Three-dimensional crustal stress intelligent prediction method fusing physical information

Abstract

The invention discloses a three-dimensional crustal stress intelligent prediction method fusing physical information, and relates to the technical field of reservoir numerical simulation. The method comprises the following steps: acquiring three-dimensional crustal stress, logging-while-drilling parameters related to the three-dimensional crustal stress and lithologic parameters of developed reservoirs in the same block, matching the three-dimensional crustal stress, the logging-while-drilling parameters and the lithologic parameters, and then performing data preprocessing on the three-dimensional crustal stress, the logging-while-drilling parameters and the lithologic parameters; on the basis of a CAE self-encoder, feature extraction is conducted on the preprocessed logging-while-drilling parameters and the preprocessed core parameters, and parameters obtained after dimension reduction are obtained; based on an AVE auto-encoder, performing data enhancement on the parameters after dimension reduction and the preprocessed three-dimensional crustal stress to generate more samples; based on a PINNs neural network, training is carried out in combination with an Adam optimizer, and a trained model is used as a prediction model. According to the method provided by the invention, the prediction precision and robustness are improved when the method faces heterogeneous and anisotropic reservoirs, so that accurate crustal stress prediction is provided for efficient development of compact conglomerate reservoirs.

Description

Technical Field

[0001] The invention relates to the technical field of reservoir numerical simulation, and in particular to a three-dimensional geostress intelligent prediction method integrating physical information. Background Art

[0002] Due to its significant anisotropy and complex geological conditions, tight conglomerate reservoirs transfer stress in different ways in different directions, increasing the complexity and uncertainty of the stress field. At the same time, the heterogeneity and structural characteristics inside the reservoir lead to local stress concentration, further exacerbating the unevenness of stress distribution. Many traditional models are based on static conditions or simplified assumptions, and it is difficult to accurately capture the stress evolution under dynamically changing environments. Especially when dealing with complex geological conditions, it brings major challenges to the prediction of ground stress.

[0003] Existing machine learning relies heavily on a large amount of high-quality training data. Insufficient or unevenly distributed data can lead to performance degradation. The ability to capture the relationship between complex anisotropic behavior and nonlinear stress is limited, and a simple mapping relationship between input and output is usually assumed. It is difficult to handle the spatial heterogeneity and anisotropy that are common in tight conglomerate reservoirs, which makes the prediction results less accurate under complex reservoir conditions. In addition, there are limitations in interpretability and accuracy, resulting in poor universality in different geological environments and difficulty in meeting actual needs. Summary of the invention

[0004] To solve at least one of the above problems, the present invention proposes a three-dimensional geostress intelligent prediction method integrating physical information, which can quickly simulate the expansion of multiple clusters of hydraulic fractures based on data of actual working conditions.

[0005] The technical solution of the present invention is: a three-dimensional geostress intelligent prediction method integrating physical information, comprising the following steps: S1. Obtain the three-dimensional geostress, the logging parameters while drilling related to the three-dimensional geostress, and the lithology parameters of the developed reservoir in the same block, match the three, and then perform data preprocessing on the three; S2. Based on the CAE autoencoder, feature extraction is performed on the pre-processed LWD parameters and core parameters to obtain the parameters after dimension reduction; S3, based on the AVE autoencoder, the reduced dimension parameters and preprocessed three-dimensional ground stress are enhanced to generate more samples; S4, based on the PINNs neural network, takes the parameters after dimension reduction in the S3 sample as input, takes the three-dimensional geostress in the S3 sample as output, and combines with the Adam optimizer for training. At the same time, during the training process, the following formula is used as the physical constraint condition of the model, and the trained model is used as the prediction model:

[0006]

[0007]

[0008] Where α is the effective stress coefficient; is the tectonic stress coefficient of the horizontal minimum principal stress; is the tectonic stress coefficient of the maximum horizontal principal stress; P s is the formation pressure; h1 is the depth; g is the gravitational acceleration; ρ b is the volume density; σ h is the minimum principal stress; σ H is the maximum principal stress; σ v is the vertical principal stress; s is Poisson's ratio.

[0009] Beneficial effects: The method of the present invention, by combining data-driven with physical constraints, enables the present invention to accurately model the complex changes of geostress under different geological conditions, especially when facing heterogeneous and anisotropic reservoirs, to improve the accuracy and robustness of prediction, thereby providing accurate geostress prediction for the efficient development of tight conglomerate reservoirs. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Figure 1 This is the original data sample compensated for neutrons in this embodiment; Figure 2 is the original data sample of volume density in this embodiment; Figure 3 is the original data sample of the acoustic wave time difference in this embodiment; Figure 4 is the original data sample of deep resistivity in this embodiment; Figure 5 is the original data sample of the well diameter in this embodiment; Figure 6 is the original data sample of natural gamma in this embodiment; Figure 7 is the original data sample of the depth in this embodiment; Figure 8 is the original data sample of the natural potential in this embodiment; Fig. 9 are the predicted value and actual value of the vertical stress in this embodiment; Fig.10 are the predicted value and actual value of the minimum principal stress in this embodiment; Fig.11 are the predicted value and actual value of the maximum principal stress in this embodiment. DETAILED DESCRIPTION

[0011] The specific implementation modes of the present invention will be described clearly and completely below with reference to examples. Obviously, the examples described are only some embodiments of the present invention, rather than all embodiments.

[0012] A three-dimensional geostress intelligent prediction method integrating physical information comprises the following steps: S1. Obtain the three-dimensional geostress, the logging parameters while drilling related to the three-dimensional geostress, and the lithology parameters of the developed reservoir in the same block, match the three, and then perform data preprocessing on the three; In this step, the method for obtaining the true value of the three-dimensional geostress includes the following steps: First, the fracture pressure and closure pressure are obtained through the construction curve sample. The closure pressure is defined as the minimum horizontal principal stress, which indicates the pressure value at which the rock formation closes under the action of the minimum principal stress. Then, through the relationship between the fracture pressure and the closure pressure, a linear equation is established to calculate the maximum principal stress.

[0013] As for the logging while drilling parameters, the inventors found after a large number of experiments that eight logging parameters, namely depth, natural gamma, natural potential, well diameter, deep resistivity, acoustic wave time difference, volume density and compensated neutron, have a higher correlation with the three-dimensional geostress. Therefore, in this embodiment, the selected logging parameters are these eight.

[0014] As for the lithological parameters, they include a variety of parameters, but considering the spatial heterogeneity and anisotropy of the dense conglomerate reservoir, therefore, in this embodiment, porosity, permeability and elastic modulus are selected as lithological parameters involved in the prediction of geostress.

[0015] In this embodiment, the reason why lithology parameters are additionally selected is that the tight conglomerate reservoir has significant spatial heterogeneity and anisotropy. Such characteristics make it difficult for the traditional prediction method that relies solely on data-driven to reflect the complex geological conditions of the tight conglomerate reservoir, which leads to distortion of the prediction results. In this embodiment, the physical properties of the reservoir and the evolution law of the geostress field are combined to provide more reliable input features, making the prediction results of the final model more accurate.

[0016] In order to make the LWD parameters and lithology parameters accurately reflect the three-dimensional geostress, after obtaining the above parameters, it is necessary to match the three of them. For example, for a certain depth, the three-dimensional geostress, logging parameters and lithology parameters at that depth need to be regarded as a set of data. In the subsequent calculation process, this set of data needs to be calculated as a whole under specific circumstances.

[0017] After obtaining the three-dimensional geostress, logging while drilling parameters and lithology parameters, the original data still needs to be preprocessed considering possible errors.

[0018] In this embodiment, the Z-Score method is first used to detect outliers in the obtained data. When the detection result is greater than a certain threshold, such as 3, the value can be regarded as an outlier. After the outliers are detected, all outliers are deleted. Subsequently, the missing values ​​in the data are combined with the KNN interpolation method to fill in the missing values, and the filled data set is the complete data set.

[0019] In order to facilitate subsequent calculations, these data need to be normalized so that each type of data is compressed and mapped to the range of [0,1].

[0020] After the above processing, we establish a characteristic matrix B based on the LWD parameters and lithology parameters, as shown below.

[0021]

[0022] In the formula, for the symbol B n8 , B represents the LWD parameters, where n represents the nth group of data, 8 represents the normalized number of the LWD parameters; the symbols ф, K and E represent the normalized porosity, permeability and elastic modulus, respectively.

[0023] S2. Based on the CAE autoencoder, feature extraction is performed on the pre-processed LWD parameters and core parameters to obtain the parameters after dimension reduction; In this step, the input parameter is the above-mentioned feature matrix B. Considering that there are relatively many parameters in the feature matrix B, in order to further extract the spatial patterns and nonlinear features of the 11 features, in this embodiment, a CAE autoencoder is used.

[0024] The autoencoder includes an encoder and a decoder. Considering that the autoencoder is a commonly used method in this field, its composition is briefly described.

[0025] Encoder: It includes convolution layer 1, maximum pooling layer 1, convolution layer 2, maximum pooling layer 2, fully connected layer and output layer in sequence, wherein the output layer is provided with 6 output channels, therefore, 11 features are compressed into 6 features through the encoder.

[0026] Decoder: It includes fully connected layer, transposed convolution layer 1, and transposed convolution layer 2 in sequence, and uses mean square error as the loss function to evaluate the extracted features.

[0027] S3, based on the AVE autoencoder, the reduced dimension parameters and preprocessed three-dimensional ground stress are enhanced to generate more samples; In this step, the six characteristic parameters output by S2 are combined with the preprocessed three-dimensional geostress and are combined in the order matched in S1 to form a new input data; through the AVE autoencoder, the new input data can be enhanced to generate more samples.

[0028] In this embodiment, the data is enhanced because the cost of obtaining the three-dimensional geostress through S1 is relatively high. In order to reduce the cost of obtaining data, data enhancement is performed.

[0029] Considering that the AVE autoencoder is a conventional means in this field, and the AVE autoencoder also includes an encoder and a decoder, only its specific operation and composition are briefly described in this step.

[0030] The input data is input into the encoder, which maps the input data into a probability distribution in the latent space, which is a Gaussian distribution; the decoder is then used to regenerate new data samples from the variables in the latent space.

[0031] The sample data volume is greatly increased after the data samples are expanded and enhanced by the AVE autoencoder.

[0032] S4, based on the PINNs neural network, takes the parameters after dimension reduction in the S3 sample as input, takes the three-dimensional geostress in the S3 sample as output, and combines with the Adam optimizer for training. At the same time, during the training process, the following formula is used as the physical constraint condition of the model, and the trained model is used as the prediction model:

[0033]

[0034]

[0035] Where α is the effective stress coefficient; is the tectonic stress coefficient of the horizontal minimum principal stress; is the tectonic stress coefficient of the maximum horizontal principal stress; P s is the formation pressure; h1 is the depth; g is the gravitational acceleration; ρ b is the volume density; σ h is the minimum principal stress; σ H is the maximum principal stress; σ v is the vertical principal stress; s is Poisson's ratio.

[0036] In this step, the structure of the PINNs neural network is as follows: an input layer, which inputs sample data expanded by the AVE autoencoder; a hidden layer, which is composed of multiple fully connected layers, each layer is linearly transformed and then connected to the ReLU activation function; an output layer, which outputs the predicted value of the three-dimensional ground stress; a physical constraint layer, which processes data-based supervised learning and also adds the above formula as an internal physical constraint condition. When applying the physical constraint condition, the difference before and after the iteration of the physical equation is added to the loss function of the neural network, so that the physical equation is also "involved" in the training process. In this way, the neural network optimizes not only the network's own loss function during training iterations, but also the difference of each iteration of the physical equation, so that the final training result can meet the laws of physics.

[0037] In particular, the loss function of PINNs is: , where λ, β and γ are weight coefficients; L total is the loss function; L data is data loss; L physics is physical loss; L spatial is the spatial consistency loss; in, , , , where N is the maximum number of training samples, and i=1,2…,N; is the prediction result of three-dimensional geostress in the i-th sample; is the true data of the three-dimensional geostress in the i-th sample; M is the maximum number of sampling points in the physical space, and j=1,2…,M; x j is the number of the sampling point in the physical space; is the residual between the predicted value and the physical constraints; and Respectively represent adjacent physical space sampling points; Indicates that in z j The predicted ground stress value at ; Indicates that in z j+1 The predicted ground stress value at .

[0038] The Adam optimizer includes the following sub-steps: Calculate the gradient of the loss function with respect to the network weights: , where is the gradient of the current step, Indicates the parameters Find the gradient, L total is the loss function, θ t Represents the parameter vector for the current step.

[0039] First moment estimate: , where β1 is the momentum coefficient; m t is the first-order moment; m t-1 is the first-order moment estimate of the previous step; Second moment estimate: , where v t is the second-order moment; β2 is the second-order moment coefficient; v t-1 is the second-order moment estimate from the previous step; Bias correction: , , where and They are the corrected first-order moment estimate and the second-order moment estimate respectively; and They represent the t-th power of β1 and β2 respectively, where t is the current iteration number; Parameter update: , where k is the learning rate; A small constant to avoid division by zero errors, usually set to 10 -3 .

[0040] After the PINNs neural network is established, it needs to be trained. During training, the data samples used are the expanded data samples in S3. One part of them is selected as the training set and the other part is selected as the test set.

[0041] The training process of the PINNs neural network is as follows: 1. Initialize network parameters: First, initialize the weights and biases of the neural network.

[0042] 2. Calculate the loss: At each forward propagation, calculate the data loss L data , physical loss L physics and spatial consistency loss L spatial , and then calculate the total loss L based on their weighted sum total .

[0043] 3. Back propagation: The gradient of the loss function with respect to the network weights is calculated through the back propagation algorithm.

[0044] 4. Optimization update: Use the Adam optimizer to update the network parameters according to the calculated gradients. Adam dynamically adjusts the learning rate of each parameter, thereby accelerating the training process and improving stability.

[0045] 5. Iterative training: Repeat steps 2-4 until the loss function converges or the set number of iterations is reached.

[0046] Through the trained model, the three-dimensional geostress of the target well in the target block can be predicted while drilling.

[0047] In order to further demonstrate the advantages of the method of the embodiment of the present invention, a specific example is used below to illustrate it.

[0048] Select a developed well in a block of Xinjiang Oilfield and obtain its initial data, such as Figures 1 to 8 shown.

[0049] Then the method of this embodiment is used to predict it, and the final prediction result is as follows: Figures 9-11 As shown in the figure, it can be seen that the method of the embodiment of the present invention has a high prediction accuracy.

[0050] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Although the present invention has been disclosed as a preferred embodiment as above, it is not intended to limit the present invention. Any technician familiar with the profession can make some changes or modifications to equivalent embodiments of equivalent changes using the technical contents disclosed above without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent change and modification made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention still falls within the scope of the technical solution of the present invention.

Claims

1. A three-dimensional geostress intelligent prediction method integrating physical information, characterized in that: The following steps are involved: S1. Obtain the three-dimensional geostress, the logging parameters while drilling related to the three-dimensional geostress, and the lithology parameters of the developed reservoir in the same block, match the three, and then perform data preprocessing on the three; S2. Based on the CAE autoencoder, feature extraction is performed on the pre-processed LWD parameters and core parameters to obtain the parameters after dimension reduction; S3, based on the AVE autoencoder, the reduced dimension parameters and preprocessed three-dimensional ground stress are enhanced to generate more samples; S4, based on the PINNs neural network, takes the parameters after dimension reduction in the S3 sample as input, takes the three-dimensional geostress in the S3 sample as output, and combines with the Adam optimizer for training. At the same time, during the training process, the following formula is used as the physical constraint condition of the model, and the trained model is used as the prediction model: Where α is the effective stress coefficient; is the tectonic stress coefficient of the horizontal minimum principal stress; is the tectonic stress coefficient of the maximum horizontal principal stress; P s is the formation pressure; h1 is the depth; g is the gravitational acceleration; ρ b is the volume density; σ h is the minimum principal stress; σ H is the maximum principal stress; σ v is the vertical principal stress; s is Poisson's ratio.

2. The method according to claim 1, characterized in that The three-dimensional geostresses are fracture pressure, closure pressure and maximum principal stress, the logging while drilling parameters are depth, natural gamma, natural potential, well diameter, deep resistivity, acoustic wave time difference, volume density and compensated neutron, and the lithology parameters are porosity, water saturation and elastic modulus.

3. The method according to claim 1, characterized in that The data preprocessing method includes the following sub-steps: performing anomaly detection on the data, and deleting the detected abnormal values; using the KNN interpolation method to fill in the missing values; and normalizing the filled data.

4. The method according to claim 1, characterized in that: In S4, the loss function of the PINNs neural network is: , where λ, β and γ are weight coefficients; L total is the loss function; L data is data loss; L physics is physical loss; L spatial is the spatial consistency loss; in, , , , where N is the maximum number of training samples, and i=1,2…,N; is the prediction result of three-dimensional geostress in the i-th sample; is the true data of the three-dimensional geostress in the i-th sample; M is the maximum number of sampling points in the physical space, and j=1,2…,M; x j is the number of the sampling point in the physical space; is the residual between the predicted value and the physical constraints; and Represent adjacent physical space sampling points respectively; Indicates that in z j The ground stress value at ; Indicates that in z j+1 The ground stress value at .

5. The method according to claim 1, characterized in that The Adam optimizer includes the following sub-steps: Calculate the gradient of the loss function with respect to the network weights: , where is the gradient of the current step, Indicates the parameters Find the gradient, L total is the loss function, θ t Represents the parameter vector of the current step; First moment estimate: , where β1 is the momentum coefficient; m t is the first-order moment estimate; m t-1 is the first-order moment estimate of the previous step; Second moment estimate: , where v t is the second-order moment estimate; β2 is the second-order moment coefficient; v t-1 is the second-order moment estimate from the previous step; Bias correction: , , where and They are the corrected first-order moment estimate and the second-order moment estimate respectively; and They represent the t-th power of β1 and β2 respectively, where t is the current iteration number; Parameter update: , where k is the learning rate; A small constant to avoid division by zero errors, usually set to 10 -3 .

Images