A method and apparatus for fluid phase prediction based on well logging data
By combining well logging data to screen sensitive attribute pairs and utilizing deep learning models, the problems of low accuracy and efficiency in gas-water boundary identification were solved, and rapid and accurate prediction of fluid phase distribution was achieved.
Patent Information
- Application Number
- CN202111071171.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-09-13
- Publication Date
- 2025-11-18
- Estimated Expiration
- 2041-09-13
AI Technical Summary
In existing technologies, gas-water boundary identification has poor accuracy and low computational efficiency in well site deployment. Single-attribute prediction methods have large errors, and deep learning models have low training efficiency under high-dimensional samples, resulting in unsatisfactory fluid phase distribution prediction.
Sensitive attribute pairs are identified by combining well logging data. Attribute pairs that can distinguish fluid phases are screened out through PCA and cluster analysis. A deep learning probabilistic model is used for iterative training to establish a nonlinear mapping relationship. Fluid phase distribution is then predicted by combining seismic data.
It improves the accuracy and computational efficiency of fluid phase prediction, enables rapid and accurate identification of gas-water boundaries during the development phase, reduces computational load, and learns the physical properties of logging data.
Smart Images
Figure CN115808718B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of seismic exploration technology for oil and gas, and specifically to a method and equipment for fluid phase prediction based on well logging data. Background Technology
[0002] In the gas reservoir development stage, a clear understanding of water and gas distribution is needed to support well site deployment. Water and gas distribution prediction aims to directly determine the distribution of gas and water in the target layer. Since the differences in gas and water response during seismic events are very subtle, gas and water identification is a highly challenging problem. Furthermore, as the understanding of regional water and gas distribution deepens with the progress of well site deployment, continuous adjustments and multiple characterizations are required. Currently, water and gas boundary identification primarily relies on various geophysical methods based on seismic data for prediction and description. The principle is that seismic data is a dataset with time-series characteristics, and the physical properties of the formation are relatively stable. Because seismic waves travel from explosive excitation to receiver reception, passing through the target layer, the seismic data implicitly contains gas and water information for that layer. Theoretically, by processing and interpreting seismic data, certain attributes with clear physical characteristics can be obtained for gas and water prediction and related work. The five basic attributes of an earthquake are amplitude, frequency, phase, frequency subdivision, and coherence. These attributes mainly reflect the interface of the reflection coefficient and are mainly affected by the reservoir and surrounding rock. However, lithological parameters such as wave impedance, velocity, and density are the attribute parameters that can reflect the lithology and fluid characteristics of the reservoir. These parameters can usually be obtained by inverting seismic data. Then, the parameters obtained by inversion are cross-analyzed with well logging data to predict the probability distribution of fluid phases.
[0003] Methods for predicting the gas-water boundary generally employ different fluid prediction methods depending on the seismic attributes used. Examples include: LFR (low-frequency resonance) method, wavelet decomposition method, gas-bearing phase prediction method, and AVO (air-velocity vortex) fluid prediction method. These methods essentially utilize a specific attribute characteristic of the seismic data. However, in practice, because the differences in water and gas identification attributes between various seismic attributes are small, it is often impossible to find a single attribute with clear physical meaning in the seismic data for gas-water identification.
[0004] Furthermore, for gas-water boundary identification during the development phase, existing technologies mainly utilize multiple inversion methods, treating information from newly developed wells as known information. However, this process involves a very large computational load and introduces many uncertainties. On the one hand, it suffers from strong ambiguity; on the other hand, the development phase is time-sensitive, making full implementation costly and time-consuming. For example, Chinese Patent Application No. 2019105710245 discloses an AVO inversion method and system based on Bayesian and series inversion theories. It improves the accuracy of pre-stack AVO analysis by introducing Bayesian and series inversion theories into AVO seismic inversion. However, it suffers from complex computational processes, excessive time costs, and is not conducive to rapid prediction during the development phase.
[0005] In recent years, with the development of artificial intelligence, existing technologies have also utilized the high computational efficiency and strong self-learning ability of deep learning models to establish probabilistic models capable of efficiently predicting fluid phase distribution. For example, Chinese patent application number 2018111017165 discloses a method for fluid prediction using seismic data based on deep learning, including: inputting all seismic data within a specified block and preprocessing it; using a first deep learning network to perform nonlinear optimization and fitting on the linear features of the seismic data; using a second deep learning network to classify the linear features of a large amount of seismic data to establish a first fluid feature model; using the residual network in the second deep learning network to perform normalization and correction on the established first fluid feature model to obtain a second fluid feature model; and using the second fluid feature model and activation function to perform matrix set calculation on the seismic data within the prediction block to obtain the probability distribution data of fluid features within the prediction block. This method configures the first deep learning model to perform nonlinear fitting on tens of thousands of features of the seismic target layer, and configures the second deep learning model to classify and predict the fitted parameters. However, in practice, if the number of sample dimensions involved in the fitting is too large, the training efficiency will be very low and it will easily fall into local convergence. At the same time, since the differences in the numerical representation of water vapor identification of various attributes in earthquake attributes are small, using only deep learning algorithms for pure mathematical statistics will lead to the results deviating from the physical laws and having certain errors, resulting in the final classification results being less than ideal. Summary of the Invention
[0006] The purpose of this invention is to overcome the problems of poor accuracy in fluid prediction using a single attribute and low computational efficiency in fitting multi-dimensional attributes in the existing technology. It provides a water-gas boundary prediction method and device based on well logging data. By combining well logging data, it determines sensitive attribute pairs that can be used to accurately distinguish fluid phases, thereby reducing the amount of computation and improving the accuracy of prediction and classification.
[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution:
[0008] A fluid phase prediction method based on well logging data includes:
[0009] Step 1: Based on geological data, determine the lithofacies classification of a certain area, perform PCA and cluster analysis on the existing well logging data in the area to identify sensitive attribute pairs that can be used to distinguish fluid phases; based on the sensitive attribute pairs, perform lithofacies classification and labeling on the existing well logging data in the area to construct training samples;
[0010] Step 2: Iteratively train the deep learning probability model using the training samples until the deep learning probability model satisfies conditional convergence;
[0011] Step 3: Calculate the values of sensitive attribute pairs of the wells to be logged in the region from the seismic data of the region using the inversion method, input the calculation results into the trained deep learning probabilistic model, and use the deep learning probabilistic model to predict the fluid phase distribution data of the wells to be logged in the region.
[0012] According to a specific implementation, in the above-mentioned fluid phase prediction method based on well logging data, step 1 includes:
[0013] Step 101: Based on geological data, determine the lithofacies classification of a certain area;
[0014] Step 102: Correct the existing logging data in the area according to the empirical formula, and select N attributes from the corrected logging data; where N≥4;
[0015] Step 103: Perform PCA analysis on the lithofacies sample with N attributes as the attributes to be analyzed. Based on the PCA analysis results, select some attributes from the N attributes for pairwise cluster cross-plot analysis. Based on the cluster cross-plot analysis results, find the sensitive attribute pairs that can be used to distinguish the fluid phase.
[0016] Step 104: Based on the sensitive attributes, perform lithofacies classification and labeling on the existing well logging data in the region to construct training samples.
[0017] According to a specific implementation, in the above-mentioned fluid phase prediction method based on well logging data, step 101, correcting historical well logging data according to empirical formulas, includes:
[0018] The Faust formula was used to correct the P-wave velocity and resistivity in the well logging data, the Gardner formula was used to correct the P-wave velocity and density in the well logging data, and the mudstone line formula was used to correct the P-wave velocity and S-wave velocity in the well logging data.
[0019] According to one specific implementation, in the above-mentioned fluid phase prediction method based on well logging data, the deep learning probability model is constructed based on a two-dimensional deconvolutional CNN neural network.
[0020] According to a specific implementation, in the above-mentioned fluid phase prediction method based on well logging data, the sensitive attribute pair is: longitudinal wave impedance and gamma.
[0021] According to a specific implementation method, in the above-mentioned fluid phase prediction method based on well logging data, the P-wave impedance value and gamma value of the region are calculated from the seismic data using the seismic inversion method. The measured P-wave impedance value and the measured gamma value are normalized, and the normalized measured P-wave impedance value and the measured gamma value are input into the deep learning probability model.
[0022] According to one specific implementation, in the above-mentioned fluid phase prediction method based on well logging data, the cross-entropy loss function is used as the loss function of the deep learning probability model.
[0023] In a further embodiment of the present invention, a water-gas boundary prediction device based on well logging data is also provided, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the above-described water-gas boundary prediction method based on a probabilistic model.
[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:
[0025] The water-gas boundary prediction method based on well logging data provided by this invention utilizes the fixed development mode of well logging data. It analyzes gas-water sensitive attribute pairs through well logging data, and then uses the sensitive attributes to train a deep learning probabilistic model to establish a nonlinear mapping from gas-water sensitive attribute pairs to seismic fluid phases. This effectively overcomes the problems of poor accuracy in fluid prediction using a single attribute and low computational efficiency in fitting multi-dimensional attributes, thus balancing computational efficiency and accuracy. At the same time, the deep learning probabilistic model can learn the physical characteristics of well logging data with a fixed development mode, enabling it to predict the probability distribution of fluid phases with physical meaning. Attached Figure Description
[0026] Figure 1 This is a block diagram illustrating the principle of the water-vapor boundary prediction method based on a probabilistic model as described in Embodiment 1 of the present invention.
[0027] Figure 2 This is a schematic diagram of cluster analysis as described in Embodiment 2 of the present invention. Figure 1 ;
[0028] Figure 3This is a schematic diagram of cluster analysis as described in Embodiment 2 of the present invention. Figure 2 ;
[0029] Figure 4 This is a schematic diagram of cluster analysis as described in Embodiment 2 of the present invention. Figure 3 ;
[0030] Figure 5 This is a schematic diagram of well logging curve calibration based on the calibration electrical phase of sensitive attribute pairs as described in Embodiment 2 of the present invention;
[0031] Figure 6 This is a schematic diagram of the predicted distribution of fluid phases in the well to be predicted as described in Embodiment 2 of the present invention;
[0032] Figure 7 This is a schematic diagram of the fluid phase prediction device based on well logging data as described in Embodiment 3 of the present invention; Detailed Implementation
[0033] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.
[0034] Example 1
[0035] Figure 1 The water-gas boundary prediction method based on a probabilistic model according to an embodiment of the present invention is shown, including:
[0036] Step 101: Collect data on the development area, including all seismic data and existing well logging data (well logging data of known wells) within the area. Based on the current geological exploration data, classify the fluid facies of the area. The usual practice is to classify the lithofacies according to development needs. Typically, three types are classified: water-bearing facies, gas-bearing facies, and non-reservoir facies. This serves as the basis for well logging data analysis and gas-water model establishment, resulting in a clear development model.
[0037] Step 102: Correct the existing logging data in the area using empirical formulas, and select N attributes from the corrected logging data; where N≥5. Specifically, the Faust formula can be used to correct the P-wave velocity and resistivity in the logging data, the Gardner formula can be used to correct the P-wave velocity and density in the logging data, and the mudstone line formula can be used to correct the P-wave velocity and S-wave velocity in the logging data. After correcting the logging data, N attributes can be selected from the various logging attributes (prioritizing attributes with higher accuracy after correction and / or easier to obtain in actual acquisition, generally gamma, water saturation, density, resistivity, potential, clay content, etc.).
[0038] Step 103: Perform PCA analysis on the lithofacies sample with N attributes as the attributes to be analyzed. Based on the PCA analysis results, select some attributes from the N attributes for pairwise cluster cross-plot analysis. Based on the cluster cross-plot analysis results, find the sensitive attribute pairs that can be used to distinguish the fluid phase.
[0039] Specifically, PCA analysis yields the attribute weight matrix for each lithofacies sample, allowing selection of attributes with larger or smaller weights (i.e., more pronounced characteristics) from N attributes. Furthermore, PCA analysis reveals linear relationships between attributes, enabling selective selection of those exhibiting linear relationships. Thus, this invention uses PCA analysis to determine the numerical sensitivity of attributes, allowing for the initial screening of relatively sensitive attributes, achieving attribute dimensionality reduction. Further, based on the PCA analysis results and combined with seismic-related attributes such as P-wave velocity, S-wave velocity, density, and Lamé constant, unsupervised clustering analysis is employed for pairwise cross-plot analysis of the attribute data. The cross-plot scatter plot identifies sensitive attribute pairs that can separate lithofacies types in a planar distribution.
[0040] Step 103: Based on the sensitive attributes, the existing target layer logging data in the region is calibrated to identify the lithofacies type in order to construct training samples. That is, the logging data is calibrated according to the development model determined in 101 and the values of the sensitive attributes to obtain the corresponding training samples.
[0041] Step 104: Construct a deep learning probability model for predicting the probability distribution of fluid phases based on the deep learning model, and use the training samples to iteratively train the deep learning probability model until the deep learning probability model converges under the condition (reaching the preset number of iterations).
[0042] Specifically, a probabilistic model is established by fitting fluid phase curves to well logging data using artificial intelligence. This primarily utilizes well logging data for nonlinear fitting and classification, and includes the following:
[0043] First, the sensitive attribute pairs in the well logging data are preprocessed and regularized. The maximum value of each attribute is 1, and the remaining attribute values are represented by the ratio of their values to the maximum value, thus obtaining normalized attribute values in the interval (0,1). The normalized attribute values are then input into the deep learning probabilistic model, which automatically outputs a fluid phase category prediction result, i.e., a vector set representing the generated fluid phase.
[0044] Furthermore, the deep learning probability model is constructed based on a two-dimensional deconvolutional CNN neural network. The specific training calculations are as follows:
[0045] The calibrated well logging data curves are used as training data for network training. The well logging data serves as the input to the network. During training, cross-entropy is mainly used as the loss function for network classification. Since the network needs to perform multi-class classification, the following equation is obtained:
[0046]
[0047] In Equation 1, x i and y i H represents the mixed data of multiple classifications and the real labeled data, respectively. D represents the cross-entropy and the probability. The model data generation network and the seismic data classification and discrimination network are built based on the Tensorflow deep learning framework and controlled by the Python programming language.
[0048] Step 105: Calculate the values of sensitive attribute pairs in the region from the seismic data of the region using methods such as inversion, input the seismic attribute volume into the trained deep learning probabilistic model, and use the deep learning probabilistic model to predict the fluid phase distribution data of the well to be logged.
[0049] In a further embodiment of the present invention, the fixed development mode and relatively stable probability model can be utilized. Based on this, the data of each new development area can be used as training samples to correct the probability model. The corrected probability model can be used for iterative prediction, which can quickly and effectively realize the water and air boundary identification in the development stage.
[0050] Example 2
[0051] In a further embodiment of the present invention, the water-gas boundary prediction method based on the probabilistic model described in Example 1 is used to analyze existing well logging data of deep clastic rocks in the Xinchang tectonic zone of the Sichuan Basin. Table 1 shows the PCA analysis results of known areas in this tectonic zone:
[0052] Table 1
[0053] Main features explain PC1 High gamma, high porosity, low resistivity mudstone PC2 Medium to low resistivity, high water saturation, and low clay content (Gas-bearing) water layer PC3 High resistivity, low water saturation, low clay content air layer PC4 High density, low natural gamma dry layer
[0054] Table 1 shows the relatively sensitive properties obtained from PCA analysis (gamma, porosity, water saturation, density); combined with seismic data such as P-wave velocity and S-wave velocity, pairwise scatter plot clustering and cross-plot analysis are performed. Figures 2-4 A schematic diagram showing the results of partial attribute clustering cross-analysis is provided: Figures 2-3 The property pairs in the model cannot distinguish the fluid phase from the plane, but Figure 4 By using P-wave impedance and natural gamma, the fluid phase corresponding to the predicted deep clastic rocks in the Xinchang tectonic zone of the Sichuan Basin can be separated from the fluid phase on a plane. Therefore, P-wave impedance and natural gamma are the sensitive attribute pairs in this analysis. Figure 5As shown, based on this sensitive attribute, the well logging curve data is calibrated (based on P-wave impedance and natural gamma for electrical phase calibration) to construct training samples, and the deep learning probabilistic model is trained. Furthermore, the trained model is used to predict the fluid phase distribution of other wells to be predicted within this structural region. Figure 6 It can be seen that the model can accurately output the fluid phase distribution data of the well to be tested.
[0055] Example 3
[0056] Figure 7 An electronic device 310 (e.g., a computer server with program execution capabilities) based on a probabilistic model for predicting the water-gas boundary is illustrated according to an exemplary embodiment of the present invention. The device includes at least one processor 311, a power supply 314, and a memory 312 and an input / output interface 313 communicatively connected to the at least one processor 311. The memory 312 stores instructions executable by the at least one processor 311, which, when executed, enable the at least one processor 311 to perform the methods disclosed in any of the foregoing embodiments. The input / output interface 313 may include a display, keyboard, mouse, and USB interface for inputting and outputting data. The power supply 314 provides power to the electronic device 310.
[0057] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media that can store program code, such as mobile storage devices, read-only memory (ROM), magnetic disks, or optical disks.
[0058] When the integrated units of this invention are implemented as software functional units and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROMs, magnetic disks, or optical disks.
[0059] The above description is merely a detailed illustration of specific embodiments of the present invention and is not intended to limit the invention. Various substitutions, modifications, and improvements made by those skilled in the art without departing from the principles and scope of the present invention should be included within the protection scope of the present invention.
Claims
1. A fluid phase prediction method based on well logging data, characterized in that, include: Step 1: Based on geological data, determine the lithofacies classification of a certain area, perform PCA analysis on the existing well logging data in the area, and perform cluster analysis based on the PCA analysis results and seismic attributes to determine the sensitive attribute pairs that can be used to distinguish fluid phases. Based on the aforementioned sensitive attributes, existing well logging data within the region are classified and labeled with lithofacies to construct training samples; Step 2: Iteratively train the deep learning probability model using the training samples until the deep learning probability model satisfies conditional convergence; Step 3: Calculate the values of sensitive attribute pairs of the wells to be logged in the region from the seismic data of the region using the inversion method, input the calculation results into the trained deep learning probability model, and use the deep learning probability model to predict the fluid phase distribution data of the wells to be logged in the region. The deep learning probability model is constructed based on a two-dimensional deconvolutional CNN neural network, and the specific training calculations are as follows: The calibrated well logging data curves are used as training data for network training. The well logging data serves as the input to the network. During training, cross-entropy is mainly used as the loss function for network classification. Since the network needs to perform multi-class classification, the following equation is obtained: In Equation 1, x i and y i These represent mixed data with multiple classifications and real labeled data, respectively, with H representing cross-entropy; D represents probability. The model data generation network and the seismic data classification and discrimination network are built based on the Tensorflow deep learning framework and controlled using the Python programming language.
2. The fluid phase prediction method based on well logging data according to claim 1, characterized in that, Step 1 includes: Step 101: Based on geological data, determine the lithofacies classification of a certain area; Step 102: Correct the existing logging data in the area according to the empirical formula, and select N attributes from the corrected logging data; where N≥5; Step 103: Perform PCA analysis on the lithofacies sample with N attributes as the attributes to be analyzed. Select some attributes from the N attributes based on the PCA analysis results, and perform pairwise cluster cross-analysis in combination with seismic attributes. Find the sensitive attribute pairs that can be used to distinguish fluid phases based on the cluster cross-analysis results. Step 104: Based on the sensitive attributes, perform lithofacies classification and labeling on the existing well logging data in the region to construct training samples.
3. The fluid phase prediction method based on well logging data according to claim 2, characterized in that, In step 101, the historical logging data is corrected according to empirical formulas, including: The Faust formula was used to correct the P-wave velocity and resistivity in the well logging data, the Gardner formula was used to correct the P-wave velocity and density in the well logging data, and the mudstone line formula was used to correct the P-wave velocity and S-wave velocity in the well logging data.
4. The fluid phase prediction method based on well logging data according to claim 1, characterized in that, The seismic properties include P-wave velocity, S-wave velocity, density, and Lamé constant.
5. The fluid phase prediction method based on well logging data according to claim 1, characterized in that, The deep learning probability model is constructed based on a two-dimensional deconvolutional CNN neural network.
6. The fluid phase prediction method based on well logging data according to claim 5, characterized in that, The cross-entropy loss function is used as the loss function for the deep learning probabilistic model.
7. The fluid phase prediction method based on well logging data according to any one of claims 1-6, characterized in that, The sensitive attribute pair is: longitudinal wave impedance and gamma.
8. The fluid phase prediction method based on well logging data according to claim 7, characterized in that, The P-wave impedance and gamma values of the region are calculated from the seismic data using the seismic inversion method. The measured P-wave impedance and gamma values are then normalized and input into the deep learning probability model.
9. A fluid phase prediction device based on well logging data, characterized in that, The method includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor to enable the at least one processor to perform the method of any one of claims 1 to 8.