A reservoir porosity prediction method based on one-dimensional convolutional gated recurrent network

Abstract

The invention discloses a reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network, and relates to the technical field of porosity prediction. The method comprises an offline part and an online part; wherein the offline part comprises: obtaining a first seismic attribute set related to reservoir parameters at a well point in a target work area; obtaining an optimal attribute combination from the first seismic attribute set; taking the optimal attribute combination as input, taking the porosity corresponding to the optimal attribute combination as output, and training a one-dimensional convolutional gated recurrent network model; the online part comprises: obtaining a second seismic attribute set related to target reservoir parameters in a well-free area in the target work area; screening out a final attribute combination from the second seismic attribute set according to the optimal attribute combination; inputting the final attribute combination into a trained one-dimensional convolutional gated recurrent network model, and outputting the target reservoir porosity; and the method improves the prediction accuracy of reservoir porosity and expands the prediction range.

Description

Technical Field

[0001] The invention relates to the technical field of porosity prediction, and more particularly to a reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network. Background Art

[0002] Porosity is one of the important parameters for reservoir characterization. It directly reflects the ability of a reservoir to contain oil and gas. Predicting it is the key to reservoir delineation, characterization, evaluation, and well deployment. Therefore, accurate assessment of reservoir porosity is of great significance.

[0003] At present, the main methods for characterizing reservoir porosity include porosity measurement methods based on core rock physics experiments, porosity prediction methods based on linear equations, and porosity prediction methods based on artificial intelligence. Since reservoir porosity is easily affected by geological factors such as structural location and lithology changes, Bakhorji et al. believe that the porosity directly obtained through rock physics experiments is the most accurate, but it is limited by the high sampling cost and the difficulty of porosity measurement through core samples, which is not suitable for large-scale application in actual work. There is a certain intrinsic connection between porosity and parameters such as pore throat radius and wave velocity, so reservoir porosity can be predicted from well logging curves. This type of method generally tends to use empirical formulas to obtain it from three well logging curves: density, acoustic wave, and compensated neutron, and then combine it with the specific conditions of the work area to obtain empirical parameters. Domestic and foreign scholars have done a lot of relevant research on this. For example, Wyllie et al. proposed to use the time-averaged equation to predict the porosity of reservoirs, but it is only applicable to areas with relatively simple geological conditions; Raymer et al. improved it to be applicable to complex geological conditions; Hampson et al. used multiple linear regression to predict the porosity of sandstone reservoirs with simple geological conditions, and the prediction accuracy was quite impressive; Pramanik et al. combined the wave impedance inversion body with seismic attributes and predicted the porosity of conventional sandstone reservoirs by stepwise multiple linear regression. However, as unconventional reservoirs are increasingly valued, it is difficult to achieve the accuracy required for current production and development by using mathematical formulas based solely on linear equations to characterize porosity. Finding a reasonable nonlinear mapping relationship is the key to solving the problem.

[0004] With the continuous development of artificial intelligence technology, machine learning algorithms are gradually being used in reservoir parameter prediction. The idea of ​​its application in reservoir porosity prediction is to learn the nonlinear relationship between reservoir porosity and other input parameters by building a machine learning network model, thereby realizing the prediction of porosity. Conventional machine learning algorithms are relatively simple and shallow structures, such as random forest algorithms, support vector machine algorithms, BP neural networks and artificial neural networks. These algorithm models have been widely used by domestic and foreign scholars in porosity prediction in the past period of time.

[0005] In recent years, researchers have gradually introduced deep learning methods into reservoir porosity prediction, and proved that deep learning algorithms are superior to traditional neural network models in reservoir porosity prediction. Since then, domestic and foreign scholars have conducted a large number of research experiments using deep learning network models. Srisutthiyakorn used rock slices combined with deep learning algorithms to estimate porosity. Feng R et al. used convolutional neural networks to predict reservoir porosity. Chen Wen, An Peng et al. used long short-term memory neural networks to predict reservoir porosity parameters from well logging curves, proving that long short-term memory neural networks have good robustness and accuracy in time series prediction. Wang Jun, Zhang Zhenhua et al. introduced gated recurrent neural networks to estimate porosity parameters from well logging curves, solving the problem of feature extraction between adjacent sampling points of well logging curves. Yang et al. used LSTM networks and Transformer models to predict reservoir porosity from well logging data. Anh et al. combined digital elevation models (DEMs) and LSTM networks to predict the porosity of river landforms. Song Jianguo et al. weighted integration of multiple different deep learning network models (DNN) to predict formation porosity from seismic attributes. However, the prediction accuracy and prediction range of the above methods are limited.

[0006] Therefore, how to further improve the prediction accuracy of reservoir porosity and expand the prediction range is an urgent problem that technicians in this field need to solve. Summary of the invention

[0007] In view of the above problems, the present invention provides a reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network to at least solve some of the technical problems mentioned in the above background technology.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] The embodiment of the present invention provides a reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network, including an offline part and an online part;

[0010] The offline part includes the following steps:

[0011] S1, obtaining multiple seismic attributes related to reservoir parameters at well points in a target work area as a first seismic attribute set;

[0012] S2. acquiring an optimal attribute combination from the first seismic attribute set;

[0013] S3, taking the optimal attribute combination as input, taking the porosity corresponding to the optimal attribute combination as output, and training a one-dimensional convolutional gated recurrent network model;

[0014] The online part includes the following steps:

[0015] P1. Acquire multiple seismic attributes related to target reservoir parameters in a well-free area in the target work area as a second seismic attribute set;

[0016] P2. According to the optimal attribute combination, a final attribute combination is selected from the second seismic attribute set;

[0017] P3. Input the final attribute combination into the trained one-dimensional convolutional gated recurrent network model to output the target reservoir porosity.

[0018] Furthermore, the step S2 specifically includes:

[0019] S21, using the Pearson correlation coefficient method to obtain the correlation between the seismic attributes in the first seismic attribute set, and sorting the seismic attributes in the first seismic attribute set according to the correlation; removing redundant seismic attributes in the first seismic attribute set according to the correlation and the sorting result, to obtain a preliminary selection result of seismic attributes;

[0020] S22, evaluating the attribute importance of each seismic attribute in the seismic attribute preliminary selection result according to the random forest algorithm, and performing secondary sorting on each seismic attribute in the seismic attribute preliminary selection result according to the attribute importance;

[0021] S23, based on the stepwise regression method, combining the seismic attributes in the seismic attribute preliminary selection results one by one according to the secondary sorting result to obtain multiple groups of attribute combinations, and combining the multiple groups of attributes to form an attribute sample set;

[0022] S24, using a cross-validation method to divide the attribute sample set into n parts, taking n-1 parts of the data as training sets and the remaining 1 part of the data as validation sets in turn, and training the random forest model;

[0023] S25. After the training is completed, the average verification error of n trainings is taken as the generalization error, and the best attribute combination is screened out according to the generalization error.

[0024] Furthermore, removing redundant seismic attributes in the first seismic attribute set according to the correlation and the primary sorting result specifically includes: screening out multiple seismic attributes with correlation higher than a preset threshold from the first seismic attribute set, screening out the seismic attribute with the highest correlation with the target reservoir porosity curve from the multiple seismic attributes and retaining it, and removing other seismic attributes.

[0025] Furthermore, the one-dimensional convolutional gated recurrent network model includes an input layer, a one-dimensional convolutional layer, a gated recurrent neural network layer, a fully connected layer and an output layer.

[0026] Furthermore, the step P3 specifically includes:

[0027] P31, inputting the final attribute combination into the trained one-dimensional convolutional gated recurrent network model through the input layer;

[0028] P32, extracting the lateral spatial features of each seismic attribute in the final attribute combination through the one-dimensional convolution layer;

[0029] P33, extracting the longitudinal time series features of each transverse spatial feature through the gated recurrent neural network layer;

[0030] P34. Integrate the lateral spatial features and the longitudinal time series features through a fully connected layer to complete the mapping of spatiotemporal features to reservoir porosity logging curve values, and output the target reservoir porosity through the output layer.

[0031] Furthermore, the one-dimensional convolutional layer is 2 layers.

[0032] Furthermore, a Dropout layer is arranged after the one-dimensional convolution layer.

[0033] Furthermore, a Dropout layer is provided after the gated recurrent neural network layer.

[0034] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network, which has the following beneficial effects:

[0035] The present invention combines a one-dimensional convolutional neural network with a GRU network to construct a 1D-CNN-GRU joint model. Based on the improved 1D-CNN-GRU porosity prediction model, the temporal and spatial characteristics of logging and seismic in the horizontal and vertical directions are integrated, and the prediction results have high resolution in both the horizontal and vertical directions.

[0036] The Pearson correlation coefficient-random forest method proposed in the embodiment of the present invention solves the analysis problem of strong nonlinear correlation between sample data, and gradually selects the seismic attributes most suitable for predicting porosity.

[0037] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0038] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0040] Figure 1 A schematic flow chart of a reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network provided in an embodiment of the present invention.

[0041] Figure 2 A schematic diagram of the optimal attribute combination acquisition process provided by an embodiment of the present invention.

[0042] Figure 3 A schematic diagram of a convolutional neural network in the prior art provided for an embodiment of the present invention.

[0043] Figure 4 A schematic diagram of the GRU unit structure in the prior art provided for an embodiment of the present invention.

[0044] Figure 5 A schematic diagram of the structure of a one-dimensional convolutional gated recurrent network model provided for an embodiment of the present invention.

[0045] Figure 6 A schematic diagram of a Pearson correlation coefficient matrix provided in an embodiment of the present invention.

[0046] Figure 7 A schematic diagram of the importance ranking of random forest attributes provided in an embodiment of the present invention.

[0047] Figure 8 A schematic diagram of a cross-validation result provided by an embodiment of the present invention.

[0048] Fig. 9 A schematic diagram comparing the prediction results of the three models 1D-CNN, GRU, and 1D-CNN-GRU provided in an embodiment of the present invention.

[0049] Fig.10 A schematic diagram comparing the prediction effects of the three models 1D-CNN, GRU, and 1D-CNN-GRU provided in an embodiment of the present invention.

[0050] Fig.11 A schematic diagram of the test results of the 1D-CNN-GRU model provided in an embodiment of the present invention.

[0051] Fig.12 A schematic diagram of the profile prediction results of the 1D-CNN-GRU model provided in an embodiment of the present invention.

[0052] Fig.13Schematic diagram of the three-dimensional porosity prediction volume of the 1D-CNN-GRU model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0053] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] See also Figure 1 As shown, the embodiment of the present invention discloses a reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network, including an offline part and an online part; wherein:

[0055] The offline part includes the following steps:

[0056] S1, obtaining multiple seismic attributes related to reservoir parameters at well points in a target work area as a first seismic attribute set;

[0057] S2, obtaining the best attribute combination from the first seismic attribute set;

[0058] S3, taking the best attribute combination as input and the porosity corresponding to the best attribute combination as output, training the one-dimensional convolutional gated recurrent network model;

[0059] The online part includes the following steps:

[0060] P1. Acquire multiple seismic attributes related to target reservoir parameters in a wellless area within the target work area as a second seismic attribute set;

[0061] P2, based on the best attribute combination, select the final attribute combination from the second earthquake attribute set;

[0062] P3. Input the final attribute combination into the trained one-dimensional convolutional gated recurrent network model and output the target reservoir porosity.

[0063] Based on a reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network provided in the present application, before predicting the reservoir porosity in each work area, a nonlinear mapping relationship between the porosity curve at the well point and the seismic attribute is first established, and a one-dimensional convolutional gated recurrent network model is trained through the nonlinear mapping relationship; then the method is extended to the well-free area of ​​the current work area to achieve high-precision prediction of the three-dimensional porosity body of the current work area; and the predicted reservoir porosity value has a higher degree of consistency with the true porosity curve, and a reservoir spatial distribution morphology and porosity distribution with good lateral continuity can be obtained, providing new ideas and methods for obtaining a wide range of reservoir porosity through seismic data.

[0064] Next, the above steps are described in detail.

[0065] In the above step S1, before performing reservoir parameter prediction, it is necessary to find multiple seismic attributes related to reservoir parameters at the well point as the first seismic attribute set; and obtain the porosity corresponding to each seismic attribute in the seismic attribute set. The introduction of seismic attributes is usually a process from many to few, that is, in the early stage of reservoir prediction, try to select a variety of seismic attributes that may be related to reservoir physical parameters, so that not only can the information contained in the seismic data be fully utilized, but also the accuracy of reservoir physical parameters can be improved and the multi-solution of reservoir prediction can be reduced.

[0066] In the above step S2, the tight sandstone reservoir has strong heterogeneity, which leads to a strong nonlinear correlation between porosity and seismic attributes. Since the nonlinear relationship between the direct analysis attributes and the porosity parameters is relatively complicated, the Pearson correlation coefficient method is first used to perform a relatively simple linear correlation analysis to preliminarily screen the relationship between each attribute and the porosity logging parameters. Nonlinear relationships are usually analyzed using a neural network model, but the direct use of a neural network to establish a mapping relationship between seismic attributes and porosity logging parameters is inefficient. Therefore, a random forest algorithm is first used to select the attributes, and then the optimal attribute combination after optimization is used as the training data for the target deep learning model. Based on this, the present invention establishes a feature selection method based on Pearson-random forest to achieve the optimization of seismic attributes. For details, see Figure 2 As shown, the following steps are included:

[0067] S21. The correlation between the seismic attributes in the first seismic attribute set is obtained by using the Pearson correlation coefficient method. The Pearson correlation coefficient method is a method for finding the relationship between two variables and is used to detect the linear correlation between the variables. The method can be expressed by a mathematical formula as follows:

[0068]

[0069] Among them, x i ,yi They represent two groups of input data samples that need to be analyzed for correlation; ρ represents the Pearson correlation coefficient; They represent the average values ​​of each group of samples that need to be analyzed for correlation. Generally speaking, the Pearson correlation coefficient is between [-1, 1]. A negative value indicates that there is a negative linear correlation between the data, and 0 indicates that there is no linear correlation between the data.

[0070] After obtaining the correlation between the seismic attributes in the first seismic attribute set, sorting the seismic attributes in the first seismic attribute set according to the correlation; screening out the seismic attributes in the first seismic attribute set whose correlation is higher than a preset threshold, and removing the seismic attributes with a low correlation with the target reservoir porosity curve from the screened seismic attributes; thereby removing the redundant seismic attributes in the first seismic attribute set and obtaining the preliminary selection result of the seismic attributes;

[0071] For example, the seismic attributes in the first seismic attribute set are sorted from large to small according to the correlation, wherein the later the sorting position is, the lower the correlation between the seismic attribute and the target reservoir porosity curve is; then, multiple seismic attributes with correlation between attributes higher than a preset threshold are screened out, and then the seismic attribute with the highest correlation with the target reservoir porosity curve is screened out from the screened multiple seismic attributes and retained, and other seismic attributes are removed; thereby, seismic attributes with low correlation between each other and high correlation with porosity are retained, and the preliminary selection results of seismic attributes are obtained; if the correlation between two seismic attributes A and B is high (where attribute A has a high correlation with the target porosity, and attribute B has a low correlation with the target porosity), it means that the influence of the two on the prediction process is similar to a high degree; and as for the selection of attributes to be removed between the two, attribute B with a low correlation with the target porosity is selected, then attribute B has a smaller influence on the network prediction of the target porosity than seismic attribute A; therefore, removing attribute B can reduce the amount of data input into the network, which can effectively improve the network training efficiency.

[0072] S22, evaluating the attribute importance of each earthquake attribute in the earthquake attribute preliminary selection result according to the random forest algorithm, and performing secondary sorting on each earthquake attribute in the earthquake attribute preliminary selection result according to the attribute importance;

[0073] The random forest is an integrated algorithm that randomly samples the sample data set by constructing multiple decision trees, learns the nonlinear relationship between the data, and votes on the prediction results of each tree to obtain the final result. The random forest algorithm can be used for feature selection because it can calculate the importance of the input attribute to the target parameter, that is, the attribute importance.

[0074] When using the random forest algorithm to solve regression problems, the attribute importance is mainly calculated by the out-of-bag error. For a data set containing m samples, the data that is not drawn by each decision tree in m random samplings with replacement is called out-of-bag data. For each decision tree i, use its out-of-bag data set OOB1 i Input the decision tree i to calculate the corresponding out-of-bag error, denoted as errorOOB1 i . In calculating the importance of attribute K I K When the out-of-bag data OOB1 i Add random noise to the data value of the attribute in to obtain the out-of-bag data OOB2 after the attribute K is disturbed i , input it into the trained random forest model, and the corresponding out-of-bag error errorOOB2 can be calculated for each tree i i

[0075] Assuming there are N decision trees in the random forest, then the importance of feature K is obtained K It is expressed as:

[0076]

[0077] Finally, the importance of all attributes is normalized to obtain the importance of each attribute.

[0078] S23, combining the seismic attributes in the seismic attribute preliminary selection results one by one according to the secondary sorting result based on the stepwise regression method to obtain multiple groups of attribute combinations, and combining the multiple groups of attributes to form an attribute sample set;

[0079] For example, assuming that the initial selection result of earthquake attributes includes N earthquake attributes, and the secondary sorting result of the N earthquake attributes is 1 attribute>2 attribute>3 attribute>...>N attribute; then take 1 attribute as the first attribute combination, take 1 attribute+2 attributes as the second attribute combination, take 1 attribute+2 attributes+3 attributes as the third attribute combination, and so on, until taking 1 attribute+2 attributes+3 attributes+...+N attributes as the Nth attribute combination, N attribute combinations can be obtained; take these N attribute combinations as attribute sample sets;

[0080] S24, using the cross-validation method to divide the attribute sample set into n parts, taking n-1 parts of the data as the training set and the remaining 1 part of the data as the validation set, and training the random forest model;

[0081] S25. After the training is completed, the average verification error of n trainings is taken as the generalization error, and the best attribute combination is selected based on the generalization error.

[0082] In the above step S3, the convolutional neural network and the gated recurrent network are briefly described:

[0083] (1) Convolutional Neural Network:

[0084] The Convolutional Neural Network (CNN), first proposed by Lecun et al. in 1998, has excellent feature extraction capabilities and is one of the most commonly used network frameworks for classification and regression prediction in deep learning. Compared with other networks, convolutional neural networks have the characteristics of weight sharing and local connection, which reduces the complexity of model training while reducing network parameters.

[0085] See also Figure 3 As shown in the figure, according to the dimension and direction of convolution, convolutional neural networks can be divided into one-dimensional convolution 1DConvolution, two-dimensional convolution 2D Convolution and three-dimensional convolution 3D Convolution. One-dimensional convolution only performs convolution operations in one direction, and so on. Two-dimensional convolution is a convolution operation performed in two directions. Currently, one-dimensional convolutional neural networks are mainly used for regression prediction.

[0086] (2) Gated recurrent network:

[0087] The Gated Recurrent Unit (GRU) neural network is an improvement of the Long Short-Term Memory (LSTM) neural network, proposed by Cho et al. in 2014. The unit structure of the LSTM neural network consists of a forget gate, an input gate, and an output gate, which can filter time series information to control the state of the cell unit. Compared with the complex network structure of the LSTM, the unit structure of the GRU neural network simplifies the three gates in the LSTM into two, the reset gate and the update gate.

[0088] The unit structure of GRU is as follows Figure 4 As shown in the figure, the reset gate r represents the degree of forgetting the information of the previous moment, which helps to capture the short-term dependencies in the sequence. The closer its value is to 1, the lower the degree of forgetting the historical information. It can be expressed by a mathematical formula:

[0089] r t =σ(W r x t +U r h t-1 +b r )

[0090] Among them, r t represents the reset gate; σ represents the sigmoid activation function; W r and U rBoth represent weight parameters; x t represents the input sample; b r Represents the bias term.

[0091] The update gate z represents the degree to which the information from the previous moment flows into the current moment, which helps to capture the long-term dependencies in the sequence. The closer its value is to 1, the higher the degree of retention of historical information. It can be expressed by mathematical formula as follows:

[0092] z t =σ(W z x t +U z h t-1 +b z )

[0093] Among them, z t represents the update gate; W z , U z All represent weight parameters; σ represents the sigmoid activation function; h t-1 represents the hidden state of the GRU unit at time t-1; b z Represents the bias term.

[0094] Since it is impossible to determine whether to update the hidden state, a candidate hidden state is introduced in GRU, which can be expressed as:

[0095]

[0096] in, represents the candidate hidden state; W and U both represent weight parameters; tanh represents the tanh activation function; b h Represents the bias term.

[0097] The candidate hidden state needs to be combined with the result of the update gate z to determine the new hidden state h t What proportion of it comes from the old state h t-1 and the new candidate hidden state Essentially, it is the hidden state h of the previous moment. t-1 and the candidate hidden state at the current moment The weighted sum of can be used to obtain the final output result of the gated recurrent unit at the current moment, which can be expressed as:

[0098]

[0099] Among them, h t Represents the output result of the GRU unit at the current moment.

[0100] GRU can save the information of the previous time step and use it as the input of the current moment. Its learning ability on time series data can capture the intrinsic connection between adjacent sampling points at different depths of the logging curve. Compared with the conventional LSTM network, the GRU network unit structure is lighter and simpler. As the network structure parameters are reduced, the training effect is improved.

[0101] The above single deep learning network model cannot fully capture the lateral feature change trend between the input attributes and the porosity value when the depth domain changes. CNN has great advantages in extracting local spatial features of input vectors due to its characteristics of weight sharing and local receptive field, while GRU can use the two gates in its network structure to control the flow and memory of information, and dynamically adjust the retention and forgetting of information according to the sequence characteristics, so as to obtain long-term dependencies in the sequence. Therefore, an embodiment of the present invention constructs a one-dimensional convolutional gated recurrent network model (i.e., 1D-CNN-GRU model), which includes an input layer, a one-dimensional convolutional layer (1D-CNN layer), a gated recurrent neural network layer (GRU layer), a fully connected layer, and an output layer; see specifically. Figure 5 As shown;

[0102] The present invention improves the traditional convolutional neural network in the 1D-CNN layer. Specifically, the role of the pooling layer in the traditional convolutional neural network is to reduce the dimension of the high-dimensional data obtained after convolution, reduce the amount of data and thus speed up the efficiency of network training. However, since it takes the maximum value or average value of the features within the time window range, it does not record the location information of the features, resulting in the loss of the original timing information of the data, making the ordered data disordered. The role of the fully connected layer is to integrate the data features reduced by the pooling layer and map them to the label parameter space, but the existence of the fully connected layer will disrupt the order sequence of the data, which is not conducive to the subsequent GRU structure to extract the timing features. In response to the above problems, the present invention discards the pooling layer and the fully connected layer in the one-dimensional convolutional neural network, and only retains the one-dimensional convolutional layer with extracted data space features, which is placed before the GRU network, responsible for extracting the multi-dimensional and complex lateral spatial features between each input feature. This improved one-dimensional convolutional model can not only extract the horizontal spatial features of the data without disrupting the temporal features of the data in the vertical direction (time or depth range), but also optimize the training attribute data to be input into the GRU network, which can greatly improve the prediction performance of the network.

[0103] In the above step P1, before predicting the target reservoir parameters in the well-free area within the target work area, it is necessary to find out the seismic attributes related to the target reservoir parameters as the second seismic attribute set; the introduction of seismic attributes is usually a process from many to few, that is, in the early stage of reservoir prediction, try to select a variety of seismic attributes that may be related to reservoir physical property parameters, so as to not only make full use of the information contained in the seismic data, but also improve the accuracy of reservoir physical property parameters and reduce the multi-solution of reservoir prediction.

[0104] In the above step P2, according to the best attribute combination obtained in the above step S2, the same attribute combination is screened out from the second seismic attribute set as the final attribute combination, and the value of the final attribute combination is used as the input of the one-dimensional convolutional gated recurrent network model.

[0105] In the above step P3, the following steps are specifically included:

[0106] P31, input the final attribute combination into the trained one-dimensional convolutional gated recurrent network model through the input layer;

[0107] P32, extract the lateral spatial features of each seismic attribute in the final attribute combination through a one-dimensional convolutional layer (1D-CNN layer); so that it can be used as input data for the GRU layer later;

[0108] In the embodiment of the present invention, the one-dimensional convolution layer is set to 2 layers to fully extract the lateral spatial features of the data; and a dropout layer (Dropout layer) is added after the convolution layer to prevent overfitting;

[0109] P33, extract the longitudinal temporal features of each transverse spatial feature through the gated recurrent neural network layer (GRU layer);

[0110] In the embodiment of the present invention, a Dropout layer is provided after the GRU layer to prevent overfitting of the neural network; in the embodiment of the present invention, the GRU layer is set to 3 layers;

[0111] P34. The lateral spatial characteristics and longitudinal time series characteristics of each seismic attribute obtained above are integrated through the fully connected layer to complete the mapping of spatiotemporal characteristics to reservoir porosity logging curve values, and the target reservoir porosity is output through the output layer.

[0112] Next, a reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network provided by the present invention is described in detail through a specific embodiment.

[0113] 1. Overview of the study area:

[0114] The W work area is located in the Shenfu block on the eastern edge of the Ordos Basin. It is a typical loess mountain with a relatively complex surface morphology and well-developed gullies. The shape of this area is mainly distributed in a north-south strip, and it contacts the Qinling Mountains and Yinshan Mountains in the north and south respectively. The east is the Jinxi uplift belt that extends in the north-south direction. Therefore, the structural deformation in the study area is mainly controlled by the north-south structure, and there is no large structural deformation as a whole. The W work area is a monocline structure with a north-south orientation and a westward inclination. On this basis, north-south folds are developed with small fold fluctuations. The structure in the area is relatively simple, the strata are relatively gentle, and faults are not developed.

[0115] 2. Attribute selection based on Pearson-Random Forest:

[0116] A total of 12 wells were collected from the W work area, named Jing-1, Jing-2, ..., Jing-12, and the 1300m-2020m depth section of the well logging deviation center section (the main target layer is the Box 7th section to the Ben 2nd section) and the corresponding seismic data between the 800ms-1160ms depth section were selected as the target layer section for prediction research. Jing-1 to Jing-9 were used as training data sets, Jing-10 to Jing-11 as validation data sets, and Jing-12 as a blind well to test the effectiveness of network prediction.

[0117] The embodiment of the present invention extracts 20 types of seismic attributes from the target area for research, and the specific contents are as follows:

[0118] (1) 10Hz single frequency body properties (Mono Freq 10Hz);

[0119] (2) 20Hz single frequency body properties (Mono Freq 20Hz);

[0120] (3) 30Hz single frequency body properties (Mono Freq 30Hz);

[0121] (4) 40Hz single frequency body attribute (Mono Freq 40Hz);

[0122] (5) 50Hz single frequency body attribute (Mono Freq 50Hz);

[0123] (6) 60Hz single frequency body attribute (Mono Freq 60Hz);

[0124] (7) 70Hz single frequency body attribute (Mono Freq 70Hz);

[0125] (8) Instantaneous amplitude (InstantAmp);

[0126] (9) Instant Phase;

[0127] (10) Instantaneous frequency (Instant Freq);

[0128] (11) Root mean square amplitude (RMSAmp);

[0129] (12) Instant Q factor (Instant Q);

[0130] (13) Sweet;

[0131] (14) Half Energy;

[0132] (15) Energy weighted instantaneous frequency;

[0133] (16) squared amplitude difference (S-varianceAmp);

[0134] (17) Difference between instantaneous frequency and amplitude-weighted instantaneous frequency (T-layer indication);

[0135] (18) Dominant Freq;

[0136] (19) Spectral absorption factor (Absorb factor);

[0137] (20) Variance

[0138] Together with the waveform indication inversion impedance body (P-impedance) and raw seismic (Raw seismic) in the study area, 22 initial attribute data are formed.

[0139] In the process of using the Pearson correlation coefficient method to calculate the linear correlation between each attribute feature and the porosity logging curve, the Pearson correlation coefficient matrix calculation results (see Figure 6 As shown in Table 1, the related attributes with correlation coefficients higher than 0.8 are screened out, and the attributes with low correlation with the porosity curve are removed from the screened attributes, and the attribute with the highest correlation with the porosity curve is retained, thereby obtaining attributes with low correlation between attributes and high correlation with porosity, and obtaining the preliminary selection results of attributes as shown in Table 1.

[0140] Table 1 Attribute preliminary selection results based on Pearson correlation coefficient method

[0141] Serial number s attribute name iCorrelation coefficient 1 b wave impedance body 0.42 2 Spectral absorption factor 0.40 3 Frequency 0.39 4 Instantaneous Q factor 0.35 5 Energy half-decay 0.33 6 Instantaneous frequency 0.28 7 Energy-weighted instantaneous frequency 0.18 8 RMS amplitude 0.17 9 10Hz single frequency body properties 0.04 10 Variance properties 0.04 11 70Hz single frequency body properties 0.03 12 Instantaneous phase 0.01 13 The difference between instantaneous frequency and amplitude-weighted instantaneous frequency 0.01 14 30Hz single frequency body properties 0.01 15 50Hz single frequency body properties 0.01 16 40Hz single frequency body properties 0.01

[0142] Through Pearson correlation analysis, it is known that the linear correlation between seismic attributes and porosity curves is not strong. Using a single attribute or fewer attributes will lead to multiple solutions for the prediction results, and adding too many attributes will make the network complicated. Therefore, on the basis of the initial selection of attributes by the Pearson correlation coefficient method, the random forest algorithm is used for attribute optimization; that is, the attribute importance of each seismic attribute in the initial selection results of seismic attributes is evaluated according to the random forest algorithm, and the seismic attributes are secondary sorted according to the attribute importance;

[0143] Figure 7 The importance ranking of attributes obtained by the random forest algorithm model after parameter optimization shows that the wave impedance body obtained by waveform indication inversion has the highest importance for porosity curve prediction, which is 0.276, followed by the root mean square amplitude and 10Hz single frequency body attributes, which are 0.08 and 0.062 respectively. The wave impedance attribute accounts for about 28% of the importance, indicating that the wave impedance attribute has a great influence on the prediction of reservoir porosity. The importance of seismic attributes for porosity prediction can be obtained from the importance ranking of attributes. Since it is impossible to define a clear importance threshold to delete attributes, the best attribute combination cannot be selected only by importance ranking, but the attribute importance ranking provides a direction for further feature combination optimization.

[0144] In order to obtain the best attribute combination, according to the stepwise regression theory, the 16 seismic attributes are combined in sequence according to the feature importance ranking generated by the random forest, that is, the P-impedance attribute ranked first is defined as attribute combination 1, the combination of P-impedance and RMS Amp is defined as attribute combination 2, and P-impedance, RMS Amp and Mono Freq10Hz are defined as attribute combination 3. The above steps are repeated to finally obtain 16 seismic attribute combinations, which are input into the random forest algorithm as data sets for training and verification. Under each combination of characteristic attributes, 70% of the attribute data are randomly selected as the training data set, and the remaining data are made into the verification data set and input into the random forest model. In order to correspond to the logging curves of the 11 wells, one of the wells is selected as the verification well in turn by 11-fold cross validation, and the rest of the wells are selected as training wells. The average value of the 11-fold root mean square error is used as the evaluation index to evaluate the quality of the attribute combination, and then the best attribute combination is selected according to the size of the average value of the 11-fold cross validation error.

[0145] Figure 8This is the error analysis diagram of the earthquake attribute combination obtained in the 11-fold cross validation. The results show that as the number of input features increases, the average root mean square error of the 11-fold cross validation has a continuous downward trend with the increase in the number of attributes in the attribute combination. When the first 9 attributes are input in order of importance, the average value of the 11-fold cross validation error is as low as 1.13 (rounded to 1.129), and there is an upward trend when the number of input features exceeds 9, and then gradually stabilizes, that is, when the number of attributes exceeds 9, the random forest begins to overfit. Too many features will affect the prediction ability and accuracy of the model, and will also consume a lot of computing resources and time. Considering accuracy and efficiency, it is decided to select the top 9 attribute features in importance, and the final attribute combination is shown in Table 2.

[0146] Table 2 Final selection of attribute combinations

[0147] Serial number Property Name Importance 1 Wave impedance 0.276 2 RMS amplitude 0.080 3 10Hz single frequency body properties 0.062 4 Variance properties 0.060 5 40Hz single frequency body properties 0.056 6 30Hz single frequency body properties 0.055 7 Energy half-decay 0.051 8 Spectral absorption factor 0.050 9 70Hz single frequency body properties 0.047

[0148] 3. Construction of one-dimensional convolutional gated recurrent network model (1D-CNN-GRU model):

[0149] The nine seismic attribute combinations obtained after attribute optimization are used as input data for deep learning, and the deep learning model selected is the 1D-CNN-GRU model.

[0150] The 1D-CNN-GRU model constructed by the present invention is based on the Pytorch framework. In order to reduce the risk of overfitting of the constructed reservoir porosity prediction model, the Dropout function is used to optimize the number of neurons in the hidden layer of the network, and the Dropout layer is set after the last layer of one-dimensional convolutional neural network and GRU layer. In order to obtain the best 1D-CNN-GRU network parameter setting, the present invention uses the Adam Optimizer optimizer based on the Adam algorithm to find the global optimal solution in the network training process. Since the optimizer can adaptively adjust the learning rate of the network, the initial learning rate can be set to a larger value to speed up the network convergence speed, and then set to reduce the learning rate according to a certain proportion every time the training reaches a certain number of times. This time, the initial learning rate is set to 0.001 to ensure that the network has a faster convergence speed during the training process and will not miss the global optimal solution. In the process of multiple training, it is found that the network begins to converge when the number of training times is between 200-300 times, so the maximum batch (Epochs) of training is set to 500. Combined with the experience of predecessors, multiple experiments were conducted on the number of neurons in each layer, the number of 1D-CNN layers, and the number of GRU hidden layers, and finally the parameter settings of the 1D-CNN-GRU reservoir porosity prediction model to be constructed in the present invention were determined as shown in Table 3.

[0151] Table 3 Hyperparameter settings based on 1D-CNN-GRU model

[0152] name parameter 1D-CNN in_channels=9; out_channels=128; stride=1; kernel_size=1 1D-CNN in_channels=128; out_channels=64; stride=1; kernel_size=1; GRU-1 input_size=20; hDidden_tsiz0e2=64; Seq_lenth=8 GRU-2 input_size=64; hidden_size=64; Seq_lenth=8 GRU-3 input_size=64; hidden_size=64; Dropout=0.4; Seq_lenth=8 FCLayer in_features=64, hidden_size=64, out_features=1

[0153] Among them, in_channels is the input neuron dimension, out_channels is the output neuron dimension, kernal_size is the convolution kernel size, Stride is the convolution step size, input_size is the input data size, Dropout is the temporary retreat regularization parameter, hidden_size is the hidden state dimension, Seq_length is the sequence length, in_features is the input feature size, and out_features is the output feature size.

[0154] 4. Experimental results and analysis:

[0155] (1) Reservoir porosity prediction results based on 1D-CNN-GRU model

[0156] The attribute optimization results are input into the 1D-CNN-GRU model for prediction. In order to verify the performance of the network constructed by the present invention, the same data are input into the 1D-CNN network and the GRU network for comparison. The comparison results are as follows: Fig. 9 shown.

[0157] Table 4 below shows the error comparison analysis results of the three models of 1D-CNN, GRU, and 1D-CNN-GRU on the training set and the validation set. The embodiment of the present invention uses the mean absolute error (MAE) and the root mean square error (RMSE) as evaluation indicators to estimate the prediction ability of the model. It can be seen from the parameters in the table that the error of the 1D-CNN-GRU model is better than that of the 1D-CNN model and the GRU model as a whole.

[0158] Table 4 Comparison of porosity prediction model results

[0159]

[0160] Fig.10 This is a scatter plot comparing the output results of the 1D-CNN-GRU model on the validation data set and the true value. It can be clearly seen that the 1D-CNN-GRU model in the present invention has better prediction effect than the single network models of 1D-CNN and GRU, which is mainly reflected in that the predicted value and the true value are more concentrated, and the prediction performance is improved in areas with large and small porosity values. The porosity predicted by the 1D-CNN-GRU network is greatly improved in the entire value range compared with the 1D-CNN and GRU models, and is closer to the true porosity logging curve value.

[0161] (2) Prediction and verification of porosity profile of tight sandstone reservoir

[0162] Well Jing-12 did not participate in the training and verification of the entire network process, and was used as a blind well to verify the prediction performance and accuracy of the network. At the location of Well Jing-12, profile data were extracted along the Xline and Inline directions, and after data preprocessing and dimension transformation, they were made consistent with the network input, and then they were input into the network in batches for training to obtain the porosity prediction profile of the study area.

[0163] Fig.11 and Fig.12 The porosity profile prediction results in the Inline direction and Xline direction at the Jing-12 well location are respectively predicted based on the 1D-CNN-GRU network. It can be seen from the figure that the predicted porosity is consistent with the porosity logging curve at the well point, which verifies the prediction performance and accuracy of the 1D-CNN-GRU reservoir porosity prediction model constructed in this study. It can be seen from the figure that the porosity profile predicted by the 1D-CNN-GRU model has good lateral spatial continuity, and the spatial distribution form and distribution characteristics of the porosity can be obtained, and then the spatial distribution form and distribution characteristics of the tight sandstone reservoir can be obtained.

[0164] (3) Porosity prediction of tight sandstone reservoirs

[0165] The 1D-CNN-GRU network model trained at the well point is extended to the well-free area, and the 3D seismic attribute volume is input to obtain the 3D porosity prediction results of the study area, such as Fig.13 The figure shows the spatial location and general distribution of the tight sandstone reservoir in the study area, which can provide a certain basis for the study of tight sandstone reservoirs and their porosity.

[0166] In summary, seismic attributes and porosity often exhibit a complex nonlinear relationship. The Pearson correlation coefficient-random forest method proposed in the embodiment of the present invention solves the analytical problem of strong nonlinear correlation between sample data, and gradually selects the seismic attributes that are most suitable for predicting porosity; in addition, the present invention combines a one-dimensional convolutional neural network with a GRU network to construct a 1D-CNN-GRU joint model. Based on the improved 1D-CNN-GRU porosity prediction model, the temporal and spatial characteristics of well logging and seismic in the horizontal and vertical directions are integrated, and the prediction results have a high resolution in the horizontal and vertical directions; compared with a single 1D-CNN network and a GRU network, the prediction error of the 1D-CNN-GRU porosity prediction model is reduced by 50% and 23%, respectively, proving that the joint network is more advantageous than a single network in porosity prediction.

[0167] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0168] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network, characterized in that: Including offline and online parts; The offline part includes the following steps: S1, obtaining multiple seismic attributes related to reservoir parameters at well points in a target work area as a first seismic attribute set; S2. acquiring an optimal attribute combination from the first seismic attribute set; S3, taking the optimal attribute combination as input, taking the porosity corresponding to the optimal attribute combination as output, and training a one-dimensional convolutional gated recurrent network model; The online part includes the following steps: P1. Acquire multiple seismic attributes related to target reservoir parameters in a wellless area in the target work area as a second seismic attribute set; P2. According to the optimal attribute combination, a final attribute combination is selected from the second seismic attribute set; P3, inputting the final attribute combination into the trained one-dimensional convolutional gated recurrent network model, and outputting the target reservoir porosity; The step S2 specifically includes: S21, using the Pearson correlation coefficient method to obtain the correlation between the seismic attributes in the first seismic attribute set, and sorting the seismic attributes in the first seismic attribute set according to the correlation; removing redundant seismic attributes in the first seismic attribute set according to the correlation and the sorting result, to obtain a preliminary selection result of seismic attributes; S22, evaluating the attribute importance of each seismic attribute in the seismic attribute preliminary selection result according to the random forest algorithm, and performing secondary sorting on each seismic attribute in the seismic attribute preliminary selection result according to the attribute importance; S23, based on the stepwise regression method, combining the seismic attributes in the seismic attribute preliminary selection results one by one according to the secondary sorting result to obtain multiple groups of attribute combinations, and combining the multiple groups of attributes to form an attribute sample set; S24, using a cross-validation method to divide the attribute sample set into n parts, taking n-1 parts of the data as training sets and the remaining 1 part of the data as validation sets in turn, and training the random forest model; S25. After the training is completed, the average verification error of n trainings is taken as the generalization error, and the best attribute combination is screened out according to the generalization error; The one-dimensional convolutional gated recurrent network model includes an input layer, a one-dimensional convolutional layer, a gated recurrent neural network layer, a fully connected layer and an output layer; The step P3 specifically includes: P31, inputting the final attribute combination into the trained one-dimensional convolutional gated recurrent network model through the input layer; P32, extracting the lateral spatial features of each seismic attribute in the final attribute combination through the one-dimensional convolution layer; P33, extracting the longitudinal time series features of each transverse spatial feature through the gated recurrent neural network layer; P34. Integrate the lateral spatial features and the longitudinal time series features through a fully connected layer to complete the mapping of spatiotemporal features to reservoir porosity logging curve values, and output the target reservoir porosity through the output layer.

2. A reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network according to claim 1, characterized in that: The removing redundant seismic attributes in the first seismic attribute set according to the correlation and the primary sorting result specifically includes: selecting a plurality of seismic attributes having a correlation higher than a preset threshold from the first seismic attribute set, selecting a seismic attribute having the highest correlation with a target reservoir porosity curve from the plurality of seismic attributes and retaining the seismic attribute, and removing other seismic attributes.

3. The reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network according to claim 1 is characterized in that: The one-dimensional convolution layer has two layers.

4. The reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network according to claim 1, characterized in that: A Dropout layer is arranged after the one-dimensional convolution layer.

5. The reservoir porosity prediction method based on a one-dimensional convolutional gated recurrent network according to claim 1, characterized in that: A Dropout layer is arranged after the gated recurrent neural network layer.

Images