RESERVOIR PARAMETER PREDICTION METHOD AND APPARATUS BASED ON GEOLOGICAL CHARACTERISTIC CONSTRAINT, AND STORAGE MEDIUM

Abstract

A method, an apparatus, a computer storage medium and a computer device for geological characteristic constraint-based reservoir parameter prediction are provided. The method includes: selecting dominant seismic attributes according to the relevance between different types of seismic attributes of a target stratum and reservoir parameters (S100); on the basis of the dominant seismic attributes, classifying seismic waveforms of the target stratum by a preset waveform classification network model and according to waveform features, so as to obtain a waveform classification result (S200); taking the waveform classification result as a constraint, and constructing different deep neural network models corresponding to different geological characteristics (S300); fusing different trained deep neural network models into a set of spatially varying neural network prediction model (S500); and predicting the reservoir parameters of the target stratum.

Description

FIELD OF TECHNOLOGY

The present disclosure relates to the technical field of geophysical exploration, and in particular, to a reservoir parameter prediction method and apparatus based on geological characteristic constraint, a computer storage medium and a computer device.

BACKGROUND

With development of prestack depth migration technology and increasingly wide application thereof in seismic data process, reservoir parameter prediction directly performed in a depth domain is of a great significance. Technologies for seismic reservoir prediction in a depth domain include three main technologies in domestic and overseas. A first is a mapping method as a representative technology in Hampson-Russell software company, Geophysical Insight company and BGP company. The technology mainly involves establishing comprehensive network mapping relationship between data of a plurality of attribute and well logging data by a neural network to directly predict reservoir parameters. However, in different sedimentation environments, using the same model to predict may result in low prediction accuracy. Moreover, due to lack of geological characteristic constraint, generalization capability of a prediction model is poor, and a prediction result is prone to overfitting and does not conform to a macroscopic geological law. A second is an inversion method based on depth-domain “wavelet” extraction, as a representative technology in Paradigm company, northwest branch of CNPC, and Shengli Oil-field. The technology mainly involves directly predicting elastic parameters by depth-domain “wavelet” extraction in combination with a conventional inversion technology. However, a theoretical model based on depth-domain data has not been established, and thus, basic theory thereof is insufficient. A third is a depth domain reservoir parameter prediction method on the basis of high-precision velocity conversion, as a representative technology in Jason company. The technology involves converting depth-domain data into time-domain data to perform conventional reservoir parameter prediction. However, depth-time conversion has a certain accumulative conversion error, is time-consuming and labor-intensive, and is not conducive to enhancement of reservoir prediction accuracy.

SUMMARY

For the above technical problems, the present disclosure provides a reservoir parameter prediction method and apparatus based on geological characteristic constraint, a computer storage medium and a computer device.

According to a first aspect of the present disclosure, a reservoir parameter prediction method based on geological characteristic constraint provided in the present disclosure includes: S100, selecting dominant seismic attributes from seismic attributes of different types according to relevance between the seismic attributes of different types of a target stratum and reservoir parameters; S200, on the basis of the dominant seismic attributes, classifying seismic waveforms of the target stratum by a preset waveform classification network model and according to waveform characteristics, so as to obtain a waveform classification result, waveforms of different types correspondingly representing different geological characteristics; S300, constructing different deep neural network models corresponding to the different geological characteristics with the seismic attributes as an input, the reservoir parameters as an output, and the waveform classification result as constraint; S400, training the different deep neural network models by seismic data and well logging data of the target stratum as training data and predicted data, so as to optimize model parameters of each of the deep neural network models; S500, fusing different trained deep neural network models into a set of spatial variation neural network prediction models; and S600, predicting the reservoir parameters of the target stratum by the set of spatial variation neural network prediction model.

According to an embodiment of the present disclosure, the above step S100 includes: determining the relevance between the seismic attributes of different types of the target stratum and reservoir parameters through cross analysis by the seismic data and the well logging data of the target stratum, and selecting seismic attributes, the relevance between which and the reservoir parameters exceeds a preset relevance threshold value, from the seismic attributes of the different types as the dominant seismic attributes, according to a magnitude of the relevance; and decomposing and reconstructing data of each of the dominant seismic attributes through singular spectrum analysis. A sequence component in a reconstructed sequence that has a contribution degree greater than a preset contribution threshold value is reserved as a dominant component of a dominant seismic attribute according to the contribution degree to the dominant seismic attribute.

According to an embodiment of the present disclosure, in the above step S200, the waveform classification network model is an SOM unsupervised network model that is designed on a basis of an SOM unsupervised clustering algorithm, and the network model includes a seismic attribute input layer and a classification result output layer.

According to an embodiment of the present disclosure, the geological characteristics include a sedimentation characteristic.

According to an embodiment of the present disclosure, in the above step S300, each of the deep neural network models is an LSTM-RNN model, and the network model includes a seismic attribute input layer, a reservoir parameter output layer and a hidden layer located between the seismic attribute input layer and the reservoir parameter output layer. The hidden layer includes: an LSTM unit configured to reserve timing characteristics of seismic data and well logging data; a full-connected layer as a classifier of a network training model; a dropout layer configured to alleviate overfitting during a network model training process; and a regression layer as an output of the network training model.

According to an embodiment of the present disclosure, the above step S400 further includes: performing smoothing process on the well logging data, such that the spectrum of the well logging data subjected to the smoothing process matches a spectrum of the seismic data; performing normalization process on the matched seismic data and well logging data; with top and bottom of the target stratum as boundaries, intercepting seismic data and well logging data within a range of the target stratum from the seismic data and the well logging data subjected to the normalization process; and training the different deep neural network models by the seismic data and the well logging data within the range of the target stratum, so as to optimize the model parameters of each of the deep neural network models.

According to an embodiment of the present disclosure, in the above step S500, a spatial variation coefficient of each of trained deep neural network models in the set of spatial variation neural network prediction models is determined on the basis of the waveform similarity and spatial distance.

According to an embodiment of the present disclosure, in the above step S600, different trained deep neural network models are fused to form the set of spatial variation neural network prediction models according to a following formula:

$V_p\left(i,j,k\right)=\{\begin{array}{c}{w}_{1,i,j}·f_1\left(x_1,x_2,\dots x_N\right)\\ w_{2,i,j}·f_2\left(x_1,x_2,\dots x_N\right)\\ ⋮\\ w_{3,i,j}·f_k\left(x_1,x_2,\dots x_N\right)\end{array}$

In the formula, V_p represents a reservoir parameter, f_k(x₁, x₂, . . . x_N) represents a deep neural network model under a k-th type of geological characteristics, w_k,i,j represents a spatial variation coefficient of the deep neural network model under the k-th type of geological characteristics, and x₁, x₂, . . . x_N represents different types of seismic attributes.

According to an embodiment of the present disclosure, a spatial variation coefficient of each of trained deep neural network models in the set of spatial variation neural network prediction models is determined according to a following formula:

$w=\lambda w_c+\left(1-\lambda \right)w_d{c}_{12}=\sqrt{\frac{{\left(v_1-v_2\right)}^T\left(v_2-v_2\right)}{v_2^T{v}_2}}{d}_{12}=\sqrt{{\left(x_{v_1}-x_{v_2}\right)}^T\left(x_{v_1}-x_{v_2}\right)}{w}_c=\exp \left(-{\alpha }_c{c}_{12}^2\right)w_d=\exp \left(-{\alpha }_d{d}_{12}^2\right)$

In the formula, w represents a spatial variation coefficient, v₁ represents a seismic trace of a constructed deep neural network model, v₂ represents a seismic trace of a deep neural network model to be constructed, w_c represents an interpolation coefficient of similarity between the seismic trace of the constructed deep neural network model and the seismic trace of the deep neural network model to be constructed, w_d represents an interpolation coefficient of a distance between the seismic trace of the constructed deep neural network model and the seismic trace of the deep neural network model to be constructed, c₁₂ represents the correlation between v₁ and v₂, d₁₂ represents a distance between the v₁ and v₂, x_v1 and x_v2 respectively represent the spatial positions of v₁ and v₂, λ represents an adjustment factor, and α_c and α_d represent exponential factors.

According to an embodiment of the present disclosure, the above reservoir parameters of the target stratum include spatial three-dimensional elastic parameter s of the target stratum, and the method further includes outputting a distribution graph of the spatial three-dimensional elastic parameters of the target stratum.

According to a second aspect of the present disclosure, the present disclosure further provides a reservoir parameter prediction apparatus based on geological characteristic constraint. The reservoir parameter prediction apparatus includes: an attribute screening module configured to analyze relevance between seismic attributes of different types of a target stratum and reservoir parameters by seismic data and well logging data of the target stratum, and to select dominant seismic attributes from the seismic attributes of the different types according to a magnitude of the relevance; a waveform classification module configured to classify, on a basis of the dominant seismic attributes, seismic waveforms of the target stratum by a preset waveform classification network model and according to waveform characteristics, so as to obtain a waveform classification result, with waveforms of different types correspondingly representing different geological characteristics; a model construction model configured to construct different deep neural network models corresponding to the different geological characteristics with the seismic attributes as an input, the reservoir parameters as an output, and the waveform classification result as constraint; a model training module configured to train the different deep neural network models by the seismic data and the well logging data of the target stratum as training data and predicted data, so as to optimize model parameters of each of the deep neural network models; a model fusion model, which is configured to fuse different trained deep neural network models into a set of spatial variation neural network prediction models; and a parameter prediction module configured to predict the reservoir parameters of the target stratum by the set of spatial variation neural network prediction models.

According to a third aspect of the present disclosure, the present disclosure further provides a computer storage medium storing thereon a computer program executable by a processor, and the computer program, when executed by the processor, is executed to implement the above reservoir parameter prediction method based on geological characteristic constraint.

According to a fourth aspect of the present disclosure, the present disclosure further provides a computer device, including a memory and a processor. The processor is used for executing a computer program that is stored in the memory, so as to implement the above reservoir parameter prediction method based on geological characteristic constraint.

Compared with the related technology, a violence video classification technology integrated with internal and external knowledge provided in the present disclosure has the following advantages or beneficial effects.

• • 1. In the present disclosure, nonlinear mapping relationship between a seismic attribute and a reservoir parameter is described by an LSTM-RNN neural network model. The vertical association in seismic data and the vertical association in well logging data are taken into consideration, and at the same time, the timing features of the seismic data and the well logging data are also taken into consideration, such that well-to-seismic mapping relationship that is more accurate than that in the related technology is established.
  • 2. In the present disclosure, waveform clustering is introduced, and the same geological characteristic (for example, a sedimentation characteristic) corresponds to one deep network model. A set of spatial variation neural network prediction models is constructed, data of a plurality of types of seismic attributes is as an input, the geological characteristic is as constraint, and different network models under different waveform features are used to perform reservoir parameter prediction, thereby effectively enhancing the prediction accuracy.
  • 3. The present disclosure relates to a nonlinear reservoir parameter prediction technology based on geological characteristic constraint. With the nonlinear reservoir parameter prediction technology, direction prediction in particular for a depth domain reservoir parameter can be realized, the accuracy and spatial stability of prediction of reservoir parameters can be enhanced, thereby contributing to further enhancing drilling success rate, reducing exploration and development costs of an oil field, and increasing production benefit of the oil field.

Other features and advantages of the present disclosure will be described in the following description, and some will become obvious from the description, or understood by implementing the present disclosure. The purpose and other advantages of the present disclosure are realized and obtained by the structures pointed out in the description, the claims and the accompanying drawings.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic flowchart of a reservoir parameter prediction method based on geological characteristic constraint provided in an embodiment one of the present disclosure;

FIG. 2 is a schematic diagram of an SOM unsupervised clustering network model in a reservoir parameter prediction method provided in an embodiment one of the present disclosure;

FIG. 3 is a schematic diagram of an LSTM-RNN model in a reservoir parameter prediction method provided in an embodiment one of the present disclosure;

FIG. 4 is a schematic diagram of a forget gate in an LSTM-RNN model provided in an embodiment one of the present disclosure;

FIG. 5 is a schematic diagram of an input gate of an LSTM-RNN model provided in an embodiment one of the present disclosure;

FIG. 6 is a schematic diagram of an output gate of an LSTM-RNN model provided in an embodiment one of the present disclosure;

FIG. 7 is a schematic diagram of a spatial variation coefficient in a constructed set of spatial variation neural network prediction models provided in an embodiment one of the present disclosure;

FIG. 8 is a schematic diagram of waveform classification input data provided in an embodiment two of the present disclosure;

FIG. 9 is a schematic diagram of a slice, taken along 072, of a waveform classification result provided in a second embodiment of the present disclosure;

FIG. 10 is a schematic diagram of input data to a deep neural network provided in an embodiment two of the present disclosure;

FIG. 11 is a schematic diagram of constructing a deep neural network model provided in an embodiment two of the present disclosure;

FIG. 12A is a schematic diagram of a slice, taken along a stratum, of a multi-model elastic parameter prediction result based on a deep neural network provided in an embodiment two of the present disclosure;

FIG. 12B is a schematic diagram of a slice, taken along a stratum, of a multi-model elastic parameter prediction result based on a deep neural network provided in an embodiment two of the present disclosure;

FIG. 12C is a schematic diagram of a slice, taken along a stratum, of a time domain inversion result provided in an embodiment two of the present disclosure;

FIG. 12D is a schematic diagram of an engineering development deployment solution provided in an embodiment two of the present disclosure;

FIG. 13A is a comparison graph of an original well logging curve of a PU_IA well and a multi-model prediction result provided in an embodiment two of the present disclosure;

FIG. 13B is a comparison graph of an original well logging curve of a PU_IB well and a multi-model prediction result provided in an embodiment two of the present disclosure;

FIG. 13C is a comparison graph of an original well logging curve of a PU_IC well and a multi-model prediction result provided in an embodiment two of the present disclosure; and

FIG. 13D is a comparison graph of an original well logging curve of a PU_IC well and a single-model prediction result provided in an embodiment two of the present disclosure.

DETAILED DESCRIPTION OF THE EMBODIMENTS

Since a common depth domain reservoir prediction method at present has defects of low prediction accuracy, low prediction efficiency, inconformity with macroscopic geological characteristics, etc., the present disclosure provides a depth domain reservoir parameter direct prediction technology based on geological characteristic constraint, so as to enhance prediction accuracy in particular of a depth domain reservoir parameter, and to increase stability of spatial prediction, thereby providing reasonable understanding and high accuracy data for subsequent drilling and reservoir simulation to support efficient exploration and development.

The core concept of the present disclosure lies in that: for the problem of relatively low prediction accuracy possibly caused by performing prediction by using the same model, for example, in different sedimentation environments, reservoir parameter prediction is performed by a deep network and by introducing waveform clustering in the same sedimentation characteristic, so as to enhance prediction accuracy. A flow of a main method is as shown in FIG. 1. First, macroscopic geological characteristic zonation is established. Depth domain attributes of a plurality of types such as dynamics, kinematics and geometry are calculated, and are compared with a macroscopic geological background and well logging data, and seismic attributes having high correlation are preferably selected as the division of and a basis for waveform characteristic classification and also as subsequent reservoir parameters for providing a data basis. Then, automatic division of waveform characteristics is realized by combining an automatic clustering algorithm based on SOM unsupervised learning, and different geological characteristics are represented by a waveform classification result as a basis for geological characteristic zonation. Next, different reservoir parameter prediction models are constructed according to different geological characteristics. Under different geological characteristic zonations, nonlinear network prediction models at different well positions are constructed by a long short term memory recurrent neural network (LSTM-RNN) in combination with well logging data and seismic attribute data. Parameter adjustment and model parameter optimization are continuously performed, such that a single-point prediction module is generalized and converges. Finally, a set of spatial variation neural network prediction models is constructed. With data of seismic attributes of a plurality of types as an input and geological characteristics as constraint, reservoir parameter prediction is performed by different network models under different waveform characteristics, so as to finally obtain a comprehensive prediction result of three-dimensional spatial parameters.

In order to make the objectives, technical solutions and advantages of the present disclosure clearer, the present disclosure will be further described in details below in combination with the embodiments and the accompanying drawings, so as to enable the implementation process of how to apply technical means for solving technical problems and achieving technical effects to be fully understood and implemented.

Embodiment One

As shown in FIG. 1, a reservoir parameter prediction method based on geological characteristic constraint provided in the present disclosure includes the following steps.

At S100, dominant seismic attributes are selected from seismic attributes of different types according to relevance between the seismic attributes of different types of a target stratum and reservoir parameters.

Specifically, at first, the relevance between the seismic attributes of different types of the target stratum and the reservoir parameters is determined through cross analysis and by using seismic data and well logging data of the target stratum; and seismic attributes, the relevance between which and the reservoir parameters exceeds a preset relevance threshold value, are selected, as dominant seismic attributes, from the seismic attributes of different types based on a magnitude of the relevance. Then, data of each of the seismic attributes is decomposed and reconstructed through singular spectrum analysis. According to contribution to a seismic attribute, a sequence component, which has a contribution degree greater than a preset contribution threshold value in a reconstructed sequence, is reserved as a dominant component of a dominant seismic attribute.

The dominant seismic attributes that have been decomposed and reconstructed will be used for waveform clustering analysis and process at step S200.

At S200, on the basis of the dominant seismic attributes, seismic waveforms of the target stratum are classified by a preset waveform classification network model and according to waveform features, so as to obtain a waveform classification result, and different waveforms correspond to different geological characteristics.

At S300, with the seismic attributes as an input, the reservoir parameters as an output, and the waveform classification result as constraint, different deep neural network models corresponding to the different geological characteristics are constructed.

At S400, the different deep neural network models are trained by the seismic data and the well logging data of the target stratum as training data and predicted data, so as to optimize model parameters of each of the deep neural network models.

In a specific application, the seismic data and the well logging data may further be preprocessed firstly in order to further enhance the accuracy of a prediction result. For example, smoothing process is performed on the well logging data, such that the spectrum of the well logging data that has been subjected to the smoothing process matches the spectrum of the seismic data. Normalization process is performed on the matched seismic data and well logging data. With top and bottom of the target stratum as boundaries, the seismic data and well logging data within a range of the target stratum are intercepted from the seismic data and the well logging data which have been subjected to the normalization process. Then, the different deep neural network models are trained by the seismic data and the well logging data within the range of the target stratum, so as to optimize the model parameters of each of the deep neural network models.

At S500, different trained deep neural network models are fused to form a set of spatial variation neural network prediction models. A spatial variation coefficient of each of the trained deep neural network models in the set of spatial variation neural network prediction models is determined by waveform similarity and a spatial distance.

At S600, the reservoir parameters of the target stratum are predicted by the set of spatial variation neural network prediction models.

The steps will be described below in details.

• • (1) Step S100 relates to a technology for preferably selecting seismic attributes on the basis of cross analysis and singular spectrum analysis. In this step, cross analysis is first performed on the seismic attributes of different types of the target stratum and the reservoir parameters, and attributes that have higher correlation or greater contribution are preferably selected. Then, the preferably selected seismic attributes are represented as one-dimensional data, and a trajectory matrix is constructed. The trajectory matrix is further decomposed and reconstructed, and different elements and different components of the attributes are re-sorted according to contribution degree. In final, dominant components of the dominant seismic attributes are determined, so as to provide a data basis for subsequent waveform classification and reservoir parameter direct prediction. An algorithm of singular spectrum analysis is as follows.

{circle around (1)}. Embedding is performed, and preferably selected attribute data is represented as one-dimensional data:

[x₁, x₂, . . . x_N]

where N represents a sequence length.

Firstly, a suitable window length L is selected, and an original time sequence is arranged in a lagging manner to obtain a trajectory matrix:

$\begin{array}{cc}X=\left[\begin{array}{cccc}{x}_1& x_2& \dots & x_{N-L+1}\\ x_2& x_3& \dots & x_{N-L+2}\\ ⋮& ⋮& & ⋮\\ x_L& x_{L+1}& \dots & x_N\end{array}\right]& \left(1\right)\end{array}$

In general, L<N/2, and provided K=N−L+1, the trajectory matrix X is a matrix of L×K:

$\begin{array}{cc}X=\left[\begin{array}{cccc}{x}_1& x_2& \dots & x_K\\ x_2& x_3& \dots & x_{K+1}\\ ⋮& ⋮& & ⋮\\ x_L& x_{L+1}& \dots & x_N\end{array}\right]& \left(2\right)\end{array}$

{circle around (2)}. Decomposition is performed, and a covariance matrix of the trajectory matrix is calculated:

S=X·X ^T  (3)

Next, eigenvalue decomposition is performed on S to obtain λ₁>λ₂> . . . λ_L≥0 and corresponding eigenvectors U₁, U₂, . . . , U_L. In this case, U=[U₁, U₂, . . . , U_L], √{square root over (λ₁)}>√{square root over (λ₂)}> . . . >√{square root over (λ_L)}≥0 is a singular spectrum of the original sequence, and:

$\begin{array}{cc}X=\sum _{m=1}^L\sqrt{{\lambda }_m}{U}_m{V}_m^T,V_m=X^T{U}_m/\sqrt{{\lambda }_m},m=1,2,\dots ,L& \left(4\right)\end{array}$

where the eigenvector U_m corresponding to λ_m reflects an evolution type of the time sequence.

{circle around (3)}. Grouping is performed, and it is assumed that all the L elements are grouped into c number of groups that are not crossed to each other based on a specific formula as follows:

$\begin{array}{cc}X=X_{i_1}+\dots +X_{i_c},\mathrm{where}{X}_i=\sum _{m\in i}\sqrt{{\lambda }_m}{U}_m{V}_m^T=\left(\sum _{m\in i}{U}_m{U}_m^T\right)X& \left(5\right)\end{array}$

{circle around (4)}. Reconstruction is performed, in which firstly a projection, on U_m, of a lagging sequence X_i is calculated:

$\begin{array}{cc}{a}_i^m=X_i{U}_m=\sum _{j=1}^L{x}_{i+j}{U}_{m,j},0\le i\le N-L& \left(6\right)\end{array}$

where X_i represents an i-th column of the trajectory matrix X, and a_i^m represents a weight of a time evolution type reflected by X_i, in time periods of the original sequence x_i+1, x_i+2, . . . , x_i+L.

Next, a single is reconstructed by a time empirical orthogonal function and a time principal element, and a specific reconstruction process is as follows.

$\begin{array}{cc}{x}_i^k=\{\begin{array}{c}\frac{1}{i}{\sum }_{j=1}^i{a}_{i-j}^k{U}_{k,j},1⩽i⩽L-1\\ \frac{1}{L}{\sum }_{j=1}^L{a}_{i-j}^k{U}_{k,j},L⩽i⩽N-L+1\\ \frac{1}{N-i+1}{\sum }_{j=i-N+L}^L{a}_{i-j}^k{E}_{k,j},N-L+2⩽i⩽N\end{array}& \left(7\right)\end{array}$

In this way, a reconstructed sequence is equal to the original sequence, that is:

$\begin{array}{cc}{x}_i=\sum _{k=1}^L{x}_i^k,i=1,2\dots ,N& \left(8\right)\end{array}$

where x_i^k represents a k-th signal sorted according to importance.

Therefore, through cross analysis and singular spectrum analysis, dominant attributes and principal components of the dominant attributes are obtained, thereby providing a data basis for subsequent reservoir prediction.

(2) On the basis of preferable selection of a plurality of attributes, automatic division of waveform features is realized through an SOM unsupervised clustering algorithm (FIG. 2). A plurality of seismic attributes are preferably selected as input data, an SOM unsupervised network training model and a topological structure are designed, and a waveform classification result is outputted, thereby providing constraint data for subsequent reservoir parameter prediction. A specific SOM algorithm is as follows.

{circle around (1)}. Initialization is performed, so as to initialize randomly, by each node, parameters of the node itself. The number of parameters of each node is the same as the dimension number of input data.

{circle around (2)}. A node that best matches respective piece of input data, is found. Assuming that D-dimensional data is input, that is, X={x_i, i=1, . . . , D}, a discrimination function may be the Euclidean distance:

d _j(x)=Σ_i=1^D(x_i−w_ji)²  (10)

{circle around (3)}. After an active node I(x) is found, nodes adjacent thereto is expected to be updated. S_ij is made to indicate the distance between a node i and a node j, and updated weights are allocated to the nodes adjacent to the node I(x):

$\begin{array}{cc}{T}_{j,I}=\exp \left(-\frac{s_{j,I\left(x\right)}^2}{2{\delta }^2}\right)& \left(11\right)\end{array}$

{circle around (4)}. Next, the parameters of the nodes are updated. Updating is performed according to a gradient descent method:

Δw_ji=η(t)*T_i,j(x)(t)*(x_i−w_ji)  (12)

Iteration is performed until convergence is realized.

Therefore, automatic division of waveform features is realized through an SOM unsupervised automatic clustering technology in combination with a technology for preferable selection of a plurality of attributes based on cross analysis and SSA singular spectrum analysis.

(3) On the basis of waveform classification, different deep neural networks are further constructed with respect to different waveform features. Since seismic data is of a timing signal feature and well logging data also has certain association in a vertical direction, a long short term memory recurrent neural network (LSTM-RNN) is preferably selected to construct a nonlinear multi-network prediction model under different geological characteristics (FIG. 3).

FIG. 3 is a schematic diagram of an LSTM-RNN model. One LSTM unit is composed of three threshold structures and one state vector transmission line. The threshold structures are three gates including a forget gate, an input gate and an output gate. The state vector transmission line is responsible for long term memory, and is merely used to perform some simple linear operations. The three gates are responsible for short term selection, and a deletion or addition operation is performed on an input vector by setting thresholds for the gates.

The forget gate (FIG. 4) is realized by a sigmoid neural layer, and the function of the forget gate is to decide which information is allowed to pass the unit. 0 indicates “no information is allowed to pass”, and 1 indicates “all information is allowed to pass”.

The function of the input gate (FIG. 5) is to decide how much new information is allowed to a unit state. The realization of the input gate needs two steps. First, a sigmoid layer is provided to decide which information needs to be updated, and a tanh layer is provided to generate candidate information for updating contents. Next, vectors of the two layers are combined by dot product, so as to update the unit state.

The function of the output gate (FIG. 6) is to output a final result. The realization of the output gate needs two steps. First, which information may be outputted is decided by the sigmoid layer. Next, a state vector is made to pass through the tanh layer, and an output of the tanh layer is then multiplied by a weight calculated by the sigmoid layer, such that the final result to be output is obtained.

The specific mathematical process in the LSTM unit is as follows:

i _t=σ(W_xix_t+W_hih_t−1+W_cic_t−1+b_i)

f _t=σ(W_xfx_t+W_hfh_t−1+W_cfc_t−1+b_f)

C _t =f _t ·C _t−1 +i _t·(W_xcx_t+W_hch_t−1+b_c)

o _t=σ(W_xox_t+W_hoh_t−1+W_coc_t−1+b_o)

h _t =o _t·tanh(C_t)  (13)

where i represents an input gate; σ represents a logic sigmoid function; W_si, W_hi and W_ci respectively represent weight matrixes between an input eigenvector and the input gate, between a hidden layer unit and the input gate, and between a unit activation vector and the input gate; b_i represents an offset of the input gate; f represents a forget gate; W_xf, W_hf and W_cf respectively represent weight matrixes between the input eigenvector and the forget gate, between the hidden layer unit and the forget gate, and the unit activation vector and the forget gate; b_f represents an offset of the forget gate; C represents the unit activation vector; W_sc and W_hc represent respectively weight matrixes between the input eigenvector and the unit activation vector and between the hidden layer unit and the unit activation vector, the weight matrixes are diagonal matrixes; b_c represents an offset of an output gate; o represents an output gate; W_xo, W_ho and W_co respectively indicate weight matrixes between the input eigenvector and the output gate ,between the hidden layer unit and the output gate, and between the unit activation vector and the output gate; b_f represents an offset of the forget gate; t, as a subscript, represents a sampling moment; and tanh represents an activation function.

(4) A deep neural network model based on an LSTM-RNN is trained, and parameters of the network model are optimized and adjusted, such that the model is generalized and converges. Well-side data of the plurality of attributes is input to an input layer of the model; corresponding well logging elastic parameters, such as a P-wave velocity, is output from an output layer of the model; and a hidden layer of the model is composed of an LSTM unit, a full-connected layer, a dropout layer and a regression layer. The LSTM unit is configured to reserve timing features of well logging data and seismic data. The full-connected layer serves as a classifier of an entire training network. The dropout layer is configured to alleviate the occurrence of overfitting during a network training process, so as to have the effect of regularization. The regression layer serves as an output of the network training model.

(5) A set of spatial variation neural network prediction models is constructed. On the basis of the construction of the set of spatial variation neural network prediction models, spatial variation coefficients of each of the neural network models are constructed by the waveform similarity and spatial distance (FIG. 7), so as to realize the construction of the set of spatial variation neural network prediction models. A construction formula for the spatial variation coefficients is:

$\begin{array}{cc}w=\lambda w_c+\left(1-\lambda \right)w_d& \left(14\right)\end{array}$ $\begin{array}{cc}{c}_{12}=\sqrt{\frac{{\left(v_1-v_2\right)}^T\left(v_1-v_2\right)}{v_2^T{v}_2}}& \left(15\right)\end{array}$ $\begin{array}{cc}{d}_{12}=\sqrt{{\left(x_{v_1}-x_{v_2}\right)}^T\left(x_{v_1}-x_{v_2}\right)}& \left(16\right)\end{array}$ $\begin{array}{cc}{w}_c=\exp \left(-{\alpha }_c{c}_{12}^2\right)& \left(17\right)\end{array}$ $\begin{array}{cc}{w}_d=\exp \left(-{\alpha }_d{d}_{12}^2\right)& \left(18\right)\end{array}$

In the formula, w represents spatial variation coefficient; v₁ represents a seismic trace of a constructed neural network model; v₂ represents a seismic trace of a neural network model to be constructed; w_c represents an interpolation coefficient of the similarity between the seismic trace of the constructed neural network model and the seismic trace of the neural network model to be constructed; w_d represents an interpolation coefficient of the distance between the seismic trace of the constructed neural network model and the seismic trace of the neural network model to be constructed; c₁₂ represents the correlation between v₁ and v₂; d₁₂ represents distance between the v₁ and v₂; x_v1 and x_v2 respectively represent the spatial positions of v₁ and v₂; λ represents an adjustment factor; and α_c and α_d represent exponential factors.

In final, the construction of a set of spatial variation neural network prediction models is realized.

(6) The spatial elastic parameters are predicted. Taking the plurality of attributes as input data, elastic parameters as output data and the waveform classification result as constraint, the prediction of the spatial elastic parameters is realized via the set of spatial variation neural network prediction models.

$\begin{array}{cc}{V}_p\left(i,j,k\right)=\{\begin{array}{c}{w}_{1,i,j}·f_1\left(x_1,x_2,\dots x_N\right)\\ x_{2,i,j}·f_2\left(x_1,x_2,\dots x_N\right)\\ ⋮\\ w_{3,i,j}·f_k\left(x_1,x_2,\dots x_N\right)\end{array}& \left(19\right)\end{array}$

where V_p represents a reservoir parameter; f_k(x₁, x₂, . . . x_N) represents a neural network prediction model corresponding to a k-th type of geological characteristics; w_k,i,j represents a spatial variation coefficient of the neural network prediction model corresponding to the k-th type of geological characteristics; and x₁, x₂, . . . x_N represents the seismic attributes of different types.

Embodiment Two

The effectiveness of the method of the present disclosure will be described below through actual data of a certain work area. The work area is of a clastic rock reservoir type, with two reservoirs developed respectively of a layer 072 and a layer 073. The layer 072 is turbidity channel sheet sandstone. The development across the work area is stable. There are three effective logging wells in total in the work area, respectively a well PU_IA, a well PU_IB and a well PU_IC. First, a plurality of attributes are preferably selected on the basis of cross analysis and singular spectrum analysis (FIG. 8). The preferably selected attributes to be input are respectively: an envelope attribute, relative wave impedance and an instantaneous amplitude attribute, and is used to perform waveform classification in combination with an SOM unsupervised clustering algorithm (FIG. 9). Waveforms are classified into three classes, namely, a waveform having relatively low energy, a waveform having medium energy, and a waveform having high energy respectively corresponding to the well PU_IC, the well PU_IB and the well PU_IA. On the basis of waveform classification, a deep neural network is trained with respect to different waveform features. Preferably, input data (FIG. 10) includes: a seismic trace, an envelope attribute, a thin-layer factor, relative wave impedance, a Hilbert attribute, an instantaneous frequency, a dominant frequency and an instantaneous phase attribute. Output data is a longitudinal wave velocity. A five-layer deep neural network (FIG. 11) is designed for the construction of a prediction model, and a waveform classification result is regarded as constraint data. In final, multi-model elastic parameter prediction based on a deep recurrent neural network is realized.

A result of multi-model elastic parameter prediction based on a deep neural network is compared with a result of single-model elastic parameter prediction based on a deep neural network, a time domain inversion result and an engineering development deployment solution (FIGS. 12A-12D). It can be seen from a planar tendency that, compared with a longitudinal wave velocity on the basis of a single-model prediction, the present disclosure has higher spatial prediction accuracy, and better reserves overall macroscopic characteristics. It can be seen from the comparison between the present disclosure and the time-domain inversion result that, since the time-domain inversion result is limited by a well and an initial model, the inversion result cannot greatly represent macroscopic geological characteristics. Through comprehensive analysis of FIGS. 12A-12D, an elastic parameter direct prediction technology based on a plurality of models has higher spatial prediction accuracy, and has better consistency with the development deployment solution.

By comparison between multi-model-based prediction results of the wells PU_IA, PU_IB and PU_ IC and original well logging curves (FIGS. 13A to 13C), it can be seen that the prediction result of the longitudinal wave velocity is consistent with an overall tendency of the original well logging curve, and is highly consistent with the well. Furthermore, a multi-model prediction result of the well PU_IC is compared with a single-model prediction result thereof. It can be seen that precision of P-wave velocity prediction based on multi-model prediction is higher than that of P-wave velocity prediction based on single-model prediction.

In summary, on the basis of the reservoir parameter direction prediction technology based on an LSTM-RNN developed in the present disclosure, the reservoir parameter direct prediction based on the set of spatial variation neural network models is realized, thereby effectively maintaining a geological stratum structure and further enhancing the accuracy of reservoir parameter prediction.

Embodiment Three

On the basis of the previous embodiment, the present embodiment provides a reservoir parameter prediction apparatus. The reservoir parameter prediction apparatus includes an attribute screening module, a waveform classification module, a model construction model, a model training module, a model fusion model and parameter prediction module.

The attribute screening module is configured to analyze the relevance between the seismic attributes of different types of the target stratum and the reservoir parameters by using seismic data and well logging data of the target stratum, and to select dominant seismic attributes from the seismic attributes of different types according to a magnitude of the relevance.

The waveform classification module is configured to classify, on the basis of the dominant seismic attributes, seismic waveforms of the target stratum by a preset waveform classification network model and according to waveform features, so as to obtain a waveform classification result, with different types of waveforms correspondingly representing different geological characteristics.

The model construction model is configured to construct different deep neural network models corresponding to the different geological characteristics with the seismic attributes as an input, the reservoir parameters as an output, and the waveform classification result as constraint.

The model training module is configured to train the different deep neural network models by using seismic data and well logging data of the target stratum as training data and predicted data, so as to optimize model parameters of each of the deep neural network models.

The model fusion model is configured to fuse different trained deep neural network models into a set of spatial variation neural network prediction models.

The parameter prediction module is configured to predict the reservoir parameters of the target stratum by using the set of spatial variation neural network prediction models.

Embodiment Four

In addition, the present embodiment provides a computer storage medium. The computer storage medium stores a computer program thereon.

The computer program, when executed by one or more computer processor, is configured to implement the above mentioned reservoir parameter prediction method.

The above mentioned storage medium may be a flash memory, a hard disk, a multimedia card, a card-type memory (e.g., an SD or DX memory, etc.), a random access memory (RAM), a static random access memory (SRAM), a read-only memory (ROM), an electrically erasable programmable read-only memory (EEPROM), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, an optical disc, a server, an application (APP) store, and the like.

Embodiment Five

In addition, the present embodiment provides a computer device. The computer device includes a memory and a processor.

The memory stores a computer program thereon, and the computer program, when executed by the processor, executes the above mentioned reservoir parameter prediction method.

The processor may be implemented as an application specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field programmable gate array (FPGA), a controller, a microcontroller, a microprocessor or other electronic elements, and the processor may be used to execute the reservoir parameter prediction method in any one of the embodiment one to the embodiment five.

The memory may be implemented as any type of volatile or non-volatile storage device or a combination thereof, for example, a static random access memory (SRAM), an electrically erasable programmable read-only memory (EEPROM), an erasable programmable read-only memory (EPROM), a programmable read-only memory (PROM), a read-only memory (ROM), a magnetic memory, a flash memory, a magnetic disk or an optical disk.

It should be understood that the apparatus and method embodiments described in the above embodiments are merely illustrative. For example, the flowcharts and the block diagrams in the accompanying drawings display a system architecture, a function and an operation that may be implemented by the apparatus, the method and the computer program product according to the various embodiments of the present disclosure. In this regard, each block in the flowcharts and the block diagrams may represent a part of a module, a program section or a code, and the part of the module, the program section or the code contains one or more executable instructions that are used for implementing a designated logic function. It should also be noted that, in some alternative implementations, a function indicated in the block may also be implemented in an order different from that indicated in the accompanying drawings. For example, two continuous blocks can actually be executed basically concurrently, and sometime, they can also be executed in an opposite order, which is determined according to a function involved. It also needs to be noted that each block in the block diagrams and/or flowcharts, and a combination of blocks in the block diagrams and/or flowcharts can also be implemented by a specific hardware-based system that executes a designated function or action, or can also be implemented by a combination of specific hardware and a computer instruction.

In addition, various functional modules in the various embodiments of the present disclosure may be integrated into one independent part, or various modules may be present separately, or two or more modules may be integrated into one independent part.

If the function is realized in the form of a software functional module, and is sold or used as an independent product, the function may be stored in a computer-readable storage medium. Based on such understanding, the technical solution of the present disclosure, in essence or the contribution to the prior art, or part of the technical solution may be embodied in the form of a software product. The computer software product is stored in a storage medium, and includes a plurality of instructions used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in various embodiments of the present disclosure.

It also needs to be noted that the foregoing description is merely preferred embodiments of the present disclosure and is not used for limiting the present disclosure, and various changes and modifications may be made to the present disclosure by those skilled in the art. Within the spirit and principle of the disclosure, any modifications, equivalent replacements, improvements, etc., shall be contained within the scope of protection of the present disclosure. It should be noted that similar reference signs and letters refer to similar items in the following drawings. Therefore, once a specific item is defined in one of the drawings, it need not be further defined and explained in subsequent drawings.

Claims

1. A reservoir parameter prediction method based on geological characteristic constraint, comprising: S100, selecting dominant seismic attributes from seismic attributes of different types according to relevance between the seismic attributes of different types of a target stratum and reservoir parameters;S200, classifying seismic waveforms of the target stratum by a preset waveform classification network model, according to waveform features and on the basis of the dominant seismic attributes, so as to obtain a waveform classification result, waveforms of different types correspondingly representing different geological characteristics;S300, constructing different deep neural network models corresponding to the different geological characteristics with the seismic attributes as an input, the reservoir parameters as an output, and the waveform classification result as constraint;S400, training the different deep neural network models by seismic data and well logging data of the target stratum, so as to optimize model parameters of each of the deep neural network models;S500, fusing different trained deep neural network models into a set of spatial variation neural network prediction models; andS600, predicting the reservoir parameters of the target stratum by the set of spatial variation neural network prediction models.

2. The reservoir parameter prediction method of claim 1, wherein the step S100 comprises: determining the relevance between the seismic attributes of different types of the target stratum and the reservoir parameters through cross analysis by the seismic data and the well logging data of the target stratum, and selecting seismic attributes, the relevance between which and the reservoir parameters exceeds a preset relevance threshold value, from the seismic attributes of different types as the dominant seismic attributes, according to a magnitude of the relevance; anddecomposing and reconstructing data of each of the dominant seismic attributes through singular spectrum analysis, wherein a sequence component in a reconstructed sequence that has a contribution degree greater than a preset contribution threshold value is reserved as a dominant component of a dominant seismic attribute according to the contribution degree to the dominant seismic attribute.

3. The reservoir parameter prediction method of claim 1, wherein in the step S200, the waveform classification network model is an SOM unsupervised network model designed on a basis of an SOM unsupervised clustering algorithm, and the waveform classification network model comprises a seismic attribute input layer and a classification result output layer.

4. The reservoir parameter prediction method of claim 1, wherein the geological characteristics comprise a sedimentation characteristic.

5. The reservoir parameter prediction method of claim 1, wherein in the step S300, each of the deep neural network models is an LSTM-RNN model, and each of the deep neural network models comprises a seismic attribute input layer, a reservoir parameter output layer and a hidden layer between the seismic attribute input layer and the reservoir parameter output layer, and the hidden layer comprises: an LSTM unit configured to reserve timing features of seismic data and well logging data;a full-connected layer as a classifier of a network training model;a dropout layer configured to alleviate overfitting during a network model training process; anda regression layer as an output of the network training model.

6. The reservoir parameter prediction method of claim 1, wherein the step S400 further comprises: performing smoothing process on the well logging data, such that a spectrum of the well logging data subjected to the smoothing process matches a spectrum of the seismic data;performing normalization process on the matched seismic data and well logging data;with top and bottom of the target stratum as boundaries, intercepting seismic data and well logging data within a range of the target stratum from the seismic data and the well logging data subjected to the normalization process; andtraining the different deep neural network models by the seismic data and the well logging data within the range of the target stratum, so as to optimize the model parameters of each of the deep neural network models.

7. The reservoir parameter prediction method of claim 1, wherein in the step S500, a spatial variation coefficient of each of trained deep neural network models in the set of spatial variation neural network prediction models is determined on the basis of waveform similarity and spatial distance.

8. The reservoir parameter prediction method of claim 1, wherein in the step S600, different trained deep neural network models are fused to form the set of spatial variation neural network prediction models according to a following formula:

9. The reservoir parameter prediction method of claim 1, wherein a spatial variation coefficient of each of trained deep neural network models in the set of spatial variation neural network prediction models is determined according to a following formula:

10. The reservoir parameter prediction method of claim 1, wherein the reservoir parameters of the target stratum comprise spatial three-dimensional elastic parameters of the target stratum, and the method further comprises outputting a distribution graph of the spatial three-dimensional elastic parameters of the target stratum.

11. A reservoir parameter prediction apparatus based on geological characteristic constraint, comprising: an attribute screening module configured to analyze relevance between seismic attributes of different types of a target stratum and reservoir parameters by seismic data and well logging data of the target stratum, and to select dominant seismic attributes from the seismic attributes of different types according to a magnitude of the relevance;a waveform classification module configured to classify, on a basis of the dominant seismic attributes, seismic waveforms of the target stratum by a preset waveform classification network model and according to waveform features, so as to obtain a waveform classification result, with waveforms of different types correspondingly representing different geological characteristics;a model construction model configured to construct different deep neural network models corresponding to the different geological characteristics with the seismic attributes as an input, the reservoir parameters as an output, and the waveform classification result as constraint;a model training module configured to train the different deep neural network models by the seismic data and the well logging data of the target stratum, so as to optimize model parameters of each of the deep neural network models;a model fusion model configured to fuse different trained deep neural network models into a set of spatial variation neural network prediction models; anda parameter prediction module configured to predict the reservoir parameters of the target stratum by the set of spatial variation neural network prediction models.

12. A computer storage medium storing thereon a computer program executable by a processor, wherein when executed by the processor, the computer program is executed to implement the reservoir parameter prediction method based on geological characteristic constraint of claim 1.

13. A computer device, comprising a memory and a processor, wherein a computer program stored on the memory is executed by the processor, and the computer program is executed to implement the reservoir parameter prediction method based on geological characteristic constraint of claim 1.