Shale gas content prediction method based on logging curve
By combining CEEMDAN decomposition and a bi-branch model, the problem of separating long-scale and short-scale information in well logging signals was solved, achieving stability and accuracy in gas content prediction, and making it suitable for gas content prediction in complex reservoirs.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-12
AI Technical Summary
Existing technologies struggle to effectively separate long-scale information reflecting the regional geological background from mid-to-high-frequency information characterizing the local structure of thin layers and laminae in well logging signals. This results in feature aliasing and unclear physical meaning, neglects geological scale constraints and vertical continuity, and limits the model's generalization ability.
The fully ensemble empirical mode decomposition method (CEEMDAN) is used to perform multi-scale adaptive decomposition of well logging curves, construct a two-branch gas content prediction model, which respectively characterizes the overall evolution trend and local variation characteristics of gas content. Geological constraints are introduced through supervised training, and a joint loss function is constructed to improve prediction stability and accuracy.
It achieves effective separation of geological background information and local heterogeneous information, improves the physical rationality and geological interpretability of logging characteristics, enhances the stability and stratigraphic consistency of prediction results, and is suitable for gas content prediction in complex reservoirs.
Smart Images
Figure CN122014239A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of well logging technology, and in particular to a method for predicting shale gas content based on well logging curves. Background Technology
[0002] Gas content is a key parameter in the evaluation of unconventional oil and gas reservoirs and the selection of favorable areas. Its accurate prediction is of great significance for the detailed characterization of reservoirs and the formulation of development plans. Existing gas content prediction methods mainly rely on core experimental analysis or statistical regression and machine learning models based on conventional logging curves. However, due to the scarcity of core data, the strong non-stationarity of logging signals, and the significant heterogeneity of reservoirs, it is difficult to balance prediction accuracy and stability.
[0003] Traditional well logging curve processing methods often employ fixed-scale filtering or empirical feature extraction, which struggle to effectively separate long-scale information reflecting the regional geological background from mid-to-high-frequency information characterizing local structures such as thin layers and laminae. This can easily lead to feature aliasing and unclear physical meaning. In recent years, although some data-driven models have achieved certain results in gas content prediction, they often neglect geological scale constraints and vertical continuity, easily generating non-physical fluctuations and limiting model generalization ability.
[0004] Therefore, there is an urgent need for a gas content prediction method that can fully exploit multi-scale information from well logging curves, take into account geological background and local heterogeneous characteristics, and possess good stability and interpretability, in order to meet the practical needs of fine evaluation of complex reservoirs. Summary of the Invention
[0005] This invention aims to provide a method for predicting shale gas content based on well logging curves. It addresses the problems in existing technologies, such as the difficulty in effectively separating long-scale information reflecting the regional geological background from mid-to-high-frequency information describing the local structure of thin layers and laminae in well logging signals, which easily leads to feature aliasing and unclear physical meaning, easy neglect of geological scale constraints and vertical continuity, easy generation of non-physical fluctuations, and limited model generalization ability.
[0006] To achieve the above objectives, the present invention provides the following method:
[0007] This invention provides a method for predicting shale gas content based on well logging curves:
[0008] S1: Obtain at least one conventional logging curve of the target well and the measured gas content data of the corresponding well section, perform depth registration, and preprocess the conventional logging curve;
[0009] S2: The preprocessed conventional logging curves are decomposed sequentially along the depth direction using the complete set empirical mode decomposition method to obtain several intrinsic mode function (IMF) components.
[0010] S3: Calculate the energy percentage of the intrinsic mode function (IMF) components, retain the effective IMF components with energy greater than a preset threshold, and remove the noise components with the lowest energy percentage.
[0011] S4: The two lowest-frequency effective IMF components in the effective IMF components are superimposed with the residual to construct a low-frequency background component that reflects the overall trend of formation change; the remaining effective IMF components are linearly superimposed and absolute value superimposed respectively to obtain medium- and high-frequency modulation components, which are used to characterize the local physical property change characteristics and laminar development intensity inside the reservoir.
[0012] S5: Calculate the correlation coefficients for the acquired low-frequency background components and mid-to-high frequency modulation components respectively, and remove feature components that are highly similar to each other.
[0013] S6: Based on the measured gas content curve, select the corresponding background window thickness according to the layer scale, and perform sliding window averaging on the measured gas content along the depth direction to obtain the background gas content target curve that characterizes the long-scale variation characteristics, which serves as the supervision target and geological constraint of the model background branch.
[0014] S7: Based on the preprocessed background feature input and mid-to-high frequency modulation feature input, a dual-branch gas content prediction model is constructed. The dual-branch gas content prediction model includes a background branch and a modulation branch. The background branch is used to establish the mapping relationship between the logging background features and the gas content background components, and outputs the non-negative constrained background gas content prediction value. The modulation branch takes the mid-to-high frequency modulation features as input and introduces the background gas content prediction value as a condition constraint to characterize the gas content residual prediction value caused by local reservoir changes and laminar flow development. The background gas content prediction value and the residual prediction value are superimposed to obtain the gas content prediction result of the target well section, thereby realizing the collaborative modeling of gas content background trends and local change characteristics.
[0015] S8: The dual-branch gas content prediction model is trained under supervision based on measured gas content data. During the training process, the main prediction loss, background gas content constraint loss, and residual supervision loss are introduced simultaneously. Through the above multi-objective joint optimization and constraint training, a joint loss function is constructed to obtain a stable and geologically significant dual-branch gas content prediction model.
[0016] Preferably, the preprocessing step for the conventional logging curves includes: uniformly performing quantile truncation processing on the conventional logging curves to remove outliers, and using wavelet denoising to suppress high-frequency noise interference and improve the signal-to-noise ratio of the curves; setting a sampling interval, and resampling and interpolating all the conventional logging curves along the depth direction so that each curve is uniformly mapped to the same depth sampling grid.
[0017] Preferably, the step of using a complete set empirical mode decomposition method to sequentially decompose each preprocessed conventional logging curve along the depth direction to obtain several intrinsic mode function (IMF) components includes:
[0018] Multiple sets of amplitude-controlled Gaussian white noise are applied to the preprocessed conventional logging curve signal to form a noise-assisted signal set:
[0019] (1)
[0020] In the formula, Here, x(z) is the noise-assisted signal, x(z) is the conventional logging curve signal, and z is the depth. The white noise sequence added for the kth time;
[0021] For each of the noise-assisted signals, empirical mode decomposition is performed to extract its first-order intrinsic mode function (EMF). The first-order CEEMDAN EEMDAN EEMD component is then obtained by ensemble averaging of all experimental results. Its expression is:
[0022] (2)
[0023] Based on this, the first-order intrinsic mode component is removed from the original conventional logging curve signal to obtain the remaining signal:
[0024] (3)
[0025] in, The remaining signal, and then the remaining signal Repeat the above noise-assisted decomposition and ensemble averaging process to extract the i-th order intrinsic mode components step by step:
[0026] (4)
[0027] And update the remaining signals:
[0028] (5)
[0029] This continues until the remaining signal no longer satisfies the intrinsic mode function criterion or exhibits monotonical changes; ultimately, the conventional logging curve signal is represented as the sum of several intrinsic mode components and the residual term, i.e.:
[0030] (6).
[0031] Preferably, the formula for calculating the energy proportion of the intrinsic mode function (IMF) components is as follows:
[0032] (7)
[0033] in, Let be the sum of squares of the i-th eigenmode components along the depth direction, and let represent the energy of the i-th eigenmode component. The sum of squares of the corresponding logging curve along the depth direction represents the total energy of the logging curve. When its ratio is less than a preset threshold, the IMF component is regarded as an invalid noise component.
[0034] Preferably, the step of constructing a dual-branch gas content prediction model based on preprocessed background feature input and mid-to-high frequency modulation feature input, wherein the dual-branch gas content prediction model includes a background branch and a modulation branch, the background branch is used to establish the mapping relationship between well logging background features and gas content background components, and outputs non-negative constrained background gas content prediction values; the modulation branch takes mid-to-high frequency modulation features as input and introduces the background gas content prediction values as conditional constraints, used to characterize the gas content residual prediction values caused by local reservoir changes and lamination development; the step of superimposing the background gas content prediction values and the residual prediction values to obtain the gas content prediction results of the target well section, thereby realizing the collaborative modeling of gas content background trends and local change characteristics, includes: the background branch in the dual-branch gas content prediction model uses low-frequency background components... As input, the predicted background gas content is calculated using a feedforward neural network. The calculation process is as follows:
[0035] (8)
[0036] in, The weights and biases obtained during training, Represents a non-linear activation function. It is a non-negativity constraint function;
[0037] Meanwhile, the modulation branch uses mid-to-high frequency modulation feature vectors As input, local variation features are extracted using a nonlinear function:
[0038] (9)
[0039] Will With corresponding depth The extended feature vector is concatenated and then further nonlinearly mapped to obtain the residual prediction value.
[0040] (10)
[0041] Finally, the predicted gas content is obtained by superimposing the predicted background gas content and the predicted residual value, and its overall prediction model can be expressed as:
[0042] (11).
[0043] Preferably, the step of supervising the training of the dual-branch gas content prediction model based on measured gas content data, simultaneously introducing the main prediction loss, background gas content constraint loss, and residual supervision loss during the training process, and constructing a joint loss function through the above-mentioned multi-objective joint optimization and constraint training to obtain a stable and geologically significant dual-branch gas content prediction model includes:
[0044] The joint loss function is constructed as follows:
[0045] (12)
[0046] in, The mean square error between the predicted and measured gas content is used to constrain the overall prediction accuracy. To approximate the long-scale gas content constraint in the background, the background branches are made to closely resemble the smoothed measured gas content. The background constraint weight coefficients, This is the value obtained after smoothing the measured gas content along the background window. It can be represented as:
[0047] (13)
[0048] The laminar correction mean constraint is used to limit the overall mean of the laminar correction value to be close to 0, so that the "average level" and long-scale variation of total gas content are mainly borne by the background branch, and the laminar branch only makes local positive and negative small corrections near the background. This represents the laminar constraint weighting coefficient. It can be represented as:
[0049] (14).
[0050] The beneficial effects of this invention are as follows: This invention utilizes the Complete Ensemble Empirical Mode Decomposition (CEEMDAN) method to perform multi-scale adaptive decomposition of well logging curves, effectively separating geological background information from local heterogeneous information, thus improving the physical rationality and geological interpretability of well logging characteristics. Based on this, a bi-branch gas-bearing prediction model combining background components and laminar modulation components is constructed to characterize the overall evolution trend and local variation characteristics of gas-bearing, respectively, ensuring both stability and accuracy in the prediction results. Background target constraints and vertical continuity constraints are introduced during model training to effectively suppress non-physical oscillations and enhance the layer consistency and generalization ability of the prediction results. This method is applicable to complex reservoirs and geological conditions with well-developed laminar patterns, enabling stable and reliable gas-bearing prediction even with limited well logging data, and possesses high engineering application value. Attached Figure Description
[0051] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.
[0052] Figure 1 A schematic flowchart illustrating a method for predicting shale gas content based on well logging curves, provided in an embodiment of the present invention;
[0053] Figure 2 The components obtained after decomposing the natural gamma (GR) curve according to the embodiments of the present invention;
[0054] Figure 3 A heatmap showing the correlation coefficient between mid-to-high frequency modulation candidate features and measured gas content provided in this embodiment of the invention;
[0055] Figure 4 This is a schematic diagram of the dual-branch gas content prediction network structure provided in an embodiment of the present invention;
[0056] Figure 5 A typical well gas content decomposition and prediction bar chart provided for embodiments of the present invention;
[0057] Figure 6 A comparison of the accuracy of two gas content prediction methods provided in the embodiments of the present invention. Detailed Implementation
[0058] To enable those skilled in the art to better understand the present invention, 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. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0059] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or end that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or ends.
[0060] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.
[0061] Traditional well logging curve processing methods often employ fixed-scale filtering or empirical feature extraction, which struggle to effectively separate long-scale information reflecting the regional geological background from mid-to-high-frequency information characterizing local structures such as thin layers and laminae. This can easily lead to feature aliasing and unclear physical meaning. In recent years, although some data-driven models have achieved certain results in gas content prediction, they often neglect geological scale constraints and vertical continuity, easily generating non-physical fluctuations and limiting model generalization ability.
[0062] Therefore, there is an urgent need for a gas content prediction method that can fully exploit multi-scale information from well logging curves, take into account geological background and local heterogeneous characteristics, and possess good stability and interpretability, in order to meet the practical needs of fine evaluation of complex reservoirs.
[0063] This invention aims to provide a method for predicting shale gas content based on well logging curves. It addresses the problems in existing technologies, such as the difficulty in effectively separating long-scale information reflecting the regional geological background from medium- and high-frequency information describing the local structure of thin layers and laminae in well logging signals, which easily leads to feature aliasing and unclear physical meaning, easy neglect of geological scale constraints and vertical continuity, easy generation of non-physical fluctuations, and limited model generalization ability.
[0064] like Figure 1 As shown in the figure, a specific embodiment of the present invention provides a method for predicting shale gas content based on well logging curves, including the following steps:
[0065] (1) Well logging curve preprocessing: Five conventional well logging curves for the target formation were acquired, including the acoustic transit time curve (AC), neutron porosity curve (CNL), density curve (DEN), natural gamma curve (GR), and deep lateral resistivity curve (LLD). Given the obvious right-skewed distribution of the deep lateral resistivity curve, a logarithmic transformation was performed on the LLD curve, replacing the original LLD value with lg(LLD) to reduce the influence of extreme values and improve data distribution characteristics. Subsequently, quantile truncation was uniformly applied to each well logging curve to remove outliers, and wavelet denoising was used to suppress high-frequency noise interference and improve the curve signal-to-noise ratio. Based on this, a sampling interval of 0.125 m was set, and all well logging curves were resampled and interpolated along the depth direction to uniformly map each curve to the same depth sampling grid, providing a consistent data foundation for subsequent multi-scale decomposition and model construction.
[0066] (2) Adaptive Multi-Scale Decomposition of Well Logging Curves: Based on the preprocessing of well logging curves, the Complete Ensemble Empirical Mode Decomposition (CEEMDAN) method is used to decompose each preprocessed well logging curve sequentially along the depth direction. An adaptive amplitude Gaussian white noise set is constructed by introducing it into the original well logging curves.
[0067] (1)
[0068] In the formula, The white noise sequence added for the kth time.
[0069] For each noise-assisted signal, empirical mode decomposition is performed to extract its first-order intrinsic mode function (EMF). The first-order CEEMDAN EEMDAN EEMD component is obtained by ensemble averaging of all experimental results. Its expression is:
[0070] (2)
[0071] Based on this, the first-order intrinsic mode components are removed from the original signal to obtain the remaining signal:
[0072] (3)
[0073] Then the remaining signals Repeat the above noise-assisted decomposition and ensemble averaging process to extract the i-th order intrinsic mode components step by step:
[0074] (4)
[0075] And update the remaining signals:
[0076] (5)
[0077] This continues until the remaining signal no longer satisfies the intrinsic mode function criterion or exhibits monotonical changes; ultimately, the logging curve is represented as the sum of several intrinsic mode components and the residual term, i.e.:
[0078] (6)
[0079] (3) IMF component screening: Calculate the ratio of the energy of each IMF component after decomposition to the energy of the original logging curve. The calculation formula is as follows:
[0080] (7)
[0081] In the formula, Let be the sum of squares of the i-th eigenmode components along the depth direction, and let represent the energy of the i-th eigenmode component. The sum of squares along the depth direction of the corresponding logging curve represents the total energy of the logging curve. When its ratio is less than a preset threshold, the IMF component is considered an invalid noise component and is not used as a feature.
[0082] (4) Constructing low-frequency background components and mid-to-high-frequency modulation components: The two lowest-frequency components among the screened IMF components are superimposed with the residual to be regarded as low-frequency background components, reflecting the overall trend of formation changes. The remaining effective IMF components are subjected to linear superposition and absolute value superposition respectively to obtain mid-to-high-frequency modulation components, which are used to characterize the local physical property changes and laminar development intensity within the reservoir. Figure 2 The low-frequency background component and mid-to-high-frequency IMF component obtained after natural gamma curve decomposition are shown.
[0083] (5) Feature screening: Calculate the correlation coefficients of the acquired low-frequency background components and mid-to-high frequency modulation components respectively, and remove the feature components that are highly similar to each other. Figure 3 The correlation between the constructed mid-to-high frequency modulation characteristics and the measured gas content was demonstrated.
[0084] (6) Background gas content target construction: Based on the measured gas content curve, the corresponding background window thickness is selected according to the layer scale, and the measured gas content is averaged along the depth direction by sliding window to obtain the background gas content target curve that represents the long-scale variation characteristics, which serves as the supervision target and geological constraint of the model background branch.
[0085] (7) Construction of gas content prediction model: Based on the preprocessed background feature input and mid-to-high frequency modulation feature input, a two-branch gas content prediction model is constructed. The model structure is as follows: Figure 4 As shown. The model includes a background branch and a modulation branch, where the background branch uses low-frequency background components. As input, the background gas content is calculated using a feedforward neural network. The calculation process is as follows:
[0086] (8)
[0087] in, The weights and biases obtained during training, Represents a non-linear activation function. It is a non-negative constraint function used to ensure the physical rationality of the predicted background gas content.
[0088] Meanwhile, the laminar modulation branch modulates the feature vector at medium to high frequencies. As input, local variation features are extracted using a nonlinear function:
[0089] (9)
[0090] Will With corresponding depth The concatenated features form an extended feature vector, which is then further nonlinearly mapped to obtain the laminar modulation residual term.
[0091] (10)
[0092] Finally, the predicted gas content is obtained by superimposing the background component and the laminar modulation residual component, and its overall prediction model can be expressed as:
[0093] (11)
[0094] (8) Model Training and Constraints: The gas content prediction model is trained under supervision based on measured gas content data. During the training process, the master prediction loss, background gas content constraint loss, and residual supervision loss are introduced simultaneously. The joint loss function is constructed as follows:
[0095] (12)
[0096] The first term represents the mean square error between the predicted and measured gas content, used to constrain the overall prediction accuracy; the second term represents the background gas content constraint that approximates the long-scale gas content, ensuring that the background branches closely approximate the smoothed measured gas content. The background constraint weight coefficients, This is the value obtained after smoothing the measured gas content along the background window. It can be represented as:
[0097] (13)
[0098] The third term is the laminar correction mean constraint, which is used to limit the overall mean of the laminar correction value to be close to 0, so that the "average level" and long-scale variation of the total gas content are mainly borne by the background branch, and the laminar branch only makes local positive and negative small corrections near the background. This represents the laminar constraint weighting coefficient. It can be represented as:
[0099] (14)
[0100] The parameters of the background prediction branch and the texture correction branch are optimized uniformly using gradient descent-type optimization algorithms, and the model complexity and weight coefficients of the loss function are adjusted through cross-validation.
[0101] (9) Typical well analysis: Figure 5The table shows the predicted gas content of a typical well in the target formation. The bar chart indicates that layers ③ and ④, as well as the upper part of layer ⑥, are located in a high background gas content range, and the local correction values at corresponding depths vary significantly. Comparative analysis of the logging curves reveals a good synergistic relationship between the local correction values and logging curves characterizing porosity, such as sonic transit time. This suggests that the local correction of total gas content is primarily controlled by reservoir properties such as porosity. Figure 6 The comparison results of the two methods in terms of gas content prediction accuracy are presented. Compared with the single-branch neural network model, the proposed dual-branch network structure not only makes the prediction results closer to the measured gas content numerically, but also reflects both the large-scale geological background control and the thin-layer gas enrichment characteristics in terms of vertical variation features. This achieves a fine decomposition of the gas content prediction results and significantly improves the geological interpretability of the prediction results.
[0102] The beneficial effects of this invention are as follows: This invention utilizes the Complete Ensemble Empirical Mode Decomposition (CEEMDAN) method to perform multi-scale adaptive decomposition of well logging curves, effectively separating geological background information from local heterogeneous information, thus improving the physical rationality and geological interpretability of well logging characteristics. Based on this, a bi-branch gas-bearing prediction model combining background components and laminar modulation components is constructed to characterize the overall evolution trend and local variation characteristics of gas-bearing, respectively, ensuring both stability and accuracy in the prediction results. Background target constraints and vertical continuity constraints are introduced during model training to effectively suppress non-physical oscillations and enhance the layer consistency and generalization ability of the prediction results. This method is applicable to complex reservoirs and geological conditions with well-developed laminar patterns, enabling stable and reliable gas-bearing prediction even with limited well logging data, and possesses high engineering application value.
[0103] The above descriptions are merely embodiments of the present invention. Commonly known technical solutions or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A method for predicting shale gas content based on well logging curves, characterized in that, The method includes: S1: Obtain at least one conventional logging curve of the target well and the measured gas content data of the corresponding well section, perform depth registration, and preprocess the conventional logging curve; S2: The preprocessed conventional logging curves are decomposed sequentially along the depth direction using the complete set empirical mode decomposition method to obtain several intrinsic mode function (IMF) components. S3: Calculate the energy percentage of the intrinsic mode function (IMF) components, retain the effective IMF components with energy greater than a preset threshold, and remove the noise components with the lowest energy percentage. S4: The two lowest-frequency effective IMF components in the effective IMF components are superimposed with the residual to construct a low-frequency background component that reflects the overall trend of formation change; the remaining effective IMF components are linearly superimposed and absolute value superimposed respectively to obtain medium- and high-frequency modulation components, which are used to characterize the local physical property change characteristics and laminar development intensity inside the reservoir. S5: Calculate the correlation coefficients for the acquired low-frequency background components and mid-to-high frequency modulation components respectively, and remove feature components that are highly similar to each other. S6: Based on the measured gas content curve, select the corresponding background window thickness according to the layer scale, and perform sliding window averaging on the measured gas content along the depth direction to obtain the background gas content target curve that characterizes the long-scale variation characteristics, which serves as the supervision target and geological constraint of the model background branch. S7: Based on the preprocessed background feature input and mid-to-high frequency modulation feature input, a dual-branch gas content prediction model is constructed. The dual-branch gas content prediction model includes a background branch and a modulation branch. The background branch is used to establish the mapping relationship between the logging background features and the gas content background components, and outputs the non-negative constrained background gas content prediction value. The modulation branch takes the mid-to-high frequency modulation features as input and introduces the background gas content prediction value as a condition constraint to characterize the gas content residual prediction value caused by local reservoir changes and laminar flow development. The background gas content prediction value and the residual prediction value are superimposed to obtain the gas content prediction result of the target well section, thereby realizing the collaborative modeling of gas content background trends and local change characteristics. S8: The dual-branch gas content prediction model is trained under supervision based on measured gas content data. During the training process, the main prediction loss, background gas content constraint loss, and residual supervision loss are introduced simultaneously. Through the above multi-objective joint optimization and constraint training, a joint loss function is constructed to obtain a stable and geologically significant dual-branch gas content prediction model.
2. The method for predicting shale gas content based on well logging curves according to claim 1, characterized in that, The step of preprocessing the conventional logging curve includes: The conventional logging curves are uniformly truncated using quantiles to remove outliers, and wavelet denoising is used to suppress high-frequency noise interference and improve the signal-to-noise ratio of the curves. Set a sampling interval and resample and interpolate all the conventional logging curves along the depth direction so that all curves are uniformly mapped to the same depth sampling grid.
3. The method for predicting shale gas content based on well logging curves according to claim 1, characterized in that, The step of using the complete ensemble empirical mode decomposition method to sequentially decompose each preprocessed conventional logging curve along the depth direction to obtain several intrinsic mode function (IMF) components includes: Multiple sets of amplitude-controlled Gaussian white noise are applied to the preprocessed conventional logging curve signal to form a noise-assisted signal set: (1) In the formula, Here, x(z) is the noise-assisted signal, x(z) is the conventional logging curve signal, and z is the depth. The white noise sequence added for the kth time; For each of the noise-assisted signals, empirical mode decomposition is performed to extract its first-order intrinsic mode function (EMF). The first-order CEEMDAN EEMDAN EEMD component is then obtained by ensemble averaging of all experimental results. Its expression is: (2) Based on this, the first-order intrinsic mode component is removed from the original conventional logging curve signal to obtain the remaining signal: (3) in, The remaining signal, and then the remaining signal Repeat the above noise-assisted decomposition and ensemble averaging process to extract the i-th order intrinsic mode components step by step: (4) And update the remaining signals: (5) This continues until the remaining signal no longer satisfies the intrinsic mode function criterion or exhibits monotonical changes; ultimately, the conventional logging curve signal is represented as the sum of several intrinsic mode components and the residual term, i.e.: (6)。 4. The method for predicting shale gas content based on well logging curves according to claim 1, characterized in that: The formula for calculating the energy proportion of the intrinsic mode function (IMF) components is as follows: (7) in, Let be the sum of squares of the i-th eigenmode components along the depth direction, and let represent the energy of the i-th eigenmode component. The sum of squares of the corresponding logging curve along the depth direction represents the total energy of the logging curve. When its ratio is less than a preset threshold, the IMF component is regarded as an invalid noise component.
5. The method for predicting shale gas content based on well logging curves according to claim 1, characterized in that, Based on the preprocessed background feature input and the mid-to-high frequency modulation feature input, a dual-branch gas content prediction model is constructed. The dual-branch gas content prediction model includes a background branch and a modulation branch. The background branch is used to establish the mapping relationship between the logging background features and the gas content background components, and outputs the non-negative constrained background gas content prediction value. The modulation branch takes mid-to-high frequency modulation characteristics as input and introduces the background gas content prediction value as a condition constraint to characterize the gas content residual prediction value caused by local reservoir changes and lamination development; the background gas content prediction value is superimposed with the residual prediction value to obtain the gas content prediction result of the target well section, thereby realizing the steps of co-modeling the background gas content trend and local change characteristics, including: In the dual-branch gas content prediction model, the background branch consists of low-frequency background components. As input, the predicted background gas content is calculated using a feedforward neural network. The calculation process is as follows: (8) in, The weights and biases obtained during training, Represents a non-linear activation function. It is a non-negativity constraint function; Meanwhile, the modulation branch uses mid-to-high frequency modulation feature vectors As input, local variation features are extracted using a nonlinear function: (9) Will With corresponding depth The extended feature vector is concatenated and then further nonlinearly mapped to obtain the residual prediction value. (10) Finally, the predicted gas content is obtained by superimposing the predicted background gas content and the predicted residual value, and its overall prediction model can be expressed as: (11)。 6. The method for predicting shale gas content based on well logging curves according to claim 1, characterized in that, The steps of supervising the training of the bibranch gas content prediction model based on measured gas content data, simultaneously introducing the main prediction loss, background gas content constraint loss, and residual supervision loss during the training process, and constructing a joint loss function through the above multi-objective joint optimization and constraint training to obtain a stable and geologically significant bibranch gas content prediction model include: The joint loss function is constructed as follows: (12) in, The mean square error between the predicted and measured gas content is used to constrain the overall prediction accuracy. To approximate the long-scale gas content constraint in the background, the background branches are made to closely resemble the smoothed measured gas content. The background constraint weight coefficients, This is the value obtained after smoothing the measured gas content along the background window. It can be represented as: (13) The laminar correction mean constraint is used to limit the overall mean of the laminar correction value to be close to 0, so that the "average level" and long-scale variation of total gas content are mainly borne by the background branch, and the laminar branch only makes local positive and negative small corrections near the background. This represents the laminar constraint weighting coefficient. It can be represented as: (14)。