Physical property parameter inversion method combining domain knowledge and deep network

Abstract

The invention discloses a physical property parameter inversion method combining domain knowledge and a deep network, and relates to the technical field of reservoir prediction. The method comprises the following steps: performing fine well-seismic calibration by using logging, earthquake, geology and other information in combination with seismic geological conditions of a target area; on the basis of model seismic response and actual reservoir earthquake and well data analysis, multiple seismic attributes are extracted, and attribute combinations which are sensitive to lithological characters and physical properties and are not highly correlated with each other are screened out; fusing multiple networks and a progressive hierarchical extraction strategy to construct a multi-task deep network inversion model, and performing model training by taking a rock physical parameter statistical relationship and low-frequency information as constraints; and performing reservoir physical property inversion by using the trained inversion model, and performing constraint improvement on an inversion result in combination with actual exploration data and prior information. Compared with the prior art, the method can solve the problems of lack of data and poor robustness and interpretability to a certain extent, and improves the physical property inversion precision and reliability.

Description

Technical Field

[0001] The present invention relates to the technical field of reservoir prediction and evaluation, and more particularly to a physical property parameter inversion method combining domain knowledge and a deep network. Background Art

[0002] Physical properties, as important parameters for measuring reservoir quality and pore-carrying fluid conditions, are crucial for reservoir characterization and prediction. Currently, the most direct and accurate method is to obtain core samples through drilling and then obtain physical property parameters through experimental measurement. However, direct core sample measurement is costly and inefficient, and the difficulty of obtaining core samples increases with increasing drilling depth. Therefore, direct measurement is limited to certain intervals, making it difficult to obtain physical property parameters for the entire work area. Indirect interpretation, which determines physical property parameters based on the relationship between geological information and wellbore seismic data, is becoming increasingly important in oil and gas exploration and development.

[0003] Directly inverting reservoir properties from well-seismic data can avoid the inversion process being affected by factors such as approximate assumptions in wave theory and rock physics theory, initial models, and accumulated errors from secondary inversion. However, this often faces the challenge of strong nonlinear relationships between the physical properties of interest and the well-seismic data. The rapid development of deep learning provides a more flexible and intelligent means of directly mapping physical properties from well-seismic data. However, in practice, we have found that for data-driven deep learning to achieve greater application and breakthroughs in the field of physical property inversion, it is inseparable from the guidance and constraints of oil and gas reservoir domain knowledge. Therefore, how to leverage domain knowledge and the advantages of deep networks to improve the accuracy of physical property inversion is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the prior art, the present invention provides a physical property parameter inversion method that combines domain knowledge and deep networks, which can improve the accuracy and reliability of reservoir physical property inversion and provide a reliable foundation for oil and gas exploration and development.

[0005] In order to achieve the above objectives and achieve the above technical effects, the present invention provides the following technical solutions.

[0006] A physical property parameter inversion method combining domain knowledge and deep network includes the following steps: (1) Obtaining and pre-processing raw reservoir data including well logging, coring, seismic and geological data in the target area; (2) Using well logging, seismic and geological information and combining it with the seismic and geological conditions of the target area, perform fine well-seismic calibration to accurately calibrate the target layer and perform scale transformation; (3) Extract multiple seismic attribute characteristics related to reservoir lithology and physical properties based on the model seismic response and actual reservoir seismic and well data analysis; (4) Combine the correlation coefficient method and the random forest nonlinear feature selection algorithm to select seismic attributes and screen out seismic attribute combinations that are sensitive to lithology and physical properties but not strongly correlated with each other; (5) Using bidirectional time-domain convolutional networks, bidirectional gated recurrent networks, and frequency-domain attention mechanisms, combined with a progressive layered extraction strategy, a multi-task deep network inversion model is constructed; (6) Construct a labeled training dataset based on the physical parameters at the well point and the preferred seismic attribute combination, and use the statistical relationship of rock physical parameters and low-frequency information as constraints to train the inversion model in stages; (7) Input the preferred seismic attribute combination of the target area into the trained inversion model, output the inversion results of physical parameters, and constrain and improve the inversion results by combining actual exploration data and domain knowledge.

[0007] Furthermore, the preprocessing described in step (1) may include steps such as outlier removal, multi-well consistency correction, noise suppression, amplitude compensation and high-resolution processing.

[0008] Furthermore, the fine borehole seismic calibration described in step (2) is based on the conventional borehole seismic calibration process, and is combined with the geological and seismic conditions of the target area and prior information to perform one-by-one verification and adjustment, so as to accurately establish the connection between the time domain seismic geophysical information and the depth domain geological, well logging and other information.

[0009] Furthermore, the scale transformation described in step (2) is to convert the target data obtained into the time domain through time-depth conversion according to the calibration layer, and after the depth-time conversion, it is smoothed and downsampled to have the same scale as the seismic data.

[0010] Furthermore, the model seismic response described in step (3) is obtained by statistically analyzing the rock physical parameters of the reservoir and non-reservoir layers based on well data, establishing a rock physical model in combination with relevant field knowledge of oil and gas reservoirs, and using the seismic response obtained by seismic forward simulation.

[0011] Furthermore, the multiple seismic attribute features described in step (3) may include but are not limited to amplitude, frequency, energy, and waveform attribute features.

[0012] Furthermore, the correlation coefficient method described in step (4) is used to perform a relatively simple correlation analysis, and the feature importance generated by the random forest is used to characterize the nonlinear linear correlation relationship between variables. The organic combination of the two can obtain a better combination of seismic attributes.

[0013] Furthermore, the bottom layer of the multi-task deep network described in step (5) obtains the input of the tower network by explicitly separating the shared network and the expert network, and then combining the gating unit containing the processing results of the shared and expert networks for a specific task, thereby removing the association between the tower network in a specific task and the expert networks in other tasks, so that different types of expert networks can learn different data characteristics.

[0014] Furthermore, the bidirectional time-domain convolutional network in the deep network inversion model described in step (5) is used to extract the spatiotemporal correlation characteristics of the data, the bidirectional gated recurrent network is used to extract the changing trend with depth and the front-back correlation characteristics, the frequency-domain attention mechanism is used to calculate the correlation function between the preferred seismic attributes in the frequency domain and obtain the weight of each variable through the activation function to reflect the importance of each attribute, and the progressive hierarchical extraction strategy gradually extracts and separates deep semantic knowledge through a progressive routing mechanism.

[0015] Furthermore, the physical property parameters described in step (6) may include but are not limited to reservoir porosity, permeability, and gas saturation.

[0016] Furthermore, the statistical relationship of the rock physical parameters described in step (6) is a mutual fitting relationship between parameters including porosity, permeability and gas saturation obtained by statistical analysis of well data in the target area.

[0017] Furthermore, the low-frequency information in step (6) is obtained from a low-frequency model constructed based on well logging data.

[0018] Furthermore, in step (7), it is necessary to calculate the mean absolute error, root mean square error, etc. between the inversion results and the actual data based on the visual analysis of the inversion results, and to constrain and improve the inversion results in combination with the logging interpretation of the wells drilled in the target area and the single well gas test and production data.

[0019] Compared with the prior art, the present invention has the following beneficial effects: The present invention uses a bidirectional time-domain convolutional network and a bidirectional gated recurrent unit network to construct a feature learning network, and uses a frequency-domain attention mechanism to screen out variables that have an important impact on the inversion of physical parameters. This can overcome the problem of insufficient data feature extraction in existing models and learn more comprehensive feature information from well seismic data.

[0020] The present invention utilizes a progressive hierarchical extraction strategy to solve the negative transfer and seesaw problems in multi-task prediction, optimizes the information sharing mechanism, and can more effectively use the domain knowledge contained in the supervisory signals of related tasks to improve the performance and efficiency of inversion.

[0021] The present invention combines the correlation coefficient method with the random forest nonlinear feature selection algorithm, and can gradually select a better combination of seismic attributes from multiple seismic attributes.

[0022] This paper introduces domain knowledge into the data-driven deep network inversion model and uses theoretical knowledge to guide model design and training. It can to a certain extent solve the problems of lack of labeled data, robustness and poor interpretability, and obtain more accurate and reliable inversion results. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 A flow chart of a physical property parameter inversion method combining domain knowledge and deep network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0024] In order to enable people skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of the specification, not all of the embodiments.

[0025] The embodiment of the present invention provides a physical property parameter inversion method combining domain knowledge and deep network, such as Figure 1 As shown, the following steps are included: Step 101, obtain the original reservoir data including logging, coring, seismic and geological data of the target area and perform preprocessing; specifically: based on the logging, coring, seismic and geological data, the relevant data are subjected to outlier removal, multi-well consistency correction, noise suppression, amplitude compensation and high-resolution processing to obtain high-quality and reliable data.

[0026] Step 102, using logging, seismic and geological information and combining it with the seismic and geological conditions of the target area, fine well-seismic calibration is performed to accurately calibrate the target layer and perform scale transformation. Specifically: on the basis of the conventional well-seismic calibration process, combined with the geological and seismic conditions of the target area and prior information, one by one is checked and adjusted to accurately establish the connection between the time domain seismic geophysical information and the depth domain geological, well logging and other information; on this basis, the target data obtained is converted to the time domain through time-depth conversion according to the calibration layer. After the depth-time conversion, it is smoothed and down-sampled to have the same scale as the seismic data.

[0027] Step 103, based on the analysis of the model seismic response and actual reservoir seismic and well data, extract multiple seismic attribute characteristics related to the reservoir lithology and physical properties. Specifically: based on the statistical analysis of the rock physical parameters of the reservoir and non-reservoir layers based on the well data, a rock physical model is established in combination with relevant field knowledge of oil and gas reservoirs, and the reservoir seismic response corresponding to the rock physical model is obtained by seismic forward simulation. On this basis, the amplitude, frequency, energy and waveform seismic attribute characteristics of the model seismic response and the actual reservoir seismic response are extracted.

[0028] Step 104, combining the correlation coefficient method and the random forest nonlinear feature selection algorithm to perform seismic attribute optimization, and screen out seismic attribute combinations that are sensitive to lithology and physical properties but not strongly correlated with each other; specifically: first, using the correlation coefficient method, a relatively simple correlation analysis is performed on the extracted multiple seismic attribute features, and the relationship between the multiple seismic attribute features and the target physical parameters is preliminarily screened. Then, the feature importance generated by the random forest algorithm is combined to analyze the nonlinear correlation between the multiple seismic attribute features and the physical parameters, and finally, the instantaneous amplitude, total energy, root mean square amplitude, maximum positive curvature, instantaneous Q factor, energy-weighted instantaneous frequency, instantaneous frequency, density and gradient are optimized as the optimal seismic attribute combination.

[0029] Step 105, using a bidirectional time-domain convolutional network, a bidirectional gated recurrent network, and a frequency-domain attention mechanism, combined with a progressive hierarchical extraction strategy to construct a multi-task deep network inversion model; specifically: first, the preferred seismic attribute combination is input into each sub-network of the expert network and the shared network, and the output of the expert network and the shared network are obtained by processing and fusing the sub-networks under different tasks; then, the output data of the expert network and the shared network are fused through a gating unit; finally, the processing results are input into different tower network layers through the gating unit, and the inversion results of porosity, permeability, and saturation are obtained through the tower network.

[0030] The expert network and shared network in the deep network inversion model are hybrid feature learning networks constructed using a bidirectional time-domain convolutional network, a bidirectional gated recurrent network, and a frequency-domain attention mechanism. The bidirectional time-domain convolution is responsible for learning the spatiotemporal correlation characteristics in the data, the bidirectional gated recurrent unit network is responsible for learning the data's changing trend with depth and the previous and next correlation characteristics, and the frequency-domain attention mechanism calculates the correlation function between the preferred seismic attributes in the frequency domain and obtains the weight of each variable through the activation function. The weight reflects the importance of each attribute, enhances useful information, and suppresses redundant information.

[0031] Step 106: construct a labeled training data set based on the physical property parameters at the well point and the preferred seismic attribute combination, and use the statistical relationship of the rock physical parameters and the low-frequency information as constraints to train the inversion model in stages. Specifically, the target physical property parameters include reservoir porosity, permeability, and gas saturation. The well data in the target area are statistically analyzed to obtain the mutual fitting relationship between reservoir porosity, permeability, and gas saturation, which is used as the statistical relationship of the rock physical parameters. A low-frequency model is constructed based on the well logging data in the target area to obtain low-frequency information. On this basis, the constructed inversion model is trained in stages.

[0032] Before using the training data set to train the model, the input data and labels need to be multi-dimensionally normalized, and the mean square error between the predicted label and the true label is used as the loss function for model training; specifically: the size and number of convolution kernels in the bidirectional time domain convolutional network module in the model are 5×1 and 32 respectively, the expansion coefficients are 1, 2, 4, and 8, and the Dropout parameter inside the module is 0.2; the number of cell units in the bidirectional gated recurrent unit network module and the number of neurons in the fully connected network module are both 32; the batch size is 10; the activation function uses LeakyRel, which is a variant of Relu. It has the advantages of Relu and alleviates the problem of neuron death caused by Relu; in addition, the model uses the AdamW optimization algorithm for gradient optimization during training to update parameters such as weights and biases. When the mean square error of the model reaches the minimum or meets the pre-set termination condition, the model is considered to have been trained and the trained inversion model is saved.

[0033] Step 107: Input the preferred seismic attribute combination of the target area into the trained inversion model, output the inversion result of the physical property parameters, and constrain and improve the inversion result in combination with actual exploration data and domain knowledge. Specifically: load the saved inversion model, input the preferred seismic attribute combination into the trained model, the hidden layer receives the data and calculates using the trained model, and passes the calculation results to the output layer to obtain the inversion results of reservoir porosity, permeability and gas saturation and perform visual analysis. On this basis, calculate the mean absolute error, root mean square error, and symmetric mean absolute percentage error between the inversion result and the target data at the well point, and constrain and improve the inversion result in combination with the logging interpretation of the wells drilled in the target area and the single well gas test and production data.

[0034] Those skilled in the art will appreciate that the above embodiments are intended only to illustrate the beneficial effects of the present invention and are not intended to be exhaustive. Any modifications, equivalent variations, improvements, etc. made without departing from the scope and spirit of the described embodiments shall not be excluded from the scope of protection of the present invention.

Claims

1. A physical property parameter inversion method combining domain knowledge and deep network, characterized in that: The following steps are involved: (1) Obtaining and pre-processing raw reservoir data including well logging, coring, seismic and geological data in the target area; (2) Using well logging, seismic and geological information and combining it with the seismic and geological conditions of the target area, perform fine well-seismic calibration to accurately calibrate the target layer and perform scale transformation; (3) Extract multiple seismic attribute characteristics related to reservoir lithology and physical properties based on the model seismic response and actual reservoir seismic and well data analysis; (4) Combine the correlation coefficient method and the random forest nonlinear feature selection algorithm to select seismic attributes and screen out seismic attribute combinations that are sensitive to lithology and physical properties but not strongly correlated with each other; (5) Using bidirectional time-domain convolutional networks, bidirectional gated recurrent networks, and frequency-domain attention mechanisms, combined with a progressive layered extraction strategy, a multi-task deep network inversion model is constructed; (6) Construct a labeled training dataset based on the physical parameters at the well point and the preferred seismic attribute combination, and use the statistical relationship of rock physical parameters and low-frequency information as constraints to train the inversion model in stages; (7) Input the preferred seismic attribute combination of the target area into the trained inversion model, output the inversion results of physical parameters, and constrain and improve the inversion results by combining actual exploration data and domain knowledge.

2. The method for inverting physical property parameters by combining domain knowledge and deep network according to claim 1, characterized in that: The preprocessing in step (1) may include steps such as outlier removal, multi-well consistency correction, noise suppression, amplitude compensation, and high-resolution processing.

3. The method for inverting physical property parameters by combining domain knowledge and deep network according to claim 1, characterized in that: The fine well-seismic calibration in step (2) is based on the conventional well-seismic calibration process, combined with the geological and seismic conditions of the target area and prior information to conduct one-by-one verification and adjustment, and accurately establish the connection between the time domain seismic geophysical information and the depth domain geological and logging information; Scaling is the process of converting the target data into the time domain through time-depth conversion according to the calibration horizon. After the depth-time conversion, the data is smoothed and downsampled to have the same scale as the seismic data.

4. The method for inverting physical property parameters by combining domain knowledge and deep network according to claim 1, characterized in that: The model seismic response in step (3) is obtained by statistically analyzing the rock physical parameters of the reservoir and non-reservoir layers based on well data, establishing a rock physical model in combination with relevant knowledge of oil and gas reservoirs, and using seismic forward simulation to obtain the seismic response; the multiple seismic attribute characteristics may include but are not limited to attribute characteristics such as amplitude, frequency, energy and waveform.

5. The method for inverting physical property parameters by combining domain knowledge and deep network according to claim 1, characterized in that: In step (4), the correlation coefficient method is used to perform a relatively simple correlation analysis, and the feature importance generated by the random forest is used to characterize the nonlinear linear correlation between variables. The organic combination of the two can obtain a better combination of seismic attributes.

6. The method for inverting physical property parameters by combining domain knowledge and deep network according to claim 1, characterized in that: The bottom layer of the multi-task deep network in step (5) is to obtain the input of the tower network by explicitly separating the shared network and the expert network, and then combining the gating unit containing the processing results of the shared and expert networks of a specific task. This removes the association between the tower network in a specific task and the expert networks in other tasks, so that different types of expert networks can learn different data characteristics.

7. The method for inverting physical property parameters by combining domain knowledge and deep network according to claim 1, characterized in that: The bidirectional time-domain convolutional network in the deep network inversion model in step (5) is used to extract the spatiotemporal correlation characteristics of the data, the bidirectional gated recurrent network is used to extract the changing trend with depth and the front-back correlation characteristics, the frequency-domain attention mechanism is used to calculate the correlation function between the preferred seismic attributes in the frequency domain and obtain the weight of each variable through the activation function to reflect the importance of each attribute, and the progressive hierarchical extraction strategy gradually extracts and separates deep semantic knowledge through a progressive routing mechanism.

8. The method for inverting physical property parameters by combining domain knowledge and deep network according to claim 1, characterized in that: The physical parameters in step (6) may include but are not limited to reservoir porosity, permeability, and gas saturation; the statistical relationship of rock physical parameters is the mutual fitting relationship between parameters including porosity, permeability, and gas saturation obtained by statistical analysis of well data in the target area; and the low-frequency information is obtained by constructing a low-frequency model based on well logging data.

Images