Physical embedded deep learning formation pressure prediction method, device, medium and equipment

Abstract

The present invention discloses a physical embedded deep learning formation pressure prediction method, device, medium and equipment, the present invention characterizes seismic attenuation by logging impedance quality factor Q, based on the Q value and rock physics model of formation pressure, the physical mechanism of this kind of certainty replace Caianiello convolution neurons of the nonlinear activation function, using the convolution neurons, build deep learning convolution neural networks (CCNNs), can greatly increase the stress inversion precision and learning efficiency, get accurate formation pressure prediction results. Compared with the prior art, the present invention uses acoustic attenuation instead of the traditional acoustic velocity to characterize formation pressure, and solves the problem that the traditional pressure prediction method based on velocity has strong multiple solutions due to high gas content and complex structure.

Claims

1. A physical embedded deep learning formation pressure prediction method, characterized in that, comprising the following steps: S1. Collection of data during oil and gas exploration and development; S2. According to the data collected in step S1, the logging impedance quality factor Q is calculated by the slippage windows method; S3. According to the data collected in step St and the logging impedance quality factor Q obtained in step S2, a Q-pressure rock physical model based on well-logging data is established and cross-verified; S4. According to the data collected in step S1, the well seismic correlation analysis is performed on the drilling position, and the best frequency component of the seismic data is determined. Based on the optimal frequency component, the Q-pressure rock physical model based on well-logging data obtained in step S3 is extended to the well-side seismic data, and the Q-pressure rock physical model based on seismic data is established; S5. The nonlinear activation function of Caianiello convolution neuron is replaced with the Q-pressure rock physics model based on seismic data obtained in Step S4, mid the deep learning convolution neural networks are constructed by using the convolution neuron to obtain the physical embedded deep learning formation pressure prediction networks CCNNs; S6. The physical embedded deep learning formation pressure prediction networks CCNNs obtained in Step S5 is used to invert formation pressure.

2. The physical embedded deep learning formation pressure prediction method according to claim 1, characterized in that, wherein the data in step S1 includes 3D seismic data, well-logging data and drilling pressure test (DST) data.

3. The physical embedded deep learning formation pressure prediction method according to claim 1, characterized in that, in step S2, the value of logging impedance quality factor Q can be obtained according to the calculation formula: $Q=\frac{\pi f-\frac{{\alpha }^2{v}^2}{4\pi f}}{\alpha v}$ Wherein f is the peak frequency, α is the attenuation coefficient, v is the acoustic velocity.

4. The physical embedded deep learning formation pressure prediction method according to claim 3, characterized in that, seismic impedance z is used to replace v, and the logging impedance quality factor Q is calculated.

5. The physical embedded deep learning formation pressure prediction method according to claim 3, characterized in that, the calculation method of the logging impedance quality factor Q is as follows: For well impedance data, the slippage windows method is used to calculate the logging impedance quality factor Q, time-frequency transformation is carried out for each slide, peak frequency and attenuation coefficient are calculated, and window length and sliding step length are adjusted to obtain a reasonable Q value.

6. The physical embedded deep learning formation pressure prediction method according to claim 1, characterized in that, in step S4, the method for determining the optimal frequency component of the well-side seismic data described is: the seismic trace is synthesized from well log data, the phase correlation between well and seismic is evaluated, and the time-depth conversion information is obtained. The well impedance is filtered to match the seismic frequency band, and the correlation analysis is conducted with the relative impedance channel of the well-side seismic. The correlation coefficient is calculated as a frequency function to determine the optimal frequency component of the well-side seismic data. The Q-pressure rock physics model based on well-logging data is extended to well-side seismic data. Signal similarity is described by correlation function.

7. The physical embedded deep learning formation pressure prediction method according to claim 1, characterized in that, the correlation function is obtained: ${\rho }_{\mathrm{xy}}\left(\tau \right)=\frac{\frac{1}{2T}{\int }_{-T}^{T}x\left(t\right)y\left(t+\tau \right)\mathrm{dt}}{\sqrt{\frac{1}{2T}{\int }_{-T}^T{x\left(t\right)}^2\mathrm{dt}}\sqrt{\frac{1}{2T}{\int }_{-T}^T{y\left(t+\tau \right)}^2\mathrm{dt}}}$ Wherein x(t) and y(t) is two signals with period 2T, τ is the shift of the signal in time, dt is the derivative with respect to time, ρ.sub.xy(τ) is the correlation coefficient between signal y(t) and signal x(t) after shift by τ.

8. The physical embedded deep learning formation pressure prediction method according to claim 1, characterized in that, in step S6, the inversion method of formation pressure includes the following steps: (1) Pre-positioning CCNNs neural wavelet by well-side deep learning; (2) Inversion of initial pressure model based on CCNNs deconvolution; (3) The initial pressure model is modified based on CCNNs input signal reconstruction method.

9. The physical embedded deep learning formation pressure prediction method according to claim 8, characterized in that, in step S6(2), according to the deep learning formation pressure prediction networks CCNNs, which is physically embedded and obtained in Step S5, the initial pressure model is obtained through deconvolution inversion.

10. The physical embedded deep learning formation pressure prediction method according to claim 9, characterized in that, in step S6(3), the specific method of modifying the initial pressure model is as follows: The initial pressure model is taken as the input signal, the actual 3D seismic data is taken as the expected output, and the initial pressure model is modified by the input signal reconstruction method to minimize the error between the neural network output and the expected output, namely, the formation pressure prediction result is obtained.

11. The physical embedded deep learning formation pressure prediction method according to claim 10, characterized in that, the input signal reconstruction formula is obtained: $\Delta o_i\left(t\right)=-\eta \left(t\right)\frac{\partial E}{\partial o_i\left(t\right)}=\underset{j}{.Math.}\eta \left(t\right){\delta }_j\left(t\right).Math.w_{ji}\left(t\right)$ Wherein Δo.sub.i(t) is the input signal correction, η(t) is the learning efficiency, E is the cost function, o.sub.i(t) is the input signal, δ.sub.j(t) is the back propagation error, w.sub.ji(t) is the neural wavelet, .Math. is the correlation operator, $\frac{\partial E}{\partial o_i\left(t\right)}$ is the partial derivatives of cost function to the input signal.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 is a schematic diagram of three-layer CCNN structure.

(2) FIG. 2 is a method flow chart of the present invention.

(3) FIG. 3 is the formation pressure sensitivity analysis diagram.

(4) FIG. 4 is the non-linear fitting diagram of quality factor and formation pressure.

(5) FIG. 5 is the Q-pressure rock physical model of six Wells in the actual work area.

(6) FIG. 6 is a schematic diagram of well-seismic correlation analysis.

(7) FIG. 7 is the inversion results of formation pressure by CCNNs—taking the Xihu depression as an example.

(8) FIG. 8 is the comparison of formation pressure prediction results of validation Wells with traditional methods.

DESCRIPTION OF THE EMBODIMENTS

(9) The content of the present invention will be further described below with reference to the accompanying drawings. The experimental methods described in the following examples are conventional methods unless otherwise specified; the reagents and materials can be obtained from commercial sources unless otherwise specified.

(10) A physical embedded deep learning formation pressure prediction method, comprising the following steps:

(11) S1. Collection of data.

(12) 3D seismic data, well-logging data and drilling pressure test (DST) data are collected during exploration and development of oil and gas-bearing basin.

(13) S2. The logging impedance quality factor Q is calculated.

(14) According to the data collected in step S1, the logging impedance quality factor Q is calculated by the slippage windows method.

(15) The value of logging impedance quality factor Q can be obtained according to the calculation formula:

(16) $Q=\frac{\pi f-\frac{{\alpha }^2{v}^2}{4\pi f}}{\alpha v}$

(17) Wherein f is the peak frequency, α is the attenuation coefficient, v is the acoustic velocity.

(18) In order to better integrate well seismic data, seismic impedance z is used to replace v and the logging impedance quality factor Q is calculated. For well impedance data, the slippage windows method is used to calculate the well impedance quality factor Q, time-frequency transformation is carried out for each slide, peak frequency and attenuation coefficient are calculated, and window length and sliding step length are adjusted to obtain a reasonable Q value.

(19) S3. A Q-pressure rock physical model based on well-logging data is established.

(20) According to the data obtained in step S1 and the logging impedance quality factor Q obtained in step S2, a Q-pressure rock physical model based on well-logging data is established, and the model is cross-verified. The inversion results are consistent with the measured pressure curve, and the error is within 10%.

(21) S4. Determine seismic data, the Q-pressure rock physical model based on seismic data is established.

(22) According to the data in step S1, the well seismic correlation in drilling position is analyzed by synthetic seismic records, and the optimal frequency component of seismic data near the well is determined. According to the optimal frequency component of the well-side seismic data obtained, the Q-pressure rock physical model based on well-logging data obtained in step S3 is extended to the well-side seismic data, and the Q-pressure rock physical model based on seismic data is established.

(23) Specifically, synthetic seismograms as a traditional tool are produced to evaluate the well-seismic correlation in phase, and the time-depth conversion information is obtained. The well impedance is filtered to match the seismic frequency band, and the correlation analysis is conducted with the relative impedance channel of the well-side seismic. The correlation coefficient is calculated as a frequency function to determine the optimal frequency component of the well-side seismic data. The Q-pressure rock physics model based on well-logging data is extended to well-side seismic data. Signal similarity is described by correlation function.

(24) Signal similarity is described by correlation function:

(25) ${\rho }_{\mathrm{xy}}\left(\tau \right)=\frac{\frac{1}{2T}{\int }_{-T}^{T}x\left(t\right)y\left(t+\tau \right)\mathrm{dt}}{\sqrt{\frac{1}{2T}{\int }_{-T}^T{x\left(t\right)}^2\mathrm{dt}}\sqrt{\frac{1}{2T}{\int }_{-T}^T{y\left(t+\tau \right)}^2\mathrm{dt}}}$

(26) Wherein X(t) and y(t) is two signals with period 2T, τ is the shift of the signal in time, di is the derivative with respect to time, ρ.sub.xy(τ) is the correlation coefficient between signal y(t) and signal x(t) after shift by T.

(27) S5. A physical embedded deep learning formation pressure prediction networks CCNNs is constructed.

(28) The nonlinear activation function of Caianiello convolution neuron is replaced by the Q-pressure rock physics model based on seismic data obtained in Step S4, and the deep learning convolution neural networks are constructed using the convolution neuron to obtain the physical embedded deep learning formation pressure prediction networks CCNNs.

(29) The Caianiello neuron activation function regulates the input and transmission of neural signals through specific physicochemical models, allowing us to embed certain geophysical mechanisms into Caianiello neurons as activation functions for their specific geophysical applications. Deterministic models with physical significance provide a physical information learning with powerful feature extraction, where consistency of information distribution is enhanced across neurons to infer information lost in data or recover information contaminated by noise. For pressure inversion, the Q-pressure rock physical model obtained in Step S3 is used.

(30) S6. Inversion of formation pressure.

(31) The physical embedded deep learning formation pressure prediction networks CCNNs obtained in Step S5 is used to invert formation pressure.

(32) The inversion method of formation pressure includes the following steps:

(33) (1) Pre-positioning CCNNs neural wavelet by well-side deep learning;

(34) (2) Inversion of initial pressure model based on CCNNs deconvolution;

(35) (3) The initial pressure model is modified based on CCNNs input signal reconstruction method.

(36) A typical CCNN has a three-layer structure of input signal, hidden layer, and expected output (FIG. 1). By selecting different input expected output and nonlinear activation functions, this neural network can be applied to most inverse problems. Neural wavelets w.sub.ji(t) need to be constructed between different layers for signal transmission. Neural wavelets are adaptive and provide CCNN with the ability to learn time-varying signals. The cost function of this problem is defined as follows:

(37) $E=\frac{1}{2}\underset{k}{.Math.}\underset{t}{.Math.}{\left[d_k\left(t\right)-o_k\left(t\right)\right]}^2$

(38) Wherein d.sub.k(t) is the desired output signal, o.sub.k(t) is the network actual output signal.

(39) By applying the back-propagation learning algorithm to minimize the cost function through each layer, the updated neural wavelet equation can be obtained. The updated equation has a general recursive form for any neuron. For example, from the hidden layer to the input layer, the neural wavelet correction can be expressed as:
Δw.sub.ji(t)=η(t)δ.sub.j(t).Math.o.sub.i(t)

(40) Wherein .Math. is the correlation operator, η(t) is the learning efficiency.

(41) Neural wavelet is adapted to the input information environment through the learning process. During the learning process, a large number of neurons with similar time spectra in the wavelet are statistically coded to improve the focusing resolution characteristics of the neurons. The Caianiello neuron equation uses the following formula:

(42) $\Delta o_i\left(t\right)=f\left(\underset{j=1}{\stackrel{N}{.Math.}}{\int }_0^t{w}_{\mathrm{ij}}\left(\tau \right)o_j\left(t-\tau \right)d\tau -{\theta }_i\left(t\right)\right)$

(43) Wherein o.sub.j(t) is the neuronal input, o.sub.i(t) is the neuronal output, θ.sub.i(t) is the neuron threshold, f(⋅) is the activation function, w.sub.ji(t) is the neural wavelet.

(44) Get physical embedded deep learning formation pressure prediction networks CCNNs after the initial pressure is obtained by deconvolution inversion model, described the initial pressure model as the input signal, the actual 3D seismic data as the desired output, fixed initial pressure model with input signal reconstruction method, the neural network output and desired output error minimization, namely the formation pressure prediction test results.

(45) More preferably, the input signal reconstruction formula is obtained:

(46) $\Delta o_i\left(t\right)=-\eta \left(t\right)\frac{\partial E}{\partial o_i\left(t\right)}=\underset{j}{.Math.}\eta \left(t\right){\delta }_j\left(t\right).Math.w_{\mathrm{ji}}\left(t\right)$

(47) Wherein Δo.sub.i(t) is the input signal correction, η(t) is the learning efficiency, E is the cost function, o.sub.i(t) is the input signal, δ.sub.j(t) is the back propagation error, w.sub.ji(t) is the neural wavelet, .Math. is the correlation operator,

(48) 0$\frac{\partial E}{\partial o_i\left(t\right)}$
is the partial derivatives of cost function to the input signal.

(49) The second object of the present invention is to provide a physical embedded deep learning formation pressure prediction device, comprising:

(50) Module 1: Used to collect and store data in the process of oil and gas exploration and development.

(51) Module 2: Used for logging impedance quality factor Q calculation, and establish Q-pressure rock physical model based on well-logging data and cross-verify the model.

(52) Module 3: Used for determining the optimal frequency components of well-side seismic data and building Q-pressure rock physics models based on seismic data.

(53) Module 4: Use to replace the nonlinear activation function of Caianiello convolution neuron with the Q-pressure rock physics model based on seismic data, and construct the deep learning convolution neural networks by using the convolution neuron to obtain the physical embedded deep learning formation pressure prediction networks CCNNs.

(54) Module 5: Use to predict formation pressure through physical embedded deep learning formation pressure prediction networks CCNNs inversion.

(55) The third object of the present invention is to provide a computer readable storage medium on which a computer program is stored, and the steps of the physical embedded deep learning formation pressure prediction method are realized when the computer program is executed by the processor.

(56) The fourth object of the present invention is to provide a computer device includes a memory processor and a computer program stored on the memory that can be run on the processor executing the computer program implementing the steps of the physical embedded deep learning formation pressure prediction method.

Example 1

(57) As shown in FIG. 2, the present invention provides a physical embedded deep learning formation pressure prediction method, which consists of the following steps:

(58) (1) Collection of seismic data, well-logging data, and drilling pressure test (DST) data collected during exploration and development of oil and gas-bearing basin.

(59) (2) Using the slippage windows method to calculate logging impedance quality factor Q, can adjust the length of the window and sliding step, to obtain a reasonable Q value, and the formation pressure curve and Q curve and other logging curves are sensitivity analysis (FIG. 3). For example, in the actual work area K1 Wells, it can be seen that well logging curve and impedance quality factor Q is negative correlation between formation pressure curve, conventional logging curves such as density, acoustic time lag, and impedance, have poor response to pressure changes.

(60) (3) Nonlinear fitting of logging impedance quality factor Q and formation pressure is carried out to obtain their rock physical correspondence (FIG. 4), and Q-pressure rock physical model is established in several Wells in the actual work area (FIG. 5). In the figure, the solid line is the well formation pressure curve and the well impedance quality factor curve, while the dotted line is the inversion pressure curve of logging impedance quality factor. It can be seen that the inversion result is consistent with the measured pressure curve, and the error is within 10%.

(61) (4) Through the synthetic seismogram analysis well seismic phase correlation, and get a deep transformation information, impedance curve of well filtering to match the seismic frequency band, and with earthquake relative impedance frequency domain correlation analysis, the seismic amplitude information (FIG. 6), reliability high correlation and well seismic data quality can ensure the rock physics model transition from well-logging data to the seismic data.

(62) (5) To embed the Q-pressure of rock physical model CCNNs networks for learning and nonlinear inversion of seismic data of the physical information, in the process of inversion, the use of each well in the study area of rock physics Q-pressure model as well in the activation of neurons in the function, the function defines the properties of neurons, actually, the neurons in the far away from the hole using interpolation adjacent holes or extrapolated from its activation function. The activation function CCNNs inversion of formation pressure mainly includes: pre-positioning CCNNs neural wavelet by well-side deep learning, inversion of initial pressure model based on CCNNs deconvolution, improvement and modification of initial pressure model based on CCNNs input signal reconstruction method.

(63) Input signal deconvolution to get initial model of formation pressure, the expected output for practical 3D seismic data, Caianiello convolution of the CCNNs input layer neurons are arranged in exactly the same configuration with seismic data of plane, namely the spatial coordinates of these neurons according to the seismic trace positioning. The initial model was improved through input signal reconstruction to minimize the error between neural network output and expected output, and finally accurate formation pressure prediction results were obtained (FIG. 7). Among them, well W2 is a verification well, and the accuracy of correlation calculation between prediction results and actual well pressure curve reaches 86%. However, the calculation results of traditional Eaton method have great errors, especially in high-pressure sections, seriously deviating from the actual formation pressure (FIG. 8). It can accurately predict the formation pressure in the oil-gas overpressure system.

(64) The above detailed description is a specific description of one of the feasible embodiments of the present invention, which is not intended to limit the scope of the present invention. Any equivalent implementation or modification that does not depart from the present invention shall be included in the present invention.