Risk stratum identification method, system and device and computer readable storage medium

Abstract

The invention relates to a risk stratum identification method, system and device and a computer readable storage medium. The risk stratum identification method comprises the steps of obtaining exploration data of a to-be-identified oil and gas well in a preset depth range; according to the exploration data, obtaining space-depth joint features of the to-be-identified oil and gas well at a deeper underground level; inputting the space-depth joint features into a preset GRU neural network for processing to obtain lithology prediction data at a deeper underground level; and outputting an underground deeper stratum identification result according to the lithology prediction data. According to the method, the identification efficiency of the risk stratum in oil-gas exploration and development can be improved, and an accurate identification result can be obtained while relatively low calculation cost and calculation complexity are realized.

Description

Technical Field

[0001] The present invention relates to the field of geology and oil and gas exploration technology, and in particular to a method, system, equipment and computer-readable storage medium for identifying risky strata. Background Art

[0002] During deepwater oil and gas exploration and development, oil and gas accumulation areas often possess highly complex geological conditions and environmental characteristics, which can easily lead to the formation of high-temperature, high-pressure risk formations, often containing abnormally high-pressure zones. Failure to accurately identify these deepwater, high-temperature, high-pressure risk formations in advance poses a serious threat to safe offshore oil and gas production and personnel safety.

[0003] However, there are few methods proposed to identify risky formations under deepwater, high-temperature, and high-pressure conditions. Traditional methods such as cuttings logging and core analysis have the problem of delayed acquisition of formation information. Using traditional methods to identify risky formations will reduce the efficiency of deep-sea oil and gas extraction. Summary of the Invention

[0004] The technical problem to be solved by the present invention is that the traditional risk formation identification method has a lag in obtaining formation information. Based on this, a risk formation identification method, system, device and computer-readable storage medium are provided.

[0005] The technical solution adopted by the present invention to solve the technical problem is to construct a risk formation identification method, including the following steps:

[0006] S1: Acquire exploration data of the oil and gas well to be identified within a preset depth range;

[0007] S2: obtaining, based on the exploration data, a spatial-depth joint feature of the oil and gas well to be identified, which indicates that the well is located at a deeper level underground;

[0008] S3: Inputting the space-depth joint feature into a preset GRU neural network for processing to obtain lithology prediction data at a deeper level underground;

[0009] S4: Outputting identification results of deeper underground strata based on the lithologic prediction data.

[0010] In some embodiments, step S2 includes the following sub-steps:

[0011] S21: Inputting the exploration data into a preset spatial distribution feature model to obtain spatial distribution features of the oil and gas wells to be identified at a deeper level underground; the spatial distribution feature model is constructed based on a Kriging model, which is configured to infer the spatial relationship between known depth positions and unknown depth positions;

[0012] S22: Inputting the exploration data into a preset depth feature model to obtain depth features of the oil and gas wells to be identified at a deeper level underground; the depth feature model is constructed based on an attention mechanism and is configured to extract depth features based on the degree of match between the input data and the prediction target within a preset depth range;

[0013] S23: Dimensionally concatenate the corresponding spatial distribution features and the depth features to obtain the spatial-depth joint features.

[0014] In some embodiments, the spatial distribution characteristics Expressed as:

[0015]

[0016] Where, Represents the spatial distribution characteristics at depth S; are the basis functions related to the x-dimension, y-dimension, and depth at depth S, respectively; D represents the total depth; S represents the current depth; x, y, and de represent the spatial coordinate axes;

[0017] The spatial coordinates are obtained through the spatial distribution feature model, and the spatial distribution feature model is specifically:

[0018]

[0019] Where S0 is the unobserved position, r(S0)=[R(θ,S1,S0),…,R(θ,S N ,S0)] T represents the correlation matrix of S0, R(θ, S i , S0) is the correlation function of the random process, which represents the spatial correlation with reservoir lithology, S i is the position coordinate of the i-th observation, θ is the hyperparameter of the Kriging model; β * =(F T R -1 F)F T R -1 P represents the weighted least squares estimate of β, F represents the observation matrix related to lithology, R -1 represents the covariance matrix associated with lithology, and P represents the measurement matrix of spatially related locations.

[0020] In some embodiments, the depth feature model is specifically:

[0021]

[0022] Where, X eWell logging depth features or logging depth features extracted based on the importance of different well logging attributes or logging attributes to the predicted data; for The normalized value of is the depth attention weight of the logging attribute or mud logging attribute within a certain depth range; v d 、w d 、u d 、b d Respectively represent the weights and biases that need to be trained through the multi-layer perceptron network; h and z represent all hidden layer states and update gate information of the GRU neural network framework; Represents the value of the mth attribute at the well depth d; sim(x m , Y) represents the similarity between the historical information of the predicted target and the input attributes.

[0023] In some embodiments, the joint spatial-depth feature is represented as:

[0024]

[0025] The step S3 is specifically as follows:

[0026] The space-depth joint feature is input into the GRU neural network; the hidden layer of the GRU neural network calculates a hidden state according to the space-depth joint feature, and the output layer of the GRU neural network converts the hidden state into the lithology prediction data.

[0027] In some embodiments, the computational process involved in the GRU neural network includes:

[0028]

[0029] Where r d 、z d 、 h d Represents reset gate, update gate, new storage layer, and hidden layer respectively; U d , W d is the weight, b d is the bias of the GRU neural network, U d 、W d and b d Here r, z, and h represent the weights corresponding to the reset gate, update gate, and new storage layer, respectively; sig() is the sigmoid activation function; and tanh() is the tanh activation function, which is used to control data in the range of (-1, 1).

[0030] In some embodiments, step S1 includes the following sub-steps:

[0031] S11: obtaining original exploration data of the oil and gas wells to be identified;

[0032] S12: Preprocessing the original exploration data to obtain preprocessed exploration data;

[0033] S13: Using the pre-processed exploration data as exploration data for obtaining space-depth joint features.

[0034] The present invention also constructs a risk formation identification system, comprising:

[0035] An acquisition module is used to acquire exploration data of the oil and gas well to be identified within a preset depth range;

[0036] An analysis module is used to obtain, based on the exploration data, a spatial-depth joint feature of the oil and gas well to be identified, which is located at a deeper level underground;

[0037] A neural network module is used to input the space-depth joint feature into a preset GRU neural network for processing to obtain lithology prediction data at a deeper level underground;

[0038] The output module is used to output the identification result of the deeper underground strata based on the lithology prediction data.

[0039] The present invention also constructs a risk formation identification device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the above-mentioned risk formation identification method.

[0040] The present invention also constructs a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the above-mentioned risk formation identification method is implemented.

[0041] The implementation of the present invention has the following beneficial effects: secondly, the risk formation identification method can improve the efficiency of identifying risk formations in oil and gas exploration and development by acquiring exploration data and using the GRU neural network to obtain lithologic prediction data to identify risk formations; secondly, the risk formation identification method can obtain accurate identification results while achieving lower computational cost and computational complexity. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0043] Figure 1 Flowchart of a method for identifying risky formations in some embodiments of the present invention;

[0044] Figure 2 A flowchart for obtaining the space-depth joint features of the oil and gas wells to be identified in some embodiments of the present invention;

[0045] Figure 3 This is a diagram of the working principle of the GRU neural network in some embodiments of the present invention. DETAILED DESCRIPTION

[0046] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0047] It should be noted that the flowcharts shown in the accompanying drawings are for illustrative purposes only and do not necessarily include all content and operations / steps, nor must they be executed in the order described. For example, some operations / steps may be decomposed, while others may be combined or partially combined. Therefore, the actual execution order may vary depending on the actual situation.

[0048] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically separate entities. That is, these functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0049] Please refer to Figure 1 The present invention constructs a risk formation identification method, which mainly includes the following steps:

[0050] S1: Acquire exploration data of the oil and gas well to be identified within a preset depth range;

[0051] S2: Based on the exploration data, the spatial-depth joint characteristics of the oil and gas wells to be identified are obtained at a deeper level underground;

[0052] S3: Input the spatial-depth joint features into the preset GRU neural network for processing to obtain lithology prediction data at deeper levels underground;

[0053] S4: Based on the lithology prediction data, output the identification results of deeper underground strata.

[0054] Understandably, the essence of this risky stratum identification method is to observe some shallower strata and then, based on this data, identify and predict deeper strata. Using these stratum identification results, personnel can predict whether deeper strata are risky during oil and gas exploration and development, thereby ensuring safe offshore oil and gas production and personnel safety.

[0055] The present invention uses a GRU neural network (gated recurrent unit neural network) to process spatial-depth joint features to participate in recognition, which can improve recognition efficiency and obtain accurate recognition results while achieving lower computational cost and computational complexity. Compared with commonly used deep learning models (such as DT, LR, RF, SVC, Convld and other models), this risk stratum identification method has better predictive ability and is suitable for situations where stratum lithology and stratigraphic positions change greatly. Secondly, using this risk stratum identification method, strata can be identified once every 10 meters or so, breaking the limitation of traditional identification methods that strata can only be identified once every 30 meters or more, thereby improving the resolution and accuracy of stratum identification.

[0056] The specific steps of this risk formation identification method are described in detail below.

[0057] In step S1, the exploration data within the preset depth range can be understood as data obtained by personnel from exploration of shallower formations. The exploration data may include well logging data and mud logging data. In some embodiments, the well logging data may include formation density, formation porosity, formation resistivity, acoustic wave, and neutron-gamma logging; and the mud logging data may include drilling rate, weight on bit, rotary table speed, pump displacement, and ECD (equivalent bottom hole density).

[0058] Preferably, step S1 may include the following sub-steps:

[0059] S11: obtaining original exploration data of the oil and gas wells to be identified;

[0060] S12: preprocessing the original exploration data to obtain preprocessed exploration data;

[0061] S13: Using the pre-processed exploration data as exploration data for obtaining space-depth joint features.

[0062] It is understood that preprocessing may include cleaning the data (removing noise, processing missing values, correcting erroneous data, etc.); by cleaning the data, the accuracy and reliability of the data can be improved. In some embodiments, cleaning the data may include deleting abnormal data, such as data with values ​​of 0, -999, etc.

[0063] In addition, preprocessing can also include normalization. Understandably, when the value ranges (scales) of sample features vary greatly, features with large value ranges will play a dominant role. Therefore, in order to improve training performance, the acquired raw exploration data can be normalized.

[0064] In some embodiments, the Min-Max method may be used to normalize the raw exploration data, specifically:

[0065]

[0066] Where x′ i and y i are the actual value and normalized value of the original exploration data at a certain depth respectively; x max 、x min They are the total average, maximum and minimum values ​​of the original exploration data at all depths.

[0067] like Figure 1 As shown, in step S2, it can be specifically:

[0068] The exploration data is fed into a pre-set mathematical model to derive spatial-depth joint characteristics of the oil and gas wells to be identified at deeper levels underground. The mathematical model can include a spatial distribution characteristic model and a depth characteristic model. Deeper levels underground can be understood as levels deeper than the pre-set depth range.

[0069] In some embodiments, reference may be made to Figure 2 , step S2 may include the following sub-steps:

[0070] S21: Inputting the exploration data into the spatial distribution characteristic model to obtain the spatial distribution characteristics of the oil and gas wells to be identified at a deeper level underground;

[0071] S22: Inputting the exploration data into the depth feature model to obtain the depth features of the oil and gas wells to be identified at a deeper level underground;

[0072] S23: Dimensionally concatenate the corresponding spatial distribution features and depth features to obtain spatial-depth joint features.

[0073] The spatial distribution feature model can be constructed based on the Kriging model. The construction process is as follows: the Kriging model is used to predict the spatial relationship between known depth locations and unknown depth locations. The Kriging model is an interpolation method based on spatial correlation. It describes spatial correlation by establishing a semivariogram and uses the values ​​and spatial locations of known observation points to predict the values ​​and spatial locations of unknown points.

[0074] In the exploration data, according to the observed spatial observation points Among them, N is the spatial coordinate attribute, d is the well depth, and the observation value at the relevant position can be obtained, which is expressed as P = [p1, p2, ..., p D ] T .

[0075] For practical approximation, the classic Kriging spatial position prediction model can be expressed as:

[0076]

[0077] Where Z(S) represents spatial correlation; F(S)=[f1(s),f2(s)...f n (s)] T Represents the regression polynomial matrix related to spatial position; β=[β1,β2...β N ] T are the coefficients of the corresponding regression polynomial; T represents the matrix transpose.

[0078] Regarding Z(S), in some embodiments, an estimate of Z(S) can be obtained by calculating a finite covariance matrix, but this calculation method has high computational difficulty and storage complexity. Therefore, preferably, Z(S) can be approximated as a finite weighted sum of multiple radial basis functions to reduce the difficulty of obtaining; the calculation is specifically:

[0079]

[0080] Where, ω i represents a pair of uncorrelated random variables, Represents the orthogonal radial basis function transformed from spatial coordinates.

[0081] Then, for the unobserved position (S0), the Kriging spatial position prediction model can be expressed as:

[0082]

[0083] In the formula, r(S0)=[R(θ,S1,S0),…,R(θ,S N ,S0)] T represents the correlation matrix of S0, R(θ, S i , S0) is the correlation function of the random process, which represents the spatial correlation with reservoir lithology, S i is the position coordinate of the i-th observation, θ is the hyperparameter of the Kriging model; β * =(F T R -1 F)F T R -1 P represents the weighted least squares estimate of β, F represents the observation matrix related to lithology, R -1 represents the covariance matrix associated with lithology, and P represents the measurement matrix of spatially related locations.

[0084] It can be understood that through this prediction model, the value of the unknown location can be predicted using the known observation value and spatial correlation. T β *is a linear prediction based on the known observation (depth) position; r(S0) T R ―1 (P―Fβ * ) is an adjustment term based on spatial correlation, which is used to modify the linear prediction to make it more consistent with the actual value at the unknown depth position.

[0085] Therefore, at the location of the oil and gas well to be identified, for any spatial coordinate, the spatial distribution characteristics can be obtained by constructing the basis function. It can be expressed as:

[0086]

[0087] Where, Represents the spatial distribution characteristics at depth S; are the basis functions related to the x-dimension, y-dimension, and depth at depth S, respectively; D represents the total depth; S represents the current depth; x, y, and de represent the spatial coordinate axes.

[0088] Understandably, in the calculation When the second term on the right side of the equation is: r(S0) T R ―1 (P―Fβ * ) is equivalent to the spatial correlation Z(S). When calculating Z(S), the traditional method relies on calculating the finite covariance matrix R -1 Obtaining Z(S) estimation is complex. Therefore, by constructing the basis function To calculate Z(S), that is The spatial distribution characteristics are a set of basis functions. Therefore, It is calculating obtained in the process of Reverse calculation possible

[0089] Next, the deep feature model can be constructed based on an attention mechanism. Preferably, the attention mechanism can be configured as a self-attention mechanism to address the complex problem of multiple input vectors during test data input. Furthermore, it can automatically extract logging and mud logging depth features based on the degree of match between the input data and the predicted target within a certain depth range. Specifically, the attention mechanism automatically extracts key features by calculating the probability distribution between the input data and the target data. The higher the probability, the more important the input data is to the predicted target. Logging and mud logging data are data sequences acquired by actual instruments that vary with reservoir depth. The interaction between different depths and the predicted data can range up to 15-25 meters, indicating a cumulative effect within a certain depth range. To account for this cumulative effect, a deep attention mechanism block is constructed based on the attention mechanism. Its sub-block can identify the relationship between the current depth and adjacent depths of logging or mud logging attributes within a certain depth range, thereby automatically extracting depth features.

[0090] In some embodiments, the depth feature X e It can be calculated according to the following equation:

[0091]

[0092] Where, X e Well logging depth features or logging depth features extracted based on the importance of different well logging attributes or logging attributes to the predicted data; for The normalized value of is the depth attention weight of the logging attribute or mud logging attribute within a certain depth range.

[0093] The depth attention weight of the logging attribute or mud logging attribute within a certain depth range can be calculated according to the following equation:

[0094]

[0095] Where, v d 、w d 、u d 、b d Respectively represent the weights and biases that need to be trained through the multilayer perceptron network; h and z represent all hidden layer states and update gate information of the GRU neural network framework; Represents the value of the mth attribute at the well depth d; sim(x m , Y) represents the similarity between the historical information of the predicted target and the input attributes; Y=(y1,y2…y d ) represents the historical information of the prediction target, where y dis the value of the i-th historical information point.

[0096] It should be noted that the attribute "value of the mth attribute at well depth d" can be a well logging attribute or a mud logging attribute. In other words, the value of a well logging attribute or a mud logging attribute refers to data such as formation density, formation porosity, formation resistivity, acoustic wave, neutron-gamma logging, drilling rate, weight on bit, rotary table speed, pump displacement, and ECD.

[0097] sim(x m , Y) can be calculated by cosine similarity metric. Specifically:

[0098]

[0099] Preferably, when training the spatial distribution model and the depth model, the Adam algorithm can be used as the training optimization algorithm. The Adam algorithm is specifically:

[0100] v t ←β1v t-1 +(1-β1)g t

[0101]

[0102] Where, v t , s t is the state variable, β1 and β2 are non-negative weighted parameters, g t is the gradient.

[0103] Understandably, v t and s t Represents the exponentially weighted moving average of the gradient and the exponentially weighted moving average of its square. These values ​​are used to estimate the mean and variance of the gradient and thus adjust the learning rate of each parameter.

[0104] Then, after obtaining the spatial distribution features and depth features, the corresponding spatial distribution features and depth features can be dimensionally spliced. The spatial-depth joint features after dimension splicing can be expressed as:

[0105]

[0106] It can be seen that the spatial-depth joint feature after dimension splicing is a data sequence.

[0107] In step S3, the spatial-depth joint feature data is used as input to the GRU neural network. The input data is processed by the GRU neural network to predict the lithologic distribution at different logging or mud logging depths. This means that multiple lithologic prediction data can be obtained to characterize strata at different depths. Then, based on the lithologic prediction data at different depths, it is possible to determine whether the strata at different depths are risky strata, thereby obtaining strata identification results at different depths for the oil and gas wells to be identified.

[0108] It's important to note that in practice, some lithologies are often directly defined as risky formations. Furthermore, some related technologies, after predicting the lithology, combine it with the characteristics of the current block to determine whether a formation is risky. Therefore, methods for determining whether a formation is risky based on lithology, or based on lithology and the characteristics of the current block, can be referenced in related technologies and will not be elaborated on here. To further clarify, a block refers to an area designated for specific geological research, resource exploration, or development activities.

[0109] The processing process of the GRU neural network is as follows: input the spatial-depth joint feature data, calculate the hidden state through the hidden layer of the GRU neural network, and then obtain the lithology prediction data through the output layer.

[0110] In some embodiments, the working principle of the GRU neural network can be referred to Figure 3 .like Figure 3 As shown, the GRU neural network can include a reset gate, an update gate, a new storage layer (candidate hidden state) and a hidden layer. Understandably, in stratum identification, the reset gate determines how much information the hidden state of the previous depth should retain at the current depth; the update gate determines how the new information of the current depth and the hidden state of the previous depth should be combined; the hidden layer can transfer information of the measured depth to the current depth, thereby helping to predict the stratum risk index of the unknown depth; the candidate hidden state is the new information of the current depth, which is obtained by adjusting the hidden state of the previous depth through the reset gate and combining it with the current input. The calculation process involved in the GRU neural network may specifically include:

[0111]

[0112] Where r d 、z d 、 h d Represents reset gate, update gate, new storage layer, and hidden layer respectively; U d represents the weight, U d The superscripts r, z, and h represent the weights corresponding to the reset gate, update gate, and new storage layer, respectively; W d represents the weight, W d The superscripts r, z, h and Ud The same reason; b d is the bias of the GRU neural network, b d The superscripts r, z, h and U d The same is true for ; sig() is the sigmoid activation function; tanh() is the tanh activation function, which is used to control data in the range of (-1, 1); represents the joint spatial-depth feature at depth d.

[0113] Secondly, the hidden layer of the GRU neural network uses the tanh activation function, with a value range of (-1, 1). This is because compared to the sigmoid activation function, the tanh activation function has an output mean of 0, making it converge faster than the sigmoid activation function and reducing the number of iterations. The tanh activation function is particularly effective when features differ significantly, and it can continuously amplify the feature effect during the loop.

[0114] Preferably, the method is further configured to update the weights and biases of the GRU neural network using a gradient descent algorithm by comparing the differences between the predicted lithologic data and the real data and calculating the loss. The real data can be provided by expert annotation or historical data. The gradient descent algorithm is an optimization algorithm used to minimize the loss function. The parameters are updated by calculating the gradient of the loss function with respect to the model parameters (weights and biases). The specific algorithm content can be referred to in related technologies and will not be detailed here.

[0115] In addition, the present invention also constructs a risk formation identification system, which mainly includes:

[0116] An acquisition module is used to acquire exploration data of the oil and gas well to be identified within a preset depth range;

[0117] An analysis module is used to obtain, based on the exploration data, a spatial-depth joint feature of the oil and gas well to be identified, which is located at a deeper level underground;

[0118] A neural network module is used to input the space-depth joint feature into a preset GRU neural network for processing to obtain lithology prediction data at a deeper level underground;

[0119] The output module is used to output the identification results of deeper underground strata based on the lithology prediction data.

[0120] In some embodiments, the acquisition module may include:

[0121] An acquisition unit, used to obtain original exploration data of the oil and gas wells to be identified;

[0122] The preprocessing unit is used to preprocess the original exploration data to obtain preprocessed exploration data, and use the preprocessed exploration data as exploration data for obtaining space-depth joint features.

[0123] In some embodiments, the analysis module may include:

[0124] A first analysis unit is configured to input the exploration data into a preset spatial distribution feature model to obtain spatial distribution features of the oil and gas wells to be identified at a deeper level underground; the spatial distribution feature model is constructed based on a Kriging model and is configured to infer the spatial relationship between known depth positions and unknown depth positions;

[0125] The second analysis unit is configured to input the exploration data into a preset deep feature model to obtain deep features of the oil and gas wells to be identified at a deeper level underground. The deep feature model is constructed based on an attention mechanism and is configured to extract deep features based on the degree of match between the input data and the prediction target within a preset depth range.

[0126] The processing unit is used to dimensionally stitch the corresponding spatial distribution features and depth features to obtain spatial-depth joint features.

[0127] It is understandable that the risk stratum identification system further includes a storage module for storing data of the spatial distribution feature model and the depth feature model.

[0128] In some embodiments, the neural network module is configured to calculate a hidden state based on the spatial-depth joint features, and convert the hidden state into lithology prediction data.

[0129] The present invention also constructs a risk formation identification device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the risk formation identification method as described above.

[0130] The present invention also constructs a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the above-mentioned risk formation identification method is implemented.

[0131] It is understandable that the above embodiments only express the preferred implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the patent scope of the present invention. It should be pointed out that for ordinary technicians in this field, without departing from the concept of the present invention, the above technical features can be freely combined, and several deformations and improvements can be made, all of which fall within the scope of protection of the present invention. Therefore, all equivalent changes and modifications made to the scope of the claims of the present invention should fall within the scope of coverage of the claims of the present invention.

Claims

1. A method for identifying risky strata, characterized in that: The following steps are involved: S1: Acquire exploration data of the oil and gas well to be identified within a preset depth range; S2: obtaining, based on the exploration data, a spatial-depth joint feature of the oil and gas well to be identified, which indicates that the well is located at a deeper level underground; S3: Inputting the space-depth joint feature into a preset GRU neural network for processing to obtain lithology prediction data at a deeper level underground; S4: Outputting identification results of deeper underground strata based on the lithologic prediction data.

2. The risk formation identification method according to claim 1, characterized in that: The step S2 includes the following sub-steps: S21: Inputting the exploration data into a preset spatial distribution feature model to obtain spatial distribution features of the oil and gas wells to be identified at a deeper level underground; the spatial distribution feature model is constructed based on a Kriging model, which is configured to infer the spatial relationship between known depth positions and unknown depth positions; S22: Inputting the exploration data into a preset depth feature model to obtain depth features of the oil and gas wells to be identified at a deeper level underground; the depth feature model is constructed based on an attention mechanism and is configured to extract depth features based on the degree of match between the input data and the prediction target within a preset depth range; S23: Dimensionally concatenate the corresponding spatial distribution features and the depth features to obtain the spatial-depth joint features.

3. The risk formation identification method according to claim 2, characterized in that: The spatial distribution characteristics Expressed as: Where, Represents the spatial distribution characteristics at depth S; are the basis functions related to the x-dimension, y-dimension, and depth at depth S, respectively; D represents the total depth; S represents the current depth; x, y, and de represent the spatial coordinate axes; The spatial coordinates are obtained through the spatial distribution feature model, and the spatial distribution feature model is specifically: Where S0 is the unobserved position, r(S0)=[R(θ,S1,S0),…,R(θ,S N ,S0)] T represents the correlation matrix of S0, R(θ, S i , S0) is the correlation function of the random process, which represents the spatial correlation with reservoir lithology, S i is the position coordinate of the i-th observation, θ is the hyperparameter of the Kriging model; β * =(F T R -1 F)F T R -1 P represents the weighted least squares estimate of β, F represents the observation matrix related to lithology, R -1 represents the covariance matrix associated with lithology, and P represents the measurement matrix of spatially related locations.

4. The risk formation identification method according to claim 2, characterized in that: The depth feature model is specifically: Where, X e Well logging depth features or logging depth features extracted based on the importance of different well logging attributes or logging attributes to the predicted data; for The normalized value of is the depth attention weight of the logging attribute or mud logging attribute within a certain depth range; v d 、w d 、u d 、b d Respectively represent the weights and biases that need to be trained through the multi-layer perceptron network; h and z represent all hidden layer states and update gate information of the GRU neural network framework; Represents the value of the mth attribute at the well depth d; sim(x m , Y) represents the similarity between the historical information of the predicted target and the input attributes.

5. The risk formation identification method according to claim 2, characterized in that: The spatial-depth joint feature is expressed as: The step S3 is specifically as follows: The space-depth joint feature is input into the GRU neural network; the hidden layer of the GRU neural network calculates a hidden state according to the space-depth joint feature, and the output layer of the GRU neural network converts the hidden state into the lithology prediction data.

6. The risk formation identification method according to claim 1, characterized in that: The computational process involved in the GRU neural network includes: Where r d 、z d 、 h d Represents reset gate, update gate, new storage layer, and hidden layer respectively; U d 、W d is the weight, b d is the bias of the GRU neural network, U d 、W d and b d Here r, z, and h represent the weights corresponding to the reset gate, update gate, and new storage layer, respectively; sig() is the sigmoid activation function; and tanh() is the tanh activation function, which is used to control data in the range of (-1, 1).

7. The risk formation identification method according to claim 1, characterized in that: The step S1 includes the following sub-steps: S11: obtaining original exploration data of the oil and gas wells to be identified; S12: Preprocessing the original exploration data to obtain preprocessed exploration data; S13: Using the pre-processed exploration data as exploration data for obtaining space-depth joint features.

8. A risk formation identification system, characterized in that: include: An acquisition module is used to acquire exploration data of the oil and gas well to be identified within a preset depth range; An analysis module is used to obtain, based on the exploration data, a spatial-depth joint feature of the oil and gas well to be identified, which is located at a deeper level underground; A neural network module is used to input the space-depth joint feature into a preset GRU neural network for processing to obtain lithology prediction data at a deeper level underground; The output module is used to output the identification result of the deeper underground strata based on the lithology prediction data.

9. A risk formation identification device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the risk formation identification method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the risk formation identification method described in any one of claims 1 to 7 is implemented.

Images