A method for predicting shear wave velocity under lithological constraint based on space-time attention
Patent Information
- Application Number
- CN202611381543.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-08
- Publication Date
- 2026-10-13
AI Technical Summary
[0006]为解决现有横波速度预测方法存在的非线性映射关系刻画不足、时空特征协同提取不充分、岩性约束缺失，以及预测结果地质合理性不足等技术问题，本发明公开了一种岩性约束下基于时空注意力的横波速度预测方法
[0032]本发明的有益效果是，该方法将岩性信息与常规测井数据联合输入网络，能够学习不同岩性条件下测井响应与横波速度之间的差异性映射关系。通过融合卷积神经网络（CNN）、门控循环单元（GRU）和时空注意机制，同时提取测井数据的局部变化特征和深度序列特征，提高了横波速度预测的精度和稳定性。岩性约束项能够减少不符合地质规律的异常预测结果，增强预测结果的地质一致性。
Smart Images
Figure CN122901940A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of oil and gas exploration and development technology, and in particular to a method for predicting shear wave velocity based on spatiotemporal attention under lithological constraints. Background Technology
[0002] In oil and gas exploration and development, shear wave velocity is a crucial parameter characterizing the elastic properties of subsurface rocks, playing a vital role in reservoir lithology identification, fluid discrimination, rock physics analysis, and seismic inversion. However, due to factors such as the high cost and limited applicability of shear wave logging, continuous and reliable shear wave velocity data are often lacking in certain well sections or areas, thus limiting the accuracy of detailed reservoir characterization and elastic parameter inversion. Therefore, accurately predicting shear wave velocity using existing conventional logging data has become a significant technical challenge in well logging interpretation and reservoir parameter prediction.
[0003] Traditional methods for predicting shear wave velocity mainly include empirical relation methods and rock physics model methods. Empirical relation methods typically estimate the missing shear wave velocity by establishing a statistical relationship between P-wave and shear wave velocities. Rock physics model methods calculate shear wave velocity based on the relationships between rock mineral composition, pore structure, and elastic parameters. For example, Greenberg and Castagna proposed a method for estimating shear wave velocity applicable to porous rocks in 1992. Kuster and Toksöz established a velocity model considering the influence of rock composition and porous media based on the theory of seismic wave propagation in two-phase media in 1974. Xu and White proposed a clay-sandstone mixture velocity model for sandstone-mudstone mixed strata in 1995. Yang et al. studied a method for estimating shear wave velocity in calcareous sandy mudstone strata in 2019. These methods have a clear rock physics basis and have achieved certain application results under specific stratigraphic conditions. However, empirical relations and rock physics models usually require certain assumptions about rock mineral composition, pore morphology, cementation methods, and fluid properties, and their model parameters are often determined based on specific regional or lithological data. When reservoir lithology varies frequently, pore structure is complex, or heterogeneity is strong, fixed empirical relationships and model parameters are insufficient to fully describe the complex nonlinear mapping relationship between conventional logging parameters and shear wave velocity, easily leading to significant deviations between predicted and measured shear wave velocities. This is particularly true in tight sandstone, thin interbedded sandstone and mudstone, and lithological transition zones, where the elastic responses of different lithologies vary significantly, greatly limiting the applicability of using uniform empirical relationships for prediction.
[0004] With the development of artificial intelligence technology, neural networks have been increasingly applied to shear wave velocity prediction. In 2020, Sun and Liu established a nonlinear mapping relationship between conventional logging parameters and shear wave velocity using gated recurrent unit networks, predicting shear wave velocity by extracting the sequence dependency features of logging data along the depth direction. In 2021, Ma et al. proposed a shear wave velocity prediction method based on a one-dimensional convolutional neural network, utilizing convolution operations to extract local variation features of logging curves. In 2022, Chen et al. combined a two-dimensional convolutional neural network with a gated recurrent unit to simultaneously extract local spatial features and depth sequence features of logging data; in the same year, Chen et al. further introduced a spatiotemporal attention mechanism into the fusion network, assigning corresponding weights based on the correlation between different features and shear wave velocity. In 2024, Yang et al. proposed a CNN-GRU shear wave velocity prediction method based on a self-attention mechanism, further enhancing the fusion network's ability to extract key logging features.
[0005] While these methods have improved predictive performance to some extent, most remain purely data-driven, lacking explicit geological constraints and physical mechanisms. For example, sandstone, mudstone, and carbonate rocks exhibit fundamental differences in pore structure, cementation characteristics, and fluid properties, leading to significantly different elastic parameter responses. Therefore, incorporating lithological constraints into shear wave velocity prediction helps ensure that the results are more consistent with rock physics principles. Summary of the Invention
[0006] To address the technical problems of existing shear wave velocity prediction methods, such as insufficient characterization of nonlinear mapping relationships, inadequate extraction of spatiotemporal features, lack of lithological constraints, and insufficient geological rationality of prediction results, this invention discloses a shear wave velocity prediction method based on spatiotemporal attention under lithological constraints.
[0007] To achieve the above objectives, the present invention adopts the following technical solution:
[0008] A method for predicting shear wave velocity based on spatiotemporal attention under lithological constraints includes the following steps:
[0009] S1. Well logging data with measured shear wave velocity is selected as training samples. The intersection relationship between different conventional well logging parameters and shear wave velocity is analyzed, and network input parameters are selected.
[0010] The logging data was cleaned, outliers and missing values were processed, and logging data of different dimensions were standardized and normalized. Then, a sample set was constructed using a "window" many-to-one format, with a "window" length of 25, and the shear wave velocity corresponding to the 13th sampling point was selected as the label. Finally, the sample set was divided into training set and test set in a 7:3 ratio.
[0011] S2 encodes the lithology category, converts discrete lithology information into continuous feature representation, and inputs it together with conventional logging data into the LCNN fusion network, enabling the LCNN fusion network to learn the nonlinear mapping relationship between conventional logging response and shear wave velocity;
[0012] S3, the loss function consists of two parts: one is the data-driven error term, which uses the L2 norm to calculate the prediction results of the fusion network under lithological constraints. Compared with the measured shear wave velocity The error is due to two factors: first, the lithological constraint term, which uses the negative logarithmic conditional probability of predicting shear wave velocity under given lithological conditions, and the constraint prediction results conform to the shear wave velocity distribution characteristics of the corresponding lithology.
[0013] S4. Input the training set data into the fusion network of 2DCNN-GRU, 2DCNN and LCNN respectively for network training until the loss function is stable and unchanged, indicating that the network has reached the optimal state.
[0014] S5, the test well data is input into the optimal 2DCNN-GRU, 2DCNN and LCNN fusion network respectively to obtain the prediction results of shear wave velocity;
[0015] S6. To verify the rationality of the attention mechanism, a visual analysis of the weights of the spatiotemporal attention mechanism was performed. It can be seen that the weights of the spatial attention layer, from highest to lowest, are: P-wave velocity, natural gamma, porosity, water saturation, clay content, and density. This order is basically consistent with the distribution of the correlation coefficient between conventional logging data and S-wave velocity. The temporal attention layer assigns different weights to different temporal characteristics of the logging data. It can be seen that the features at the label location receive the largest weight assigned by the temporal attention layer, indicating that the data features at this location have the greatest impact on S-wave velocity prediction. As the distance from the label location gradually increases to both sides, the temporal attention weight generally shows a gradually decreasing trend. This change pattern is basically consistent with the distribution of autocorrelation in conventional logging data. This verifies the rationality of incorporating the spatiotemporal attention mechanism in this invention.
[0016] S7, using the coefficient of determination The root mean square error (RMSE) is used to compare the predicted shear wave velocity with the measured value, verifying the effectiveness and robustness of the proposed LCNN fusion network under complex geological conditions.
[0017] Furthermore, in step S1, the network input parameters include longitudinal wave velocity, density, porosity, natural gamma, clay content, and water saturation.
[0018] Furthermore, in step S2, based on the lithological interpretation results, the discrete lithological labels for each depth sampling point are represented as follows:
[0019] ;
[0020] The lithological labels are converted into continuous features through embedding encoding, represented as follows:
[0021] ;
[0022] The lithological coding features are concatenated with conventional well logging data and used as input to the neural network.
[0023] ;
[0024] Among them, among them, Indicates word embedding, Indicates lithological tag sequence, Indicates lithological coding characteristics, This represents conventional well logging data.
[0025] Furthermore, in step S3, the loss function is:
[0026] ;
[0027] in, This represents the transverse wave velocity predicted by the network. This represents the measured shear wave velocity. This represents the conditional probability distribution for predicting shear wave velocity under given lithological conditions. and These represent the weighting coefficients for shear wave velocity data loss and lithological constraint loss, respectively.
[0028] Furthermore, in step S7, the coefficient of determination The root mean square error (RMSE) and the root mean square error (RMSE) are expressed as follows:
[0029] ;
[0030] ;
[0031] in, This represents the measured shear wave velocity. This represents the mean of the true shear wave velocity. This represents the predicted shear wave velocity value, where m is the number of samples.
[0032] The beneficial effects of this invention are that by jointly inputting lithological information and conventional logging data into a network, this method can learn the differential mapping relationship between logging response and shear wave velocity under different lithological conditions. By fusing convolutional neural networks (CNN), gated recurrent units (GRU), and spatiotemporal attention mechanisms, it simultaneously extracts local variation features and depth sequence features from logging data, thereby improving the accuracy and stability of shear wave velocity prediction. The lithological constraint term can reduce anomalous prediction results that do not conform to geological laws, enhancing the geological consistency of the prediction results.
[0033] The predicted shear wave velocity curves of this method are in better agreement with the measured curves, and have better predictive performance in the lithological variation range and the rapid variation range of shear wave velocity. Attached Figure Description
[0034] Figure 1 This is a schematic diagram of the workflow of the present invention;
[0035] Figure 2 This is a diagram of the LCNN fusion network structure in an embodiment of the present invention;
[0036] Figure 3 This is a cross-plot of conventional logging data and shear wave velocity in an embodiment of the present invention.
[0037] Figure 4 This is an autocorrelation analysis of conventional well logging data in this embodiment of the invention;
[0038] Figure 5 These are the prediction results of different methods in the embodiments of the present invention on the test set;
[0039] Figure 6 This is a visualization analysis of the weights of the spatiotemporal attention mechanism in an embodiment of the present invention. Detailed Implementation
[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 a part of the embodiments of the present invention, and not all of them. 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.
[0041] This invention discloses a method for predicting shear wave velocity based on spatiotemporal attention under lithological constraints, such as... Figure 1As shown, firstly, lithological categories are encoded, converting discrete lithological information into continuous feature representations, which are then input into a neural network along with conventional well logging data. Subsequently, a convolutional neural network (CNN) is used to extract local variation features from the well logging data as input to a spatial attention layer, where weights are assigned. The weighted features are then input to a gated recurrent unit (GRU) for temporal feature extraction, which is then used as input to a temporal attention layer, where weights are assigned again. Figure 2 Furthermore, a lithology constraint term is introduced based on the data-driven error term to construct a loss function. An adaptive weighting strategy based on uncertainty is then employed to dynamically balance the data fitting error and the lithology constraint term. Finally, the trained network is used to establish a mapping relationship between conventional logging data and shear wave velocity, enabling continuous prediction of missing shear wave velocity curves. The specific process is as follows:
[0042] 1. Optimization of logging parameters and data preprocessing.
[0043] Well logging data with measured shear wave velocities were selected as training samples. The cross-sectional relationships between different conventional well logging parameters and shear wave velocities were analyzed. Figure 3 The longitudinal wave velocity, density, porosity, natural gamma, clay content, and water saturation were selected as network input parameters. Figure 4 The autocorrelation coefficient of conventional well logging data varies with hysteresis. It can be seen that the autocorrelation gradually decreases as the hysteresis increases, indicating that the well logging data are correlated along the sequence direction. The well logging data undergoes data cleaning, outliers and missing values are handled, and well logging data of different dimensions are standardized and normalized. Subsequently, a sample set is constructed using a "window" many-to-one approach. In this invention, the "window" length is 25, and the shear wave velocity corresponding to the 13th sampling point is selected as the label to construct the sample set. Finally, the data is divided into training and test sets in a 7:3 ratio.
[0044] 2. Lithology coding.
[0045] Based on the lithological interpretation results, the discrete lithological labels for each depth sampling point are represented as follows:
[0046] ;
[0047] Transforming lithological labels into continuous features through embedding encoding:
[0048] ;
[0049] The lithological coding features are concatenated with conventional well logging data and used as input to the neural network.
[0050] ;
[0051] Among them, among them, Indicates word embedding, Indicates lithological tag sequence, Indicates lithological coding characteristics, This represents conventional well logging data.
[0052] This process allows lithological information to be used in shear wave velocity prediction, establishing a mapping relationship between logging response and shear wave velocity under different lithological conditions.
[0053] 3. Construction of loss function.
[0054] The loss function of the LCNN fusion network consists of two parts: the first part is the data-driven error term, which uses the L2 norm to calculate the prediction results of the fusion network under lithological constraints. Compared with the measured shear wave velocity The first part is the error between the two; the second part is the lithological constraint term, which uses the negative logarithmic conditional probability of the predicted shear wave velocity under given lithological conditions to constrain the prediction results to conform to the shear wave velocity distribution characteristics of the corresponding lithology, thereby reducing abnormal prediction results that do not conform to geological laws and improving the physical rationality of the prediction results.
[0055] The final loss function is expressed as follows:
[0056] ;
[0057] in, This represents the transverse wave velocity predicted by the network. This represents the measured shear wave velocity. This represents the conditional probability distribution for predicting shear wave velocity under given lithological conditions. and These represent the weighting coefficients for shear wave velocity data loss and lithological constraint loss, respectively, used to balance the relationship between data fitting ability and lithological constraint terms.
[0058] To avoid the subjectivity introduced by manually assigning loss weights, an adaptive weighting strategy based on uncertainty is adopted to dynamically balance the data fitting term and the lithological constraint term during training.
[0059] 4. Input the training set data into the fusion networks of 2DCNN-GRU, 2DCNN, and LCNN respectively for network training until the loss function is stable and remains unchanged, indicating that the network has reached the optimal state;
[0060] 5. The test well data were input into the optimal 2DCNN-GRU, 2DCNN, and LCNN fusion networks respectively to obtain the predicted shear wave velocity. Figure 5As shown, (a) LCNN fusion network; (b) 2DCNN-GRU; (c) 2DCNN. It can be seen that the LCNN fusion network has a better prediction effect on sandstone locations than 2DCNN-GRU and 2DCNN.
[0061] 6. To verify the rationality of the attention mechanism, a visual analysis of the weights of the spatiotemporal attention mechanism is performed, such as... Figure 6 Among them, (a) spatial attention layer weight visualization; (b) temporal attention layer weight visualization, it can be seen that the weights of the spatial attention layer from high to low are P-wave velocity, natural gamma, porosity, water saturation, clay content, and density. This order is consistent with... Figure 3 The correlation coefficient distribution between conventional logging data and shear wave velocity is basically consistent. The temporal attention layer assigns different weights to different temporal features of the logging data. It can be seen that the features at the label location receive the largest weight assigned by the temporal attention layer, indicating that the data features at this location have the greatest impact on shear wave velocity prediction. As the distance from the label location gradually increases to both sides, the temporal attention weight generally shows a gradual decreasing trend. This change pattern is consistent with... Figure 4 The distribution of autocorrelation in conventional well logging data is basically consistent. This verifies the rationality of incorporating a spatiotemporal attention mechanism in this invention.
[0062] 7. Network evaluation.
[0063] Using the coefficient of determination ( The predicted shear wave velocity and the root mean square error (RMSE) were compared with the measured values to verify the effectiveness and robustness of the LCNN fusion network under complex geological conditions (Table 1).
[0064] Coefficient of determination The root mean square error (RMSE) and the root mean square error (RMSE) are expressed as follows:
[0065] ;
[0066] ;
[0067] in, This represents the measured shear wave velocity. This represents the mean of the true shear wave velocity. This represents the predicted shear wave velocity value, where m is the number of samples.
[0068] Table 1. Quantitative Evaluation of Shear Wave Velocity Prediction Results Using Different Methods
[0069] 2DCNN 0.040 0.90 2DCNN-GRU 0.035 0.92 LCNN 0.021 0.95
[0070] As can be seen from the table above, the LCNN fusion network... The accuracy is 5.0% and 3.0% higher than 2DCNN and 2DCNN-GRU, respectively, which means that the LCNN fusion network has higher prediction accuracy and generalization.
[0071] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. A method for predicting shear wave velocity based on spatiotemporal attention under lithological constraints, characterized in that, Includes the following steps: S1. Well logging data with measured shear wave velocity is selected as training samples. The intersection relationship between different conventional well logging parameters and shear wave velocity is analyzed, and network input parameters are selected. Data cleaning is performed on the logging data, outliers and missing values are handled, and logging data of different dimensions are standardized and normalized. A sample set of logging data was constructed using a window-to-one approach with a window length of 25, and the shear wave velocity corresponding to the 13th sampling point was selected as the label to construct the sample set. The sets were divided into training and testing sets in a 7:3 ratio. S2 encodes lithology categories, converts discrete lithology information into continuous feature representations, and inputs them together with conventional logging data into the LCNN fusion network, enabling the network to learn the nonlinear mapping relationship between conventional logging response and shear wave velocity; S3, the loss function consists of two parts: one is the data-driven error term, which uses the L2 norm to calculate the prediction results of the fusion network under lithological constraints. Compared with the measured shear wave velocity The error is due to two factors: first, the lithological constraint term, which uses the negative logarithmic conditional probability of predicting shear wave velocity under given lithological conditions to constrain the prediction results to conform to the shear wave velocity distribution characteristics of the corresponding lithology; S4. Input the training set data into the fusion network of 2DCNN-GRU, 2DCNN and LCNN respectively for network training until the loss function is stable and unchanged, indicating that the network has reached the optimal state. S5. The test well data is input into the optimal 2DCNN-GRU, 2DCNN and LCNN fusion network respectively to obtain the prediction results of shear wave velocity; S6. In order to verify the rationality of the attention mechanism, the weights of the spatiotemporal attention mechanism were visualized and analyzed. The weights of the spatial attention layer from high to low are longitudinal wave velocity, natural gamma, porosity, water saturation, clay content and density. The features at the location of the label obtained the maximum weights assigned to the temporal attention layer. S7, using the coefficient of determination The root mean square error (RMSE) is used to compare the predicted shear wave velocity with the measured value, verifying the effectiveness and robustness of the LCNN fusion network under complex geological conditions.
2. The shear wave velocity prediction method based on spatiotemporal attention under lithological constraints as described in claim 1, characterized in that, In step S1, the network input parameters include longitudinal wave velocity, density, porosity, natural gamma, clay content, and water saturation.
3. The shear wave velocity prediction method based on spatiotemporal attention under lithological constraints as described in claim 2, characterized in that, In step S2, based on the lithological interpretation results, the discrete lithological labels for each depth sampling point are represented as follows: ； Lithological information is converted into continuous features through embedding encoding, represented as follows: ； The lithological coding features are concatenated with conventional well logging data and used as input to the neural network. ； in, Indicates word embedding, Indicates lithological tag sequence, Indicates lithological coding characteristics, This represents conventional well logging data.
4. The shear wave velocity prediction method based on spatiotemporal attention under lithological constraints as described in claim 3, characterized in that, In step S3, the loss function is: ； in, This represents the transverse wave velocity predicted by the network. This represents the measured shear wave velocity. This represents the conditional probability distribution for predicting shear wave velocity under given lithological conditions. and These represent the weighting coefficients for shear wave velocity data loss and lithological constraint loss, respectively.
5. The shear wave velocity prediction method based on spatiotemporal attention under lithological constraints as described in claim 4, characterized in that, In step S7, the coefficient of determination The root mean square error (RMSE) and the root mean square error (RMSE) are expressed as follows: ； ； in, This represents the measured shear wave velocity. This represents the mean of the true shear wave velocity. This represents the predicted shear wave velocity value, where m is the number of samples.