Method for establishing pore pressure prediction model based on multi-scale information fusion

Abstract

The invention discloses a pore pressure prediction model establishment method based on multi-scale information fusion, and belongs to the field of oil and gas resource exploration and development, and the method comprises the following steps: collecting a seismic data volume, determining an input feature set, and constructing a label data set, a label pore pressure data set and a non-label data set; constructing a label training dataset, a label verification dataset, a label test dataset, a label training pore pressure dataset, a label verification pore pressure dataset, a label test pore pressure dataset, a non-label training dataset and a non-label verification dataset, and determining a convolutional neural network hyper-parameter based on the label datasets; the number of layers of full connection, the batch size of single training and the number of steps of single training are determined; performing virtual confrontation training on hyper-parameters; according to the method, the convolutional neural network model is constructed, the performance of the convolutional neural network model on the label test data set is tested, the logic is simple, accurate and reliable, and the method can be suitable for three-dimensional pore pressure prediction of a deep complex structure.

Description

Technical Field

[0001] The present invention relates to the field of oil and gas resource exploration and development, and in particular to a method for establishing a pore pressure prediction model based on multi-scale information fusion. Background Art

[0002] Formation pore pressure is a key basic parameter for oil and gas resource evaluation, drilling and completion, reservoir transformation, and drainage. Formation pore pressure testing based on fracturing engineering is the most accurate method, but the cost is extremely high and the amount of data obtained is very small, which cannot meet the needs of engineers for understanding the longitudinal continuous pore pressure distribution characteristics of large areas. In order to solve this problem, oil and gas engineers and geophysicists have jointly developed formation pore pressure prediction technology based on geophysical information. As oil and gas exploration and development gradually move towards deep and complex structural areas, conventional three-dimensional pore pressure prediction methods based on elastic wave theory, equivalent depth method, Eaton method, etc. have poor adaptability to deep and complex structural formations. In order to solve this problem, some experts and scholars have carried out relevant research based on machine learning, but in actual application, they are mainly based on labeled data constructed by field measurements or forward simulations and carry out prediction work, which ignores the information contained in a large amount of unlabeled data, and the data scale is single, which seriously restricts the prediction accuracy. Summary of the invention

[0003] In response to the above problems, the present invention provides a method for establishing a pore pressure model based on multi-scale information fusion, which integrates the measured pore pressure, the pore pressure predicted by well logging, and the three-dimensional pore pressure of the three-dimensional seismic body's own attributes. It has simple logic, is accurate and reliable, and fully mines the data laws contained in the engineering and geophysical information in the oil and gas field, constructs a deep learning network structure that integrates points, lines, surfaces, and bodies, and determines the hyperparameter set based on virtual adversarial training. It can be applied to the prediction of three-dimensional pore pressure in deep complex structural areas.

[0004] The technical solution of the present invention is:

[0005] A method for establishing a pore pressure prediction model based on multi-scale information fusion includes the following steps:

[0006] S1, collect seismic data volume and determine input feature set;

[0007] S2. Construct a labeled dataset, a labeled pore pressure dataset, and an unlabeled dataset based on the input feature set;

[0008] S3, constructing a labeled training dataset, a labeled verification dataset, a labeled test dataset, a labeled training pore pressure dataset, a labeled verification pore pressure dataset, a labeled test pore pressure dataset, an unlabeled training dataset, and an unlabeled verification dataset according to the labeled dataset and the unlabeled dataset;

[0009] S4. Determine the hyperparameters of the convolutional neural network based on the labeled dataset;

[0010] S5. Determine the number of fully connected layers, the size of a single training batch, and the number of single training steps;

[0011] S6. Determine the virtual adversarial training hyperparameters;

[0012] S7. Build a convolutional neural network model and test the performance of the convolutional neural network model on the labeled test data set;

[0013] In step S1, the feature data set is a set consisting of original post-stack 3D seismic blocks, 3D frequency attribute blocks, amplitude differences, fractal amplitude density, fractal frequency density and phase;

[0014] The convolutional neural network model formula is:

[0015] PP=FC(CNN(Cube_s,Cube_f),Diff_am,Roh_fa,Roh_ff,Phase);

[0016] Among them, CNN is convolution calculation, FC is full connection calculation, PP is pore pressure, Cube_s is the original post-stack 3D seismic block, Cube_f is the frequency attribute 3D block, Diff_am is the amplitude difference, Roh_fa is the fractal amplitude density, Roh_ff is the fractal frequency density, and Phase is the phase.

[0017] In step S2, the method for determining the label data set and the label pore pressure data set is as follows:

[0018] For the original post-stack 3D seismic blocks and 3D frequency attribute blocks, original post-stack seismic small blocks with a size of 5×5×17 and frequency attribute small blocks with a size of 25×25×25 were extracted at the locations with measured pore pressure along the well trajectory; for amplitude difference, fractal amplitude density, fractal frequency density and phase, single-point data were randomly extracted at the locations with measured pore pressure along the well trajectory.

[0019] In step S2, the method for determining the unlabeled data set is as follows:

[0020] 1) For the original post-stack 3D seismic blocks and 3D frequency attribute blocks, small original post-stack seismic blocks with a size of 5×5×17 and small frequency attribute blocks with a size of 25×25×25 were randomly extracted at locations without measured pore pressure along the well trajectory; for amplitude difference, fractal amplitude density, fractal frequency density and phase, single point data were randomly extracted at locations without measured pore pressure along the well trajectory;

[0021] 2) For the original post-stack 3D seismic blocks and 3D frequency attribute blocks, small original post-stack seismic blocks with a size of 5×5×17 and small frequency attribute blocks with a size of 25×25×25 were randomly extracted at locations without wellbore. For amplitude difference, fractal amplitude density, fractal frequency density and phase, single-point data were randomly extracted at locations without wellbore.

[0022] In step S4, the hyperparameters of the convolutional neural network are the convolution kernel, the size of the pooling layer sliding window, the number of network layers and the number of neurons.

[0023] In step S4, the method for determining the combination of hyperparameters of the convolutional neural network is as follows:

[0024] 1) Analyze the relative error values ​​of the convolutional neural network model on the training data set and the validation data set;

[0025] 2) Analyze the change rules of the relative error curve of the convolutional neural network model on the training data set and the verification data set;

[0026] 3) Optimize hyperparameters.

[0027] The beneficial effects of the present invention are:

[0028] The data used in the present invention is multi-scale data, which integrates logging, seismic and measured pore pressure information to establish a convolutional neural network model. The model training process adopts virtual counter-disturbance, with simple logic, accuracy and reliability. It can fully explore the data laws contained in engineering and geophysical information in the oil and gas field, and can be applied to the three-dimensional pore pressure prediction of deep complex structures. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a flow chart of a method for establishing a pore pressure prediction model based on multi-scale information fusion according to an embodiment of the present invention;

[0030] Figure 2 is the Pearson correlation coefficient between the pore pressure and each characteristic parameter in the method for establishing a pore pressure prediction model based on multi-scale information fusion described in an embodiment of the present invention;

[0031] Figure 3 is the relative error of the model on the training set and the validation set under the condition of gradually adding the number of input features in the method for establishing a pore pressure prediction model based on multi-scale information fusion described in an embodiment of the present invention;

[0032] Figure 4 is the error of the model on the training set and the validation set under the same convolution kernel and sliding window size conditions of the method for establishing a pore pressure prediction model based on multi-scale information fusion described in an embodiment of the present invention;

[0033] Figure 5 It is the relative error (F1 group) under different fully connected layer numbers, single training batch sizes and single training step numbers of a pore pressure prediction model establishment method based on multi-scale information fusion described in an embodiment of the present invention;

[0034] Figure 6 It is the relative error (F2 group) under different fully connected layer numbers, single training batch sizes and single training step numbers of a pore pressure prediction model establishment method based on multi-scale information fusion described in an embodiment of the present invention;

[0035] Figure 7 It is the relative error (F3 group) under different fully connected layer numbers, single training batch sizes and single training step numbers of a pore pressure prediction model establishment method based on multi-scale information fusion described in an embodiment of the present invention;

[0036] Figure 8 It is the prediction result of the deep learning model of the method for establishing a pore pressure prediction model based on multi-scale information fusion described in an embodiment of the invention on the test set;

[0037] Fig. 9 It is a three-dimensional pore pressure prediction result data body of a pore pressure prediction model establishment method based on multi-scale information fusion described in an embodiment of the invention;

[0038] Fig.10 It is a comparison chart of the pore pressure of a single well extracted along the well and the pore pressure calculated based on logging information in a method for establishing a pore pressure prediction model based on multi-scale information fusion described in an embodiment of the invention. DETAILED DESCRIPTION

[0039] The embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

[0040] Example:

[0041] like Figure 1 As shown, a method for establishing a pore pressure prediction model based on multi-scale information fusion includes the following steps:

[0042] S1, collect seismic data volume and determine input feature set;

[0043] Based on the original acquired seismic data body, the original post-stack seismic body (three-dimensional), frequency seismic body (three-dimensional), phase, relative wave impedance, fractal amplitude density, fractal frequency density, amplitude difference, original amplitude, root mean square amplitude, sweet spot, and the imaginary part of the signal after the phase is rotated 90° can be processed. These features are all affected by pore pressure. Finally, the set consisting of the original post-stack seismic block (three-dimensional), frequency attribute block (three-dimensional), amplitude difference, fractal amplitude density, fractal frequency density and phase is determined as the input feature set.

[0044] The input feature set is determined as follows:

[0045] (1) Based on general experience, the set consisting of the original post-stack seismic volume (3D) and the frequency seismic volume (3D) is recorded as feature set 0;

[0046] (2) The phase, relative wave impedance, fractal amplitude density, fractal frequency density, amplitude difference, original amplitude, root mean square amplitude, sweet spot, and the absolute value of the Pearson correlation coefficient between the imaginary part of the signal after phase rotation by 90° and the pore pressure are calculated according to the following formula: Figure 2 ), and sorted according to the absolute value of the Pearson correlation coefficient. The sorting result is: amplitude difference > fractal amplitude density > fractal frequency density > root mean square amplitude > original amplitude > sweet spot > relative wave impedance > phase > imaginary part of the signal after phase rotation of 90°;

[0047]

[0048] Among them, r is the correlation coefficient, x i With y i For the analyzed data pair, and are the average values ​​of data x and y respectively;

[0049] (3) The amplitude difference, fractal amplitude density, fractal frequency density, root mean square amplitude, original amplitude, sweet spot, relative wave impedance, phase, and imaginary part of the signal after phase rotation by 90° are placed into feature set 0 in order of precedence and recorded as feature set 1 to feature set 9 respectively. Corresponding to each feature set, the corresponding label data set is extracted and recorded as label feature data set 1 to label feature data set 9 respectively. At the same time, each label feature data set is divided into a training set, a validation set, and a test set according to a data volume ratio of 7:2:1. Label feature data sets 0 to label feature data sets 9 are loaded into the convolutional neural network model established by the present invention respectively, and the relative error of each model on the training set and the validation set in the label feature data set can be calculated ( Figure 3 ).analyze Figure 3It can be seen that when label feature data set 1, label feature data set 2, label feature data set 3, and label feature data set 8 are loaded into the convolutional neural network model, the relative errors in the training set and the verification set are reduced compared with the previous feature set loaded; when label feature data set 4, label feature data set 6, and label feature data set 9 are loaded into the convolutional neural network model, the relative errors in the training set and the verification set are increased or basically unchanged compared with the previous feature set loaded; when label feature data set 5 is loaded into the convolutional neural network model, the relative error in the training set is slightly increased, and the relative error in the verification set is slightly reduced. In summary, the characteristic parameters contained in the final feature set of the present invention are determined to be: original post-stack three-dimensional seismic block, frequency attribute three-dimensional block, amplitude difference, fractal amplitude density, fractal frequency density, and phase.

[0050] S2. Construct a labeled dataset, a labeled pore pressure dataset, and an unlabeled dataset based on the input feature set;

[0051] The extraction rules are as follows:

[0052] (1) The extraction rules of the label input data set and the label pore pressure data set are as follows: for the original post-stack seismic block (3D) and the frequency attribute block (3D), extract the original post-stack seismic small block of size 5×5×17 and the frequency attribute small block of size 25×25×25 at the location with measured pore pressure along the well trajectory; for the amplitude difference, fractal amplitude density, fractal frequency density and phase, randomly extract single point data at the location with measured pore pressure along the well trajectory; the set of all the data extracted above is the label input data set. At the same time, extract the measured pore pressure data along the well trajectory, and the set of all the extracted data is the label pore pressure data set. The label data set has a one-to-one correspondence with the label input data set.

[0053] (2) Rules for extracting unlabeled input data sets: The unlabeled data consists of two parts. In the first part, for the original post-stack seismic block (3D) and the frequency attribute block (3D), a small original post-stack seismic block with a size of 5×5×17 and a small frequency attribute block with a size of 25×25×25 are randomly extracted along the well trajectory at locations without measured pore pressure; for amplitude difference, fractal amplitude density, fractal frequency density and phase, single point data are randomly extracted along the well trajectory at locations without measured pore pressure. In the second part, for the original post-stack seismic block and the frequency attribute block, a small original post-stack seismic block with a size of 5×5×17 and a small frequency attribute block with a size of 25×25×25 are randomly extracted at locations without wellbore; for amplitude difference, fractal amplitude density, fractal frequency density and phase, single point data are randomly extracted at locations without wellbore. The set of all the extracted data is the unlabeled input data set.

[0054] S3, constructing a labeled training dataset, a labeled verification dataset, a labeled test dataset, a labeled training pore pressure dataset, a labeled verification pore pressure dataset, a labeled test pore pressure dataset, an unlabeled training dataset, and an unlabeled verification dataset according to the labeled dataset and the unlabeled dataset;

[0055] The division requirements are as follows;

[0056] (1) Divide the labeled input dataset into a labeled training dataset, a labeled verification dataset, and a labeled test dataset in a ratio of 7:2:1;

[0057] (2) The labeled pore pressure dataset has a one-to-one correspondence with the labeled input dataset. Therefore, by dividing the labeled pore pressure dataset in the ratio of 7:2:1, we can obtain the labeled training pore pressure dataset, the labeled verification pore pressure dataset, and the labeled test pore pressure dataset. The labeled training dataset corresponds to the labeled training pore pressure dataset, the labeled verification dataset corresponds to the labeled verification pore pressure dataset, and the labeled test dataset corresponds to the labeled test pore pressure dataset.

[0058] (3) The unlabeled input dataset is divided into an unlabeled training dataset and an unlabeled validation dataset in a ratio of 7:3. The labeled training dataset and the unlabeled training dataset are merged into the training dataset, the labeled validation dataset and the unlabeled validation dataset are merged into the validation dataset, and the labeled test dataset is used as the test dataset.

[0059] S4. Determine the hyperparameters of the convolutional neural network based on the labeled dataset;

[0060] Set the convolution kernel, pooling layer sliding window size, network layer number and neuron number as shown in Table 1 and Table 2. Figure 4 As shown, the errors of the model on the training data set and the validation data set under different conditions are tested and analyzed.

[0061] Table 1 Convolution kernel and pooling layer sliding window settings

[0062]

[0063] Table 2 Convolutional neural network structure and neuron number settings

[0064]

[0065] According to Table 1, Table 2 and Figure 4 , determine the best combination of parameters, the specific method is as follows;

[0066] (1) Analyze the relative error values ​​of the convolutional neural network model on the training data set and the validation data set. Figure 4It can be seen that the error of the convolutional neural network model on the training data set is small, and the error on the validation data set is large, indicating that the performance of the convolutional neural model at this time is poor and fails to fit the data well.

[0067] (2) Analyze the change rules of the relative error curve of the convolutional neural network model on the training data set and the verification data set. Figure 4 By comparing the performance of the convolutional neural model in different groups, it can be seen that under the C7 group condition, the two relative error curves have similar changing trends, indicating that the performance of the convolutional neural model on the training data set and the validation data set changes in a similar pattern with the number of neurons. At this time, the convolutional neural model has relatively good robustness, so the convolutional neural network hyperparameters are determined to be the parameters corresponding to the C7 group.

[0068] (3) Further optimize the hyperparameters within the determined hyperparameter group. Figure 3 By comparing the error data within group C7, it is not difficult to see that under the conditions corresponding to group number C-5, the relative error of the convolutional neural model on the training data set and the validation data set is relatively low, and the difference between the relative error on the validation data set and the error on the training data set is relatively the lowest. At this time, the convolutional neural network model performs relatively well. Therefore, the number of neurons in the convolutional neural network is determined to be the parameter corresponding to group number C-5 in group C7.

[0069] S5. Determine the number of fully connected layers, the size of a single training batch, and the number of single training steps;

[0070] The specific optimization method is as follows:

[0071] (1) Calculation of relative error under different conditions. Three groups of parameters were set to calculate the relative error of the convolutional neural network model on the training data set and the validation data set. The number of neurons in the input layer of group F1 was 36, and the number of neurons in the output layer was 1. The number of neurons in the input layer of group F2 was 36, the number of neurons in hidden layer 1 was 18, and the number of neurons in the output layer was 1. The number of neurons in the input layer of group F3 was 36, the number of neurons in hidden layer 1 was 24, the number of neurons in hidden layer 2 was 12, and the number of neurons in the output layer was 1. In groups F1, F2, and F3, there were nine combinations of single training batch size and single training step number, namely (2,11), (3,11), (4,8), (5,7), (6,6), (7,5), (8,4), (12,3), and (16,2), which were respectively recorded as groups F-1 to F-9. The relative error calculation results are shown in the figure. Figure 5-7 shown.

[0072] (2) Carry out relative error calculation and analysis under different conditions, which is divided into three steps:

[0073] ① Analyze the overall law of relative error changes of convolutional neural networks on data sets under three groups of conditions. Figure 5-7 It can be seen that under the three groups of conditions, the error of the convolutional neural network on the training data set gradually decreases slowly, and the relative error on the validation data set shows a change pattern of first decreasing and then increasing, and the relative error of the convolutional neural network model on the validation data set reaches the minimum value at the group number F-2.

[0074] ②Analyze the difference in the minimum relative error of the convolutional neural network model on the validation data set under different conditions and the difference in the two relative errors at the corresponding group numbers. Comparing the error values ​​at the position of the group number F-2 under three different conditions, it can be seen that the error of the model at the F-2 position in the F2 group on the validation set is the smallest, and at this time the error difference between the model on the validation set and the training set is relatively small, and the model performance is relatively optimal.

[0075] ③ Determine the optimal parameters by combining the analysis results of ① and ②. Determine the number of fully connected layers of the invented convolutional neural network model by combining the analysis results of ① and ② as the parameters of group F2, and the size of a single training batch and the number of single training steps as the parameters corresponding to the number F-2 in group F2. That is, the number of neurons in the input layer is 36, the number of neurons in the hidden layer is 18, the size of a single training batch is 3, and the number of single training steps is 11.

[0076] S6. Determine the virtual adversarial training hyperparameters;

[0077] The specific method is as follows;

[0078] (1) Label training data set (symbol: x l ) and the unlabeled training dataset (x ul ) are input into the convolutional neural network model respectively, and the corresponding pore pressure value PP is calculated respectively. l With PP ul , PP l With PP ul All are collections;

[0079] (2) Use PP l Training pore pressure data with labels (symbol: PP lt ) set is calculated according to the following formula to obtain the loss function value of the label training data set on the convolutional neural network model;

[0080]

[0081] Among them, L1 is the loss function value, pp lti Training pore pressure data for labels (symbol: PP lt ), the i-th element in li For PP l The i-th element in ;

[0082] (3) Using PP ul With unlabeled training data (symbol: x ul ) Calculate the perturbation in the loss function value according to the following formula;

[0083]

[0084] Among them, r qadv is a virtual adversarial perturbation, θ is a model parameter, d is a random unit vector, ξ is a constant that controls the initial perturbation vector, r = ξd is the initial perturbation vector, ε is the canonical constraint coefficient that determines the adversarial direction, and L is the loss function.

[0085] (4) r qadv +x ul Re-enter the convolutional neural network model as a new parameter to calculate the new pore pressure value PP ulr , and calculate the loss function value caused by the addition of virtual adversarial disturbance according to the following formula;

[0086]

[0087] Among them, L2 is the loss function value, pp uli For PP ul The i-th element in , pp ulri For PP ulr The i-th element in ;

[0088] (5) Based on the above two loss function values ​​L1 and L2, the convolutional neural network model parameter θ is optimized according to the following formula;

[0089]

[0090] Wherein, D is a non-negative metric function, L is the MSE loss function, and according to the literature, when α=1, a better application effect can be achieved, therefore, the present invention takes α=1;

[0091] (6) Repeat (1) to (5) until loss Less than 0.001 or the total number of calculation steps is greater than 1000;

[0092] (7) Set the value of ξ to 10-8, 10-6, 10-4, and 10-2, respectively, for a total of four different conditions. Set the value of ε to 0.5-5 for each ξ value condition, and the interval between two adjacent data points is 0.5. Compare the relative errors of the convolutional neural network model on the labeled training data set and the unlabeled training data set with the labeled verification data set and the unlabeled verification data set under different conditions, and select the optimal ξ and ε values ​​of 10-6 and 2.5, respectively. The optimization criteria are: ① The relative error of the convolutional neural network model on the labeled training data set and the unlabeled training data set is less than 10%; ② The relative error of the convolutional neural network model on the labeled verification data set and the unlabeled verification data set is less than 10%; ③ The sum of the relative errors of the convolutional neural network model on the labeled training data set and the unlabeled training data set and the labeled verification data set and the unlabeled verification data set is relatively the lowest.

[0093] S7. Build a convolutional neural network model and test the performance of the convolutional neural network model on the labeled test data set.

[0094] like Figure 8 As shown, the convolutional neural network is a feedforward neural network with convolution and sampling as its core. According to the researchers' ideas, a complete convolutional neural network model can be built by combining these network structures in a certain order. The convolutional neural network model formula of the present invention is:

[0095] PP=FC(CNN(Cube_s,Cube_f),Diff_am,Roh_fa,Roh_ff,Phase)

[0096] Among them, CNN is convolution calculation, FC is full connection calculation, PP is pore pressure, Cube_s is the original post-stack seismic block (three-dimensional), Cube_f is the frequency attribute block (three-dimensional), Diff_am is the amplitude difference, Roh_fa is the fractal amplitude density, Roh_ff is the fractal frequency density, and Phase is the phase. During the model calculation process, the size of Cube_s corresponding to each data point is 5×5×17, the size of Cube_s is 25×25×25, and the size of the remaining feature parameters is 1×1×1, that is, the fusion convolution of multi-scale information is realized.

[0097] Input the labeled test data set into the convolutional neural network model and calculate the corresponding pore pressure value PP le , and combined with the labeled test pore pressure dataset to calculate the relative error of the convolutional neural network model on the labeled test dataset ( Figure 8 ).from Figure 8 It is not difficult to see that, on the whole, the predicted pore pressure is in good agreement with the measured pore pressure values, the error value is relatively stable, and the average relative error is 9.15%, indicating that the model has good prediction performance.

[0098] Fig. 9 The figure shows the three-dimensional pore pressure prediction results based on the prediction method established in this patent in the MX lower than DY formation in the Sichuan Basin. Fig.10 Shown from Fig. 9 The 3D pore pressure data volume shown is a comparison of the single well pore pressure data extracted along the well trajectory and the pore pressure calculated based on the logging information. Fig. 9 , Fig.10 It was found that the three-dimensional pore pressure data body predicted by the method proposed in this technology reflects the spatial variation characteristics of the pore pressure, and the predicted pore pressure is highly consistent with the pore pressure obtained based on well logging information, indicating that the three-dimensional pore pressure method based on multi-scale information fusion convolution virtual confrontation established by the present invention can achieve better prediction effect.

[0099] The above-mentioned embodiments only express the specific implementation of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the present invention. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A method for establishing a pore pressure prediction model based on multi-scale information fusion, characterized in that: The steps include: S1, collect seismic data volume and determine input feature set; S2. Construct a labeled dataset, a labeled pore pressure dataset, and an unlabeled dataset based on the input feature set; S3, constructing a labeled training dataset, a labeled verification dataset, a labeled test dataset, a labeled training pore pressure dataset, a labeled verification pore pressure dataset, a labeled test pore pressure dataset, an unlabeled training dataset, and an unlabeled verification dataset according to the labeled dataset and the unlabeled dataset; S4. Determine the hyperparameters of the convolutional neural network based on the labeled dataset; S5. Determine the number of fully connected layers, the size of a single training batch, and the number of single training steps; S6. Determine the virtual adversarial training hyperparameters; S7. Build a convolutional neural network model and test the performance of the convolutional neural network model on the labeled test data set; In step S1, the feature data set is a set consisting of original post-stack 3D seismic blocks, 3D frequency attribute blocks, amplitude differences, fractal amplitude density, fractal frequency density and phase; The convolutional neural network model formula is: PP=FC(CNN(Cube_s,Cube_f),Diff_am,Roh_fa,Roh_ff,Phase) Among them, CNN is convolution calculation, FC is full connection calculation, PP is pore pressure, Cube_s is the original post-stack 3D seismic block, Cube_f is the frequency attribute 3D block, Diff_am is the amplitude difference, Roh_fa is the fractal amplitude density, Roh_ff is the fractal frequency density, and Phase is the phase.

2. The method for establishing a pore pressure prediction model based on multi-scale information fusion according to claim 1, characterized in that: In step S2, the method for determining the label data set and the label pore pressure data set is as follows: For the original post-stack 3D seismic blocks and 3D frequency attribute blocks, original post-stack seismic small blocks with a size of 5×5×17 and frequency attribute small blocks with a size of 25×25×25 were extracted at the locations with measured pore pressure along the well trajectory; for amplitude difference, fractal amplitude density, fractal frequency density and phase, single-point data were randomly extracted at the locations with measured pore pressure along the well trajectory.

3. The method for establishing a pore pressure prediction model based on multi-scale information fusion according to claim 1, characterized in that: In step S2, the method for determining the unlabeled data set is as follows: 1) For the original post-stack 3D seismic blocks and 3D frequency attribute blocks, small original post-stack seismic blocks with a size of 5×5×17 and small frequency attribute blocks with a size of 25×25×25 were randomly extracted at locations without measured pore pressure along the well trajectory; for amplitude difference, fractal amplitude density, fractal frequency density and phase, single point data were randomly extracted at locations without measured pore pressure along the well trajectory; 2) For the original post-stack 3D seismic blocks and 3D frequency attribute blocks, small original post-stack seismic blocks with a size of 5×5×17 and small frequency attribute blocks with a size of 25×25×25 were randomly extracted at locations without wellbore. For amplitude difference, fractal amplitude density, fractal frequency density and phase, single-point data were randomly extracted at locations without wellbore.

4. The method for establishing a pore pressure prediction model based on multi-scale information fusion according to claim 1, characterized in that: In step S4, the hyperparameters of the convolutional neural network are the convolution kernel, the size of the pooling layer sliding window, the number of network layers and the number of neurons.

5. The method for establishing a pore pressure prediction model based on multi-scale information fusion according to claim 4, characterized in that: In step S4, the method for determining the combination of hyperparameters of the convolutional neural network is as follows: 1) Analyze the relative error values ​​of the convolutional neural network model on the training data set and the validation data set; 2) Analyze the change rules of the relative error curve of the convolutional neural network model on the training data set and the verification data set; 3) Optimize hyperparameters.

Images