A multi-task reservoir parameter prediction method based on well logging data

Abstract

The present invention provides a multi-task reservoir parameter prediction method based on well logging data, which belongs to the technical field of oil and gas field development. It includes obtaining well logging curve data and preprocessing the well logging curve data; constructing an LSTM-CNN curve reconstruction model and reconstructing the preprocessed well logging missing curve data; constructing a Transformer multi-task model, and predicting reservoir parameters using the Transformer multi-task model based on the reconstructed well logging missing curve data, and completing the prediction of multi-task reservoir parameters. The present invention solves the problem that traditional methods fail to fully utilize the trend of well logging data changing with formation depth.

Description

Technical Field

[0001] The invention belongs to the technical field of oil and gas field development, and in particular relates to a multi-task reservoir parameter prediction method based on well logging data. Background Art

[0002] Well logging, also known as geophysical logging, is a method of measuring geophysical parameters using the geophysical properties of rock formations, such as electrochemical properties, electrical conductivity, acoustic properties, and radioactivity. It is one of the applied geophysical methods. Using well logging data to predict reservoir parameters is an important part of oil and gas field development.

[0003] Porosity (POR), permeability (PERM) and water saturation (SW) are important physical parameters of reservoirs. In order to provide a reliable reference for reservoir evaluation, they must be accurately calculated. At present, there are two main methods for determining reservoir physical parameters in the field of oilfield exploration and development: one is the direct method, which is to use instruments and equipment to directly perform physical analysis and testing on the rocks after drilling and coring to obtain parameter results. However, due to the limited acquisition of core samples in actual engineering, it is impossible to accurately estimate the physical parameters of the entire work area, especially the incomplete logging series data of old oil and gas fields and old wells, which makes reservoir evaluation difficult; the other is the indirect method, which is to establish mathematical equations or empirical formulas based on the geophysical characteristics of rock formations according to logging data to indirectly calculate reservoir physical parameters. It can be divided into two categories: conventional logging interpretation methods and machine learning methods. However, the empirical formulas and statistical models constructed by this method have poor portability and do not fully consider the complex nonlinear characteristic mapping relationship and spatial continuity between reservoir parameters and logging data.

[0004] With the advent of the data-driven era and the popularization of digital oil fields, a large number of rock characteristic parameters can be obtained through logging technology. There are extremely complex nonlinear relationships between them. There are obviously many constraints in calculating reservoir parameters using direct methods and conventional logging interpretation methods. Some experts and scholars have begun to use machine learning algorithms to solve problems such as inaccurate reservoir parameter prediction. These machine learning algorithms perform better in prediction than traditional rock physics models and geological parameter modeling. However, there are also some limitations. First, the generalization ability of these models in complex geological blocks will be subject to certain constraints; second, most algorithms do not fully consider the time series characteristics of logging data and reservoir parameters, and ignore the changing trend and correlation of logging data with formation depth, so the prediction accuracy needs to be further improved.

[0005] Existing research on geophysical logging machine learning is mainly based on single-task learning. The limitation of single-task learning is that when faced with a complex problem, it needs to be decomposed into multiple single independent sub-problems, which are solved one by one and then summarized and merged to obtain the solution to the original complex problem. However, many complex problems in geophysical logging are internally interrelated and cannot be decomposed into single independent sub-problems. In addition, if reservoir parameter prediction is decomposed into single task processing, the correlation information between reservoir parameters will be ignored. Compared with single-task machine learning, multi-task learning has more advantages in reservoir parameter prediction. Summary of the invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a multi-task reservoir parameter prediction method based on well logging data. The present invention can fully capture the complex relationship between reservoir parameters and well logging data, and solves the problem that traditional methods fail to fully utilize the trend of well logging data changing with formation depth.

[0007] In order to achieve the above objectives, the technical solution adopted by the present invention is: a multi-task reservoir parameter prediction method based on well logging data, comprising the following steps:

[0008] S1, obtaining well logging curve data and preprocessing the well logging curve data;

[0009] S2. Build an LSTM-CNN curve reconstruction model and reconstruct the preprocessed missing logging curve data;

[0010] S3. Construct a Transformer multi-task model, and use the Transformer multi-task model to predict reservoir parameters based on the reconstructed logging missing curve data to complete the prediction of multi-task reservoir parameters.

[0011] The beneficial effects of the present invention are as follows: the present invention can be applied in the old well scenario, the present invention utilizes the LSTM-CNN curve reconstruction model to reconstruct the missing logging curve data, and ensures the integrity of subsequent prediction data on the basis of improving the accuracy, and adopts the Transformer multi-task model to predict reservoir parameters. The Transformer multi-task model fully captures the complex relationship between reservoir parameters and logging data. Compared with traditional methods, the root mean square error of the average predicted reservoir parameters is reduced by about 15%, and the correlation coefficient is increased by about 11%, which solves the problem that traditional methods fail to fully utilize the trend of logging data changing with formation depth.

[0012] Furthermore, the S1 comprises the following steps:

[0013] S101, collecting well logging curve data and conducting well-by-well inspection to unify the stitching curve units and perform depth correction on the well position curve;

[0014] S102, dividing the checked logging curve data into characteristic curve data and curve data to be reconstructed to obtain logging curve data;

[0015] S103, using a K-nearest neighbor classification algorithm to fill in the missing parts in the acquired logging curve data;

[0016] S104, according to the well logging curve data after missing filling, the well logging curve data exceeding the preset range is replaced with a valid value;

[0017] S105, according to the replaced logging curve data, for the logging curve data with the largest difference between the upper and lower limits, making it fall within a preset range;

[0018] S106, performing normalization processing on each sequence of the well logging curve data processed by S105;

[0019] S107, according to the normalized logging curve data, the original data is divided into a plurality of subsequences through a sliding window to complete the preprocessing of the logging curve data.

[0020] The beneficial effects of the above further scheme are: the present invention ensures the consistency and reliability of logging curve data through unit unification, depth correction and outlier processing; enhances the integrity of logging curve data and reduces the impact of missing and abnormal data on analysis through missing value filling and outlier replacement; makes the scale of logging curve data consistent through normalization processing, improves the training effect of the model; and enhances the ability to capture local features and improves the learning effect of the model by dividing the data through sliding windows. Overall, these steps work together to significantly improve the quality and availability of logging curve data, providing a high-quality data foundation for subsequent analysis and modeling.

[0021] Furthermore, the S2 comprises the following steps:

[0022] S201, build LSTM-CNN curve reconstruction model;

[0023] S202, dividing the preprocessed logging curve data into a training set, a test set and a validation set, wherein the training set is used to train the LSTM-CNN curve reconstruction model, the validation set is used to evaluate the performance of the LSTM-CNN curve reconstruction model during the training process, and the test set is used to evaluate the performance of the LSTM-CNN curve reconstruction model, and does not participate in the training of the LSTM-CNN curve reconstruction model;

[0024] S203, using mean square error as the loss function of the LSTM-CNN curve reconstruction model;

[0025] S204. Evaluate the accuracy of the LSTM-CNN curve reconstruction model using mean absolute percentage error.

[0026] S205, after the accuracy reaches the requirement, the logging curve data to be reconstructed is input into the LSTM-CNN curve reconstruction model, and the prediction output is denormalized to obtain the prediction data with the same scale as the corresponding real data;

[0027] S206: Fill in the missing parts of the reconstructed logging curve data using the predicted data with the same scale as the corresponding real data, and complete the reconstruction of the preprocessed missing logging curve data.

[0028] The beneficial effects of the above further scheme are: the LSTM-CNN curve reconstruction model combines the advantages of the long short-term memory network LSTM and the convolutional neural network CNN, and can simultaneously capture the long-term and short-term dependencies and local features of time series data, thereby improving the reconstruction accuracy; through reasonable data set division, the generalization ability and reliability of the LSTM-CNN curve reconstruction model are ensured to prevent overfitting; the mean square error and mean absolute percentage error are used as loss functions and evaluation indicators to effectively measure and improve the reconstruction performance of the model; the original scale of the predicted data is restored through denormalization processing, making the reconstruction results more reliable in practical applications; the reconstructed data fills in the missing parts, enhances the continuity and integrity of the data, and provides a high-quality data basis for subsequent geological analysis and decision-making.

[0029] Furthermore, the LSTM-CNN curve reconstruction model includes a convolutional network and a long short-term memory network LSTM;

[0030] The convolutional network is used to extract the features of the preprocessed missing well logging curve data;

[0031] The long short-term memory network LSTM is used to predict the well logging curve sequence according to the extracted features, and after denormalization, fill in the missing parts of the reconstructed well logging curve data with the predicted data of the same scale as the corresponding real data.

[0032] The beneficial effect of the above further scheme is that the LSTM-CNN curve reconstruction model combines the advantages of the long short-term memory network LSTM and the convolutional neural network CNN, and can effectively capture the long-term and short-term dependencies and local features of time series data. The long short-term memory network LSTM is used to process time series data and capture long-term dependency information; the convolutional neural network CNN is used to extract local features and improve the LSTM-CNN curve reconstruction model's ability to recognize complex patterns. The LSTM-CNN curve reconstruction model structure improves the accuracy and effectiveness of curve reconstruction.

[0033] Furthermore, the S3 comprises the following steps:

[0034] S301, build Transformer multi-task model;

[0035] S302, using the reconstructed missing logging curve data as input data, and mapping the input data to a high-dimensional embedding space through linear transformation: ;in, represents the well logging curve data after high-dimensional embedding, Represents input data, represents the weight matrix, represents the bias vector;

[0036] S303. According to the mapping results, the constructed Transformer multi-task model is used to predict the reservoir parameters, and the prediction of the multi-task reservoir parameters is completed. In the multi-task learning, for porosity and water saturation, the mean square error is used as the loss function, and for permeability, the smooth L1 loss function SmoothL1Loss in the following formula is used as the loss function. In the multi-task learning, the uncertainty weighted loss function is used to dynamically adjust the loss weight of each task: ;in, represents the value of the loss function, Represents the difference between the true value and the predicted value. When , the smooth L1 loss function SmoothL1Loss degenerates into square loss, When the smooth L1 loss function SmoothL1Loss degenerates into absolute loss, Indicates taking the absolute value;

[0037] The expression of the uncertainty weighted loss function is as follows: ;in, represents the uncertainty weighted loss function, i Indicates i tasks, represents the uncertainty parameter of the task, Indicates i Loss of a mission.

[0038] The beneficial effects of the above further scheme are: the Transformer multi-task model is good at processing long time series data and can capture the complex feature relationships in the data; the multi-task learning method utilizes the correlation between tasks to improve the overall prediction performance and efficiency. The Transformer multi-task model can optimize multiple reservoir parameter prediction tasks at the same time and provide more comprehensive prediction results; the input data is mapped to a high-dimensional embedding space through linear transformation, so that the Transformer multi-task model can better represent and learn data features and improve prediction accuracy; the uncertainty weighted loss function dynamically adjusts the loss weight of each task, making the Transformer multi-task model more flexible during the training process, and can adaptively adjust according to the difficulty of each task, thereby improving the robustness of the Transformer multi-task model; combined with the above technologies, the Transformer multi-task model can process and predict reservoir parameters more efficiently, improve the overall prediction accuracy, and provide high-quality data support for subsequent geological analysis and decision-making.

[0039] Furthermore, the Transformer multi-task model includes several encoders and several decoders:

[0040] The encoder is used to extract the features of the input data according to the mapping results through the following formula : ;in, Represents the encoder of the Transformer multi-task model;

[0041] Decoder, used to extract features and target sequence, predict reservoir parameters and generate prediction output , and use the following formula to output the prediction Mapped to the output dimension corresponding to the task: ; ;in, Indicates i The output of a task, and Respectively represent i The weight matrix and bias matrix of the task decoder, represents the decoder of the Transformer multi-task model, Represents the target sequence.

[0042] The beneficial effect of the above further scheme is that by introducing the detailed architecture of the encoder and decoder, the capabilities of the Transformer multi-task model in feature extraction and prediction generation have been significantly enhanced, providing a more advanced and reliable technical means for the efficient prediction and analysis of reservoir parameters.

[0043] Furthermore, the training process of the Transformer multi-task model is as follows:

[0044] A1. Initialize the Transformer multi-task model, including Transformer multi-task model parameters and uncertainty parameters;

[0045] A2. For each task, pass the input data through the Transformer multi-task model to get the predicted output, and calculate the loss of each task separately;

[0046] A3. Use the uncertainty weighted loss function to perform weighted summation of the loss functions of each task;

[0047] A4. Through back propagation, calculate the gradient of the uncertainty weighted loss function with respect to the Transformer multi-task model parameters and the gradient of the uncertainty parameters respectively;

[0048] A5. Use the Adam optimizer to update the Transformer multi-task model parameters and uncertainty parameters respectively.

[0049] A6. Repeat A2 to A5 until the uncertainty weighted loss function converges to complete the training of the Transformer multi-task model.

[0050] The beneficial effects of the above further scheme are: initializing the Transformer multi-task model parameters and uncertainty parameters, effectively avoiding the gradient vanishing or exploding problem in the early stage of training, and accelerating the convergence of the Transformer multi-task model; calculating the loss for each task separately, so that the Transformer multi-task model can accurately capture the specific requirements and difficulty of each task, which helps to improve the accuracy of multi-task learning; using the uncertainty weighted loss function, dynamically adjusting the loss weight according to the prediction uncertainty of each task, balancing the contribution of each task, and avoiding overfitting or underfitting; calculating the gradient through back propagation to ensure the direction and amplitude of parameter update, so that the Transformer multi-task model can gradually reduce the loss and optimize the performance; using the Adam optimizer to update the Transformer multi-task model parameters and uncertainty parameters, accelerate the training process, improve the convergence speed and stability of the Transformer multi-task model, and enhance the flexibility and adaptability of the training process. Overall, through the detailed description of the training process, the Transformer multi-task model has been significantly improved in terms of training efficiency, convergence, prediction accuracy and flexibility, providing more advanced and reliable technical means for the efficient prediction and analysis of reservoir parameters.

[0051] Furthermore, the gradient expression of the Transformer multi-task model parameters is as follows: ;in, Represents the Transformer multi-task model parameters;

[0052] The expression of the gradient of the uncertainty parameter is as follows: .

[0053] The beneficial effects of the above further scheme are: initializing the Transformer multi-task model parameters and uncertainty parameters, effectively avoiding the gradient vanishing or exploding problem in the early stage of training, and accelerating the convergence of the Transformer multi-task model; calculating the loss for each task separately, so that the Transformer multi-task model can accurately capture the specific requirements and difficulty of each task, which helps to improve the accuracy of multi-task learning; using the uncertainty weighted loss function to dynamically adjust the loss weight according to the prediction uncertainty of each task, balance the contribution of each task, and avoid overfitting or underfitting; calculating the gradient through back propagation to ensure the direction and amplitude of parameter update, so that the model can gradually reduce the loss and optimize performance; using the Adam optimizer to update the model parameters and uncertainty parameters, accelerate the training process, improve the model convergence speed and stability, and enhance the flexibility and adaptability of the training process.

[0054] Furthermore, the updating of the Transformer multi-task model parameters and uncertainty parameters is specifically as follows:

[0055] Using the Adam optimizer, the Transformer multi-task model parameters and uncertainty parameters are updated by calculating the gradient. The update rule of the Adam optimizer is as follows: ;in, and Represent the updated parameters and the parameters before the update respectively. represents the learning rate, and denote the bias correction of momentum estimation and second-order momentum estimation, respectively, Indicates a minimum value.

[0056] The beneficial effect of the above further scheme is that the Adam optimizer combines the advantages of momentum and adaptive learning rate adjustment, and can update the parameters of the Transformer multi-task model efficiently and robustly. Using the Adam optimizer, the training process can be accelerated, the convergence speed of the Transformer multi-task model can be improved, and the stability of parameter updates can be ensured. The update of uncertainty parameters enables the Transformer multi-task model to dynamically adjust the loss weights of each task, improving the flexibility and adaptability of the training process. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0058] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0059] Example

[0060] Before describing the present invention, the following terms are explained:

[0061] Depth correction: Due to errors in measurement or splicing, the logging curves of the same well may have some obvious peak misalignment problems, which need to be corrected.

[0062] Curve reconstruction: Select several logging curve data as the basis, and perform regression prediction on the missing curves through empirical formulas or machine learning.

[0063] Multi-task learning: The model shares representation information and learns multiple related tasks at the same time, making these tasks achieve better results than training one task alone. This can alleviate the overfitting of the model to a certain extent and improve the generalization ability of the model.

[0064] like Figure 1 As shown, the present invention provides a multi-task reservoir parameter prediction method based on well logging data, and its implementation method is as follows:

[0065] S1. Obtain logging curve data and preprocess the logging curve data. The implementation method is as follows:

[0066] S101, collecting well logging curve data and conducting well-by-well inspection to unify the stitching curve units and perform depth correction on the well position curve;

[0067] S102, dividing the checked logging curve data into characteristic curve data and curve data to be reconstructed to obtain logging curve data;

[0068] S103, using a K-nearest neighbor classification algorithm to fill in the missing parts in the acquired logging curve data;

[0069] S104, according to the well logging curve data after missing filling, the well logging curve data exceeding the preset range is replaced with a valid value;

[0070] S105, according to the replaced logging curve data, for the logging curve data with the largest difference between the upper and lower limits, making it fall within a preset range;

[0071] S106, performing normalization processing on each sequence of the well logging curve data processed by S105;

[0072] S107, according to the normalized logging curve data, the original data is divided into a plurality of subsequences through a sliding window to complete the preprocessing of the logging curve data.

[0073] In this embodiment, logging curve data are collected, and specific curves include: acoustic wave (AC), well diameter (CAL), density (DEN), neutron (CNL), natural gamma (GR), flushing zone formation resistivity (RXO), resistivity (RT), porosity (POR), permeability (PERM), water saturation (SW), and then the collected logging curve data are checked well by well: the spliced ​​curve units are unified, and the key curves of the well position are depth corrected.

[0074] In this embodiment, due to the incomplete logging data of old wells and old areas, some curves are missing in segments, and it is extremely difficult to directly predict reservoir parameters. In order to carry out subsequent work, the logging curve data after inspection are divided into characteristic curve data: acoustic wave (AC), wellbore (CAL), natural gamma (GR), flushing zone formation resistivity (RXO), resistivity (RT), curve data to be reconstructed: density (DEN), neutron (CNL) and label curve data: porosity (POR), permeability (PERM), water saturation (SW).

[0075] S2. Build an LSTM-CNN curve reconstruction model and reconstruct the preprocessed missing logging curve data. The implementation method is as follows:

[0076] S201, build LSTM-CNN curve reconstruction model;

[0077] S202, dividing the preprocessed logging curve data into a training set, a test set and a validation set, wherein the training set is used to train the LSTM-CNN curve reconstruction model, the validation set is used to evaluate the performance of the LSTM-CNN curve reconstruction model during the training process, and the test set is used to evaluate the performance of the LSTM-CNN curve reconstruction model, and does not participate in the training of the LSTM-CNN curve reconstruction model;

[0078] S203, using mean square error as the loss function of the LSTM-CNN curve reconstruction model;

[0079] S204. Evaluate the accuracy of the LSTM-CNN curve reconstruction model using mean absolute percentage error.

[0080] S205, after the accuracy reaches the requirement, the logging curve data to be reconstructed is input into the LSTM-CNN curve reconstruction model, and the prediction output is denormalized to obtain the prediction data with the same scale as the corresponding real data;

[0081] S206: Fill in the missing parts of the reconstructed logging curve data using the predicted data with the same scale as the corresponding real data, and complete the reconstruction of the preprocessed missing logging curve data.

[0082] In this embodiment, the LSTM-CNN curve reconstruction model includes a convolutional network and a long short-term memory network LSTM;

[0083] The convolutional network is used to extract the features of the preprocessed missing well logging curve data, capture the local pattern of the data and remove noise;

[0084] The long short-term memory network LSTM is used to obtain predicted data with the same scale as the corresponding real data by denormalizing the predicted output according to the extracted features, and fill in the missing parts of the reconstructed logging curve data with the predicted data with the same scale as the corresponding real data to capture time dependency and perform curve sequence prediction.

[0085] In this embodiment, the missing values ​​in the data are filled by the K-nearest neighbor classification algorithm, and the processing formula is: ; ;in, represents missing values, Indicates nearby The value of the data. k Represents the total number of data near the missing value, Indicates The weight of the data, Indicates The Euclidean distance between the data and the missing value.

[0086] In this embodiment, after filling the missing values, there are still a few logging curve data that exceed the industry standard range (i.e., the preset range), and this part of the logging curve data is replaced with a valid value. The replacement method is: replace it with the upper or lower limit of the valid value close to the value. For the logging curve data with a large gap between the upper and lower limits: GR, RXO, RT, take log10 of the data to make it fall within the preset range.

[0087] In this embodiment, in order to obtain the data change trend in the subsequence, each sequence is normalized separately, and the processing formula is: ;in, represents the normalized data, Represents the original data, Indicates the minimum value, In order to obtain the time information in the data, the original data is divided into several sequences through a sliding window, and the processing formula is: ;in, N represents the total number of sequences, l represents the sequence length, m represents the sliding window size, s Indicates the sliding window moving step size.

[0088] In this embodiment, an LSTM-CNN curve reconstruction model is constructed, which is composed of a convolutional neural network (CNN) and a long short-term memory (LSTM) network. The convolution layer includes two convolutional networks, and the hidden layer includes three LSTM layers (the convolution layer first processes the original input data to extract local features; then, the extracted features are passed to the LSTM layer of the hidden layer; the LSTM layer of the hidden layer receives the features output by the convolution layer as input, and further processes these features to capture the temporal relationship. (2) In the LSTM-CNN curve reconstruction model, the role of the convolutional network is to extract local features, and the role of the long short-term memory (LSTM) network is to model the long-term dependency in the temporal data). The tanh activation function is used, and the random inactivation Dropout layer is used to randomly delete the information of each layer of the long short-term memory (LSTM) network according to a preset ratio to prevent overfitting. The output layer uses a fully connected layer. Feature extraction is performed through the combination of the long short-term memory (LSTM) network and the convolutional neural network (CNN), and the final result is output through the fully connected layer.

[0089] In this embodiment, the characteristic curve data is processed in all S1, and the reconstructed curve data is processed in S106 and S107. The training data is divided into a training set, a validation set, and a test set in a ratio of 80%, 10%, and 10%, wherein: the training set is used to train the parameters and weights of the LSTM-CNN curve reconstruction model. The LSTM-CNN curve reconstruction model adjusts its own parameters by learning the sample data in the training set to minimize the loss function, so as to complete the training of the LSTM-CNN curve reconstruction model; the validation set is used to evaluate the performance of the LSTM-CNN curve reconstruction model during the training process and perform parameter tuning; the test set does not participate in the training of the LSTM-CNN curve reconstruction model, and is used to evaluate the performance of the LSTM-CNN curve reconstruction model.

[0090] In this embodiment, the mean square error (MSE) is used as the loss function. The mean square error measures the difference between the predicted value and the true value of the LSTM-CNN curve reconstruction model. The smaller the value, the better the performance of the LSTM-CNN curve reconstruction model. The calculation formula is: ;in, Indicates the total number of logging curve data. Indicates The true value of the well logging curve data, Indicates The true value of the well logging curve data, Indicates The predicted value of the well logging data.

[0091] In this embodiment, the mean absolute percentage error (MAPE) is used to evaluate the accuracy of the reconstructed logging curve data, and its calculation formula is: ;in, Indicates the total number of logging curve data. Indicates The true value of the well logging curve data, Indicates The true value of the well logging curve data, Indicates The predicted value of the well logging data.

[0092] In this embodiment, after the accuracy reaches the standard, the characteristic curve data of the data to be reconstructed is input into the LSTM-CNN curve reconstruction model, and the predicted output is denormalized to obtain the predicted data with the same scale as the corresponding real data. The predicted data is used to fill the missing part of the well logging curve to be reconstructed, so as to achieve the purpose of reconstructing the missing well logging curve data.

[0093] S3. Construct a Transformer multi-task model, and use the Transformer multi-task model to predict reservoir parameters based on the reconstructed missing logging curve data to complete the prediction of multi-task reservoir parameters. The implementation method is as follows:

[0094] S301, build Transformer multi-task model;

[0095] S302, using the reconstructed missing logging curve data as input data, and mapping the input data to a high-dimensional embedding space through linear transformation: ;in, represents the well logging curve data after high-dimensional embedding, Represents input data, represents the weight matrix, represents the bias vector;

[0096] S303. According to the mapping results, the constructed Transformer multi-task model is used to predict the reservoir parameters, and the prediction of the multi-task reservoir parameters is completed. In the multi-task learning, for porosity and water saturation, the mean square error is used as the loss function, and for permeability, the smooth L1 loss function SmoothL1Loss in the following formula is used as the loss function. In the multi-task learning, the uncertainty weighted loss function is used to dynamically adjust the loss weight of each task: ;in, represents the value of the loss function, Represents the difference between the true value and the predicted value. When , the smooth L1 loss function SmoothL1Loss degenerates into square loss, When the smooth L1 loss function SmoothL1Loss degenerates into absolute loss, Indicates taking the absolute value;

[0097] The expression of the uncertainty weighted loss function is as follows: ;in, represents the uncertainty weighted loss function, i Indicates i tasks, represents the uncertainty parameter of the task, Indicates i Loss of a mission.

[0098] In this embodiment, the Transformer multi-task model includes several encoders and several decoders:

[0099] The Transformer multi-task model includes several encoders and several decoders:

[0100] The encoder is used to extract the features of the input data according to the mapping results through the following formula : ;in, Represents the encoder of the Transformer multi-task model;

[0101] Decoder, used to extract features and target sequence, predict reservoir parameters and generate prediction output , and use the following formula to output the prediction Mapped to the output dimension corresponding to the task: ; ; in, Indicates i The output of a task, and Respectively representi The weight matrix and bias matrix of the task decoder, represents the decoder of the Transformer multi-task model, Represents the target sequence.

[0102] In this embodiment, the training process of the Transformer multi-task model is as follows:

[0103] A1. Initialize the Transformer multi-task model, including Transformer multi-task model parameters and uncertainty parameters;

[0104] A2. For each task, pass the input data through the Transformer multi-task model to get the predicted output, and calculate the loss of each task separately;

[0105] A3. Use the uncertainty weighted loss function to perform weighted summation of the loss functions of each task;

[0106] A4. Through back propagation, calculate the gradient of the uncertainty weighted loss function with respect to the Transformer multi-task model parameters and the gradient of the uncertainty parameters respectively;

[0107] A5. Use the Adam optimizer to update the Transformer multi-task model parameters and uncertainty parameters respectively.

[0108] A6. Repeat A2 to A5 until the uncertainty weighted loss function converges to complete the training of the Transformer multi-task model.

[0109] In this embodiment, the reconstructed logging curve data is input , abbreviated as ,in, n Indicates the total number of logging curve data. d Represents the dimension of the data, Represents each logging curve data, Represents the set of real numbers.

[0110] Map the input data into a high-dimensional embedding space through a linear transformation: ;in, represents the well logging curve data after high-dimensional embedding, Represents input data, represents the weight matrix, Represents the bias vector.

[0111] In this embodiment, the Transformer multi-task model consists of multiple encoder layers and decoder layers. The encoder layer extracts the features of the input data through the self-attention mechanism and feedforward neural network: ;in, represents the encoder of the Transformer multi-task model, represents the output of the encoder, t Represents the time step.

[0112] In this embodiment, the decoder layer is used to generate the prediction output. The input of the decoder is the output of the encoder and the target sequence (the actual label during training and the prediction of the previous step during prediction): ;in, is the output of the decoder and T represents the target sequence.

[0113] In order to share information between different tasks, the model introduces an additional multi-head attention mechanism: ;in, represents the output of the shared attention mechanism, represents the multi-head attention mechanism, , , Represents the encoder output for the porosity (POR), permeability (PERM), and water saturation (SW) tasks, respectively.

[0114] Each task has a separate decoder that maps the output of the Transformer multi-task model to task-specific output dimensions: ;in, Indicates i The output of a task, represents the output dimension, and Respectively represent i The weight matrix and bias matrix of the task decoder.

[0115] In this embodiment, in multi-task learning, each task needs to set a loss function according to its characteristics. For the tasks of porosity (POR) and water saturation (SW), the mean square error MSE is used as the loss function; and for the permeability PERM task, the smooth L1 loss function SmoothL1Loss is used, which can effectively handle the large difference between the predicted value and the true value, while avoiding numerical instability when it is close to zero. Its calculation formula is: ;in, represents the value of the loss function, Represents the difference between the true value and the predicted value. When , the smooth L1 loss function SmoothL1Loss degenerates into square loss, When the smooth L1 loss function SmoothL1Loss degenerates into absolute loss, Indicates taking the absolute value.

[0116] In this implementation, in multi-task learning, a weighting function can be used to balance the importance of different tasks in order to better optimize the Transformer multi-task model. The uncertainty weighting function is used to assign a weight to each task, which represents the importance of the task in the overall optimization. The uncertainty weighted loss function is used to dynamically adjust the loss weight of each task. The weight is controlled by the uncertainty parameter of the task, and its calculation formula is: ;in, represents the uncertainty weighted loss function, i Indicates i tasks, represents the uncertainty parameter of the task, Indicates i Loss of a mission.

[0117] In this embodiment, in the training process of the Transformer multi-task learning model, back propagation and parameter updating are key steps. By calculating the gradient and updating the Transformer multi-task learning model parameters, the loss function is gradually reduced. The dynamic weighted total loss is obtained (i.e., uncertainty weighted loss function), and then calculate the total loss to model parameters by back propagation (all trainable and updateable parameters) and uncertainty parameters The gradient of the Transformer multi-task model parameters : ;

[0118] For the uncertainty parameters : ;

[0119] In this embodiment, the Adam optimizer is used to update the model parameters and uncertainty parameters according to the calculated gradients. The update rule of the Adam optimizer is: ;in, and Represent the updated parameters and the parameters before the update respectively. represents the learning rate, and denote the bias correction of momentum estimation and second-order momentum estimation, respectively, Indicates the minimum value, which is added to prevent the denominator from being zero.

[0120] In this embodiment, when using the Transformer multi-task model to predict reservoir parameters, the input data does not use sliding window processing, because the input of the complete sequence retains the natural dependency of the sequence data, and can fully utilize the model's attention mechanism to capture global information and enhance the synergy between different tasks in multi-task learning. These advantages enable the Transformer multi-task model to better understand and process complex geological data and improve the accuracy and reliability of predictions. Therefore, the input data only needs to be processed from S104 to S106 before it can be input into the LSTM-CNN curve reconstruction model for training and prediction.

[0121] In this example, the root mean square error (RMSE), mean absolute error (MAE) and correlation coefficient R are calculated using the real data of reservoir parameters and the predicted data output by the Transformer multi-task model: ; ; ;in, represents the mean of the true value of the logging data, Represents the mean of the predicted values.

[0122] The performance of the Transformer multi-task model is judged based on the above evaluation indicators. The smaller the root mean square error RMSE and the mean absolute error MAE, the more accurate the prediction results of the reservoir parameters by Transformer multi-task learning. The correlation coefficient R describes the degree of fit between the true value and the predicted value of the reservoir parameter. The closer it is to 1, the better.

[0123] In this embodiment, other techniques such as wavelet transform, principal component analysis (PCA) and independent component analysis (ICA) can also be added to the data preprocessing to reduce noise and extract key features. In addition to the LSTM-CNN model, other deep learning models such as GRU (Gated Recurrent Unit) or BiLSTM (Bidirectional LSTM) can also be used for missing data reconstruction. In addition to the Transformer model, other advanced deep learning models such as BERT (Bidirectional Encoder Representation Transformer), XLNet (an improved autoregressive language pre-training model) or hybrid models (such as the combination of LSTM and Transformer) can also be used for multi-task prediction.

Claims

1. A multi-task reservoir parameter prediction method based on well logging data, characterized in that: The following steps are involved: S1, obtaining well logging curve data and preprocessing the well logging curve data; S2. Build an LSTM-CNN curve reconstruction model and reconstruct the preprocessed missing logging curve data; The S2 comprises the following steps: S201, build LSTM-CNN curve reconstruction model; S202, dividing the preprocessed logging curve data into a training set, a test set and a validation set, wherein the training set is used to train the LSTM-CNN curve reconstruction model, the validation set is used to evaluate the performance of the LSTM-CNN curve reconstruction model during the training process, and the test set is used to evaluate the performance of the LSTM-CNN curve reconstruction model, and does not participate in the training of the LSTM-CNN curve reconstruction model; S203, using mean square error as the loss function of the LSTM-CNN curve reconstruction model; S204. Evaluate the accuracy of the LSTM-CNN curve reconstruction model using mean absolute percentage error. S205, after the accuracy reaches the requirement, the logging curve data to be reconstructed is input into the LSTM-CNN curve reconstruction model, and the prediction output is denormalized to obtain the prediction data with the same scale as the corresponding real data; S206, using the predicted data with the same scale as the corresponding real data to fill in the missing parts of the reconstructed logging curve data, and complete the reconstruction of the preprocessed missing logging curve data; The LSTM-CNN curve reconstruction model includes a convolutional network and a long short-term memory network LSTM; The convolutional network is used to extract the features of the preprocessed missing well logging curve data; The long short-term memory network LSTM is used to predict the well logging curve sequence according to the extracted features, and after denormalization, fill in the missing parts of the reconstructed well logging curve data with the predicted data of the same scale as the corresponding real data; S3, constructing a Transformer multi-task model, and using the Transformer multi-task model to predict reservoir parameters based on the reconstructed missing logging curve data, completing the prediction of multi-task reservoir parameters; The S3 comprises the following steps: S301, build Transformer multi-task model; S302, using the reconstructed missing logging curve data as input data, and mapping the input data to a high-dimensional embedding space through linear transformation: in, represents the well logging curve data after high-dimensional embedding, Represents input data, represents the weight matrix, represents the bias vector; S303. According to the mapping results, the constructed Transformer multi-task model is used to predict the reservoir parameters, and the prediction of the multi-task reservoir parameters is completed. In the multi-task learning, for porosity and water saturation, the mean square error is used as the loss function, and for permeability, the smooth L1 loss function SmoothL1Loss in the following formula is used as the loss function. In the multi-task learning, the uncertainty weighted loss function is used to dynamically adjust the loss weight of each task: in, represents the value of the loss function, Represents the difference between the true value and the predicted value. When , the smooth L1 loss function SmoothL1Loss degenerates into square loss, When the smooth L1 loss function SmoothL1Loss degenerates into absolute loss, Indicates taking the absolute value; The expression of the uncertainty weighted loss function is as follows: in, represents the uncertainty weighted loss function, i Indicates i tasks, represents the uncertainty parameter of the task, Indicates i Loss of a mission.

2. The multi-task reservoir parameter prediction method based on well logging data according to claim 1, characterized in that: The S1 comprises the following steps: S101, collecting well logging curve data and conducting well-by-well inspection to unify the stitching curve units and perform depth correction on the well position curve; S102, dividing the checked logging curve data into characteristic curve data and curve data to be reconstructed to obtain logging curve data; S103, using a K-nearest neighbor classification algorithm to fill in the missing parts in the acquired logging curve data; S104, according to the well logging curve data after missing filling, the well logging curve data exceeding the preset range is replaced with a valid value; S105, according to the replaced logging curve data, for the logging curve data with the largest difference between the upper and lower limits, making it fall within a preset range; S106, performing normalization processing on each sequence of the well logging curve data processed by S105; S107, according to the normalized logging curve data, the original data is divided into a plurality of subsequences through a sliding window to complete the preprocessing of the logging curve data.

3. The multi-task reservoir parameter prediction method based on well logging data according to claim 1, characterized in that: The Transformer multi-task model includes several encoders and several decoders: The encoder is used to extract the features of the input data according to the mapping results through the following formula : in, Represents the encoder of the Transformer multi-task model; Decoder, used to extract features and target sequence, predict reservoir parameters and generate prediction output , and use the following formula to output the prediction Mapped to the output dimension corresponding to the task: in, Indicates i The output of a task, and Respectively represent i The weight matrix and bias matrix of the task decoder, represents the decoder of the Transformer multi-task model, Represents the target sequence.

4. The multi-task reservoir parameter prediction method based on well logging data according to claim 1, characterized in that: The training process of the Transformer multi-task model is as follows: A1. Initialize the Transformer multi-task model, including Transformer multi-task model parameters and uncertainty parameters; A2. For each task, pass the input data through the Transformer multi-task model to get the predicted output, and calculate the loss of each task separately; A3. Use the uncertainty weighted loss function to perform weighted summation of the loss functions of each task; A4. Through back propagation, calculate the gradient of the uncertainty weighted loss function with respect to the Transformer multi-task model parameters and the gradient of the uncertainty parameters respectively; A5. Use the Adam optimizer to update the Transformer multi-task model parameters and uncertainty parameters respectively; A6. Repeat A2 to A5 until the uncertainty weighted loss function converges to complete the training of the Transformer multi-task model.

5. The multi-task reservoir parameter prediction method based on well logging data according to claim 4, characterized in that: The gradient expression of the Transformer multi-task model parameters is as follows: in, Represents the Transformer multi-task model parameters; The expression of the gradient of the uncertainty parameter is as follows: 。 6. The multi-task reservoir parameter prediction method based on well logging data according to claim 5, characterized in that: The updating of the Transformer multi-task model parameters and uncertainty parameters is specifically as follows: Using the Adam optimizer, the Transformer multi-task model parameters and uncertainty parameters are updated by calculating the gradient. The update rule of the Adam optimizer is as follows: in, and Represent the updated parameters and the parameters before the update respectively. represents the learning rate, and denote the bias correction of momentum estimation and second-order momentum estimation, respectively, Indicates a minimum value.

Images