Stratum pressure prediction method based on artificial intelligence

Abstract

The invention discloses a formation pressure prediction method based on artificial intelligence, and relates to the technical field of oil exploration and development, and the method comprises the steps: 1, carrying out the spatial dimension increase of feature parameters of attribute constraints, and constructing a dimension increase sample set, 2, carrying out the pre-training of a network model, 3, carrying out the optimization and updating of the network model, 4, carrying out the prediction of the network model, inputting a post-stack seismic record in a mature network model, and carrying out the prediction of the formation pressure. High-sensitivity characteristic parameters are output and obtained, a formation pressure coefficient is calculated according to the high-sensitivity characteristic parameters, and finally the predicted formation pressure is obtained according to the formation pressure coefficient. According to the method, the formation pressure coefficient characteristic sample under the seismic data geologic body constraint is obtained through pre-stack inversion, the data is optimized through the convolutional network, and the weight and the characteristic vector in the layer are adjusted to be optimal. The real well data is reasonably added into a subsequent training network for error reverse transmission, so that a well-trained network is obtained and applied to formation pressure prediction, and the problem that a traditional formation pressure prediction method is insufficient in performance stability is solved.

Description

Technical Field

[0001] This invention relates to the field of petroleum exploration and development technology, specifically to an artificial intelligence-based method for predicting formation pressure. Background Technology

[0002] Formation pressure refers to the pressure exerted by fluids such as water, oil, and gas within the pores of formation rocks. Typically, formation pressure exhibits both abnormally high and abnormally low pressure phenomena.

[0003] When the formation pressure exceeds the hydrostatic column pressure, an abnormally high pressure phenomenon will occur in the formation. In the fields of oil and gas geological exploration and oil and gas well engineering, an abnormally high pressure formation will not only seriously affect the wellbore stability during the drilling process, but may also increase the risk of drilling accidents such as well kicks and blowouts.

[0004] Previous studies on the causes of abnormal low pressure are relatively limited. It is generally believed that abnormal low pressure mainly develops in some tight gas-bearing sandstone layers and basins with strong erosion. The causes of abnormal low pressure are mostly related to factors such as formation uplift and erosion, and the release of light hydrocarbons. Abnormal low pressure will lead to a rapid decline in oil and gas production in the later stages.

[0005] Practice has shown that both abnormally high and abnormally low pressures can significantly affect natural gas production capacity. Therefore, accurately predicting formation pressure is a powerful guarantee for safe production and efficient development.

[0006] Currently, the spatial distribution of formation pressure can be predicted using seismic data. Commonly used seismic data prediction methods include the Eaton method and the effective stress method. However, in actual research and production, the accuracy of these commonly used prediction methods is relatively limited, especially in the context of abnormally low pressure, where the applicability of these methods is even more limited and the accuracy is even lower.

[0007] Therefore, it is necessary to improve a new prediction technology that combines artificial intelligence to solve the technical problem of low prediction accuracy of traditional prediction methods under abnormal stress. Summary of the Invention

[0008] The purpose of this invention is to overcome the aforementioned problems in existing technologies and provide an artificial intelligence-based method for predicting formation pressure. This invention utilizes pre-stack inversion to obtain formation pressure coefficient feature samples under the constraints of geological bodies in seismic data. It then uses a convolutional network to optimize the data, adjusting the weights and feature vectors within each layer to achieve the optimal result. Real well data is then appropriately incorporated into the subsequent training of the network for error backpropagation, thereby obtaining a maturely trained network that can be applied to formation pressure prediction, thus solving the problem of insufficient performance stability in traditional formation pressure prediction methods.

[0009] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0010] An artificial intelligence-based method for predicting formation pressure includes the following steps:

[0011] Step 1: Dimensionalization of the feature parameter space of attribute constraints to construct an expanded sample set

[0012] S11: Using pre-stack seismic data and real well data, reservoir attribute data are obtained through simultaneous pre-stack inversion;

[0013] S12: Combining rock physics analysis, characteristic parameters related to formation pressure coefficient are determined, and parameters with high sensitivity to formation pressure coefficient are selected as high-sensitivity characteristic parameters for real wells;

[0014] S13: Set up several points around the real well as pseudo wells. Use the vertical and horizontal variations of the geological body represented by the reservoir attribute data as spatial distribution constraints. Dimensionally increase the high-sensitivity characteristic parameters of the real wells through interpolation extrapolation to obtain the high-sensitivity characteristic parameters of all pseudo wells.

[0015] S14: Merge the high-sensitivity feature parameters of real wells and pseudo wells into an augmented sample set;

[0016] Step 2: Network Model Pre-training

[0017] Post-stack seismic records are obtained from pre-stack seismic data, and a nonlinear mapping relationship between post-stack seismic records and highly sensitive feature parameters is established. Then, the two-dimensional convolutional neural network model is pre-trained using post-stack seismic records and an augmented sample set.

[0018] Step 3: Optimize and update the network model

[0019] The highly sensitive feature parameters of the real well are used to optimize and update the pre-trained two-dimensional convolutional neural network model. After training, a mature network model is obtained.

[0020] Step 4: Predict using network models

[0021] The model inputs post-stack seismic records into a mature network model, outputs highly sensitive feature parameters, calculates formation pressure coefficients based on these parameters, and finally obtains the predicted formation pressure based on the formation pressure coefficients and hydrostatic column pressure.

[0022] In S11, the real well data includes the P-wave and S-wave velocities and density of the real well, and the reservoir attribute data includes the P-wave and S-wave velocities, density, lithology, and reservoir physical property data of the reservoir.

[0023] In S12, the characteristic parameters related to the formation pressure coefficient include acoustic transit time parameter, neutron porosity parameter, density parameter, natural gamma parameter, and resistivity parameter.

[0024] In S13, the method of using the vertical and lateral variations of geological bodies characterized by reservoir attribute data as spatial distribution constraints, and adding dimensionality to the highly sensitive characteristic parameters of real wells through interpolation extrapolation, is as follows: Based on seismic attributes, using a weight function of the similarity between real and pseudo wells and the well point location, real wells with high correlation to the location to be interpolated are selected for interpolation. The formula for calculating the weight function is:

[0025]

[0026] In the formula, λ represents the weighting factor that balances distance weighting and similarity weighting, and its value ranges from 0 to 1; α c This is the similarity radius factor, typically ranging from 1 to 10; c ij This represents the waveform difference between the actual well-side seismic record and the seismic record to be interpolated. A greater difference indicates a lower similarity between the two seismic records, and vice versa. ij =dis i -dis j This represents the distance between the actual well and the location to be interpolated.

[0027] In step two, the established nonlinear mapping relationship between the post-stack seismic record and the highly sensitive characteristic parameters is as follows:

[0028] y = Ω(x)

[0029] In the formula, Ω represents the nonlinear mapping relationship between the post-stack seismic record x and the highly sensitive characteristic parameter y.

[0030] Step two, pre-training the two-dimensional convolutional neural network model, refers to: using post-stack seismic records as input and high-sensitivity feature parameters from the augmented sample set as output; extracting features using the convolutional layers of the convolutional neural network; and adjusting the weights and feature vectors within each layer to optimize the mean squared error between the estimated values ​​of the high-sensitivity feature parameters obtained from the seismic records and the true values ​​in the augmented sample set.

[0031] In step two, the calculation formula for the two-dimensional convolutional neural network model is as follows:

[0032]

[0033] In the formula, l represents the number of network layers; u and v are the spatial dimensions of the convolution kernels, respectively; and These represent the j-th features of the l-th and (l-1)-th layers, respectively; The convolutional kernel connects two layers. The convolutional kernel is equivalent to several filters. In this paper, the convolutional kernel size is set to 3×3 and the number of convolutional kernel layers is 10. denoted by , which represents the bias corresponding to the j-th feature of the l-th layer; f is the activation function, and in this paper, the Exponential Linear Unit (ELU), a variant of the ReLU function, is chosen as the activation function.

[0034] The activation function incorporates exponentiation, calculated using the following formula:

[0035]

[0036] In the formula, u is the output of the previous layer of the network; f(u) is the output of the current network layer. When u>0, the output of the activation function is equal to the input; otherwise, the output value is close to 0. Therefore, the output will not experience gradient explosion due to the increase of the input, or the model will fail to learn because the weights cannot be updated when the input is less than 0.

[0037] In step two, the method for calculating the mean of the root mean square deviation is as follows:

[0038]

[0039] In the formula, θ is the mean of the root mean square error; x i (i = 1, 2, 3, ..., N) represents the training samples of N input post-stack seismic data; Ω(x i Let y be the estimated values ​​of N highly sensitive feature parameters. i (i = 1, 2, 3, ..., N) are the training samples for the N highly sensitive feature parameters of the output; It is the Frobenius norm.

[0040] In step three, optimizing and updating the pre-trained two-dimensional convolutional neural network model means: inputting the highly sensitive feature parameters of the real well into the two-dimensional convolutional neural network model to carry out cumulative learning, and backpropagating the error of the network until it gradually approaches the true value in the increased-dimensional sample set.

[0041] In step four, the method for calculating the formation pressure coefficient is as follows:

[0042] Y i = a1p1 + a2p2 + ... + a i p i +b

[0043] in,

[0044] P i =[p1,p2,...,p i ]

[0045] In the formula, Y i P is the formation pressure coefficient. i Highly sensitive feature parameters; a i,b i These are the weights and constants of the formation pressure coefficient, respectively, which can be calculated based on known well logging data and linear regression prediction models.

[0046] The advantages of using this invention are:

[0047] 1. The formation pressure prediction method of this invention includes four steps: constructing a sample set by increasing the dimensionality of the feature parameter space with attribute constraints, pre-training the network model, optimizing and updating the network model, and using the network model for prediction.

[0048] The advantage of step one is that it increases the size of the sample set by using seismic data constraints. The increase in the number of samples in the sample set can significantly improve the stability and generalization ability of the training model.

[0049] The advantage of step two is that it establishes a nonlinear relationship between post-stack seismic records and formation pressure coefficients, which allows for the prediction of formation pressure in three-dimensional space across the entire work area using seismic data, rather than being limited to well locations.

[0050] The advantage of step three is that it uses real wells to optimize and update the model, further improving the model's prediction accuracy.

[0051] In summary, this invention utilizes pre-stack inversion to obtain formation pressure coefficient feature samples under the geological constraints of seismic data, optimizes the data using a convolutional network, and adjusts the weights and feature vectors within each layer to achieve the optimal result. Real well data is then appropriately incorporated into the subsequent network training for error backpropagation, resulting in a maturely trained network that can be applied to formation pressure prediction, effectively addressing the performance instability issues of traditional formation pressure prediction methods.

[0052] 2. This invention significantly improves the performance and accuracy of traditional formation pressure prediction methods, enabling them to be widely applied to different lithological conditions and different formation pressure coefficient backgrounds. It ensures the performance stability of earthquake formation pressure prediction methods and accurately solves the "pain points" of traditional formation pressure prediction methods.

[0053] 3. The present invention has strong expansion capability. After obtaining the relationship between the seismic data of the work area and the formation pressure coefficient through the present invention, the formation pressure coefficient can be directly calculated using the model obtained by the present invention after obtaining seismic data in the adjacent areas of the subsequent work area, and finally the formation pressure can be calculated, which is more convenient to use. Attached Figure Description

[0054] Figure 1 This is a flowchart of the present invention;

[0055] Figure 2 This is a schematic diagram of spatial dimensionality enhancement for highly sensitive feature parameters;

[0056] Figure 3 This is a graph of the ELU activation function;

[0057] Figure 4 This is a distribution map of formation pressure coefficients in the No. 6 sand group of the JQ5H well area;

[0058] Figure 5 This is an inversion profile of the P-wave and S-wave velocity ratios through well JQ511;

[0059] Figure 6 This is a P-wave velocity inversion profile of well JQ511;

[0060] Figure 7 This is a density inversion profile of well JQ511;

[0061] Figure 8 It is a lithofacies probability PDF diagram;

[0062] Figure 9 This is a predicted distribution map of the pressure coefficient of the 6th sand group in the second section of the JQ5H well area;

[0063] Figure 10 This is a comparison chart of the measured formation pressure coefficient and the formation pressure coefficient predicted by this invention. Detailed Implementation

[0064] Example 1

[0065] like Figure 1 As shown, this invention provides an artificial intelligence-based method for predicting formation pressure, which includes the following steps:

[0066] Step 1: Dimensionalization of the feature parameter space of attribute constraints to construct an expanded sample set

[0067] S11: Acquire pre-stack seismic data and well data. The acquired well data includes the P-wave and S-wave velocities and densities of the wells. Then, using the pre-stack seismic data and well data, reservoir attribute data is obtained through simultaneous pre-stack inversion. Specifically, the reservoir attribute data includes the reservoir's P-wave and S-wave velocities, density, lithology, and reservoir physical properties.

[0068] S12: Combining rock physics analysis, identify characteristic parameters in the well that have clear rock physics significance and are related to the formation pressure coefficient. Then, select the parameters with high sensitivity to the formation pressure coefficient as the high-sensitivity characteristic parameters of the actual well.

[0069] Typically, characteristic parameters related to formation pressure coefficient include acoustic transit time (AC), neutron porosity (CNL), density (DEN), natural gamma ray (GR), and resistivity (RT). Highly sensitive characteristic parameters to formation pressure coefficient are those in known wells where statistical analysis of known characteristic parameters and known formation pressure coefficients reveals that even small changes in the characteristic parameter can cause significant changes in the formation pressure coefficient.

[0070] It should be noted that, given the limited number of known wells, the sample size of formation pressure coefficients is small and does not match the spatial scale of seismic data. Using only the formation pressure coefficient parameters of well points will miss samples from other high-pressure development zones, making it difficult to meet the completeness requirements of deep learning. Therefore, this embodiment selects parameters with high sensitivity to formation pressure coefficients as highly sensitive feature parameters for actual wells, which improves the completeness of the samples, thereby satisfying the subsequent deep learning requirements of the model and improving its accuracy.

[0071] S13: As Figure 2 As shown, several points are set as pseudo-wells around the real well. The vertical and horizontal variations of the geological body characterized by the reservoir attribute data are used as spatial distribution constraints. The high-sensitivity characteristic parameters of the real well are augmented by interpolation extrapolation to obtain the high-sensitivity characteristic parameters of all pseudo-wells.

[0072] Specifically, the method of using the vertical and lateral variations of geological bodies represented by reservoir attribute data as spatial distribution constraints, and then using interpolation extrapolation to increase the dimensionality of highly sensitive characteristic parameters of real wells, is as follows: Based on seismic attributes, using a weighting function of the similarity between real and pseudo-wells and the well point location, real wells with high correlation to the location to be interpolated are selected for interpolation. The formula for calculating the weighting function is as follows:

[0073]

[0074] In the formula, λ represents the weighting factor that balances distance weighting and similarity weighting, and its value ranges from 0 to 1; α c This is the similarity radius factor, typically ranging from 1 to 10; c ij This represents the waveform difference between the actual well-side seismic record and the seismic record to be interpolated. A greater difference indicates a lower similarity between the two seismic records, and vice versa. ij =dis i -dis j This represents the distance between the actual well and the location to be interpolated.

[0075] This step augments the dimensionality of the high-sensitivity feature parameters of the real wells through interpolation and extrapolation, thereby obtaining the high-sensitivity feature parameters of all pseudo wells. On the one hand, this increases the sample data of the real wells, which is beneficial for deep learning and establishing a stable network model; on the other hand, it can improve the problem of low inversion accuracy caused by uneven distribution of real well sample data.

[0076] S14: Merge the high-sensitivity feature parameters of real wells and pseudo wells into an augmented sample set.

[0077] Step 2: Network Model Pre-training

[0078] Post-stack seismic records are obtained from pre-stack seismic data, and a nonlinear mapping relationship between post-stack seismic records and highly sensitive feature parameters is established. Then, a two-dimensional convolutional neural network model is pre-trained using post-stack seismic records and an augmented sample set.

[0079] Specifically, the established nonlinear mapping relationship between post-stack seismic records and highly sensitive characteristic parameters is as follows:

[0080] y = Ω(x)

[0081] In the formula, Ω represents the nonlinear mapping relationship between the post-stack seismic record x and the highly sensitive characteristic parameter y.

[0082] Pre-training a two-dimensional convolutional neural network model refers to taking post-stack seismic records as input, using highly sensitive feature parameters from the augmented sample set as output, extracting features using the convolutional layers of the convolutional neural network, and adjusting the weights and feature vectors within each layer to optimize the mean squared error between the estimated values ​​of the highly sensitive feature parameters obtained from the seismic records and the true values ​​in the augmented sample set.

[0083] It should be noted that the true values ​​in the augmented sample set can be obtained through actual measurement using well logging methods.

[0084] The calculation formula for the two-dimensional convolutional neural network model is as follows:

[0085]

[0086] In the formula, l represents the number of network layers; u and v are the spatial dimensions of the convolution kernels, respectively; and These represent the j-th features of the l-th and (l-1)-th layers, respectively; The convolutional kernel connects two layers. The convolutional kernel is equivalent to several filters. In this paper, the convolutional kernel size is set to 3×3 and the number of convolutional kernel layers is 10. The bias represents the j-th feature of the l-th layer; f is the activation function, such as... Figure 3As shown, in this embodiment, the Exponential Linear Unit (ELU), a variant of the ReLU function, is selected as the activation function;

[0087] The activation function incorporates exponentiation, calculated using the following formula:

[0088]

[0089] In the formula, u is the output of the previous layer of the network; f(u) is the output of the current network layer. When u>0, the output of the activation function is equal to the input; otherwise, the output value is close to 0. Therefore, the output will not experience gradient explosion due to the increase of the input, or the model will fail to learn because the weights cannot be updated when the input is less than 0.

[0090] The method for calculating the mean of the root mean square deviation is as follows:

[0091]

[0092] In the formula, θ is the mean of the root mean square error; x i (i = 1, 2, 3, ..., N) represents the training samples of N input post-stack seismic data; Ω(x i Let y be the estimated values ​​of N highly sensitive feature parameters. i (i = 1, 2, 3, ..., N) are the training samples for the N highly sensitive feature parameters of the output; It is the Frobenius norm.

[0093] Step 3: Optimize and update the network model

[0094] The highly sensitive feature parameters of the real well are used to optimize and update the pre-trained two-dimensional convolutional neural network model. After training, a mature network model is obtained.

[0095] Specifically, optimizing and updating the pre-trained two-dimensional convolutional neural network model means: inputting the highly sensitive feature parameters of the real well into the two-dimensional convolutional neural network model to carry out cumulative learning, and backpropagating the network error until it gradually approaches the true value in the augmented sample set.

[0096] Step 4: Predict using network models

[0097] The model inputs post-stack seismic records into a mature network model, outputs highly sensitive feature parameters, calculates formation pressure coefficients based on these parameters, and finally obtains the predicted formation pressure based on the formation pressure coefficients and hydrostatic column pressure.

[0098] Specifically, the calculation method for the formation pressure coefficient is as follows:

[0099] Y i= a1p1 + a2p2 + ... + a i p i +b

[0100] in,

[0101] P i =[p1,p2,...,p i ]

[0102] In the formula, Y i P is the formation pressure coefficient. i Highly sensitive feature parameters; a i ,b i These are the weights and constants of the formation pressure coefficient, respectively, which can be calculated based on known well logging data and linear regression prediction models.

[0103] It should be noted that the post-stack seismic records in step two only include seismic records at the well location, while the post-stack seismic records input in this step include not only seismic records at the well location but also seismic records at other locations outside the well, resulting in a larger data volume.

[0104] This invention can effectively predict formation pressure based on the formation pressure coefficient, significantly improving the performance accuracy of traditional formation pressure prediction methods. It enables formation pressure prediction methods to be widely applied to different lithological conditions and different pressure coefficient backgrounds, ensuring the performance stability of earthquake formation pressure prediction methods.

[0105] Example 2

[0106] This embodiment selects a seismic data volume from a certain work area to conduct a prediction of the formation pressure of the Sha-2 section of the river channel within the area to further illustrate the present invention. Considering the different geomechanical environments of different fault blocks, all wells are located in the same fault block in actual use. Taking a certain fault block as an example, some wells are selected as training data, and a certain number of wells are reserved not to participate in the training for the purpose of verifying the prediction results.

[0107] like Figure 4 As shown, based on the statistical results of measured pressure points in the study area, there are 27 wells in Sand Formation No. 6. 21 wells (78%) have a formation pressure coefficient greater than 0.8, while 6 wells (22%) have a formation pressure coefficient less than 0.8. This indicates that the overall formation pressure coefficient of Channel No. 6 is relatively high and changes rapidly.

[0108] in addition, Figure 5-7The inversion profiles of the P-wave velocity ratio, P-wave velocity, and density from well JQ511 are shown. The values ​​next to the well in the inversion profile are the square-wave converted values ​​of the corresponding curves. Warm colors in the inversion profile represent relatively low velocity ratio, low velocity, and low density. It can be seen that sandstone exhibits low velocity, low density, and low velocity ratio characteristics, which correspond well to the well logging results, indicating reliable inversion results. Based on the pre-stack simultaneous inversion, and according to geological analysis and rock physics statistical analysis, a reasonable probability density function for the intersection of P-wave impedance and Vp / Vs in sandstone and mudstone is obtained, as shown below. Figure 8 As shown, through lithofacies probability analysis, a sandstone probability volume is obtained, which provides a phase-controlled constraint basis for subsequent phase-controlled pore pressure prediction.

[0109] Figure 9 , 10 The distribution map of the measured formation pressure coefficient and the predicted formation pressure coefficient using the present invention is shown for the 6th sandstone group of the second sandstone section in the JQ5H well area. The measured and predicted formation pressure coefficients for some wells are now compared, as shown in the table below:

[0110]

[0111]

[0112] from Figure 10 Based on the prediction results in the table above, the distribution pattern of the pressure coefficient in channel 6 of the JQ5H well area is basically consistent with the geological analysis results: The western shallow 3 well of sand group 6 exhibits multi-stage superimposed channels, with good quantitative parameters of sand body connectivity at individual well points. The channel sections show strong seismic amplitude and good continuity of reflection phase axes in both plan and profile views, indicating an overall connected channel section. Connected channels generally have higher pressure coefficients. The two intersecting channels in the central part have high pressure coefficients, generally greater than 0.8, consistent with geological understanding. Tectonic structure also affects the pressure coefficient. Well Jinhua 201 is located in a slight anticline area, and the formation pressure coefficient of sand group 6 is higher than the surrounding area, consistent with the prediction results and geological understanding. Influenced by the natural gas escape channel provided by the western fault, the formation pressure coefficient of sand group 6 in wells JQ517 and eastwards is all less than 0.8, indicating a low formation pressure coefficient, consistent with the prediction results and geological understanding. Overall, the No. 6 sand group mainly exhibits a favorable combination of source-faulted sandstone formations with low-pressure to normal-pressure systems, while locally it exhibits an unfavorable combination of source-faulted sandstone formations with abnormally low-pressure systems. The artificial intelligence pore pressure prediction results effectively characterize this pattern, and the results are reliable. Furthermore, a quantitative comparison between the actual formation pressure coefficient and the formation pressure coefficient predicted by this method shows that the overall average error (absolute prediction error / measured pressure coefficient) of the final predicted formation pressure is small, and the prediction accuracy can reach over 90%, meeting the prediction requirements.

[0113] The above description is merely a specific embodiment of the present invention. Any feature disclosed in this specification may be replaced by other equivalent or similar features unless otherwise specified. All features or steps in the disclosed methods or processes may be combined in any way, except for mutually exclusive features and / or steps.

Claims

1. An artificial intelligence-based formation pressure prediction method, characterized by The method comprises the following steps: Step one: attribute constraint characteristic parameter space dimension increasing construction of dimension increasing sample set S11: obtaining attribute data of the reservoir by prestack simultaneous inversion using prestack seismic data and true well data; S12: determining characteristic parameters related to the formation pressure coefficient in combination with rock physical analysis, and selecting parameters with high sensitivity to the formation pressure coefficient as high sensitivity characteristic parameters of the true well; S13: setting a plurality of points as pseudo wells around the true well, using longitudinal and lateral changes of the geological body represented by the attribute data of the reservoir as spatial distribution phase constraints, and increasing the dimension of the high sensitivity characteristic parameters of the true well by interpolation extrapolation to obtain high sensitivity characteristic parameters of all the pseudo wells; S14: combining the high sensitivity characteristic parameters of the true well and the high sensitivity characteristic parameters of the pseudo wells into a dimension increasing sample set; Step two: network model pre-training obtaining poststack seismic records according to the prestack seismic data, establishing a nonlinear mapping relationship between the poststack seismic records and the high sensitivity characteristic parameters, and pre-training a two-dimensional convolutional neural network model using the poststack seismic records and the dimension increasing sample set; Step three: network model optimization and updating optimizing and updating the pre-trained two-dimensional convolutional neural network model using the high sensitivity characteristic parameters of the true well, and obtaining a mature network model after the training is completed; Step four: network model prediction inputting the poststack seismic records into the mature network model, outputting the high sensitivity characteristic parameters, calculating the formation pressure coefficient according to the high sensitivity characteristic parameters, and finally obtaining the predicted formation pressure according to the formation pressure coefficient and the static column pressure.

2. The method of predicting formation pressure based on artificial intelligence according to claim 1, characterized in that: In the S11, the true well data includes longitudinal and transverse wave velocities and density of the true well, and the attribute data of the reservoir includes longitudinal and transverse wave velocities, density, lithology and reservoir physical property data of the reservoir.

3. The method of predicting formation pressure based on artificial intelligence according to claim 1, characterized in that: In the S12, the characteristic parameters related to the formation pressure coefficient include acoustic time difference parameters, neutron porosity parameters, density parameters, natural gamma parameters and resistivity parameters.

4. The method of claim 1, wherein: In the S13, the method for increasing the dimension of the high sensitivity characteristic parameters of the true well by interpolation extrapolation using the longitudinal and lateral changes of the geological body represented by the attribute data of the reservoir as spatial distribution phase constraints is as follows: according to the seismic attribute, the similarity between the true well and the pseudo well and the weight function of the well point position are used to select the true well with high correlation degree to the position to be interpolated for interpolation, wherein the calculation formula of the weight function is: where λ represents a weighting factor for balancing the distance weighting and the similarity weighting, and λ is in the range of 0-1; a c is a similarity radius factor, and a is usually in the range of 1-10; c ij represents the waveform difference between the well seismic record of the true well and the seismic record to be interpolated; d ij = dis i -dis j represents the distance between the true well and the position to be interpolated.

5. The method of predicting formation pressure based on artificial intelligence according to any one of claims 1-4, characterized in that: In step two, the nonlinear mapping relationship between the poststack seismic records and the high sensitivity characteristic parameters is: y = Ω (x) In the formula, Ω represents the nonlinear mapping relationship between the poststack seismic records x and the high sensitivity characteristic parameters y.

6. The method of predicting formation pressure based on artificial intelligence according to claim 5, characterized in that: In step two, the pre-training of the two-dimensional convolutional neural network model means that the poststack seismic records are taken as input, the high sensitivity characteristic parameters in the dimension increasing sample set are taken as output, the convolutional layer of the convolutional neural network is used to extract features, the weight values and feature vectors in the layer are adjusted layer by layer, and the mean square error average value of the estimated value of the high sensitivity characteristic parameters obtained from the seismic records and the true value in the dimension increasing sample set is optimized 7. The method of predicting formation pressure based on artificial intelligence according to claim 6, characterized in that: In step two, the calculation formula of the two-dimensional convolutional neural network model is: In the formula, l represents the number of network layers; u, v are the spatial dimension size of the convolution kernel respectively; and represents the jth feature of the lth layer and the l-1th layer respectively; is the convolution kernel connecting two layers, the convolution kernel is equivalent to a number of filters, the size of the convolution kernel is set to 3*3 in this paper, and the number of convolution kernel layers is 10; represents the bias corresponding to the jth feature of the lth layer; f is an activation function; In the formula, the activation function introduces power calculation, and the calculation formula is: In the formula, u is the output of the previous network layer; f(u) is the output of the current network layer, and when u>0, the activation function outputs the same as the input, otherwise the output value is close to 0.

8. The method of predicting formation pressure based on artificial intelligence according to claim 7, characterized in that: In step two, the calculation method of the mean square error average value is: where θ is the mean of the mean square deviation; x i (i = 1, 2, 3,..., N) are training samples of N input stacked seismic data; Ω(x i ) are estimated values of N high-sensitive feature parameters, y i (i = 1, 2, 3,..., N) are training samples of N output high-sensitive feature parameters; is the Frobenius norm.

9. The method of predicting formation pressure based on artificial intelligence according to claim 1, characterized in that: In step three, the optimization and update training of the pre-trained two-dimensional convolutional neural network model refers to: inputting the high sensitivity characteristic parameters of the true well into the two-dimensional convolutional neural network model to carry out cumulative learning, and performing error back propagation on the network until gradually approaching the true value in the dimension-increased sample set.

10. The method of claim 8, wherein: In step four, the calculation method of the formation pressure coefficient is: Y i = a1p1+ a2p2+... +a i p i +b wherein, P i = [p1, p2,..., p i ] where Y i is the formation pressure coefficient; P i is the high sensitivity characteristic parameter; a i ,b i are the weight and constant term of the formation pressure coefficient, respectively, which can be calculated according to known logging data and linear regression prediction model.

Images