficient related to lithology; The cementation index; This is the saturation index.

[0085] In some feasible implementations, lithology-related coefficients Cementation index and saturation index It was obtained through core experiments. The true resistivity of the formation was obtained by inverting the apparent resistivity obtained from multi-depth array resistivity scanning, which is a conventional existing method and will not be described in detail in this embodiment.

[0086] In step S200, N depth sampling points are set, and corresponding parameters are collected to finally obtain the reference formation water resistivity of each depth segment in the entire well section of the target well. The value serves as the baseline label for subsequent machine learning model training.

[0087] Step S300: Process the conventional well logging data to construct a supervised dataset for machine learning training.

[0088] The specific implementation process of this step is as follows:

[0089] Step S301: Select the subset of curves from the set of conventional well logging curves that has the strongest interpretability of formation water resistivity.

[0090] The specific implementation process is as follows:

[0091] Different curve combinations are generated based on the curve set acquired by conventional well logging as curve subsets; the curve set includes: gamma ray (GR) curve, deep resistivity (RT) curve, acoustic transit time (AC) curve, density (DEN) curve, neutron porosity (CNL) curve, and porosity (Ф) curve.

[0092] Based on measurements at the same depth, the curve response values ​​of each curve combination are used as independent variables, along with the reference formation water resistivity. Using the dependent variable, construct a multiple linear regression model;

[0093] Calculate the coefficient of determination of the multiple linear regression model for each curve combination. The ability of quantitative curve combinations to explain the overall changes in formation water resistivity;

[0094] Selecting the coefficient of determination The highest curve combination is taken as the optimal logging curve combination and used as a feature for machine learning.

[0095] This embodiment applies the multiple linear regression model in statistics to the selection of curve combinations, thereby completing feature dimensionality reduction based on the explanatory power of the combination, finding the most effective features and eliminating redundancy, thus completing feature dimensionality reduction based on the explanatory power of the combination, and the selected optimal combination constitutes the feature index evaluation system of this block.

[0096] Please refer to Figure 2 It can be seen that the determining factor of the combination of curves consisting of the gamma ray (GR) curve, deep resistivity (RT) curve, acoustic transit time (AC) curve, density (DEN) curve, and neutron porosity (CNL) curve is... The maximum value is found in the combination of logging curves, making it the optimal combination.

[0097] Step S302: Based on the selected optimal logging curve combination, perform sampling at each depth Z point. i Extract feature data at three scales.

[0098] The specific implementation process of step S302 is as follows: extract fine-scale feature data, mesoscale feature data, and coarse-scale feature data;

[0099] Fine-scale: Directly reading depth sampling points All conventional logging curve values;

[0100] Mesoscale: Based on the detection of morphological abrupt changes in conventional logging curves that are sensitive to lithology, the system automatically divides the data into small-scale units and calculates the statistical characteristics of each curve in the small-scale unit where the depth sampling point Zi is located, such as mean, standard deviation, skewness, and kurtosis.

[0101] In some feasible embodiments, the method for automatically dividing into sub-layers is as follows: Select a pair of curves and calculate their first or second derivative. Set a threshold; when the absolute value of the derivative exceeds the threshold, it is considered that a lithological abrupt change has occurred at that point, thus dividing the continuous segment between two abrupt change points into a relatively homogeneous sub-layer. Alternatively, methods such as sliding window analysis of variance and abrupt change point detection algorithms can be used. For a divided sub-layer (containing multiple continuous depth points), calculate the statistical characteristics of each logging curve within it.

[0102] Coarse scale: Global statistics of all curves within the entire geological stratum.

[0103] Specifically, the entire geological stratum usually refers to the geological stratum determined in advance by geologists based on regional geological knowledge, seismic data, and well logging responses, and calculates the global statistics (such as mean, maximum, minimum, range, etc.) of each well logging curve in the geological stratum where the depth sampling point Zi is located.

[0104] It is worth noting that fine-scale feature data (single sampling point), mesoscale feature data (single lithological layer), and coarse-scale feature data (geological segment) reflect geological characteristics in different dimensions.

[0105] Step S303: Apply the minimum-maximum scaling method to linearly normalize the dataset using the feature data at the three scales:

[0106]

[0107] in, These are the original eigenvalues. For the normalized result, These are the maximum and minimum values ​​of the feature observed throughout the entire training sample set, respectively.

[0108] Step S304: Compare the normalized feature data with the corresponding... Value pairing is used to construct the final supervised learning dataset. .

[0109] Step S400: Construct a multi-branch attention network model based on feature data at three scales, train the network model using a composite loss function constraint, and establish an evaluation model for the heterogeneity of formation water resistivity.

[0110] In step S400, the specific implementation process of constructing a multi-branch attention network model based on feature data at three scales is as follows:

[0111] Step S411: For the fine-scale, mesoscale, and coarse-scale feature data representing different dimensional information of the strata, three sub-network branches with identical structures but independent parameters are constructed respectively. Each sub-network branch consists of two fully connected layers, with ReLU activation function used for nonlinear transformation between layers, thereby mapping the features of the corresponding scale input to a high-dimensional abstract feature vector, and outputting fine-scale feature vectors respectively. Mesoscale eigenvectors and coarse-scale eigenvectors .

[0112] Specifically, the first layer maps the input features to a 64-dimensional hidden space, and the second layer further abstracts them into 32-dimensional feature vectors.

[0113] Step S412: Extract the fine-scale feature vectors from the outputs of the sub-network branches. Mesoscale eigenvectors and coarse-scale eigenvectors The data is concatenated, input into the attention computation unit, and the attention computation unit outputs the fused features. ;

[0114]

[0115] Specifically, the attention computation unit generates a set of normalized attention weights through a fully connected layer and a Softmax activation function, namely: fine-scale attention weights. Mesoscale attention weights Coarse-scale attention weights Attention weights satisfy: The weights dynamically represent the relative importance of features at different scales to the final prediction in the current sample.

[0116] The branch outputs are weighted and summed using these weights to obtain the fused comprehensive feature representation. :

[0117]

[0118] Step S413: Merge features The input is fed into a fully connected output layer, directly mapped to the final predicted formation water resistivity value. .

[0119] In step S400, the network model is trained using a composite loss function constraint. The specific implementation process for establishing the formation water resistivity heterogeneity evaluation model is as follows:

[0120] Step S421: To ensure the constraint model possesses both high accuracy and geological plausibility, construct a composite loss function. :

[0121] ;

[0122] In the formula, Mean squared error loss: used to constrain the predicted formation water resistivity. Compared with the benchmark value The overall deviation between them ensures prediction accuracy; Gradient consistency loss is used to constrain the predicted formation water resistivity. The gradient of change in the vertical direction is made to maintain physical consistency with the gradient of change in the deep resistivity curve in terms of trend; To weigh the importance of the two losses, the hyperparameter is set to 0.1.

[0123] Step S422: Using the supervised learning dataset described above, the network parameters of the network model are iteratively trained using a backpropagation algorithm based on an adaptive optimizer to minimize the composite loss function. When the composite loss function no longer decreases significantly over multiple consecutive training cycles on the validation set, training is stopped, resulting in an evaluation model for the heterogeneity of formation water resistivity.

[0124] In this embodiment, a formation water resistivity heterogeneity evaluation model is established by constructing a three-branch deep network. This network first sets up a parallel branch structure to learn logging features at different scales, and then designs a cross-branch attention module to adaptively calculate and fuse the weights of the features of each branch. Based on this, a composite loss function combining mean square error and a gradient penalty term for heterogeneity based on depth variation is defined. Finally, the network is trained with this loss function as the optimization objective to establish a deep learning model that can quantitatively evaluate the heterogeneity of formation water resistivity.

[0125] Step S500: Input the dataset of characteristic indicators of the well to be evaluated, predict and verify it through the formation water resistivity heterogeneity evaluation model, and output the verified formation water resistivity.

[0126] The specific implementation of step S500 is as follows:

[0127] Step S501: Process the conventional logging data of the well to be evaluated, extract fine-scale feature data, meso-scale feature data, and coarse-scale feature data, and perform normalization processing to obtain feature data. Input the feature data into the trained formation water resistivity heterogeneity evaluation model to obtain the formation water resistivity continuous prediction curve.

[0128] Step S502: Calculate the gradient of the continuous prediction curve in the depth direction (first-order difference), and at the same time calculate the gradient of the input deep resistivity RT curve. Compare and analyze the two gradient curves to determine whether they maintain physical and logical consistency in the main trend of change. If they are consistent, the geological and physical consistency check is passed.

[0129] In this step, determining whether the two maintain physical logic consistency in their main trends of change mainly refers to determining whether there is a systematic contradiction in their trends that contradicts the basic physical relationship of Archie's formula. For example, in suspected oil and gas strata where the deep resistivity RT increases significantly, there should not be a synchronous significant increase in formation water resistivity.

[0130] Step S503: The validated continuous prediction curve is used as the final quantitative evaluation result of formation water resistivity heterogeneity. Based on the measured oil saturation data, the Archie formula is used for inversion verification and accuracy evaluation.

[0131] Please refer to Figure 3 This is a depth profile of an oil well, which uses conventional logging techniques to collect natural gamma curves, density curves, neutron porosity curves, sonic transit time curves, and deep resistivity curves.

[0132] Based on the formation water resistivity heterogeneity evaluation model trained in the example, the predicted values ​​of formation water resistivity and the corresponding continuous prediction curves were output. The predicted values ​​of formation water resistivity were then substituted into Archie's formula to calculate the oil saturation prediction curve.

[0133] from Figure 3 As can be seen from the data, the oil saturation prediction curve and the core experimental data of oil saturation show a good trend and similar values, indicating that the model described in the embodiment accurately predicts the formation water resistivity. At the same time, the oil saturation prediction curve and the core measured data are highly consistent, which means that the present invention can obtain oil saturation with the same accuracy as core analysis using only conventional well logging data, and realizes accurate data acquisition under nuclear magnetic resonance-free conditions.

[0134] Furthermore, comparing the calculated formation water resistivity with the predicted formation water resistivity output by the model, it can be seen that the two are in good agreement, indicating the reliability of the prediction model.

[0135] In this invention, the predicted values ​​of formation water resistivity and oil saturation are consistent with the corresponding values ​​obtained through experimental data. This indicates that, starting from conventional logging data, the model intelligently calculates and outputs a prediction of formation water properties that conforms to physical laws, and ultimately calculates an oil saturation that can withstand core testing.

[0136] The various embodiments of the present invention have now been described in detail. To avoid obscuring the concept of the invention, some details known in the art have not been described. Those skilled in the art will fully understand how to implement the technical solutions of this invention based on the above description, and the scope of the invention is defined by the appended claims.

Claims

1. A method for quantitatively evaluating the heterogeneity of formation water resistivity using well logging integration, characterized in that, The implementation process is as follows: Feature data is extracted from formation signal data collected by conventional well logging and input into the formation water resistivity prediction model, and the predicted formation resistivity value is output; the predicted formation resistivity values ​​at multiple depths constitute a continuous prediction curve of formation water resistivity. Calculate the gradient of the continuous prediction curve of formation water resistivity in the depth direction, and at the same time calculate the gradient of the input deep resistivity curve. Compare and analyze the two curves to determine whether they maintain physical and logical consistency in the main trend of change. If they are consistent, the geological and physical consistency check is passed. The validated continuous prediction curve of formation water resistivity is used as the final quantitative evaluation result of formation water resistivity heterogeneity. The input deep resistivity curve is obtained from formation signal data acquired through conventional well logging; The process of constructing the formation water resistivity prediction model is as follows: The baseline formation water resistivity was calculated based on formation signal data acquired from various well logging sources. Multiple logging methods are available, including conventional logging and nuclear magnetic resonance logging. A machine learning model is constructed based on feature data extracted from formation signal data acquired through conventional well logging, with a benchmark formation water resistivity. The target variable of this machine learning model is used to train the machine learning model to obtain a formation water resistivity prediction model; the supervised dataset for training the machine learning model is the feature data extracted from formation signal data acquired by conventional well logging. The feature data extracted from conventional well logging includes fine-scale feature data, meso-scale feature data, and coarse-scale feature data that characterize different dimensions of geological information. The specific implementation process of building a machine learning model based on feature data extracted from formation signal data acquired through conventional well logging is as follows: For fine-scale, medium-scale, and coarse-scale feature data, three structurally identical but parameter-independent sub-network branches are constructed respectively. Each sub-network branch consists of at least two fully connected layers, with non-linear transformations performed between layers, and the outputs are fine-scale feature vectors. Mesoscale eigenvectors and coarse-scale eigenvectors ; The fine-scale feature vectors of the output of the sub-network branches Mesoscale eigenvectors and coarse-scale eigenvectors The data is concatenated, input into the attention computation unit, and the attention computation unit outputs the fused features. ; ； In the formula, Represents fine-scale attention weights, Represents mesoscale attention weights, This represents coarse-scale attention weights; ; Fusion features The input is fed into a fully connected output layer, directly mapped to the final predicted formation water resistivity value. ; The process of training a machine learning model to obtain a formation water resistivity prediction model is as follows: Construct a composite loss function to train the constrained model; The backpropagation algorithm based on the adaptive optimizer uses a supervised learning dataset to iteratively train the network parameters of the machine learning model in order to minimize the composite loss function. When the composite loss function no longer decreases significantly over multiple consecutive training cycles on the validation set, training is stopped, and the formation water resistivity prediction model is obtained. Composite loss function for: ； In the formula, Mean squared error loss: used to constrain the predicted formation water resistivity. Compared with the reference formation water resistivity The overall deviation between them ensures prediction accuracy; Gradient consistency loss is used to constrain the predicted formation water resistivity. The gradient of change in the vertical direction is made to maintain physical consistency with the gradient of change in the deep resistivity curve in terms of trend; To weigh the importance of the two losses, the hyperparameter is set to 0.

1.

2. The method for quantitatively evaluating the heterogeneity of formation water resistivity using well logging integration according to claim 1, characterized in that, The process of acquiring fine-scale feature data is as follows: read depth sampling points Conventional well logging curve values ​​are used as fine-scale feature data; The process of acquiring mesoscale feature data is as follows: based on the morphological abrupt change detection of conventional logging curves sensitive to lithology, small-layer units are divided, and depth sampling points are calculated. The statistical characteristics of each logging curve in the sub-layer unit are used as mesoscale characteristic data; The process of obtaining coarse-scale feature data is as follows: based on the geological strata pre-divided by humans, depth sampling points are calculated. The global statistics of each logging curve in the geological interval are used as coarse-scale characteristic data.

3. The method for quantitative evaluation of formation water resistivity heterogeneity using well logging integration according to claim 2, characterized in that, The baseline formation water resistivity was calculated based on formation signal data acquired from various well logging sources. The implementation process is as follows: The true resistivity of the formation is obtained from formation signal data acquired through conventional well logging. Based on the formation signal data acquired by nuclear magnetic resonance logging, oil-water signal decoupling is performed using the nuclear magnetic multi-relaxation mechanism to calculate the oil saturation of the formation. Based on Archie's formula, the baseline formation water resistivity is calculated by substituting the formation true resistivity, porosity, rock electrical parameters, and oil saturation. : ； In the formula, Porosity, % Oil saturation, % The true resistivity of the formation is given in Ω·m. A coefficient related to lithology; The cementation index; It is the saturation index; The above lithology-related coefficients Cementation index and saturation index It was obtained through core experiments.

4. The method for quantitative evaluation of formation water resistivity heterogeneity using well logging integration according to claim 2, characterized in that, The supervised dataset is obtained as follows: From the set of conventional well logging curves, the subset of curves with the strongest interpretability of formation water resistivity was selected. Based on the selected optimal combination of logging curves, sampling points at each depth... Extract fine-scale feature data, meso-scale feature data, and coarse-scale feature data; The dataset is linearly normalized using the minimum-maximum scaling method for fine-scale, medium-scale, and coarse-scale feature data. Normalized feature data Compared with the corresponding baseline formation resistivity prediction value Pairing is performed to construct the final supervised learning dataset. .

5. The method for quantitative evaluation of formation water resistivity heterogeneity using well logging integration according to claim 4, characterized in that, The process of selecting the subset of curves with the strongest interpretability for formation water resistivity from the set of conventional well logging curves is as follows: Different curve combinations are generated based on the curve set acquired from conventional well logging; the curve set includes: gamma curve, deep resistivity curve, sonic transit time curve, density curve, and neutron porosity curve; Based on the same depth measurement, a multiple linear regression model is constructed with the curve response value of each curve combination as the independent variable and the reference formation water resistivity as the dependent variable. Calculate the coefficient of determination of the multiple linear regression model corresponding to each curve combination, and select the curve combination with the highest coefficient of determination as the subset of curves with the strongest explanatory power for formation water resistivity.

6. The method for quantitatively evaluating the heterogeneity of formation water resistivity using well logging integration according to claim 1, characterized in that, The validated continuous prediction curve is used as the final quantitative evaluation result of formation water resistivity heterogeneity. Based on the measured oil saturation data, the Archie formula is used for inversion verification and accuracy evaluation.

7. The method for quantitative evaluation of formation water resistivity heterogeneity using well logging integration according to claim 1, characterized in that, The process for calculating oil saturation is as follows: Determining the effective porosity of formations Cutoff value, obtain greater than The cutoff value is a fluid signal, where fluid includes recoverable oil and / or water; ； In the formula, Indicates temporary Cutoff value, which is from arrive Dynamic changes cause the measured porosity to differ from the effective porosity. The mean square error is minimized at this point. That is, the effective porosity Cutoff value ; express Distribution function; Determining oil and water Determine the cutoff value and calculate the oil-bearing porosity; ； In the formula, Indicates oil-bearing porosity; The lower limit of integration represents the oil-water content. The cutoff value was obtained through experimental calibration. Indicates the maximum value of the points; express Distribution function; Calculate the oil saturation of the formation ; 。

Images