Missing well log curve prediction method and device

Abstract

The application discloses a kind of missing well logging curve prediction method and device, the method includes: obtaining the well logging curve of study area, horizon data, well site data and geological information;According to the correlation between the preset to be predicted curve of well logging curve complete well section of study area and other well logging curves, the combination of well logging curves for optimizing the establishment of machine learning network model;According to the well logging curve of study area, horizon data, well site data and geological information, determine the weight of sample well data for machine learning model training;Machine learning network model is constructed;Sample well data is used to train and verify machine learning network model;According to the known well logging curve of well to be processed in study area, use the trained and verified machine machine learning network model to predict, obtain the missing well logging curve of well to be processed.The present application can improve the precision of missing curve prediction using machine learning method.

Description

Technical Field

[0001] This invention relates to the field of well logging curve prediction technology for complex lithological reservoirs such as carbonate rocks, volcanic rocks, and shale, and particularly to a method and device for predicting missing well logging curves. Background Technology

[0002] This section is intended to provide background or context for the embodiments of the invention set forth in the claims. The description herein is not an admission that it is prior art simply because it is included in this section.

[0003] Well logging is a crucial technical means for oil and gas exploration and development. By processing and analyzing the downhole acoustic, radiometric, and electrical rock geophysical characteristic curves obtained from logging, qualitative identification of oil and gas reservoirs and quantitative calculation of parameters can be achieved, providing key data support for comprehensive reservoir evaluation. However, due to the complexity of underground conditions and various unpredictable and unavoidable problems during the measurement process, such as wellbore enlargement and instrument malfunctions, as well as human factors such as improper logging operations and considerations of economic factors, distorted or missing logging data often occurs in some well sections in practical applications. These missing portions, or even the entire logging curve, pose a significant challenge to reservoir logging evaluation. Curve prediction is a commonly used technical method to solve these problems.

[0004] Traditional methods for predicting missing logging curves primarily rely on the inherent relationships between various logging data. For example, methods like cross-plots and multiple regression are used to determine the empirical relationship between the curve to be predicted and one or more known curves. However, due to the complexity and heterogeneity of subsurface conditions, logging data often exhibits strong nonlinear relationships, and the mapping relationships between data are extremely complex, resulting in poor practical application performance. In recent years, with the widespread application of machine learning methods in science and engineering, many researchers have suggested using data-driven methods to solve geological problems. For example, methods such as Support Vector Machines (SVM), Fuzzy Logic Models (FLM), and Artificial Neural Networks (ANN) are used for logging curve prediction. However, these methods essentially construct a point-to-point or depth sequence mapping relationship, without considering the correlation and differences between the sample data used to build the prediction model and the well to be predicted in terms of reservoir geological structure, stratigraphic lithology, etc. This contradicts practical geological analysis experience and geological principles, thus leading to low accuracy in predicting and generating logging curves. Summary of the Invention

[0005] This invention provides a method for predicting missing logging curves to improve the accuracy of predicted logging curves. The method includes:

[0006] Acquire well logging curves, stratigraphic data, well location data, and geological information for the study area, including geological structures, sedimentary facies, or lithofacies paleogeography;

[0007] Based on the correlation between the preset prediction curve of the complete well logging section in the study area and other well logging curves, the combination of well logging curves used to establish the machine learning network model is selected.

[0008] Based on the well logging curves, stratigraphic data, well location data and geological information of the study area, the weights of the sample well data used for training the machine learning model are determined. The sample wells are wells with complete well logging curves in the study area, and the sample well data are the well logging curve values ​​at different depths in the sample wells.

[0009] Build machine learning network models;

[0010] The constructed machine learning network model was trained and validated using sample well data;

[0011] Based on the known logging curves of the wells to be treated in the study area, the missing logging curves of the wells to be treated are predicted using a trained and validated machine learning network model.

[0012] This invention also provides a device for predicting missing logging curves to improve the accuracy of predicted logging curves. The device includes:

[0013] The information acquisition module is used to acquire well logging curves, stratigraphic data, well location data and geological information of the study area, including geological structures, sedimentary facies or lithofacies paleogeography;

[0014] The curve combination priority module is used to select the best combination of logging curves for building a machine learning network model based on the correlation between the preset curve to be predicted in the complete well section of the logging curve in the study area and other logging curves.

[0015] The weight determination module is used to determine the weights of sample well data for training machine learning models based on well logging curves, stratigraphic data, well location data and geological information of the study area. The sample wells are wells with complete well logging curves in the study area, and the sample well data are well logging curve values ​​at different depths in the sample wells.

[0016] The machine learning network model building module is used to build machine learning network models;

[0017] The training and validation module is used to train and validate machine learning network models using sample well data;

[0018] The prediction module is used to predict the missing logging curves of the wells to be treated in the study area based on the known logging curves of the wells to be treated, using a trained and validated machine learning network model.

[0019] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for predicting missing logging curves.

[0020] This invention also provides a computer-readable storage medium storing a computer program that performs the above-described method for predicting missing logging curves.

[0021] In this embodiment of the invention, compared with the prior art, the present invention proposes a method for weighting the sample dataset used to establish a well logging curve prediction model. Based on the well logging curves, stratigraphic data, well location data and geological information of the study area, the weights of the sample well data used for training the machine learning model are determined, thereby increasing the contribution of the training sample well data in the prediction model training and significantly improving the accuracy of predicting missing curves using machine learning methods. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0023] Figure 1 This is a flowchart of the missing logging curve prediction method in an embodiment of the present invention;

[0024] Figure 2 This is a more detailed flowchart of the missing logging curve prediction process in an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the LSTM unit structure in an embodiment of the present invention;

[0026] Figure 4 This is a diagram illustrating the data transmission process corresponding to the five-layer LSTM in this embodiment of the invention.

[0027] Figure 5 This is a schematic diagram illustrating the weight calculation based on well distance in an embodiment of the present invention;

[0028] Figure 6 This is a schematic diagram of weight calculation based on geological attributes in an embodiment of the present invention;

[0029] Figure 7 This is a diagram illustrating the effect of using the present invention for missing curve prediction in an embodiment of the present invention;

[0030] Figure 8This is a comparison chart of the prediction effects of the method of the present invention and the traditional method in an embodiment of the present invention;

[0031] Figure 9 This is a structural block diagram of the missing logging curve prediction device in an embodiment of the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0033] To achieve the above objectives, the present invention provides the following technical solutions, such as... Figure 1 and Figure 2 As shown:

[0034] Step 102: Obtain well logging curves, stratigraphic data, well location data, and geological information for the study area. The geological information includes geological structures, sedimentary facies, or lithofacies paleogeography. It may also include data such as core experimental analysis.

[0035] Specifically, logging curves may include a series of logging data such as natural gamma, gamma spectrum, resistivity, and three-porosity; well location data refers to the coordinates of each well in the work area, which can be used to determine the location of each well and the distance between wells; stratigraphic data refers to the geological stratification of each well; geological structure / sedimentary facies / lithologic paleogeography refers to the geological structure map, sedimentary facies map, and lithologic paleogeographic map of each stratum drawn through geological research in the study area; core experimental analysis data can further verify the results of this invention.

[0036] Step 104: Based on the correlation between the preset prediction curve of the complete well section of the well logging curve in the study area and other well logging curves, select the best combination of well logging curves for establishing the machine learning network model.

[0037] Specifically, commonly used logging curves include sonic transit time (AC), neutron density (CNL), density (DEN), natural gamma ray (GR), uranium-reduced gamma ray (CGR), deep resistivity (RD), shallow resistivity (RS), and photoelectric adsorption cross-sectional index (PE), denoted as... Let the curve be... i1 represents the curve order (1≤i1≤n, where n is the total number of logging curves), and j1 also represents the curve order (1≤j1≤k1, where k1 is the total number of curves to be predicted).

[0038] In a well section with complete logging curves, assume any X i and Y j The Pearson correlation coefficient If it exists, then use ρ ijAn n×k1 matrix with n elements is called the correlation matrix of the random vector of that dimension, denoted as R, and expressed as:

[0039]

[0040] in, and express and The expected value of the variable, D(X) i ) and D(Y j )express and The variance of the variable.

[0041] The correlation coefficients of corresponding rows and columns in the correlation matrix can be used to intuitively and quantitatively evaluate the correlation between each logging curve and the curve to be predicted. Generally, when the correlation coefficient is... This indicates a low correlation between the two curves; when the correlation coefficient is low... This indicates a moderate correlation between the two curves, where the correlation coefficient is... This indicates a high correlation between the two curves. By setting a correlation coefficient threshold, highly correlated logging combinations can be optimally selected for predicting each target curve. A correlation coefficient threshold of 0.3 can be set.

[0042] Step 106: Based on the logging curves, stratigraphic data, well location data and geological information of the study area, determine the weights of the sample well data used for training the machine learning model, wherein the sample well data is a preferred combination of logging curves.

[0043] Existing methods for missing curve prediction using machine learning mainly focus on machine learning techniques and network structures, without considering the correlation and differences between the sample data used to build the prediction model and the wells to be predicted in terms of reservoir geological structure, stratigraphic lithology, etc. Therefore, their applicability is limited and their prediction accuracy is low. This invention addresses this problem by proposing a method for variable training sample weights. Optionally, one or more of the following methods can be used to determine the weights of each data point in each sample well used for training the machine learning model.

[0044] The sample wells mentioned below refer to typical wells among the wells with complete logging curves in the study area, while the sample well data refer to the logging curve values ​​at different depths in the sample wells.

[0045] Wells awaiting processing refer to wells with missing logging curves.

[0046] 1) Determine the weight of each data point in the sample well by using the distance between the well to be processed and the sample well.

[0047] In a typical oilfield block, the distance between wells generally reflects the similarity between them, especially in oilfields with relatively underdeveloped structures and stable sediments. The smaller the distance between two wells, the smaller the differences in lithology, physical properties, and hydrocarbon potential within the same formation. Therefore, when predicting missing curves, sample data from adjacent wells of the well to be treated can be given a higher weight, while sample data from wells farther away from the well to be treated can be given a lower weight. Let the coordinates of well A (x0, y0) be (x0, y0), and the coordinates of sample well B (x0, y0) be (x0, y ... b ,y b If the weight coefficients of the sample data points of well B are uniformly set to:

[0048]

[0049] Among them, w db The weighting coefficients for the data points in well B are 0 ≤ w db ≤1; a and b are the weight calculation coefficients for the work area, generally set a=1 and b=0; L m This represents the maximum well spacing in the work area.

[0050] The weight coefficients w of all sample wells involved in training are calculated. d Then, the weights of all sample data points corresponding to each sample well can be set to w. d .

[0051] 2) Determine the weight of each data point in the sample well by using the geological attributes of the well to be processed and the sample well.

[0052] In oilfield blocks with stable sedimentary formations and minimal structural changes, the distance between wells can usually represent their similarity. However, in blocks with significant sedimentary variations and complex tectonic movements, even wells very close to each other may exhibit significant changes in their stratigraphic characteristics and logging response. Their similarity is also influenced by sedimentary facies distribution and regional tectonics. Therefore, when predicting missing curves, sample data with similar geological attributes to the well in question can be assigned a higher weight, while sample data with significantly different geological attributes can be assigned a lower weight.

[0053] Suppose that the sedimentary facies of a certain stratum in the block can be divided into N categories. First, select several typical wells in different sedimentary zones within the work area. The number of typical wells included in each sedimentary facies zone is denoted as num. i(Where i represents the sedimentary facies zone number). The sedimentary facies zone where the well to be treated is located is designated as the reference zone. The average similarity of the preset target curves corresponding to typical wells in other sedimentary facies zones and typical wells in the reference zone is compared. The average similarity of the preset target curves is determined by selecting typical wells in each sedimentary facies zone, as the preset target curves are already present in these typical wells. The similarity coefficient of this curve in these wells is used to determine the weight. The similarity between two curves can be evaluated by calculating the cumulative distance γ using the Dynamic Time Warping (DTW) method. The smaller the cumulative distance γ, the higher the similarity between the two curves.

[0054] Let s be the similarity of the preset target curves between wells. Here, the reciprocal of the power function with base e is taken as the similarity evaluation index for the cumulative distance γ:

[0055] s = e -γ ;

[0056] The average similarity between a certain sedimentary facies zone and the reference area is:

[0057]

[0058] Where, num refer Indicates the number of typical wells in the reference area; num i This indicates the number of typical wells in a given sedimentary facies zone; s lk This indicates the similarity between the l-th well in the reference area and the k-th well in a certain sedimentary facies zone; This represents the average similarity between a sedimentary facies zone and a reference area; i represents the sedimentary facies zone number; l and k are well numbers. Where s lk According to the formula s = e -γ calculate.

[0059] Therefore, the weighting coefficient can be set according to the sedimentary facies region to which a certain layer of the sample well belongs. If the sample well and the well to be treated belong to the same sedimentary facies region, then w f =1. When calculating the weight coefficient w of a certain layer in all samples participating in training... f Then, the weights of all sample data points corresponding to a certain layer in each sample well can be set to w. f .

[0060] The geological structure or lithofacies paleogeography is similar, and the average similarity between typical wells in the reference area and typical wells in other blocks is determined based on the block where the well to be treated is located.

[0061] 3) Determine the weight of each data point in the sample well by using the curve similarity between the well to be processed and the sample well.

[0062] In addition to determining the weight of each data point in the sample well based on the distance between the well to be processed and the geological attribute relationship between the well to be processed and the sample well, the similarity of the corresponding input curve combinations of the well to be processed and the sample well can also be used.

[0063] Let X be the combination of well logging curves used to establish the machine learning model as determined in step 104. i′ Let X be the combination of input logging curves for a certain sample well A. Ai′ The combination of logging curves corresponding to the well to be treated is denoted as X. 0i′ Referring to the above, the similarity of the corresponding curve in a certain segment is calculated using the dynamic time warping method, denoted as s. i′ Then the weighting coefficient of this section in well A is:

[0064]

[0065] Where, ρ i′ The specified weight of the i′-th curve is given based on the correlation coefficient between each logging curve and the curve to be predicted in step 104; s i′ Let w represent the similarity between the i′-th curve of well A and the corresponding curve of the well to be processed; As The weighting coefficient for well A in this layer is determined; N is the total number of input logging curve combinations.

[0066] The weight coefficients w of all sample wells involved in training are calculated. s Then, the weights of all sample data points in the corresponding layers of all sample wells can be set to w. s .

[0067] 4) Determine the weight of each sample well data point by combining the weights calculated using the methods described above.

[0068] Let w be the weighting coefficient determined by the distance between the well to be processed and the sample well for a certain well. d The weighting coefficient w is determined by the geological attributes of the well to be treated and the sample well. f The weighting coefficient w is determined by the curve similarity between the well to be processed and the sample well. s The final weighting coefficient w for the sample data points corresponding to the well can be determined based on a combination of weighting coefficients determined by one or more methods. c :

[0069] w c =(w d +w f +w s ) / 3.

[0070] The above formula assumes all three options are selected. If only two options are selected, the corresponding weights are added together and then divided by 2.

[0071] Step 108: Build a machine learning network model.

[0072] Specifically, commonly used machine learning models include linear regression, support vector machines (SVM), and multilayer neural networks (MLP), whose model structure can be represented as an input layer, hidden layers, and an output layer. The input layer is responsible for receiving signals, the hidden layers are responsible for decomposing and processing the data, and the final results are integrated into the output layer.

[0073] Optionally, this invention uses a Long Short-Term Memory (LSTM) network to establish a curve prediction model. LSTM is an improved recurrent neural network that can solve the problem of traditional recurrent neural networks being unable to handle long-distance dependencies, making it suitable for well logging curve prediction problems with depth sequence significance.

[0074] The model uses a 7-layer network structure, and the LSTM model is built using TensorFlow, where:

[0075] The first layer is the input layer, used for inputting sample point data. The data structure is a three-dimensional tensor of [Batch_size, Sequence_lenth, Data_dim]. Here, Batch_size is the size of the input data blocks, equal to the total number of sample points. Sequence_lenth is the length of the sequence information contained in each data block, its size related to the sampling interval of the input logging curves, typically between 10 and 80. The larger the sampling interval of the input logging curves, the smaller Sequence_lenth. Data_dim is the curve dimension of the data points, i.e., the number of curves included in the logging curve combination determined in step 2. Thus, each data block actually contains a Sequence_lenth × Data_dim data volume, and each data block corresponds to a target output value Y_output and a target weight Weight_Y for the curve to be predicted.

[0076] Layers 2 through 6 of the network are standard LSTM layers, which will be further abstracted based on the principle of stacked recurrent neural networks. Each LSTM unit is as follows: Figure 3 As shown. Figure 4 It is the data information transmission process corresponding to the five-layer LSTM.

[0077] The LSTM unit has several internal thresholds, and the working process of each gate can be expressed as follows:

[0078] Unit input formula: z t =g(W z ×x t +R z y t-1 +b z );

[0079] Input gate formula: i t =σ(W i ×x t +R i y t-1 +p i ec t-1 +b i );

[0080] Forget gate formula: f t =σ(W f ×x t +R f y t-1 +p f ec t-1 +b f );

[0081] Neuron cell state formula: c t =i t ez t +f t ec t-1 ;

[0082] Output gate formula: o t =σ(W o x t +R o y t-1 +p o ec t +b o );

[0083] Unit output formula: y t =o t eh(c t ).

[0084] Among them, z t Indicates the input module of the LSTM unit, x t W represents the feature data input at time t. z This represents the parameter matrix between the input data and the input module, y t-1 and y t The table shows the outputs at times t-1 and t (representing the hidden states in practice), g is the activation function of the input module of the LSTM unit, and R... z R i R f R o The weights of the hidden units represent the utilization rate of the prediction results from the previous depth point; b z b i b f b o Both represent bias parameters, p ip f p o c t-1 Let c be the vector value of the memory unit at time t-1. t For neuronal cell states, i t f t o t Let W be the activation vector values ​​of the input gate, forget gate, and output gate of a node in an LSTM neural network at time t. i W represents the parameter matrix between the input module and the hidden layer cell unit. f represents the parameter matrix between the input data and the forgetting module cell unit, and e represents the dot product between the two vectors.

[0085] The seventh layer of the network is the output layer. Each unit in the hidden layer and the output layer has many inputs, which are connected to the next layer of neurons until the output layer calculates the network's prediction.

[0086] Step 110: Train and validate the constructed machine learning network model using sample well data.

[0087] Specifically, due to differences in logging principles and physical response characteristics, different logging curves have different dimensions and response values. Therefore, the input curves used in the model computation should be normalized before training, testing, and applying the machine learning model. The target normalization interval is [0,1], and the conversion method is shown in the following formula:

[0088]

[0089] Where, x i This represents the value of a curve at a certain depth point before normalization; x i * This represents the normalized value; x max x min These represent the maximum and minimum values ​​of the response across the entire well section via the logging curve, respectively.

[0090] After data normalization, the machine learning model is trained and validated using sample data from complete well logging sections. The model training process essentially involves using known data of a target curve and other well logging data to determine the parameters of the machine learning model (i.e., establishing a correspondence between the target curve data and other well logging data), thereby enabling the prediction of reasonable curve results using other well logging curves even when the target curve data is missing.

[0091] During model training, a certain percentage η of the training data can be randomly extracted to verify the convergence of the model (generally η = 0.2). For neural network models, the backpropagation algorithm is usually used for gradient updates to obtain the weights and bias parameters in the hidden and output layers of the model, minimizing the difference between the measured target curve and the model's predicted curve. The objective function can be expressed as:

[0092]

[0093] Among them, Loss(Y) pred ,Y data Y represents the cost function for the difference between the actually measured target curve and the curve predicted by the model; pred Y represents the curve value predicted by the model. data w represents the actual measured target curve value. t This represents the weight coefficient of each sample point determined according to step 106.

[0094] Typically, if Loss(Y) pred ,Y data If the values ​​are below a certain threshold and do not decrease with increasing training iterations, then the model is considered to have converged. The weights and bias parameters obtained at this point can be considered as the optimal parameters obtained from model training.

[0095] In the model validation phase, the target curve predicted by the model is obtained using the model weights and bias parameters trained above. Then, the difference between the target curve and the actual well logging curve is calculated. When the difference is within an acceptable range and the trend is similar to that in the training phase, the model parameters are considered stable and no overfitting has occurred. The model can then be used for subsequent prediction and processing of actual data.

[0096] Step 112: Based on the known logging curves of the wells to be treated in the study area, use the trained and validated machine learning network model to predict the missing logging curves and obtain the missing logging curves of the wells to be treated.

[0097] Specifically, in the well section to be processed, the trained machine learning model is used to obtain the predicted curve results. The operation is similar to the model verification stage in step 110. At this point, other well logging curve values ​​are known and input into the trained machine learning model to obtain the corresponding target curve prediction results. It is important to note that the well logging curve data input into the machine learning model also requires the normalization processing described in step 110. The predicted results output by the model need to undergo inverse normalization processing to restore the predicted curve to the normal target well logging response distribution range. The inverse normalization processing formula is as follows:

[0098]

[0099] in, This refers to the prediction result of a target curve at a certain depth point based on the output of a machine learning model; y max y min The maximum and minimum response values ​​of the logging curve given in step 110; The predicted curve value is the value that is restored to the normal well logging response distribution range after inverse normalization.

[0100] Example

[0101] 1. Taking the prediction of shear wave transit time using conventional logging curves in a shale oil reservoir in a block of Daqing Oilfield as an example, the conventional logging curves already available in this block include acoustic longitudinal wave transit time (AC), neutron (CNL), density (DEN), natural gamma (GR), deep resistivity (RD), shallow resistivity (RS), etc. The curve to be predicted is acoustic shear wave transit time (DTS).

[0102] 2. Based on the six conventional logging curves for this well section—acoustic longitudinal wave transit time (AC), neutron (CNL), density (DEN), natural gamma (GR), deep resistivity (RD), and shallow resistivity (RS)—and the acoustic transverse wave transit time (DTS) obtained from well processing of array acoustic logging data, the correlation coefficient matrix of each curve is calculated according to the formula described in step 104, as shown in Table 1 below:

[0103] Table 1

[0104] Correlation coefficient DTS AC 0.79 CNL 0.78 GR 0.4 RS -0.38 RD -0.42 DEN -0.52

[0105] According to the description in step 104, the correlation coefficient threshold is set to 0.3. As can be seen from the table, the absolute values ​​of the correlation coefficients between the six curves (AC, CNL, GR, deep resistivity (RD), shallow resistivity (RS), and density (DEN) and the acoustic shear wave transit time (DTS) are 0.79, 0.78, 0.4, 0.38, 0.42, and 0.52, respectively, all greater than the correlation coefficient threshold of 0.3. This indicates that the six curves (AC, CNL, GR, RD, RS, and DEN) and the acoustic shear wave transit time (DTS) have good correlations. Therefore, the preferred conventional logging combination for predicting the acoustic shear wave transit time (DTS) is the six curves: AC, CNL, GR, RD, RS, and DEN.

[0106] 3. Determine the weights of each data point in each sample well used for training the machine learning model.

[0107] 1) Determine the weight of each data point in the sample well by using the distance between the well to be processed and the sample well.

[0108] like Figure 5 The diagram shown illustrates the weighting calculation based on well distance. Figure 5 Well A is the target well, with coordinates (18551675, 3349782). Wells B, C, and D are sample wells, with coordinates (18555039, 3351527), (18550893, 3358441), and (18541819, 3354836), respectively. We set a = 1, b = 0, and define the maximum distance L in the work area. m If the distance is 20km, then according to the formula in step 106, we can obtain:

[0109]

[0110] w dc =0.565;

[0111] w dd =0.446;

[0112] Therefore, the weighting coefficients for all sample data points in sample well B can be set to 0.811, the weighting coefficients for all sample data points in sample well C to 0.565, and the weighting coefficients for all sample data points in sample well D to 0.446.

[0113] 2) Determine the weight of each data point in the sample well by using the geological attributes of the well to be processed and the sample well.

[0114] like Figure 6 The diagram shows a sedimentary facies map of a certain area. The sedimentary facies of a certain stratum in this area can be divided into four main categories: shallow lacustrine, littoral lacustrine, interdistributary bay, and basaltic river. Assuming well A to be processed is located in the basaltic river facies zone, and this zone is designated as the reference zone, typical wells A1, A2, and A3 are selected within this zone; typical wells B1, B2, and B3 are selected in the shallow lacustrine facies zone; typical wells C1 and C2 are selected in the littoral lacustrine facies zone; and typical well D1 is selected in the interdistributary bay facies zone. The similarity of the shear wave transit time (DTS) curves of wells B1, B2, B3, C1, C2, and D1 with those of wells A1, A2, and A3 in this stratum is compared. The similarity between two curves can be evaluated by calculating the cumulative distance γ using the Dynamic Time Warping (DTW) method. Let s denote the similarity between wells in a certain stratum. Here, the reciprocal of a power function to the base e is used as the similarity evaluation index for the cumulative distance γ. Taking the similarity calculation of the shear wave time difference (DTS) curves of wells B1 and A1 in a certain formation as an example, assuming the cumulative distance γ calculated through dynamic time warping is 0.89, the similarity between the two curves is:

[0115] s b1a1 =e -0.89 =0.41;

[0116] Similarly, the similarity between B1, B2, B3 and A1, A2, A3 can be calculated as follows:

[0117]

[0118] The similarity assessment between the shallow lacustrine facies zone and the reference zone is as follows:

[0119]

[0120] Similarly, the similarity between the lakeside facies zone and the reference zone can be determined as follows: The similarity between the interdiversion bay and the reference zone is:

[0121] Therefore, the weights of the sample data points of the sample well in this layer segment can be set, such as... Figure 6 Well B, located in the shallow lacustrine facies zone, can be assigned a weighting factor of 0.43. The weighting factor for the sample data points in this layer of Well C is 0.67, and the weighting factor for the sample data points in this layer of Well D is 0.52.

[0122] 3) Determine the weight of each data point in the sample well by using the curve similarity between the well to be processed and the sample well.

[0123] Taking a block in Daqing Oilfield as an example, the input logging curve combination for establishing the machine learning model is the sonic P-wave transit time (AC), neutron (CNL), natural gamma (GR), deep resistivity (RD), shallow resistivity (RS), and density (DEN). Referring to the above, the similarity s between the curves of well B and the well to be processed in a certain layer is calculated using the dynamic time warping method. AC =0.83, s CNL =0.66, s GR =0.81, s RD =0.76, s RS =0.75, s DEN =0.61, then the weighting coefficient of this section of well B is:

[0124] w Bs = (0.79×0.83+0.78×0.66+0.4×0.81+0.38×0.76+0.42×0.75+0.52×0.61) / 6 = 0.43

[0125] Similarly, the weights of data points can be determined by calculating the similarity between the curves of other sample wells and the well to be processed in a certain layer.

[0126] 4) Determine the weight of each sample well data point by combining the weights calculated using the methods described above.

[0127] After calculating the weighting coefficients of the sample well data using the methods described above, the final weighting coefficients for the corresponding sample data points of the well can also be determined based on a combination of weighting coefficients determined by one or more methods. For example, well B has a weighting coefficient of 0.81 determined by the distance between the well and the sample well, a weighting coefficient of 0.43 determined by the geological attributes of the well to be processed and the sample well, and a weighting coefficient of 0.43 determined by the curve similarity between the well to be processed and the sample well. Therefore, the final weighting coefficients for the corresponding sample data points of this well are:

[0128] w cb = (0.81 + 0.43 + 0.43) / 3 = 0.56.

[0129] 4. Build a machine learning network model

[0130] This invention employs a Long Short-Term Memory (LSTM) network to establish a curve prediction model. The model uses a 7-layer network structure and is built using TensorFlow. The model hyperparameters are set as follows:

[0131] Input the number of logging curves: data_dim = 6;

[0132] The sequence length seq_length = 20;

[0133] The number of neurons in each hidden layer of the network is hidden_dim = 49;

[0134] Output curve dimension output_dim = 1;

[0135] The number of LSTM layers, n_layers, is 5.

[0136] Dropout rate: dropout_rate = 0.2;

[0137] The learning rate is 0.005.

[0138] Batch size BATCH_SIZE = 640;

[0139] The total number of data iteration rounds is EPOCHS = 30.

[0140] 5. Use sample well data to train and validate machine learning models.

[0141] 6. In the well section to be processed, the predicted curve results are obtained using the machine learning model trained in step 108.

[0142] Using the machine learning model trained above, the predicted curve result is obtained in the well section to be processed. The operation is similar to the model verification stage in step 110. At this time, other well logging curve values ​​are known and input into the trained machine learning model to obtain the corresponding target curve prediction result.

[0143] Figure 7 This paper compares the prediction results of the present invention with those of traditional methods by setting sample data point weights. Both methods are based on Long Short-Term Memory (LSTM) networks. As described in step 108, the hyperparameters of the network models remain consistent, ensuring the comparability of the prediction results. In the figure, the first line is the wellbore diameter curve, the second is the natural gamma curve, the third is the depth curve, the fourth is the formation level, the fifth is the deep and shallow resistivity curves, the sixth is the density curve, the seventh is the compensated neutron curve, the eighth is the P-wave transit time curve, the ninth is the shear wave transit time curve measured by array acoustic logging, the tenth is a comparison between the shear wave transit time curve predicted by the traditional method with equal sample data point weights and the measured shear wave transit time curve; the tenth line is a comparison between the shear wave transit time curve predicted by the present invention after averaging the weights of the three methods as described in step 106 (i.e., method 4 in step 106) and the measured shear wave transit time curve. As can be seen from the figure, the shear wave time difference curve predicted by the method of the present invention is significantly more consistent with the measured shear wave time difference curve in terms of curve change trend and curve amplitude, and the prediction accuracy is higher.

[0144] Figure 8 This paper compares the prediction results of the method of this invention with those of traditional methods. The left figure shows a comparison between the predicted shear wave time difference curve and the measured shear wave time difference curve using the traditional method with equal weights of sample data points. The right figure shows a comparison between the predicted shear wave time difference curve and the measured shear wave time difference curve using the method described in step 106 (i.e., method 4 in step 106) with the sample data point weights set according to the method of this invention. The comparison shows that the data points of the predicted results and the measured shear wave time difference curves of the traditional method are scattered, with a correlation coefficient of about 0.37. In contrast, the data points of the predicted results and the measured shear wave time difference curves of the method of this invention are more concentrated, with a correlation coefficient of about 0.50, indicating higher accuracy.

[0145] This invention also provides a missing logging curve prediction device, as described in the following embodiments. Since the principle by which this device solves the problem is similar to that of the missing logging curve prediction method, the implementation of this device can refer to the implementation of the missing logging curve prediction method; repeated details will not be elaborated further.

[0146] Figure 9 This is a structural block diagram of the missing well logging curve prediction device in an embodiment of the present invention, such as... Figure 9 As shown, the device includes:

[0147] Information acquisition module 02 is used to acquire well logging curves, stratigraphic data, well location data and geological information of the study area, wherein the geological information includes geological structure or sedimentary facies or lithofacies paleogeography;

[0148] The curve combination optimization module 04 is used to optimize the combination of logging curves used to establish a machine learning network model based on the correlation between the preset curve to be predicted in the complete well section of the logging curve in the study area and other logging curves.

[0149] The weight determination module 06 is used to determine the weight of the sample well data used for training the machine learning model based on the logging curves, stratigraphic data, well location data and geological information of the study area. The sample wells are wells with complete logging curves in the study area, and the sample well data are logging curve values ​​at different depths in the sample wells.

[0150] Machine Learning Network Model Building Module 08 is used to build machine learning network models;

[0151] Training and validation module 10 is used to train and validate machine learning network models using sample well data;

[0152] The prediction module 12 is used to predict the missing logging curves of the wells to be treated based on the known logging curves of the wells to be treated in the study area, using a trained and validated machine learning network model.

[0153] In this embodiment of the invention, the curve combination optimization module is specifically used for:

[0154] Calculate the correlation between the preset prediction curve and other logging curves in the well section with complete logging curves in the study area;

[0155] The correlation is compared with a preset correlation coefficient threshold. When the correlation exceeds the preset correlation coefficient threshold, the corresponding logging curves are selected to form a logging curve combination.

[0156] In this embodiment of the invention, the weight determination module is specifically used for:

[0157] The weights of the sample well data used for training machine learning models are determined using one or more combinations of the following:

[0158] Based on well location data, determine the weights of the sample well data used for training the machine learning model;

[0159] Alternatively, based on the geological information, well logging curves, and stratigraphic data of the study area, determine the weights of the sample well data used for training the machine learning model;

[0160] Alternatively, the weights of sample well data used for training machine learning models can be determined based on well logging curves and stratigraphic data of the study area.

[0161] In this embodiment of the invention, the weight determination module is specifically used for:

[0162] The weights of the sample well data used for training the machine learning model are determined based on well location data and stratigraphic data as follows:

[0163] Based on the well location data of the well to be processed and the well location data of the sample well, the distance between the well to be processed and the sample well is calculated, and the weight of the sample well data used for training the machine learning model is determined based on the distance.

[0164] In this embodiment of the invention, the weight determination module is specifically used for:

[0165] The weights of the sample well data used for training the machine learning model are determined according to the distance using the following formula:

[0166]

[0167] Among them, w db These are the weighting coefficients for the sample well data, 0 ≤ w db ≤1; a and b are the weight calculation coefficients for the work area, generally set a=1 and b=0; L m The maximum well spacing in the work area; (x0, y0) are the well location coordinates of the well to be processed, (x b ,y b () represents the well location coordinates of the sample well.

[0168] In this embodiment of the invention, the weight determination module is specifically used for:

[0169] The weights of the sample well data used for training the machine learning model are determined based on the geological information, well logging curves, and stratigraphic data of the study area as follows:

[0170] The study area was divided into multiple blocks based on geological information.

[0171] Select multiple typical wells in each block;

[0172] The block containing the well to be processed is used as a reference area to determine the average similarity between typical wells in the reference area and typical wells in other blocks;

[0173] The average similarity is used as the weight of the sample well data used for training the machine learning model.

[0174] In this embodiment of the invention, the weight determination module is specifically used for:

[0175] The block containing the well to be processed is used as a reference area, and the average similarity between typical wells in the reference area and typical wells in other blocks is determined as follows:

[0176]

[0177] Where, num refer Indicates the number of typical wells in the reference area; num i s represents the number of typical wells in the i-th sedimentary facies zone; lk This indicates the similarity between the l-th well in the reference area and the k-th well in the i-th sedimentary facies zone; This represents the average similarity between the i-th sedimentary facies zone and the reference area; i represents the number of the sedimentary facies zone; l and k are the well numbers.

[0178] In this embodiment of the invention, the weight determination module is specifically used for:

[0179] The weights of the sample well data used for training the machine learning model are determined based on the well logging curves and stratigraphic data of the study area as follows:

[0180] Determine the optimal logging curve combination for the sample wells and the optimal logging curve combination for the wells to be treated;

[0181] Based on the optimal combination of logging curves from the sample wells and the optimal combination of logging curves from the wells to be processed, the similarity of logging curves corresponding to each layer is determined.

[0182] The weights of the sample well data used for training the machine learning model are determined based on the similarity.

[0183] In this embodiment of the invention, the weight determination module is specifically used for:

[0184] The weights of the sample well data used for training the machine learning model are determined based on the similarity using the following formula:

[0185]

[0186] Where, ρ i′ s represents the specified weight of the i′-th curve; i′ The similarity between the i′-th curve of a certain layer in the sample well and the corresponding curve of the well to be processed; w As is the weighting coefficient of the sample well in the corresponding layer; N is the total number of curves in the well logging curve combination.

[0187] This invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the above-described method for predicting missing logging curves.

[0188] This invention also provides a computer-readable storage medium storing a computer program that performs the above-described method for predicting missing logging curves.

[0189] In this embodiment of the invention, compared with previous processing methods, the well logging curve prediction method proposed in this invention can perform weighted processing on the sample dataset used to establish the well logging curve prediction model. Accordingly, the weights of data points from different wells and different layers in the sample dataset can be set according to well location distance, reservoir geological model, curve morphology characteristics, etc., during model training. The data of sample wells that are close to the well to be processed, have similar geological characteristics, and similar curve morphology are given higher weights, thereby increasing the contribution of these sample well data to the prediction model training and significantly improving the accuracy of predicting missing curves using machine learning methods.

[0190] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0191] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0192] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0193] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0194] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting missing well logging curves, characterized in that, include: Acquire well logging curves, stratigraphic data, well location data, and geological information for the study area, including geological structures, sedimentary facies, or lithofacies paleogeography; Based on the correlation between the preset prediction curve of the complete well logging section in the study area and other well logging curves, the combination of well logging curves used to establish the machine learning network model is selected. Based on the well logging curves, stratigraphic data, well location data and geological information of the study area, the weights of the sample well data used for training the machine learning model are determined. The sample wells are wells with complete well logging curves in the study area, and the sample well data are the well logging curve values ​​at different depths in the sample wells. Build machine learning network models; The constructed machine learning network model was trained and validated using sample well data; Based on the known logging curves of the wells to be treated in the study area, the missing logging curves of the wells to be treated are predicted using a trained and validated machine learning network model. Based on well logging curves, stratigraphic data, well location data, and geological information of the study area, the weights of the sample well data used for training the machine learning model are determined, including: using one or more combinations of the following to determine the weights of the sample well data used for training the machine learning model: Based on the well location data of the well to be processed and the well location data of the sample well, the distance between the well to be processed and the sample well is calculated, and the weight of the sample well data used for training the machine learning model is determined based on the distance. The study area is divided into multiple blocks based on geological information; multiple typical wells are selected in each block; the block where the well to be processed is located is used as a reference area, and the average similarity between typical wells in the reference area and typical wells in other blocks is determined; the average similarity is used as the weight of the sample well data for training the machine learning model. Alternatively, determine the preferred combination of logging curves for the sample wells and the preferred combination of logging curves for the wells to be processed; based on the preferred combination of logging curves for the sample wells and the preferred combination of logging curves for the wells to be processed, determine the similarity of the logging curves corresponding to each layer; and based on the similarity, determine the weights of the sample well data used for training the machine learning model. The objective function of the machine learning network model is as follows: Among them, Loss(Y) pred ,Y data Y represents the cost function representing the difference between the actually measured target curve and the curve predicted by the machine learning network model; pred Y represents the curve value predicted by the model. data w represents the actual measured target curve value. t represents the weighting coefficient of the t-th sample well data, and n represents the number of sample wells.

2. The method for predicting missing logging curves as described in claim 1, characterized in that, Based on the correlation between the preset prediction curve of the complete well logging section in the study area and other well logging curves, the optimal combination of well logging curves for establishing the machine learning network model is selected, including: Calculate the correlation between the preset prediction curve and other logging curves in the well section with complete logging curves in the study area; The correlation is compared with a preset correlation coefficient threshold. When the correlation exceeds the preset correlation coefficient threshold, the corresponding logging curves are selected to form a logging curve combination.

3. The method for predicting missing logging curves as described in claim 1, characterized in that, The weights of the sample well data used for training the machine learning model are determined according to the distance using the following formula: Among them, w db These are the weighting coefficients for the sample well data, 0 ≤ w db ≤1; a and b are the weight calculation coefficients for the work area, set a=1 and b=0; L m The maximum well spacing in the work area; (x0, y0) are the well location coordinates of the well to be processed, (x b ,y b () represents the well location coordinates of the sample well.

4. The method for predicting missing logging curves as described in claim 1, characterized in that, The block containing the well to be processed is used as a reference area, and the average similarity between typical wells in the reference area and typical wells in other blocks is determined as follows: Where, num refer Indicates the number of typical wells in the reference area; num i Indicates the number of typical wells in the i-th sedimentary facies zone; s lk This indicates the similarity between the l-th well in the reference area and the k-th well in the i-th sedimentary facies zone; This represents the average similarity between the i-th sedimentary facies zone and the reference area; i represents the number of the sedimentary facies zone; l and k are the well numbers.

5. The method for predicting missing logging curves as described in claim 1, characterized in that, The weights of the sample well data used for training the machine learning model are determined based on the similarity using the following formula: Where, ρ i′ s represents the specified weight of the i′-th curve; i′ The similarity between the i′-th curve of a certain layer in the sample well and the corresponding curve of the well to be processed; w As is the weighting coefficient of the sample well in the corresponding layer; N is the total number of curves in the well logging curve combination.

6. A device for predicting missing logging curves, characterized in that, include: The information acquisition module is used to acquire well logging curves, stratigraphic data, well location data and geological information of the study area, including geological structures, sedimentary facies or lithofacies paleogeography; The curve combination optimization module is used to optimize the well logging curve combination for building a machine learning network model based on the correlation between the preset curve to be predicted in the complete well section of the well logging curve in the study area and other well logging curves. The weight determination module is used to determine the weights of sample well data for training machine learning models based on well logging curves, stratigraphic data, well location data and geological information of the study area. The sample wells are wells with complete well logging curves in the study area, and the sample well data are well logging curve values ​​at different depths in the sample wells. The machine learning network model building module is used to build machine learning network models; The training and validation module is used to train and validate machine learning network models using sample well data; The prediction module is used to predict the missing logging curves of the wells to be treated based on the known logging curves of the wells to be treated in the study area, using a trained and validated machine learning network model. The weight determination module is specifically used for: The weights of the sample well data used for training machine learning models are determined using one or more combinations of the following: The weights of the sample well data used for training the machine learning model are determined based on the well location data as follows: the distance between the well to be processed and the sample well is calculated based on the well location data of the well to be processed and the sample well, and the weights of the sample well data used for training the machine learning model are determined based on the distance. The weights of the sample well data used for training the machine learning model were determined based on the geological information, well logging curves, and stratigraphic data of the study area as follows: the study area was divided into multiple blocks based on the geological information; multiple typical wells were selected in each block; the block where the well to be processed was located was used as a reference area, and the average similarity between typical wells in the reference area and typical wells in other blocks was determined. The average similarity is used as the weight of the sample well data for training the machine learning model; Alternatively, the weights of sample well data used for training the machine learning model can be determined based on the well logging curves and stratigraphic data of the study area as follows: determine the preferred combination of well logging curves for the sample wells and the preferred combination of well logging curves for the wells to be processed; determine the similarity of the well logging curves corresponding to each stratigraphic layer based on the preferred combination of well logging curves for the sample wells and the preferred combination of well logging curves for the wells to be processed; and determine the weights of the sample well data used for training the machine learning model based on the similarity. The objective function of the machine learning network model is as follows: Among them, Loss(Y) pred ,Y data Y represents the cost function representing the difference between the actually measured target curve and the curve predicted by the machine learning network model; pred Y represents the curve value predicted by the model. data w represents the actual measured target curve value. t represents the weighting coefficient of the t-th sample well data, and n represents the number of sample wells.

7. The missing logging curve prediction device as described in claim 6, characterized in that, The curve combination optimization module is specifically used for: Calculate the correlation between the preset prediction curve and other logging curves in the well section with complete logging curves in the study area; The correlation is compared with a preset correlation coefficient threshold. When the correlation exceeds the preset correlation coefficient threshold, the corresponding logging curves are selected to form a logging curve combination.

8. The missing logging curve prediction device as described in claim 6, characterized in that, The weight determination module is specifically used for: The weights of the sample well data used for training the machine learning model are determined according to the distance using the following formula: Among them, w db These are the weighting coefficients for the sample well data, 0 ≤ w db ≤1; a and b are the weight calculation coefficients for the work area, set a=1 and b=0; L m The maximum well spacing in the work area; (x0, y0) are the well location coordinates of the well to be processed, (x b ,y b () represents the well location coordinates of the sample well.

9. The missing logging curve prediction device as described in claim 6, characterized in that, The weight determination module is specifically used for: The block containing the well to be processed is used as a reference area, and the average similarity between typical wells in the reference area and typical wells in other blocks is determined as follows: Where, num refer Indicates the number of typical wells in the reference area; num i Indicates the number of typical wells in the i-th sedimentary facies zone; s lk This indicates the similarity between the l-th well in the reference area and the k-th well in the i-th sedimentary facies zone; This represents the average similarity between the i-th sedimentary facies zone and the reference area; i represents the number of the sedimentary facies zone; l and k are the well numbers.

10. The missing logging curve prediction device as described in claim 6, characterized in that, The weight determination module is specifically used for: The weights of the sample well data used for training the machine learning model are determined based on the similarity using the following formula: Where, ρ i′ s represents the specified weight of the i′-th curve; i′ The similarity between the i′-th curve of a certain layer in the sample well and the corresponding curve of the well to be processed; w As is the weighting coefficient of the sample well in the corresponding layer; N is the total number of curves in the well logging curve combination.

11. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 5.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 5.

Images