Lithology prediction method based on prior information constraint and time-frequency feature fusion

Abstract

The invention discloses a lithology prediction method based on prior information constraint and time-frequency feature fusion, and relates to the technical field of reservoir prediction. The method comprises the following steps: performing well seismic calibration according to seismic geological conditions of a target area; seismic inversion and seismic attribute extraction are carried out on the pre-stack and post-stack seismic data, and seismic multi-information is determined in combination with seismic attribute optimization; building a time domain stratigraphic framework by using seismic horizon information and dividing seismic facies through seismic waveform clustering analysis; jointly applying a multi-scale convolutional network, a bidirectional Mama network, a bidirectional long-short-term memory network and a frequency domain attention mechanism to construct a lithology prediction model, and training by adopting a generative adversarial mode; and performing reservoir lithology prediction by using the trained prediction model, and performing constraint improvement on a prediction result in combination with actual exploration data. Compared with the prior art, the lithology prediction precision and reliability can be improved, and oil-gas exploration and development can be guided.

Description

Technical Field

[0001] The present invention relates to the technical field of reservoir prediction and evaluation, and in particular to a lithology prediction method integrating prior information constraints with time-frequency characteristics. Background Art

[0002] Lithology prediction is an essential component of reservoir characterization in oil and gas exploration and development, and is crucial for subsequent reservoir characterization and evaluation. With the continuous advancement of oil and gas exploration and development, reservoir types and lithologic variations are becoming increasingly complex and diverse, making accurate and reliable lithologic prediction extremely difficult. Complex lithologic prediction has become a new challenge. Traditional lithologic prediction methods rely primarily on expert experience and knowledge, making it difficult to accurately identify complex lithologies and susceptible to subjective factors.

[0003] In recent years, numerous classic machine learning algorithms, such as support vector machines, decision trees, and random forests, as well as deep learning algorithms such as deep belief networks, convolutional neural networks, and long-short-term memory networks, have been applied to lithologic prediction tasks, achieving superior results compared to traditional methods. However, these methods still struggle to fully mine and utilize the seismic and geological information implicit in seismic data, and rarely consider adding prior information constraints, making them inadequate for the exploration of complex oil and gas reservoirs. Summary of the Invention

[0004] In response to the above-mentioned deficiencies in the existing technology, the present invention provides a lithology prediction method that integrates prior information constraints with time-frequency characteristics to solve the problem that existing models are difficult to fully mine and utilize the seismic geological information implicit in seismic data, improve the accuracy and reliability of complex lithology prediction, and thus reduce the risks of oil and gas exploration and development.

[0005] The purpose of the present invention can be achieved through the following technical solutions.

[0006] A lithology prediction method integrating prior information constraints with time-frequency characteristics includes the following steps: S1. Acquire well logging and seismic data, and perform well-seismic calibration based on the seismic geology of the target area, accurately establishing the connection between time-domain seismic geophysical information and depth-domain geological and well logging information; S2. Based on the statistical analysis of well data and rock physics analysis results, seismic inversion and seismic attribute extraction are performed on pre-stack and post-stack seismic data to obtain seismic parameter inversion results and multiple seismic attributes related to reservoir lithology; S3. Using the recursive feature elimination method and combining expert experience to select seismic attributes, we screen out seismic attribute combinations that are sensitive to rock properties and have weak correlation with each other, and then combine them with pre-stack and post-stack seismic data and seismic inversion results to form seismic multi-information; S4. Use seismic horizon information to build a time-domain stratigraphic framework, use a deep embedded self-organizing feature map network to cluster seismic waveforms, and analyze the clustering results in combination with well logging data and geological prior information of the target area to divide seismic phases; S5. Combined application of multi-scale convolutional networks, bidirectional Mamba networks, bidirectional long short-term memory networks, and frequency-domain attention mechanism algorithms to construct a time-frequency feature fusion lithology prediction model that can effectively extract both time-domain and frequency-domain features of seismic data. S6. Based on the well logging interpretation and production data of the drilled wells, a sample dataset is constructed by labeling the wellside seismic multi-information with corresponding lithologic labels using one-hot encoding. Simultaneously, prior information such as stratigraphic framework and seismic phases is used as constraints, and a generative adversarial approach is used for model training. S7. Input the target area seismic data and stratigraphic framework information into the trained prediction model to obtain the reservoir lithology distribution prediction results, and constrain and improve the prediction results by combining actual exploration data and prior information.

[0007] Furthermore, the seismic inversion described in step S2 may include but is not limited to pre-stack and post-stack seismic inversion methods such as post-stack sparse pulse inversion, pre-stack simultaneous inversion and pseudo-acoustic inversion.

[0008] Furthermore, the seismic inversion results described in step S2 may include but are not limited to seismic inversion parameters such as longitudinal wave impedance, density, and gamma-ray pseudo-acoustic impedance.

[0009] Furthermore, the seismic attributes described in step S2 may include, but are not limited to, seismic attribute features such as amplitude, instantaneous, spectrum, wavelet coefficient, and waveform structure.

[0010] Furthermore, the recursive feature elimination method used in step S3 is a feature selection method based on random forest, which realizes the importance ranking and selection of features by training the random forest model and gradually eliminating the least important features, and finally selects a combination of seismic attributes that are sensitive to rock properties and have low correlation with each other.

[0011] Furthermore, the seismic multi-information described in step S3 may include but is not limited to sensitive seismic attributes such as seismic attenuation gradient, instantaneous amplitude, root mean square amplitude, pre-stack and post-stack seismic data, and seismic inversion parameters such as longitudinal wave impedance, density and gamma pseudo-acoustic impedance.

[0012] Furthermore, the deep embedded self-organizing map network adopted in step S4 is a seismic phase visualization analysis method that simultaneously completes the self-encoding feature extraction process and the self-organizing map feature clustering, and improves the stability of the seismic phase classification results by adding sparse constraints, so that the visualization results can show more details.

[0013] Furthermore, the lithologic prediction model described in step S5 is composed of time-domain feature learning, frequency-domain feature learning, time-frequency domain feature fusion, and a fully connected network module. Specifically, the model first combines the advantages of a multi-scale convolutional network and a bidirectional Mamba network to extract the time-domain feature information of the data. This information is then used as input to a bidirectional long-short-term memory network to mine the contextual correlation characteristics of the data's trend changes. Simultaneously, the seismic data is converted from the time domain to the frequency domain, and a frequency-domain attention mechanism is employed in the frequency domain to effectively capture frequency-domain features. Finally, the extracted time-domain and frequency-domain features are fused and mapped to the output through a fully connected network module.

[0014] Furthermore, the generative adversarial approach adopted in step S6 for model training means that during the training process, the generator generates samples similar to real data by learning the characteristics of data distribution; the discriminator is trained using real data and data generated by the generator to distinguish between the data generated by the generator and real data, and strengthens the generator's ability to generate higher quality samples through back propagation. The training process ends when the samples generated by the generator cannot be accurately distinguished by the discriminator.

[0015] Compared with the prior art, the present invention has at least the following beneficial effects: The present invention combines the advantages of multi-scale convolutional networks, bidirectional Mamba networks, bidirectional long short-term memory networks and frequency domain attention mechanisms to construct a lithologic prediction model, which can effectively extract the time domain characteristics and frequency domain characteristic information of seismic data and integrate the vertical trend information of the strata. At the same time, the present invention adds prior information such as stratigraphic framework and seismic phases to the lithologic prediction model, which can enable seismic multi-information and lithology to be learned and trained in the same layer and similar phase belt, thereby improving the accuracy and reliability of lithologic prediction. In addition, the present invention adopts a generative adversarial approach to train the lithologic prediction model, which can gradually generate high-quality samples through the synergistic effect of the generator and the discriminator, thereby alleviating the problem of the scarcity of lithologic labels and improving the training effect and the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 Flow chart of the method of the present invention. DETAILED DESCRIPTION

[0017] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0018] like Figure 1 As shown, the present invention provides a lithology prediction method that integrates prior information constraints with time-frequency features. The specific implementation steps are as follows: Step S1: Acquire well logging and seismic data, and perform well-seismic calibration based on the seismic geology of the target area to accurately establish the connection between time-domain seismic geophysical information and depth-domain geological and well logging information; Specifically, the well logging and seismic data of the target area are collected and organized, and the well-seismic calibration is carried out using the conventional well-seismic calibration process. The data are checked and adjusted one by one in combination with the geological and seismic conditions of the target area. The influence of wavelets is fully considered in the calibration results, and the method of extracting the optimal wavelet of multiple wells is used to maintain the consistency of wavelet parameters, thereby accurately establishing the connection between time-domain seismic geophysical information and depth-domain geological, logging and other information.

[0019] Step S2: Based on the statistical analysis of well data and the results of rock physics analysis, seismic inversion and seismic attribute extraction are performed on the pre-stack and post-stack seismic data to obtain seismic parameter inversion results and multiple seismic attributes related to reservoir lithology; Specifically, statistical analysis and rock physical analysis are performed on the actual well data in the target area. According to the analysis results, three parameters commonly used for lithology classification, namely gamma value, longitudinal wave impedance and density, are selected. The inversion results of gamma value, longitudinal wave impedance and density are obtained by performing post-stack sparse pulse inversion, pre-stack simultaneous inversion and pseudo-acoustic inversion on the pre-stack and post-stack seismic data. At the same time, seismic attributes are extracted from the pre-stack and post-stack seismic data to obtain seismic attribute characteristics of amplitude, instantaneous, spectrum, wavelet coefficient and waveform structure.

[0020] Step S3: using a recursive feature elimination method combined with expert experience to select seismic attributes, screening out seismic attribute combinations that are sensitive to rock properties and have weak correlation with each other, and combining pre-stack and post-stack seismic data and seismic inversion results to form seismic multi-information; Specifically, the importance of features is ranked and selected by training a random forest model and gradually eliminating the least important features. Furthermore, combined with expert experience, seismic attributes such as seismic attenuation gradient, instantaneous amplitude, and root mean square amplitude that are sensitive to rock properties are selected. On this basis, the seismic inversion results of gamma value, longitudinal wave impedance and density parameters, the selected seismic attributes, and pre-stack and post-stack seismic data are combined to form seismic multi-information.

[0021] Step S4: Using the seismic horizon information to build a time-domain stratigraphic framework, using a deep embedded self-organizing feature mapping network to cluster seismic waveforms, and combining the well logging data and geological prior information of the target area to analyze the clustering results and divide the seismic phases; Specifically, the seismic data is input into a deep embedded self-organizing map network for unsupervised seismic waveform clustering, and it is divided into categories with different structures according to the similarity of the seismic waveforms. On this basis, the clustering results are analyzed in combination with the well logging data and geological prior information of the target area to obtain the seismic phase division results.

[0022] Step S5: jointly applying a multi-scale convolutional network, a bidirectional Mamba network, a bidirectional long short-term memory network, and a frequency domain attention mechanism algorithm to construct a time-frequency feature fusion lithology prediction model that can simultaneously and effectively extract the time domain and frequency domain features of seismic data; Specifically, the advantages of multi-scale convolutional networks and bidirectional Mamba networks are first combined to extract the time domain feature information of the data, which is then used as the input of a bidirectional long short-term memory network to mine the before-after correlation characteristics of the data trend changes; at the same time, the seismic data is converted from the time domain to the frequency domain, and the frequency domain attention mechanism is used in the frequency domain to effectively capture the frequency domain features; then, the extracted time domain features and frequency domain features are fused and mapped to the output through a fully connected network module.

[0023] Step S6: Based on the well logging interpretation and production data of the drilled wells, the sample dataset is constructed by labeling the corresponding lithologic labels for the near-well seismic multi-information using one-hot encoding. At the same time, the stratigraphic framework and seismic phase prior information are used as constraints, and the model is trained using a generative adversarial approach; Specifically, during the training process, the generator generates samples similar to real data by learning the characteristics of data distribution; the discriminator is trained using real data and data generated by the generator to distinguish the data generated by the generator from real data, and strengthens the generator's ability to generate higher quality samples through backpropagation. When the samples generated by the generator cannot be accurately distinguished by the discriminator, the training process is considered to be over and the trained model is saved.

[0024] Specifically, during the training process, the activation function is LeakyRelu, which is a variant of Relu and has the advantages of Relu; the optimization algorithm is AdamW; the convolution kernel sizes are 3, 5, and 7 respectively, the learning rate is 0.001, the dropout rate is 0.2, and other model parameters such as the hidden layer neurons are set to 64.

[0025] Step S7: Input the target area seismic data and stratigraphic framework information into the trained prediction model to obtain the reservoir lithology distribution prediction results, and constrain and improve the prediction results in combination with actual exploration data and prior information.

[0026] Specifically, the saved trained prediction model is loaded, and the seismic data and stratigraphic framework information of the area to be predicted are input into the trained prediction model. The hidden layer receives the data and uses the trained model for calculation, and the calculation results are passed to the output layer to obtain the reservoir lithology distribution prediction results and perform visual analysis. On this basis, the prediction results are constrained and improved according to the logging interpretation of the wells drilled in the target area and the single well gas test and production data.

[0027] Those skilled in the art should understand that the above content is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent variations, and improvements to the above embodiments that do not depart from the technical solution of the present invention and are based on the technical essence of the present invention are within the scope of protection of the present invention.

Claims

1. A lithologic prediction method integrating prior information constraints with time-frequency features, characterized in that: The following steps are involved: (1) Obtain well logging and seismic data, and perform well-seismic calibration based on the seismic geological conditions of the target area, accurately establishing the connection between time-domain seismic geophysical information and depth-domain geological and well logging information; (2) Based on the statistical analysis of well data and the results of rock physics analysis, seismic inversion and seismic attribute extraction are performed on pre-stack and post-stack seismic data to obtain seismic parameter inversion results and multiple seismic attributes related to reservoir lithology; (3) Using the recursive feature elimination method and combining it with expert experience to select seismic attributes, we can select attribute combinations that are sensitive to rock properties and have weak correlation with each other, and then combine them with pre-stack and post-stack seismic data and seismic inversion results to form seismic multi-information; (4) Using seismic horizon information to build a time-domain stratigraphic framework, a deep embedded self-organizing feature mapping network is used to cluster seismic waveforms, and the clustering results are analyzed and divided into seismic phases in combination with well logging data and geological prior information of the target area; (5) By combining the multi-scale convolutional network, bidirectional Mamba network, bidirectional long short-term memory network, and frequency domain attention mechanism algorithm, a time-frequency feature fusion lithology prediction model was constructed that can effectively extract both time and frequency domain features of well seismic data. (6) Based on the well logging interpretation and production data of the drilled wells, a sample data set is constructed by using one-hot encoding to label the corresponding lithologic labels for the wellside seismic multi-information. At the same time, the stratigraphic framework and seismic phase prior information are used as constraints, and the model is trained using a generative adversarial approach. (7) Input the target area seismic data and stratigraphic framework information into the trained prediction model to obtain the reservoir lithology distribution prediction results, and then constrain and improve the prediction results by combining actual exploration data and prior information.

2. The lithology prediction method integrating prior information constraints with time-frequency features according to claim 1 is characterized by: The seismic inversion described in step (2) may include but is not limited to post-stack sparse pulse inversion, pre-stack simultaneous inversion and pseudo-acoustic inversion; the seismic inversion results may include but are not limited to longitudinal wave impedance, density and gamma pseudo-acoustic impedance; the seismic attributes may include but are not limited to amplitude, instantaneous, spectrum, wavelet coefficient and waveform structure attributes.

3. The lithology prediction method integrating prior information constraints with time-frequency characteristics according to claim 1 is characterized by: The lithologic prediction model described in step (5) combines the advantages of multi-scale convolutional networks and bidirectional Mamba networks to extract the time domain feature information of the data, and then uses it as the input of the bidirectional long short-term memory network to mine the before and after correlation characteristics of the trend change of the data; at the same time, the seismic data is converted from the time domain to the frequency domain, and the frequency domain attention mechanism is used in the frequency domain to effectively capture the frequency domain features; The extracted time domain and frequency domain features are then fused and output through a fully connected network mapping.

4. The lithology prediction method integrating prior information constraints with time-frequency features according to claim 1 is characterized by: The generative adversarial training method described in step (6) means that during the training process, the generator generates samples similar to the real data by learning the data distribution characteristics; The discriminator is trained using real data and data generated by the generator to distinguish between the data generated by the generator and real data, and strengthens the generator's ability to generate higher quality samples through backpropagation. The training process ends when the samples generated by the generator cannot be accurately distinguished by the discriminator.

Images