Deep learning reservoir parameter prediction method based on rock physical model driving in offshore less-well area
By constructing a deep learning method based on a rock physics model, the problem of uncertainty in reservoir parameter prediction in offshore oil and gas exploration is solved, and high-precision reservoir parameter inversion is achieved, which is applicable to the exploration and development of oil and gas fields in offshore areas with few wells.
Patent Information
- Application Number
- CN202511908431.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-02-27
AI Technical Summary
In offshore oil and gas exploration, the low drilling density and limited logging sample points lead to strong uncertainty in reservoir parameter prediction results. Conventional methods lack rock physical significance and have low prediction accuracy.
A model is constructed based on the rock physics equivalent medium theory. Parameters are calibrated using well logging curves, and forward modeling of the rock physics model is performed to obtain a large sample set. A mapping relationship between elastic parameters and reservoir parameters is established using deep neural network training. Combined with pre-stack elastic parameter inversion, reservoir parameter inversion is achieved.
It effectively expands the sample points with rock physical significance, reduces the uncertainty of reservoir parameter prediction results, improves prediction accuracy, and is suitable for oil and gas field exploration and development in offshore areas with few wells.
Smart Images
Figure CN121578375A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the technical field of seismic exploration of oil and gas, and more specifically, to a deep learning-based method for predicting reservoir parameters in offshore areas with few wells, driven by a rock physics model. Background Technology
[0002] Accurate prediction of reservoir parameters provides strong support for reservoir evaluation and is a crucial step in oil and gas exploration and development. It not only helps explorers pinpoint oil and gas enrichment areas, optimize well location selection and development plan design, and improve exploration success rates and development efficiency, but also provides a scientific basis for oil and gas field management and potential tapping, ensuring long-term stable production and efficient development. As oil and gas exploration and development progresses, the geological conditions become increasingly complex, demanding higher accuracy in reservoir parameter prediction. However, the complexity and heterogeneity of reservoirs, as well as the limitations of prediction technologies themselves, remain the main factors restricting the accuracy of reservoir parameter prediction.
[0003] Conventional reservoir parameter prediction methods establish linear or nonlinear fitting or mapping relationships between well logging elastic parameter curves and reservoir parameter curves, and then apply these fitting or mapping relationships to the elastic parameter volume obtained from pre-stack inversion to obtain the reservoir parameter volume. However, this method is a mathematical statistical representation and lacks rock physical significance. Furthermore, in practical applications, especially in offshore oil and gas exploration and development, drilling density is often low due to drilling and completion cost constraints, resulting in a small number of actual logging sample points. This leads to poor representativeness of the mathematical statistical representation and high uncertainty in the reservoir parameter prediction results. Summary of the Invention
[0004] The purpose of this invention is to overcome the problems of poor rock physical significance in conventional methods and the high uncertainty of reservoir parameter prediction results due to the limited number of sample points in offshore low-well areas. This invention provides a deep learning reservoir parameter inversion method based on a rock physical model driven for offshore low-well areas. This method can greatly expand the number of sample points with certain rock physical significance, reduce the uncertainty of reservoir parameter prediction results, and provide support for the exploration and development of oil and gas fields in offshore low-well areas.
[0005] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a deep learning reservoir parameter prediction method based on rock physics model driven in offshore areas with few wells, comprising the following steps: S1. Constructing a rock physics model based on the rock physics equivalent medium theory; S2. The rock physical model is calibrated using well logging curves to determine the parameters of the rock physical model; S3. Input reservoir parameters within any range, and perform forward modeling calculations of the rock physics model based on the calibrated rock physics model to obtain a large sample set of elastic parameters varying with reservoir parameters. S4. Train the sample set obtained in step S3 based on a deep neural network to obtain the deep neural network mapping relationship between elastic parameters and reservoir parameters. S5. Use seismic data to perform pre-stack elastic parameter inversion to obtain the elastic parameter volume; S6. Apply the elastic parameter volume obtained by pre-stack simultaneous inversion to the deep neural network mapping relationship constructed in step S4 to obtain the reservoir parameter inversion volume.
[0006] Preferably, in step S1, the rock physical equivalent medium theory refers to describing the rock as an equivalent body consisting of a solid mineral framework plus pores and fluids when describing rock characteristics. Step S1 specifically includes the following steps: S11. Calculate the solid mineral framework modulus based on Hill's average theory; S12. Based on Krief theory, pores are added, and the dry rock skeleton modulus is calculated. S13. Calculate the bulk modulus of the mixed fluid using Wood's formula; S14. Using Gassmann theory, fill the pores with the mixed fluid and calculate the modulus of the saturated rock.
[0007] Preferably, in step S11, the Hill average theoretical calculation formula for the solid mineral framework modulus is as follows: (1) Where, in the formula The modulus of a solid mineral framework is divided into bulk modulus. and shear modulus ; It is the first Volume fraction of solid-like minerals It is the first Modulus of solid-like minerals It represents the total number of categories of solid minerals.
[0008] Preferably, in step S12, the formula for calculating the dry rock skeleton modulus using Krief theory is as follows: (2) (3) In the formula , These are the bulk modulus and shear modulus of the dry rock skeleton, respectively. Porosity , These are the bulk modulus and shear modulus of the solid mineral framework, respectively. Preferably, in step S13, the Wood formula for calculating the bulk modulus of the mixed fluid is as follows: (4) In the formula For the bulk modulus of the mixed fluid, For gas saturation, , These are the bulk moduli of gas and water, respectively.
[0009] Preferably, in step S14, the formula for calculating the saturated rock modulus using Gassmann theory is as follows: (5) (6) In the formula , These are the bulk modulus and shear modulus of saturated rock, respectively.
[0010] Preferably, in step S2, the rock physical model calibration specifically includes the following steps: S21. Input the logging elastic parameter curves, including P-wave velocity, S-wave velocity, and density curves; S22. Based on the rock physics model established in step S1, forward model the reservoir parameter curves. S23. Compare the reservoir parameter curves calculated by forward modeling in step S22 with the well logging reservoir parameter curves, and continuously adjust the rock physics model parameters to make them match.
[0011] Preferably, in step S4, the deep neural network includes multiple interconnected hidden layers, in which the first... neurons in the layer The input and output are defined as follows: (7) (8) In the formula For the first neurons in the layer Input, For the first neurons in the layer The output, This is the output vector of the previous layer. Indicates from the first neurons in the layer To the neurons in the layer The weight, Indicates the first Layer neurons The bias, Indicates the first The number of neurons in the layer This represents the total number of neurons. The activation function is defined as follows: (9) In the formula To obtain and The maximum of the two, i.e., if the input parameter If the input parameter is less than or equal to 0, the output is 0. If the value is greater than 0, the output will be equal to the parameter itself.
[0012] Preferably, in step S4, the loss function during the deep neural network training process is defined as: (10) In the formula For loss function, Represents real data. This represents the output of the final layer L. Indicates the first There are N training sample points, where N represents the total number of training sample points.
[0013] Preferably, in step S5, the pre-stack elastic parameter inversion method is a simultaneous pre-stack inversion, and the objective function for the inversion is: (11) in, It is the objective function for the reflection coefficient. It is the objective function of the seismic data. It is the objective function for low-frequency change trends. It is the objective function for spatial change trend. It is the objective function of elastic impedance. yes Constrain the objective function, It is the objective function of the P-wave and S-wave velocity relationship. yes Objective function.
[0014] Compared with the prior art, the beneficial effects of the present invention are as follows: The present invention is based on the rock physics equivalent medium theory, and performs forward modeling calculations of the rock physics model after calibration, obtaining a large number of sample sets of elastic parameters that vary with reservoir parameters in the study area, laying a data foundation for deep neural network training and learning, solving the problem of insufficient training and learning of deep neural networks caused by the small number of sample points in offshore areas with few wells, and effectively improving the accuracy of reservoir parameter prediction. Attached Figure Description
[0015] Figure 1 This is a flowchart of the deep learning reservoir parameter inversion method based on rock physics model driven by the present invention for offshore areas with few wells; Figure 2 Comparison diagram of rock physical model calibration in the embodiments; Figure 3 This is a well profile diagram obtained by porosity inversion using a conventional method in the embodiments. Figure 4 This is a well profile diagram for porosity inversion in this embodiment; Figure 5 This is a cross-plot of porosity inverted using conventional methods and measured porosity in the examples. Figure 6 This is a cross-plot of the porosity inverted by the method in this embodiment and the measured porosity. Detailed Implementation
[0016] The present invention will be further described below with reference to specific embodiments.
[0017] Example 1 like Figure 1 As shown, a deep learning-based method for predicting reservoir parameters in offshore areas with few wells, driven by a rock physics model, includes the following steps: S1. Constructing a rock physics model based on the rock physics equivalent medium theory; S2. The rock physical model is calibrated using well logging curves to determine the parameters of the rock physical model; S3. Input reservoir parameters within any range, and perform forward modeling calculations of the rock physics model based on the calibrated rock physics model to obtain a large sample set of elastic parameters varying with reservoir parameters. S4. Train the sample set obtained in step S3 based on a deep neural network to obtain the deep neural network mapping relationship between elastic parameters and reservoir parameters. S5. Use seismic data to perform pre-stack elastic parameter inversion to obtain the elastic parameter volume; S6. Apply the elastic parameter volume obtained by pre-stack simultaneous inversion to the deep neural network mapping relationship constructed in step S4 to obtain the reservoir parameter inversion volume.
[0018] Example 2 Example 2 uses a gas field in the Yinggehai Basin of the South China Sea as an example. Accurate prediction of reservoir parameters provides strong support for gas field reservoir evaluation and is a crucial step in oil and gas exploration and development. However, due to the cost and risk constraints of offshore drilling, drilling density is often low. Under these circumstances, conventional reservoir parameter prediction methods suffer from high uncertainty and low accuracy, hindering the progress of gas field exploration and development. For example... Figure 1The diagram shows a flowchart of a deep learning-based reservoir parameter inversion method driven by a rock physics model for offshore areas with few wells, provided by this invention. This method can solve the problem of insufficient training and learning of deep neural networks caused by the small number of sample points in offshore areas with few wells, effectively improving the accuracy of reservoir parameter prediction and laying a good foundation for reservoir evaluation. The method includes the following steps: Step S1: Construct a rock physics model based on the rock physics equivalent medium theory; Step S2: Use logging curves (elastic parameter curves and reservoir parameter curves) to calibrate the rock physical model and determine the rock physical model parameters; Step S3: Input reservoir parameters within any range, and perform forward modeling calculations of the rock physics model based on the calibrated rock physics model to obtain a large sample set of elastic parameters that vary with reservoir parameters; Step S4: Train the large sample set obtained in step S3 based on the deep neural network to obtain the deep neural network mapping relationship between elastic parameters and reservoir parameters. Step S5: Use seismic data to perform pre-stack elastic parameter inversion to obtain the elastic parameter volume; Step S6: Apply the elastic parameter volume obtained by pre-stack simultaneous inversion to the deep neural network mapping relationship constructed in step S4 to obtain the reservoir parameter inversion volume.
[0019] It should be noted that the logging curves in step S2 are elastic parameter curves and reservoir parameter curves. The elastic parameter curves include P-wave velocity, S-wave velocity, and density curves, while the reservoir parameter curves include porosity, clay content, and gas saturation curves. The reservoir parameter inversion volume in step S6 is a porosity volume, a clay content volume, and a gas saturation volume.
[0020] Figure 3 To invert well profiles based on porosity predicted using conventional methods, Figure 4 To utilize the porosity predictions obtained using this method to invert well profiles, compared to conventional methods, the porosity predictions obtained using this method show better agreement with the actual drilling data (the well column in the figure represents the porosity curve, and the color scale matches the inversion body). For example, in wells W4 and W5, the actual drilling porosity is good (represented by warm colors on the color scale), while the porosity prediction results using conventional methods are poor (represented by cool colors on the color scale). The porosity prediction results obtained using this method are better (represented by warm colors on the color scale), showing a better match with the actual drilling results. In addition, the porosity prediction results obtained using this method have a higher signal-to-noise ratio than conventional methods. This is attributed to the fact that this method has more sample points, allowing for more thorough training and learning of the deep neural network, resulting in better convergence of the inversion process.
[0021] Figure 5 The results of cross-plotting the predicted porosity using conventional methods and the measured porosity at all well points are presented. Figure 6The intersection of the predicted porosity and the measured porosity at all well points is presented. Theoretically, if the prediction accuracy is 100%, meaning the predicted results are completely consistent with the measured results, then the intersection point of the predicted and measured results should all fall within the intersection point. Figure 4 On a straight line at a 15-degree angle (with consistent x and y coordinates), in practical production applications, the prediction accuracy cannot reach 100%. The closer the intersection point is to the 45-degree angle line, the higher the prediction accuracy; conversely, the further away the intersection point is from the 45-degree angle line, the lower the prediction accuracy. (Comparison) Figure 5 and Figure 6 It can be seen that, overall, this method converges to a straight line at a 45-degree angle more closely than the conventional method, indicating that the prediction accuracy of this method is better than that of the conventional method, and it can better guide the subsequent reservoir parameter evaluation work in the study area.
[0022] In this embodiment, in step S1, the rock physical equivalent medium theory refers to describing the rock as an equivalent body consisting of a solid mineral framework plus pores and fluids when describing rock characteristics. Step S1 specifically includes the following steps: S11. Calculate the solid mineral framework modulus based on Hill's average theory; S12. Based on Krief theory, pores are added, and the dry rock skeleton modulus is calculated. S13. Calculate the bulk modulus of the mixed fluid using Wood's formula; S14. Using Gassmann theory, fill the pores with the mixed fluid and calculate the modulus of the saturated rock.
[0023] In addition, in step S11, the formula for calculating the solid mineral framework modulus using Hill's average theory is as follows: (1) Where, in the formula The modulus of a solid mineral framework is divided into bulk modulus. and shear modulus ; It is the first Volume fraction of solid-like minerals It is the first Modulus of solid-like minerals It represents the total number of categories of solid minerals.
[0024] In step S12, the formula for calculating the dry rock skeleton modulus using Krief theory is as follows: (2) (3) In the formula , These are the bulk modulus and shear modulus of the dry rock skeleton, respectively. Porosity , These are the bulk modulus and shear modulus of the solid mineral framework, respectively. In addition, in step S13, the Wood formula for calculating the bulk modulus of the mixed fluid is as follows: (4) In the formula For the bulk modulus of the mixed fluid, For gas saturation, , These are the bulk moduli of gas and water, respectively.
[0025] In step S14, the formula for calculating the saturated rock modulus using Gassmann theory is as follows: (5) (6) In the formula , These are the bulk modulus and shear modulus of saturated rock, respectively.
[0026] In addition, in step S2, the rock physics model is calibrated using logging curves (elastic parameter curves and reservoir parameter curves): First, the logging elastic parameter curves of a real drilled well are input, including P-wave velocity, S-wave velocity, and density curves; then, the reservoir parameter curves are calculated using forward modeling based on the rock physics model established in step S1; finally, the forward-modeled reservoir parameter curves and the measured reservoir parameter curves of the real drilled well are compared, and the rock physics model parameters (modulus and density of various solid minerals) are continuously adjusted to achieve a match. Figure 2 As shown, the forward modeling curves of the rock physics model agree well with the measured curves, indicating that the calibration results are reliable. It should be noted that the adjusted rock physics model parameters are the modulus and density of various solid minerals.
[0027] In step S3 of this embodiment, reservoir parameters in any range from 0 to 1 are input, and forward modeling calculations of the rock physics model are performed based on the calibrated rock physics model to obtain a large sample set of elastic parameters that vary with reservoir parameters.
[0028] Example 3 The difference from Embodiment 2 is that in this embodiment, step S4 trains a deep neural network on the large sample set obtained in step S3 to obtain the deep neural network mapping relationship between elastic parameters and reservoir parameters. The deep neural network consists of multiple interconnected hidden layers (10 hidden layers in this embodiment). In the network, the first... neurons in the layer The input and output are defined as follows: (7) (8) In the formula For the first neurons in the layer Input, For the first neurons in the layer The output, This is the output vector of the previous layer. Indicates from the first neurons in the layer To the neurons in the layer The weight, Indicates the first Layer neurons The bias, Indicates the first The number of neurons in the layer This represents the total number of neurons. The activation function is defined as follows: (9) In the formula To obtain and The maximum of the two, i.e., if the input parameter If the input parameter is less than or equal to 0, the output is 0. If the value is greater than 0, the output will be equal to the parameter itself.
[0029] Furthermore, in step S4, the loss function during the deep neural network training process is defined as: (10) In the formula For loss function, Represents real data. This represents the output of the final layer L. Indicates the first There are 1800000 training sample points, where N represents the total number of training sample points. In this embodiment, N is 1800000. In step S5 of this embodiment, pre-stack elastic parameter inversion is performed using seismic data to obtain the elastic parameter volume (including P-wave impedance, P-wave / S-wave velocity ratio, and density). The pre-stack elastic parameter inversion method is simultaneous pre-stack inversion, and the objective function for the inversion is: (11) in, It is the objective function for the reflection coefficient. It is the objective function of the seismic data. It is the objective function for low-frequency change trends. It is the objective function for spatial change trend. It is the objective function of elastic impedance. yes Constrain the objective function, It is the objective function of the P-wave and S-wave velocity relationship. yes Objective function.
[0030] Finally, the elastic parameter volume obtained by pre-stack simultaneous inversion is applied to the deep neural network mapping relationship constructed in step S4 to obtain the reservoir parameter inversion volume.
[0031] In the specific implementation of the above embodiments, the technical features can be combined in any non-contradictory way. For the sake of brevity, not all possible combinations of the above technical features are described. However, as long as the combination of these technical features is not contradictory, it should be considered to be within the scope of this specification.
[0032] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art can make other variations or modifications based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the claims of the present invention.
Claims
1. A deep learning-driven method for predicting reservoir parameters in offshore areas with few wells, characterized in that: Includes the following steps: S1. Constructing a rock physics model based on the rock physics equivalent medium theory; S2. The rock physical model is calibrated using well logging curves to determine the parameters of the rock physical model; S3. Input reservoir parameters within any range, and perform forward modeling calculations of the rock physics model based on the calibrated rock physics model to obtain a large sample set of elastic parameters varying with reservoir parameters. S4. Train the sample set obtained in step S3 based on a deep neural network to obtain the deep neural network mapping relationship between elastic parameters and reservoir parameters. S5. Use seismic data to perform pre-stack elastic parameter inversion to obtain the elastic parameter volume; S6. Apply the elastic parameter volume obtained by pre-stack simultaneous inversion to the deep neural network mapping relationship constructed in step S4 to obtain the reservoir parameter inversion volume.
2. The deep learning reservoir parameter prediction method based on rock physics model driven for offshore low-well areas according to claim 1, characterized in that, In step S1, the rock physical equivalent medium theory refers to describing rocks as an equivalent body consisting of a solid mineral framework plus pores and fluids when describing rock characteristics. Step S1 specifically includes the following steps: S11. Calculate the solid mineral framework modulus based on Hill's average theory; S12. Based on Krief theory, pores are added, and the dry rock skeleton modulus is calculated. S13. Calculate the bulk modulus of the mixed fluid using Wood's formula; S14. Using Gassmann theory, fill the pores with the mixed fluid and calculate the modulus of the saturated rock.
3. The deep learning-driven reservoir parameter prediction method based on a rock physics model for offshore low-well areas according to claim 2, characterized in that, In step S11, the Hill average theoretical calculation formula for the solid mineral framework modulus is as follows: （1） Where, in the formula The modulus of a solid mineral framework is divided into bulk modulus. and shear modulus ; It is the first Volume fraction of solid-like minerals It is the first Modulus of solid-like minerals It represents the total number of categories of solid minerals.
4. The deep learning-driven reservoir parameter prediction method based on a rock physics model for offshore low-well areas according to claim 2, characterized in that, In step S12, the formula for calculating the dry rock skeleton modulus using Krief theory is as follows: （2） （3） In the formula , These are the bulk modulus and shear modulus of the dry rock skeleton, respectively. Porosity , These are the bulk modulus and shear modulus of the solid mineral skeleton, respectively.
5. The deep learning-driven reservoir parameter prediction method based on a rock physics model for offshore low-well areas according to claim 2, characterized in that, In step S13, the Wood formula is used to calculate the bulk modulus of the mixed fluid as follows: （4） In the formula For the bulk modulus of the mixed fluid, For gas saturation, , These are the bulk moduli of gas and water, respectively.
6. The deep learning-driven reservoir parameter prediction method based on a rock physics model for offshore low-well areas according to claim 2, characterized in that, In step S14, the formula for calculating the saturated rock modulus using Gassmann theory is as follows: （5） （6） In the formula , These are the bulk modulus and shear modulus of saturated rock, respectively.
7. The deep learning-driven reservoir parameter prediction method based on a rock physics model for offshore low-well areas according to claim 1, characterized in that, In step S2, the rock physical model calibration specifically includes the following steps: S21. Input the logging elastic parameter curves, including P-wave velocity, S-wave velocity, and density curves; S22. Based on the rock physics model established in step S1, forward model the reservoir parameter curves. S23. Compare the reservoir parameter curves calculated by forward modeling in step S22 with the well logging reservoir parameter curves, and continuously adjust the rock physics model parameters to make them match.
8. The deep learning-driven reservoir parameter prediction method based on a rock physics model for offshore low-well areas according to claim 1, characterized in that, In step S4, the deep neural network includes multiple interconnected hidden layers. In the network, the first... neurons in the layer The input and output are defined as follows: （7） （8） In the formula For the first neurons in the layer Input, For the first neurons in the layer The output, This is the output vector of the previous layer. Indicates from the first neurons in the layer To the neurons in the layer The weight, Indicates the first Layer neurons The bias, Indicates the first The number of neurons in the layer This represents the total number of neurons. The activation function is defined as follows: （9） In the formula To obtain and The maximum of the two, i.e., if the input parameter If the input parameter is less than or equal to 0, the output is 0. If the value is greater than 0, the output will be equal to the parameter itself.
9. The deep learning-driven reservoir parameter prediction method based on a rock physics model for offshore low-well areas according to claim 1, characterized in that, In step S4, the loss function during the deep neural network training process is defined as: （10） In the formula For loss function, Represents real data. This represents the output of the final layer L. Indicates the first There are N training sample points, where N represents the total number of training sample points.
10. The deep learning reservoir parameter prediction method based on rock physics model driven for offshore low-well areas according to claim 1, characterized in that, In step S5, the pre-stack elastic parameter inversion method is a simultaneous pre-stack inversion, and the objective function for the inversion is: （11） in, It is the objective function for the reflection coefficient. It is the objective function of the seismic data. It is the objective function for low-frequency change trends. It is the objective function for spatial change trend. It is the objective function of elastic impedance. yes Constrain the objective function, It is the objective function of the P-wave and S-wave velocity relationship. yes Objective function.