A method and apparatus for identifying fluid types in a tight sand gas reservoir

Abstract

The embodiment of the present application relates to the technical field of oil and gas field exploration and development geology, and discloses a method and equipment for identifying fluid type of tight sandstone gas reservoir by well logging, which comprises the following steps: obtaining sensitive well logging curves of blind wells in the tight sandstone gas reservoir; the fluid type of the blind well is unknown; identifying the fluid type of the blind well according to the sensitive well logging curves of the blind well and a pre-generated fluid determination model; the fluid determination model is generated by a pre-generated convolutional neural network and an integrated algorithm model, the convolutional neural network is used to extract first feature data of the sensitive well logging curves of the blind well, the integrated model is composed of multiple machine learning models and is used to identify the fluid type of the blind well according to the first feature data, and the convolutional neural network comprises a fully connected layer. The method provided by the present application avoids the significant difference of the well logging identification effect of a single algorithm model on the fluid type of the tight sandstone gas reservoir, and significantly improves the overall identification stability and accuracy of the model.

Description

Technical Field

[0001] This invention relates to the field of geological technology for oil and gas field exploration and development, and in particular to a logging identification method, apparatus, equipment, medium, and product program for fluid types in tight sandstone gas reservoirs. Background Technology

[0002] Currently, there have been numerous attempts and research results in the study of conventional logging identification methods for reservoir fluids in tight sandstone gas reservoirs. Conventional logging identification methods include: curve overlay method (Li Yunsheng et al., 2003; Mao Keyu, 2016; Yang Junru, 2021) and intersection chart method (Li Yunsheng et al., 2003; Tang Yongmei, 2004; Liu Xingjun, 2015; Mao Keyu, 2016; Yao Yabin, 2018; Zhu Wenqing, 2019; Yang Junru, 2021), etc. The advantages of these methods are that they are fast, intuitive, and simple to implement, and are widely used. However, their limitations are that they are greatly affected by human factors, the degree of automation of interpretation results is low, and the identification accuracy in tight sandstone reservoirs is low.

[0003] In addition, array acoustic logging and nuclear magnetic resonance logging are also commonly used identification methods. For example, nuclear magnetic resonance imaging logging identification method (Hook et al., 1994; Liu Yan, 2005; Zhu Wenqing et al., 2019; Yang Junru, 2021) uses shear wave data to calculate relevant parameters such as Poisson's ratio, volume compressibility coefficient, equivalent elastic modulus, and P-wave and S-wave transit time, and uses the comparison of relevant parameters to identify gas layers (Huang Xiaodong et al., 2005; Wang Guiqing et al., 2008; Bian Huiyuan et al., 2010; Ma Long et al., 2011; Xin Yi, Zhao Hui et al., 2012; Wang Wenwen et al., 2020). In general, array acoustic logging technology and nuclear magnetic resonance logging technology have greatly improved the accuracy compared with conventional logging identification methods, but they still have problems such as high cost, difficulty in obtaining logging curve data, and complex construction of rock physics models with low calculation accuracy. Summary of the Invention

[0004] The purpose of this invention is to provide at least one well logging method and device for identifying fluid types in tight sandstone gas reservoirs. The aim is to overcome the following technical problems in the existing technology of identifying fluid types in tight sandstone gas reservoirs using well logging data through machine learning methods: sensitive curves are only constructed based on the characteristics of core well curves, which cannot be widely applied; the integrated algorithm model adopts a cascaded approach, with large differences between different modules; different algorithms and parameters need to be continuously adjusted and verified; and the automaticity of interpretation is poor.

[0005] To address the aforementioned technical problems, at least one embodiment of this application provides a well logging method for identifying fluid types in tight sandstone gas reservoirs, including:

[0006] Sensitive logging curves of blind wells in the tight sandstone gas reservoir are obtained; wherein the fluid type of the blind well is unknown.

[0007] The fluid type of the blind well is identified based on the sensitive logging curve of the blind well and a pre-generated fluid determination model. The fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model. The convolutional neural network is used to extract the first feature data of the sensitive logging curve of the blind well. The ensemble algorithm model consists of multiple machine learning models and is used to identify the fluid type of the blind well based on the first feature data. The convolutional neural network includes a fully connected layer.

[0008] In some embodiments, obtaining the sensitive logging curves of blind wells in the tight sandstone gas reservoir includes:

[0009] The fluid type of the known wells is determined based on fluid test data and production data from known wells within the tight sandstone gas reservoir;

[0010] Calculate the mutual information between each logging curve of the known well and its corresponding fluid type;

[0011] Among the logging curves, the sensitive logging curve for the blind well is selected based on the mutual information.

[0012] In some embodiments, the step of generating the convolutional neural network includes:

[0013] The second feature data of the sensitive logging curve of the known well and the third feature data of the corresponding fluid type are extracted through the convolutional layer of the initial model of the convolutional neural network.

[0014] The initial model is forward-propagated and backward-propagated using the second feature data and the third feature data to generate the convolutional neural network.

[0015] In some embodiments, the second feature data of the sensitive logging curve of the known well and the third feature data of the corresponding fluid type are extracted through the convolutional layers of the initial model of the convolutional neural network, including:

[0016] The convolutional layer performs a convolution operation on the sensitive logging curves of the known well and the corresponding fluid type to extract the second feature data and the third feature data.

[0017] In some embodiments, identifying the fluid type of the blind well based on the sensitive logging curves of the blind well and a pre-generated fluid determination model includes:

[0018] The first feature data of the sensitive logging curve of the blind well is extracted by the convolutional neural network;

[0019] The fluid type of the blind well is identified using the first feature data and the integrated algorithm model.

[0020] In some embodiments, the step of generating the ensemble algorithm model includes:

[0021] Training data is generated based on the second feature data and the third feature data;

[0022] The training data is input into the multiple machine learning models to generate multiple fluid type recognition results;

[0023] The results of the multiple fluid type identifications are input into a weak classifier to calculate the errors of each of the multiple machine learning models.

[0024] The classifier weights of the multiple machine learning models are modified based on the error.

[0025] The ensemble algorithm model is generated based on the multiple machine learning models and their respective classifier weights.

[0026] In some embodiments, the sensitive logging curve includes at least one of the following: sonic transit time curve, deep lateral resistivity curve, shallow lateral resistivity curve, compensated neutron logging curve, and natural gamma logging curve.

[0027] At least one embodiment of this application also provides a logging identification device for fluid types in tight sandstone gas reservoirs, comprising:

[0028] A sensitive logging curve acquisition module is used to acquire sensitive logging curves of blind wells in the tight sandstone gas reservoir; wherein the fluid type of the blind well is unknown.

[0029] A fluid type identification module is used to identify the fluid type of the blind well based on the sensitive logging curve of the blind well and a pre-generated fluid determination model. The fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model. The convolutional neural network is used to extract the first feature data of the sensitive logging curve of the blind well. The ensemble algorithm model consists of multiple machine learning models and is used to identify the fluid type of the blind well based on the first feature data. The convolutional neural network includes a fully connected layer.

[0030] At least one embodiment of this application also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the above-described well logging identification method for fluid types in tight sandstone gas reservoirs.

[0031] At least one embodiment of this application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method for identifying fluid types in tight sandstone gas reservoirs.

[0032] An embodiment of this application provides a well logging method for identifying fluid types in tight sandstone gas reservoirs, comprising: first, acquiring sensitive logging curves of blind wells in tight sandstone gas reservoirs; wherein the fluid type of the blind wells is unknown; then, identifying the fluid type of the blind wells based on the sensitive logging curves of the blind wells and a pre-generated fluid determination model; wherein the fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model, the convolutional neural network is used to extract first feature data from the sensitive logging curves of the blind wells, the ensemble model consists of multiple machine learning models, and is used to identify the fluid type of the blind wells based on the first feature data, and the convolutional neural network includes a fully connected layer.

[0033] This invention avoids the significant differences in logging identification effects of individual algorithm models for fluid types in tight sandstone gas reservoirs, and significantly improves the overall identification stability and accuracy of the model. Attached Figure Description

[0034] One or more embodiments are illustrated by way of example with reference to the accompanying drawings, and these illustrative descriptions do not constitute a limitation on the embodiments.

[0035] Figure 1 This is a schematic flowchart of a well logging method for identifying fluid types in tight sandstone gas reservoirs, provided in one embodiment of this application.

[0036] Figure 2 This is a flowchart illustrating step 100 provided in one embodiment of this application;

[0037] Figure 3 This is a schematic diagram of the process for generating the convolutional neural network provided in one embodiment of this application;

[0038] Figure 4 This is a flowchart illustrating step 200 provided in one embodiment of this application;

[0039] Figure 5 This is a schematic diagram of the process for generating the ensemble algorithm model provided in one embodiment of this application;

[0040] Figure 6 This is a flowchart illustrating the well logging identification method for fluid types in tight sandstone gas reservoirs provided in a specific embodiment of this application;

[0041] Figure 7This is a flowchart illustrating the technical route of the well logging identification method for fluid types in tight sandstone gas reservoirs provided in a specific embodiment of this application.

[0042] Figure 8 This is a schematic diagram of the process for extracting CNN convolutional features according to a specific implementation of this application;

[0043] Figure 9 This is a comparison chart showing the accuracy of cross-validation of blind wells using the algorithm model provided in the specific implementation of this application.

[0044] Figure 10 This is a flowchart illustrating the selection and optimization model provided in a specific implementation of this application;

[0045] Figure 11 This is a schematic diagram illustrating the construction process of the algorithm model provided in a specific implementation of this application;

[0046] Figure 12 This is a comparison chart of the accuracy of cross-validation of blind wells for the algorithm model provided in the specific implementation of this application;

[0047] Figure 13 This is a schematic diagram illustrating the blind well identification effect of the algorithm model provided in a specific implementation of this application;

[0048] Figure 14 This is a schematic diagram of a logging identification device for fluid types in tight sandstone gas reservoirs, provided in one embodiment of this application;

[0049] Figure 15 This is a schematic diagram of the structure of an electronic device provided in another embodiment of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the various embodiments of this application will be described in detail below with reference to the accompanying drawings. However, those skilled in the art will understand that many technical details have been provided in the various embodiments of this application to help readers better understand this application. However, the technical solutions claimed in this application can be implemented even without these technical details and various changes and modifications based on the following embodiments. The division of the various embodiments below is for the convenience of description and should not constitute any limitation on the specific implementation of this application. The various embodiments can be combined with and referenced by each other without contradiction.

[0051] To facilitate understanding of the embodiments of this application, the technical challenges of well logging identification of fluid types in tight sandstone gas reservoirs will be introduced first.

[0052] Understandably, reservoir fluid logging interpretation refers to the comprehensive analysis of various conventional logging curves to determine the fluid type and logging response characteristics of the reservoir, and then identifying the reservoir fluids through the logging curves. Because logging curves represent a comprehensive response of the geological body, and most tight sandstone gas reservoirs, as unconventional oil and gas reservoirs, are characterized by large burial depths, ultra-tight reservoirs, low porosity, low permeability, strong reservoir heterogeneity, and complex gas-water distribution, the logging response characteristics of tight sandstone reservoir fluids are complex, making accurate identification and evaluation of reservoir fluids challenging.

[0053] With the rapid development of artificial intelligence technology, industry professionals are attempting to utilize this technology to solve the aforementioned problems. Specifically, intelligent interpretation of reservoir fluid logging refers to using AI-related technologies to mine logging curve information and identify and interpret different types of reservoir fluids. This is an interdisciplinary field between earth science and information science, currently in its initial and exploratory stages. Its basic process includes: obtaining the main fluid types that reflect the true condition of the underground reservoirs in the study area based on gas testing data and production data; collecting logging data from the target layers of the gas testing wells and non-gas testing wells; establishing a mapping model between fluid types and logging curves; evaluating the model; and then applying the non-gas testing well model to interpret reservoir fluids.

[0054] As the difficulty of logging interpretation in unconventional reservoirs continues to increase, scholars have successively carried out research on logging interpretation methods for unconventional reservoirs based on machine learning-based intelligent algorithms (Mao Ningbo, 1996; Zou Changchun, 1997; Yang Bin, 2005; Zhang Yinde, 2008; Wang Peng, 2014; Zhao Jun et al., 2015; Yi Jun et al., 2022). Compared with traditional logging identification methods for reservoir fluids, intelligent logging identification technology for reservoir fluids can identify fluid types in non-gas-testing wells and has higher efficiency and accuracy. However, overall, the intelligent identification technology for tight sandstone reservoir fluids is still underdeveloped. It only constructs sensitive curves based on the characteristics of cored well curves, which cannot be widely applied. The integrated algorithm model adopts a cascaded approach, with significant differences between different modules. Different algorithms and parameters need to be continuously adjusted and verified, resulting in poor automaticity of interpretation.

[0055] Example 1:

[0056] For the reasons stated above, the well logging identification method for fluid types in tight sandstone gas reservoirs in this embodiment can be applied to electronic devices with communication, computing, and data storage capabilities. The specific process can be as follows: Figure 1 As shown, it includes:

[0057] Step 100: Obtain sensitive logging curves for blind wells in the tight sandstone gas reservoir; wherein the fluid type of the blind well is unknown;

[0058] Step 200: Identify the fluid type of the blind well based on the sensitive logging curve of the blind well and the pre-generated fluid determination model; wherein, the fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model, the convolutional neural network is used to extract the first feature data of the sensitive logging curve of the blind well, the ensemble algorithm model is composed of multiple machine learning models, and is used to identify the fluid type of the blind well based on the first feature data, the convolutional neural network includes a fully connected layer.

[0059] This application provides a well logging method for identifying fluid types in tight sandstone gas reservoirs, comprising: first, acquiring sensitive logging curves of blind wells in tight sandstone gas reservoirs; wherein the fluid type of the blind wells is unknown; then, identifying the fluid type of the blind wells based on the sensitive logging curves and a pre-generated fluid determination model; wherein the fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model, the convolutional neural network is used to extract first feature data from the sensitive logging curves of the blind wells, the ensemble model consists of multiple machine learning models, and is used to identify the fluid type of the blind wells based on the first feature data; the convolutional neural network includes a fully connected layer. Specifically, this invention has the following beneficial effects:

[0060] In summary, the method provided by the embodiments of the present invention avoids significant differences in the logging identification effect of a single algorithm model for fluid types in tight sandstone gas reservoirs, and significantly improves the overall identification stability and accuracy of the model.

[0061] For step 100, the response characteristics between each fluid type and the logging curve can be analyzed based on the fluid type and the logging curve. Preferably, the SelectKBest method can be used to select sensitive logging curves or combinations of multiple sensitive logging curves.

[0062] Regarding step 200, it should be noted that the convolutional neural network is only used for feature extraction of sensitive logging curves and does not perform fitting operations, so only one fully connected layer (Flatten layer) is used.

[0063] Furthermore, for the ensemble algorithm module in step 200, based on the Boosting concept, multiple machine learning models are applied to construct an ensemble algorithm intelligent interpretation model. Specifically, firstly, the training set samples and initial weight decomposition are input into multiple machine learning models. The recognition results of each machine learning model are then fed into a weak classifier. The weak classifier calculates the error between the recognition result and the true result, as well as the classifier weights. Simultaneously, the weights of the input samples are adjusted based on the classifier weights until the minimum error is determined. Finally, based on the classifier weights of each machine learning model, a weighted combination strategy is used to obtain a strong learner, i.e., the ensemble algorithm module, which is used to output the final recognition result.

[0064] Boosting is an ensemble learning method that aims to create a strong learner by combining multiple weak learners (preferably decision trees). Its basic idea is to build the model incrementally, with each step focusing on correcting the errors of the previous step. The basic principles of Boosting are as follows:

[0065] Weak learners: Boosting uses weak learners that are slightly better than random guessing. These models may perform poorly individually, but by combining multiple weak learners, the overall performance can be significantly improved.

[0066] Iterative training: Boosting is an iterative method in which a new weak learner is added to the model in each iteration. Each new learner focuses on samples that the previous learner misclassified.

[0067] Weighted strategy: In each round, Boosting adjusts the weights of samples based on their error rate. Misclassified samples receive higher weights, causing subsequent learners to pay more attention to these difficult-to-classify samples.

[0068] Weighted voting or weighted averaging: The final prediction model is a weighted combination of all weak learners. Weighted voting can be used for classification problems.

[0069] Example 2:

[0070] In some examples, see Figure 2 Step 100 includes:

[0071] Step 101: Determine the fluid type of the known wells based on the fluid test data and production data of the known wells in the tight sandstone gas reservoir;

[0072] Specifically, based on the gas testing data of the study area, the possible fluid types in the study area are determined. Since the testing time of a single well is relatively short, it cannot fully reflect the actual situation of the underground reservoir fluids. Combined with production data, including single-well test gas and water data, single-well daily production, and cumulative gas and water production data, a comprehensive judgment is made to finally determine the main fluid types of the underground reservoir in the study area.

[0073] Single-well test production gas and water data refers to the amount of gas and water extracted from a single oil and gas well through test production operations within a certain period. This data is collected during the initial assessment phase of the well to understand its production potential. Specifically, it includes the following factors: Gas production: measured in cubic meters per day (m³ / day). 3 / d) or 1000 cubic meters / day (×1000m 3 / d). Water production: The unit is usually cubic meters per day (m³ / day). 3 / d) or tons / day (t / d). Trial mining time: The duration of trial mining is usually recorded in days.

[0074] Single-well daily production refers to the daily production of gas and water from a single oil and gas well within a specific time period. Data types include: Daily gas production: unit is cubic meters per day (m³ / day). 3 / d) or 1000 cubic meters / day (×1000m 3 / d). Daily water production: in cubic meters per day (m³ / day) 3 / d) or tons / day (t / d).

[0075] Cumulative gas and water production data refers to the total amount of gas and water extracted since the well began production. This data is used to assess the overall economic value and production potential of oil and gas wells. Data type: Cumulative gas production: unit is cubic meters (m³). 3 ) or 1,000 cubic meters (×1000m 3 Cumulative water production: in cubic meters (m³) 3 ) or tons (t).

[0076] Step 102: Calculate the mutual information between each logging curve of the known well and its corresponding fluid type;

[0077] Mutual information measures the dependency between two variables. In feature selection, mutual information can be used to assess the correlation between a feature and the target variable. If a feature has a high mutual information value with the target variable, it is considered an important feature.

[0078] Step 103: Select the sensitive logging curve of the blind well from the logging curves based on the mutual information.

[0079] Preferably, the top five logging curves with the highest mutual information are selected as the sensitive logging curves.

[0080] In some examples, see Figure 3 The steps for generating the convolutional neural network include:

[0081] Step 301: Extract the second feature data of the sensitive logging curve of the known well and the third feature data of the corresponding fluid type through the convolutional layer of the initial model of the convolutional neural network;

[0082] Convolutional Neural Networks (CNNs) are a class of deep learning models that effectively extract local features from data through local connections and shared weights, and combine them layer by layer into more abstract global features.

[0083] The structure of a CNN includes:

[0084] Convolutional Layer: The core operation is convolution, which uses convolution kernels (filters) to slide across the input data and extract local features. Each convolution kernel produces one feature.

[0085] Activation functions, such as the ReLU (Rectified Linear Unit) function, set negative values ​​in the input to zero while keeping positive values ​​unchanged. Activation functions increase the non-linearity of the model, enabling it to learn complex patterns.

[0086] Pooling Layer: Reduces the volume of data through downsampling operations (such as max pooling or average pooling), thereby reducing computational load and the number of parameters.

[0087] Fully Connected Layer: Similar to the fully connected layer in a traditional neural network, it flattens the features from the previous layer into vectors for classification or regression tasks. The fully connected layer is located at the end of a CNN and is used to generate the final output.

[0088] Output Layer: Based on the specific task (such as classification or regression), it uses an appropriate activation function (such as Softmax for classification and linear activation for regression) to output the results.

[0089] In the convolutional neural network in step 301, convolutional layers and pooling layers are stacked alternately. By extracting and combining local features of the second and third feature data layer by layer, a high-level abstract representation is gradually formed.

[0090] Step 302: Perform forward and backward propagation on the initial model using the second feature data and the third feature data to generate the convolutional neural network.

[0091] The forward propagation process involves processing data from the input layer of a convolutional neural network through various convolutional layers, pooling layers, activation functions, and finally reaching the output layer. The core purpose of forward propagation is to extract features through layer-by-layer computation and ultimately obtain the network's output (such as classification labels or regression values).

[0092] The gradient of the loss function with respect to each parameter is calculated using the backpropagation algorithm, and the parameters are updated using the gradient descent algorithm to gradually optimize the convolutional neural network.

[0093] In some examples, step 301 specifically refers to:

[0094] The convolutional layer performs a convolution operation on the sensitive logging curves of the known well and the corresponding fluid type to extract the second feature data and the third feature data.

[0095] Specifically, a CNN (Convolutional Neural Network) mainly consists of two parts: an encoder and a decoder. The encoder primarily comprises convolutional layers, along with additional modules such as pooling layers, and is the core module for data extraction. The decoder, mainly composed of multiple fully connected layers, is responsible for fitting the data and outputting the results. Here, we mainly use CNN to extract features without performing fitting operations, so only one fully connected layer (Flatten layer) is used. During the forward propagation of the input data in the CNN, the data stream is continuously compressed and dimensionality reduced in the encoder, and its channels (feature maps) continuously increase. Each channel can be seen as a response to a specific abstract feature; the data stream output from the encoder represents the extraction of the abstract features.

[0096] In some examples, see Figure 4 Step 200 includes:

[0097] Step 201: Extract the first feature data of the sensitive logging curve of the blind well using the convolutional neural network;

[0098] Step 202: Identify the fluid type of the blind well using the first feature data and the integrated algorithm model.

[0099] In some examples, see Figure 5 The steps for generating the ensemble algorithm model include:

[0100] Step 401: Generate training data based on the second feature data and the third feature data;

[0101] Step 402: Input the training data into the multiple machine learning models to generate multiple fluid type recognition results;

[0102] This approach borrows from the Boosting concept and applies multiple machine learning models to construct an ensemble algorithm model (ensemble algorithm intelligent interpretation model).

[0103] Preferably, the multiple machine learning models include: Fully Convolutional Networks (FCN), XGBoost (Extreme Gradient Boosting), CatBoost (Categorical Boosting), Decision Tree (DT), and Random Forest (RF) algorithms.

[0104] FCN is an extension of Convolutional Neural Network (CNN). Unlike traditional CNN, FCN uses convolutional layers instead of fully connected layers, allowing the network to accept inputs of arbitrary size and generate outputs with the same spatial dimensions.

[0105] XGBoost is a high-efficiency gradient boosting algorithm widely used for classification and regression tasks on structured data. XGBoost generates decision trees iteratively, with each tree attempting to correct the errors of the previous one, thus progressively improving the model's prediction accuracy. XGBoost incorporates numerous optimizations, such as improvements to the loss function and computation, and regularization to prevent overfitting. Its advantages include: high efficiency, support for parallel computation, handling of large-scale data, and prevention of overfitting (through regularization).

[0106] CatBoost is a gradient boosting tree-based algorithm specifically optimized for categorical features. It effectively handles categorical features without requiring complex preprocessing (such as one-hot encoding). A key feature of CatBoost is its automatic handling of categorical variables and its unique symmetric tree structure, which improves model performance and training speed. Its advantages include optimized handling of categorical features, avoidance of overfitting, and fast training speed. It is particularly suitable for classification and regression problems with a large number of categorical features, a characteristic also found in sensitive logging curves.

[0107] Decision trees (DTs) are a supervised learning algorithm used for classification and regression tasks. They work by partitioning data into distinct subsets, forming a tree-like structure where each internal node represents a feature split, and each leaf node represents a class label (for classification problems) or a predicted value (for regression problems). Decision trees are intuitive and easy to understand, but they are prone to overfitting.

[0108] Random Forest (RF) is an ensemble learning method that improves model accuracy and robustness by constructing multiple decision trees and integrating their results (through voting or averaging). A key feature of Random Forest is that each tree is built using randomly selected samples from the training data, and a subset of features is randomly selected at each node split. This increases model diversity and reduces the risk of overfitting. Its advantages include: integrating multiple decision trees, reducing overfitting, strong robustness, and strong ability to handle missing data.

[0109] Step 403: Input the multiple fluid type identification results into a weak classifier to calculate the error of each of the multiple machine learning models;

[0110] Step 404: Modify the classifier weights of the plurality of machine learning models according to the error;

[0111] Step 405: Generate the ensemble algorithm model based on the multiple machine learning models and their respective classifier weights.

[0112] In steps 403 to 405, the training set samples and initial weights w are first input into the FCN, XGBoost, CatBoost, DT, and RF algorithm modules, respectively. The recognition results of each algorithm module are then input into the weak classifier C. i Then, the error r between the recognition result and the true result is calculated using a weak classifier. i and classifier weights α i At the same time, based on α i Adjust the input sample weights w of the model i until the minimum error r is determined. i Finally, based on the classifier weights of each module, a strong learner is obtained through a weighted combination strategy, which is used to output the final recognition result.

[0113] In some examples, a well logging identification method for fluid types in tight sandstone gas reservoirs further includes cross-validation, Bayesian hyperparameter optimization, blind well testing, and model output after step 405, specifically:

[0114] Cross-validation: To improve the reliability and generalization ability of the model, cross-validation is used to test the model's performance.

[0115] Bayesian hyperparameter optimization: Since there are multiple algorithm modules in the ensemble algorithm model, the model parameters are huge. Manual parameter tuning is inefficient and has poor results. Therefore, the Bayesian hyperparameter intelligent optimization algorithm is used to optimize the model parameters.

[0116] Bayesian hyperparameter optimization is an optimization method based on Bayesian statistical principles, used for hyperparameter optimization and other black-box functions requiring optimization. The advantage of Bayesian optimization lies in its ability to find the optimal hyperparameter configuration with fewer experiments or evaluations, making it particularly effective when the objective function is computationally expensive.

[0117] Bayesian optimization guides the search process by constructing a probabilistic model of the objective function. Unlike traditional grid search and random search, Bayesian optimization does not traverse all possible hyperparameters, but rather infers the possible optimal parameters based on existing information (historical experimental results) and makes decisions accordingly.

[0118] The core idea of ​​Bayesian optimization is to use a surrogate model (preferably a Gaussian process) to approximate the true objective function, and to use the surrogate model to predict the function values ​​in unexplored regions. Then, the hyperparameter combination most likely to bring better results is selected for testing. By continuously iterating and updating the model, the optimal hyperparameter configuration is eventually found.

[0119] Specifically, the main steps of the Bayesian optimization described above include:

[0120] First, a surrogate model needs to be selected, with a Gaussian Process (GP) being the preferred choice. The surrogate model is used to approximate the objective function and estimate the uncertainty at each parameter point. By assuming the objective function is sampled from a Gaussian process, the GP provides a probability distribution for the objective function. It not only predicts the value of the objective function but also gives the uncertainty (variance) of the prediction. The goal of the surrogate model is to fit the objective function to historical data and make predictions for unexplored hyperparameter regions.

[0121] The acquisition function measures which hyperparameter configurations are most worthy of evaluation under the current surrogate model. Its purpose is to strike a balance between exploration and exploitation. Regions with higher uncertainty, i.e., where the model is uncertain, are selected to try new hyperparameter configurations. Regions known to provide good results are selected, i.e., hyperparameter configurations most likely to yield the best results are chosen based on the current knowledge of the surrogate model. After selecting a suitable acquisition function, Bayesian optimization searches the hyperparameter space to find an optimal hyperparameter configuration. This search process is based on the surrogate model's predictions and selects the next evaluation point by optimizing the acquisition function.

[0122] Evaluating the objective function involves running the selected hyperparameter combination on the actual machine learning model and obtaining the model's performance results (such as the error on the validation set). This result is then added to historical data as input for Bayesian optimization. Based on the newly obtained evaluation results, the surrogate model is updated. In this way, the model's predictions and uncertainties gradually become more accurate, and the optimization process becomes more efficient.

[0123] Repeat the above steps, each time selecting the hyperparameter configuration most likely to improve the objective function through the surrogate model for evaluation, until the stopping condition is met (such as reaching the maximum number of evaluations or the performance improvement is no longer significant).

[0124] Blind well testing: In order to evaluate the actual application effect of the model, blind wells in the application research area that were not involved in the training were used to test the algorithm model to ensure that the model meets the application requirements.

[0125] Model output. Output the algorithm model that meets the requirements.

[0126] In some examples, the sensitive logging curve includes at least one of the following: sonic transit time curve, deep lateral resistivity curve, shallow lateral resistivity curve, compensated neutron logging curve, and natural gamma logging curve.

[0127] The sonic log (or acoustic log) is a method of inferring information such as porosity and elastic properties (e.g., elastic modulus) of rock formations by measuring the propagation time of sound waves through them. The sonic log uses sensors placed near the wellbore to send and receive sound signals, thus measuring the propagation time of sound waves in different rock layers. Its functions include:

[0128] Determining porosity: Sound waves propagate more slowly in rock strata with larger pores; therefore, the sound wave transit time is related to the porosity of the rock strata.

[0129] Lithological analysis can help determine the nature of rock strata, such as sandstone and shale.

[0130] Measuring the elasticity of rock strata: The elastic modulus of rock strata can be estimated by measuring the speed of sound wave propagation.

[0131] The deep lateral resistivity log (DLR) is used to infer characteristics of subsurface rock formations, such as water content, porosity, and mineral composition, by measuring their resistivity. In resistivity logging, current flows through the resistance between the wellbore and the rock formation, generating a voltage change that reflects the resistivity of the formation. The DLR measures the resistivity at greater depths to reflect the properties of deeper rocks. Its functions include:

[0132] Determining the water content of rock strata: Resistivity is closely related to the water saturation of rock strata.

[0133] Lithological analysis: related to the mineral composition, porosity and other characteristics of rock strata.

[0134] Shallow lateral resistivity logs are similar to deep lateral resistivity logs, but they measure the resistivity of shallower layers. Due to the shallower measurement depth, they primarily reflect the resistivity characteristics of shallower rock formations. Their main advantage is that they can capture areas closer to the wellbore, thus providing more information for assessing the lithology of groundwater layers or near-wellbore areas. They serve the following purposes:

[0135] Distinguishing between aquifers and oil-bearing layers allows for a more accurate understanding of the water content near the wellbore.

[0136] Short-term reservoir assessment: For shallow oil and gas reservoirs, it can provide more information on lithology and water content.

[0137] Compensated neutron logging assesses the hydrogen content of subsurface rock formations by detecting the backscattering intensity of neutrons after their reaction with hydrogen atoms. Since hydrogen content is highly correlated with water content, compensated neutron logging can effectively determine the water saturation and porosity of rock formations. Its functions include:

[0138] Porosity calculation: Compensated neutron logging is very sensitive to porosity measurement, and is especially suitable for determining the porosity of oil and gas reservoirs.

[0139] Aquifer assessment: It can distinguish between oil and water layers because water and oil have different hydrogen atom contents.

[0140] Natural gamma ray logs (NGR) study the composition of underground rock formations by measuring the intensity of natural gamma rays (gamma rays produced by the decay of radioactive elements in the rock, such as uranium, thorium, and potassium). The intensity of gamma rays is closely related to the radioactive elements in the rock formations; different types of rocks (such as sandstone, shale, and limestone) will have different gamma ray intensities. Its functions include:

[0141] Lithological identification: Shale has a higher natural gamma-ray intensity, while sandstone, carbonate rocks and other rocks have a lower gamma-ray intensity.

[0142] Porosity analysis: It can be used in conjunction with other logging curves (such as sonic transit time curves) to further infer the porosity of rock formations.

[0143] Stratigraphic division: Analysis of gamma-ray data can help determine the interfaces and thicknesses of different rock layers.

[0144] This application provides a well logging method for identifying fluid types in tight sandstone gas reservoirs, comprising: confirming the fluid type in the target area; preprocessing the well logging curves; optimizing a combination of sensitive well logging curves; extracting data features using a CNN; and applying an ensemble algorithm to train, optimize, and select a model for fluid type and well logging curves, ultimately outputting a qualified model. This invention avoids significant differences in the recognition performance of individual algorithm models, significantly improving the overall stability and accuracy of the model's recognition.

[0145] Example 3:

[0146] For further explanation of the plan, see Figure 6 as well as Figure 7 The present invention also provides a specific implementation method for a well logging identification method for fluid types in tight sandstone gas reservoirs, taking data from the tight sandstone gas reservoir in the Xujiahe Formation of a gas field in Sichuan as an example. The specific implementation method includes the following:

[0147] Current traditional methods for identifying fluid types in tight sandstone reservoirs rely too heavily on human factors, resulting in low identification efficiency and poor interpretation. Newer logging technologies such as array acoustic logging and nuclear magnetic resonance logging are expensive, and data acquisition is difficult. Intelligent reservoir fluid identification technology is still underdeveloped, relying solely on core well curve features to construct sensitive curves, limiting its widespread application. Furthermore, integrated algorithm models use a cascaded approach, leading to significant differences between different modules. Specifically, existing technologies for identifying fluid types in tight sandstone gas reservoirs suffer from the following technical problems:

[0148] (1) Conventional logging curve characteristic analysis methods are greatly affected by human factors, resulting in low interpretation accuracy and efficiency. Analysis methods based on array acoustic logging and nuclear magnetic resonance logging cannot be widely applied in the study area due to the high cost of data acquisition and the difficulty in obtaining logging data. The logging identification method for fluid types in tight sandstone gas reservoirs proposed in this invention is easy to acquire, has high identification accuracy and efficiency, and can be widely applied in the study area.

[0149] (2) Existing intelligent reservoir fluid identification methods construct new curves as feature data based on the interpretation results of cored wells, which enhances the identification accuracy of cored wells. However, they overfit the curve features of the cored wells, reducing the generalization ability of the model and making it unsuitable for widespread application. This invention proposes a "CNN feature extraction method," which uses a convolutional neural network to extract deep-level features of the curves based on logging curves, while retaining the original geological features, thereby improving the accuracy of subsequent predictions and the generalization ability of the model.

[0150] (3) Existing integrated intelligent identification methods mostly use a cascaded approach for algorithm modules and weak classifiers, which cannot update the weights of the current algorithm modules in a timely manner, resulting in significant differences in the performance of different modules and failing to meet production requirements. This invention proposes a logging identification method for fluid types in tight sandstone gas reservoirs, integrating algorithm models such as FCN, XGBoost, DT, CatBoost, and RF, and improving the weak classifier to a parallel approach, thereby improving the accuracy of individual modules as well as the accuracy and robustness of the entire model.

[0151] S1: Analyze the fluid type in the target area;

[0152] Specifically, based on the gas testing data of the study area, the possible fluid types in the study area are determined. Since the testing time of a single well is relatively short, it cannot fully reflect the actual situation of the underground reservoir fluids. Combined with production data, including single-well test gas and water data, single-well daily production, and cumulative gas and water production data, a comprehensive judgment is made, and finally the main fluid types of the underground reservoir in the study area are determined.

[0153] First, based on the gas testing data of the study area, the possible fluid types in the study area are determined. Since the testing time of a single well is relatively short, it cannot fully reflect the actual situation of the underground reservoir fluids. Combined with production data, including single-well test gas and water data, single-well daily production, and cumulative gas and water production data, the main fluid types of the underground reservoir in the study area are determined, including gas layers, differential gas layers, gas and water co-existing layers, water layers, and dry layers.

[0154] S2: Preprocess the logging curves;

[0155] Due to geological factors and the influence of logging instruments, there are often slight depth deviations between logging curves and core samples. Furthermore, there are minor differences in the same logging curve data from different wells, which can affect the accuracy of subsequent identification. Therefore, the logging curve data in the study area were standardized and subjected to consistency correction to eliminate the influence of logging instruments and the measurement environment on the logging data.

[0156] Specifically, based on well logging and core data, data processing is performed on the well logging curves, including rock electrical correction, outlier and missing value detection, and standardization. Then, the SelectKBest method is used to screen combinations of well logging sensitivity curves.

[0157] S3: Optimize logging curves and their combinations.

[0158] Based on the fluid type and logging curves, the response characteristics between each fluid type and the logging curves are analyzed, and the SelectKBest method is applied to optimize the combination of sensitive logging curves. Preferably, the combination of sensitive curves consists of five logging curves: AC, RD, RS, CNL, and GR.

[0159] S4: Extract CNN convolutional features.

[0160] A CNN (Convolutional Neural Network) mainly consists of two parts: an encoder and a decoder. The encoder primarily comprises convolutional layers, along with additional modules such as pooling layers, and is the core module for data extraction. The decoder mainly consists of multiple fully connected layers, responsible for fitting the data and outputting the results. This patent primarily uses CNN for feature extraction without performing fitting operations, so only one fully connected layer (Flatten layer) is used. During the forward propagation of the input data in the convolutional neural network, the data stream is continuously compressed and dimensionality reduced in the encoder, and its channels (feature maps) continuously increase. Each channel can be seen as a response to a specific abstract feature; the data stream output from the encoder represents the extraction of the abstract feature. The technical roadmap for CNN convolutional feature extraction is as follows: Figure 8 As shown, the specific contents include the following:

[0161] ① Input Layer: Optimized well logging curve data is used as the input layer data. ② Convolutional Layer Data Feature Extraction: The optimized well logging curves and corresponding fluid types are input into the CNN model, and feature extraction is completed through convolution operations. ③ Pooling Layer Data Feature Optimization: Pooling layers are used to reduce parameter size, lower computational complexity, and optimize data features. ④ Fully Connected Layer Data Feature Mapping: Distributed features of the data are mapped to the sample label space, outputting one-dimensional feature data. ⑤ Feature Data Output: The feature data extracted by the fully connected layer (Flatten layer) is output and used as input to the integrated algorithm model.

[0162] Specifically, a CNN (Convolutional Neural Network) is applied for feature extraction. The optimized logging curves and their corresponding fluid types are used as input data. A sliding window of size t is used to slide downwards, resulting in single-channel two-dimensional data of size 5*t, where t is an odd number. This constructs the two-dimensional data. The data is then fed into the CNN model, sequentially passing through convolutional layers (kernel size 3), pooling layers, and fully connected layers to complete the extraction, optimization, and output of data features. This process uncovers deeper features in the data, improving prediction accuracy. Compared to existing conventional algorithms that directly input logging curve data, using a CNN to extract features from the sample data before inputting them improves the overall recognition rate by an average of 4.61%. Figure 9 ).

[0163] S5: Select and optimize the model.

[0164] Machine learning algorithms were applied to train, optimize, and select models for fluid types and well logging curves. The technical roadmap for algorithm model construction and optimization is shown below. Figure 10 As shown, the specific contents include the following:

[0165] ① Input Data. The well logging curve features and fluid type data extracted by the CNN are used as the model input dataset. ② Model Construction ( Figure 11 Based on the Boosting concept, an integrated algorithm intelligent interpretation model is constructed using algorithms such as FCN, XGBoost, CatBoost, DT, and RF. First, the training set samples and initial weights w are decomposed and input into the FCN, XGBoost, CatBoost, DT, and RF algorithm modules. The recognition results of each algorithm module are then fed into the weak classifier C. i Then, the error r between the recognition result and the true result is calculated using a weak classifier. i and classifier weights α i At the same time, based on α i Adjust the weights w of the model input samples i until the minimum error r is determined. iFinally, based on the classifier weights of each module, a strong learner is obtained through a weighted combination strategy, which is used to output the final recognition result.

[0166] Specifically, model training is completed based on features extracted by CNN. This example model primarily employs five algorithm modules: FCN, XGBoost, CatBoost, DT, and RF. First, the feature data extracted by CNN is input into the ensemble algorithm model, along with the training set data and initialized sample weights w. i First, the system is trained using five algorithm modules, including parameter optimization and cross-validation of the modules. Then, the prediction results from each module are fed into the next-level weak classifier, which is then evaluated based on the prediction error r. i Adjust its own weight α i With sample weight w i This process is repeated until the minimum error of the model is found; finally, the minimum error is determined based on the weights α of each weak classifier. i The prediction results are output through a strong classifier.

[0167] S6: Output model.

[0168] Output a qualified fluid type identification ensemble algorithm model. Then apply the trained ensemble algorithm model to other wells in the study area and verify it through blind wells to determine the model's identification effect and generalization ability.

[0169] Cross-validation with blind wells in the study area showed that the final model achieved an average recognition accuracy of 85.7%, with the highest accuracy exceeding 90%, significantly higher than existing AdaBoost ensemble models of the same type, effectively improving recognition accuracy. Furthermore, the cross-validation results were consistently good, demonstrating strong generalization ability and making the model widely applicable to reservoir fluid interpretation in non-gas-testing zones within the study area (see [reference]). Figure 12 as well as Figure 13 ).

[0170] This application provides a well logging method for identifying fluid types in tight sandstone gas reservoirs, comprising: first, acquiring sensitive logging curves of blind wells in tight sandstone gas reservoirs; wherein the fluid type of the blind wells is unknown; then, identifying the fluid type of the blind wells based on the sensitive logging curves and a pre-generated fluid determination model; wherein the fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model, the convolutional neural network is used to extract first feature data from the sensitive logging curves of the blind wells, the ensemble model consists of multiple machine learning models, and is used to identify the fluid type of the blind wells based on the first feature data; the convolutional neural network includes a fully connected layer. Specifically, this invention has the following beneficial effects:

[0171] (1) This invention uses a CNN (Convolutional Neural Network) to extract features from sample data, effectively uncovering the complex mapping relationship between well logging curves and reservoir fluids, obtaining real and effective feature information, and further improving the accuracy of identification. In this study area, compared with existing conventional algorithm models that directly input well logging curve data, using a CNN to extract features from sample data before inputting the data improves the overall recognition rate by an average of 4.61%. Figure 9 ).

[0172] (2) This invention proposes a method for identifying fluid logging in tight sandstone gas reservoirs based on a CNN feature engineering fusion Boosting integrated algorithm model. This method can effectively address the identification problem of different reservoir fluid types. Compared with existing research schemes, the identification accuracy of a single algorithm module is higher, avoiding significant differences in the identification effect of a single algorithm module. In addition, the overall average identification rate of the model is also higher than that of the existing AdaBoost integrated algorithm model, which significantly improves the overall identification effect of the model. It can be widely applied to the interpretation of reservoir fluids in non-testing gas sections of the study area.

[0173] Example 4:

[0174] Another embodiment of this application relates to a logging identification device for fluid types in tight sandstone gas reservoirs. The implementation details of this logging identification device for fluid types in tight sandstone gas reservoirs are described below. The following details are provided for ease of understanding and are not essential for implementing this solution. A schematic diagram of the logging identification device for fluid types in tight sandstone gas reservoirs in this embodiment can be seen as follows: Figure 14 As shown, there is a sensitive logging curve acquisition module 801 and a fluid type identification module 802.

[0175] Sensitive logging curve acquisition module 801 is used to acquire sensitive logging curves of blind wells in the tight sandstone gas reservoir; wherein the fluid type of the blind well is unknown;

[0176] The fluid type identification module 802 is used to identify the fluid type of the blind well based on the sensitive logging curve of the blind well and a pre-generated fluid determination model; wherein, the fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model, the convolutional neural network is used to extract the first feature data of the sensitive logging curve of the blind well, the ensemble algorithm model is composed of multiple machine learning models, and is used to identify the fluid type of the blind well based on the first feature data, and the convolutional neural network includes a fully connected layer.

[0177] In some embodiments, the sensitive logging curve acquisition module 801 includes:

[0178] The known well fluid determination unit is used to determine the fluid type of the known well based on the fluid test data and production data of the known well in the tight sandstone gas reservoir.

[0179] A mutual information calculation unit is used to calculate the mutual information between each logging curve of the known well and its corresponding fluid type.

[0180] A sensitive logging curve acquisition unit is used to select the sensitive logging curve of the blind well based on the mutual information among the logging curves.

[0181] In some embodiments, a logging identification device for fluid types in tight sandstone gas reservoirs further includes:

[0182] A neural network generation module is used to generate the convolutional neural network; the neural network generation module includes:

[0183] The second and third feature data extraction units are used to extract the second feature data of the sensitive logging curve of the known well and the third feature data of the corresponding fluid type through the convolutional layer of the initial model of the convolutional neural network.

[0184] The neural network generation unit is used to perform forward and backward propagation on the initial model using the second feature data and the third feature data to generate the convolutional neural network.

[0185] In some embodiments, the second and third feature data extraction units include:

[0186] The second and third feature data extraction subunits are used to perform convolution operations on the sensitive logging curves of the known well and the corresponding fluid types through the convolution layer to extract the second and third feature data.

[0187] In some embodiments, the fluid type identification module 802 includes:

[0188] The first feature data extraction unit is used to extract the first feature data of the sensitive logging curve of the blind well through the convolutional neural network;

[0189] A blind well fluid identification unit is used to identify the fluid type of the blind well using the first feature data and the integrated algorithm model.

[0190] In some embodiments, a logging identification device for fluid types in tight sandstone gas reservoirs further includes:

[0191] An ensemble algorithm model generation module is used to generate the ensemble algorithm model; the ensemble algorithm generation module includes:

[0192] A training data generation unit is used to generate training data based on the second feature data and the third feature data;

[0193] A multi-recognition result generation unit is used to input the training data into the multiple machine learning models to generate multiple fluid type recognition results;

[0194] A multi-error calculation unit is used to input the multiple fluid type identification results into a weak classifier to calculate the errors of each of the multiple machine learning models;

[0195] A classifier weight modification unit is used to modify the classifier weights of the plurality of machine learning models according to the error;

[0196] An ensemble algorithm generation unit is used to generate the ensemble algorithm model based on the multiple machine learning models and their respective classifier weights.

[0197] In some embodiments, the sensitive logging curve includes at least one of the following: sonic transit time curve, deep lateral resistivity curve, shallow lateral resistivity curve, compensated neutron logging curve, and natural gamma logging curve.

[0198] This application provides a well logging identification device for fluid types in tight sandstone gas reservoirs, comprising: a sensitive logging curve acquisition module for acquiring sensitive logging curves of blind wells in tight sandstone gas reservoirs, wherein the fluid type of the blind wells is unknown; and a fluid type identification module for identifying the fluid type of the blind wells based on the sensitive logging curves and a pre-generated fluid determination model; wherein the fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model, the convolutional neural network is used to extract first feature data from the sensitive logging curves of the blind wells, the ensemble model consists of multiple machine learning models, and is used to identify the fluid type of the blind wells based on the first feature data; the convolutional neural network includes a fully connected layer. This invention has the following beneficial effects:

[0199] (1) This invention uses a CNN convolutional neural network to extract features from sample data, which can effectively uncover the complex mapping relationship between well logging curves and reservoir fluids, obtain real and effective feature information, and further improve the accuracy of recognition. In this study area, compared with the existing conventional algorithm model that directly inputs well logging curve data, the overall recognition rate of the model is significantly improved by using a CNN convolutional neural network to extract features from the sample data before input.

[0200] (2) This invention can effectively address the identification problem of different reservoir fluid types. Compared with existing research schemes, the identification accuracy of a single algorithm module is higher, avoiding significant differences in the identification effect of a single algorithm module. In addition, the overall average identification rate of the model is also higher than that of the existing AdaBoost ensemble algorithm model, which significantly improves the overall identification effect of the model. It can be widely applied to the interpretation of reservoir fluids in non-test gas layers in the research area.

[0201] It is worth mentioning that all modules involved in this embodiment are logical modules. In practical applications, a logical unit can be a physical unit, a part of a physical unit, or a combination of multiple physical units. Furthermore, to highlight the innovative aspects of this application, this embodiment does not introduce units that are not closely related to solving the technical problems proposed in this application; however, this does not mean that other units are absent in this embodiment.

[0202] Example 5:

[0203] Another embodiment of this application relates to an electronic device, such as... Figure 15 As shown, the electronic device specifically includes the following:

[0204] Processor 1201, memory 1202, communications interface 1203, and bus 1204;

[0205] The processor 1201, memory 1202, and communication interface 1203 communicate with each other via bus 1204; the communication interface 1203 is used to realize information transmission between server-side devices and user-side devices and other related devices.

[0206] The processor 1201 is used to call the computer program in the memory 1202. When the processor executes the computer program, it implements all the steps in the well logging identification method for fluid types in tight sandstone gas reservoirs in the above embodiments. For example, when the processor executes the computer program, it implements the following steps:

[0207] Sensitive logging curves of blind wells in the tight sandstone gas reservoir are obtained; wherein the fluid type of the blind well is unknown.

[0208] The fluid type of the blind well is identified based on the sensitive logging curve of the blind well and a pre-generated fluid determination model. The fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model. The convolutional neural network is used to extract the first feature data of the sensitive logging curve of the blind well. The ensemble algorithm model consists of multiple machine learning models and is used to identify the fluid type of the blind well based on the first feature data. The convolutional neural network includes a fully connected layer.

[0209] In some embodiments, obtaining the sensitive logging curves of blind wells in the tight sandstone gas reservoir includes:

[0210] The fluid type of the known wells is determined based on fluid test data and production data from known wells within the tight sandstone gas reservoir;

[0211] Calculate the mutual information between each logging curve of the known well and its corresponding fluid type;

[0212] Among the logging curves, the sensitive logging curve for the blind well is selected based on the mutual information.

[0213] In some embodiments, the step of generating the convolutional neural network includes:

[0214] The second feature data of the sensitive logging curve of the known well and the third feature data of the corresponding fluid type are extracted through the convolutional layer of the initial model of the convolutional neural network.

[0215] The initial model is forward-propagated and backward-propagated using the second feature data and the third feature data to generate the convolutional neural network.

[0216] In some embodiments, the second feature data of the sensitive logging curve of the known well and the third feature data of the corresponding fluid type are extracted through the convolutional layers of the initial model of the convolutional neural network, including:

[0217] The convolutional layer performs a convolution operation on the sensitive logging curves of the known well and the corresponding fluid type to extract the second feature data and the third feature data.

[0218] In some embodiments, identifying the fluid type of the blind well based on the sensitive logging curves of the blind well and a pre-generated fluid determination model includes:

[0219] The first feature data of the sensitive logging curve of the blind well is extracted by the convolutional neural network;

[0220] The fluid type of the blind well is identified using the first feature data and the integrated algorithm model.

[0221] In some embodiments, the step of generating the ensemble algorithm model includes:

[0222] Training data is generated based on the second feature data and the third feature data;

[0223] The training data is input into the multiple machine learning models to generate multiple fluid type recognition results;

[0224] The results of the multiple fluid type identifications are input into a weak classifier to calculate the errors of each of the multiple machine learning models.

[0225] The classifier weights of the multiple machine learning models are modified based on the error.

[0226] The ensemble algorithm model is generated based on the multiple machine learning models and their respective classifier weights.

[0227] In some embodiments, the sensitive logging curve includes at least one of the following: sonic transit time curve, deep lateral resistivity curve, shallow lateral resistivity curve, compensated neutron logging curve, and natural gamma logging curve.

[0228] The memory and processor are connected via a bus, which can include any number of interconnecting buses and bridges, connecting various circuits of one or more processors and memories. The bus can also connect various other circuits, such as peripheral devices, voltage regulators, and power management circuits, which are well known in the art and will not be described further herein. The bus interface provides an interface between the bus and the transceiver. The transceiver can be a single element or multiple elements, such as multiple receivers and transmitters, providing a unit for communicating with various other devices over a transmission medium. Data processed by the processor is transmitted over the wireless medium via an antenna, which further receives data and transmits it to the processor.

[0229] The processor manages the bus and general processing, and also provides various functions, including timing, peripheral interfaces, voltage regulation, power management, and other control functions. Memory is used to store data used by the processor during operation.

[0230] Example 6:

[0231] Another embodiment of this application relates to a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps in the above-described embodiment of the well logging identification method for fluid types in tight sandstone gas reservoirs, the steps including:

[0232] Sensitive logging curves of blind wells in the tight sandstone gas reservoir are obtained; wherein the fluid type of the blind well is unknown.

[0233] The fluid type of the blind well is identified based on the sensitive logging curve of the blind well and a pre-generated fluid determination model. The fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model. The convolutional neural network is used to extract the first feature data of the sensitive logging curve of the blind well. The ensemble algorithm model consists of multiple machine learning models and is used to identify the fluid type of the blind well based on the first feature data. The convolutional neural network includes a fully connected layer.

[0234] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. In particular, hardware + program embodiments are relatively simple in description because they are fundamentally similar to method embodiments; relevant parts can be referred to the descriptions in the method embodiments.

[0235] The foregoing has described specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims may be performed in a different order than that shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are possible or may be advantageous.

[0236] While this application provides method operation steps as shown in the embodiments or flowcharts, more or fewer operation steps may be included based on conventional or non-inventive labor. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only execution order. In actual device or client product execution, the method can be executed sequentially as shown in the embodiments or drawings, or in parallel (e.g., in a parallel processor or multi-threaded processing environment).

[0237] 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 1One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0238] 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.

[0239] 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.

[0240] Specific embodiments have been used to illustrate the principles and implementation methods of this invention. The descriptions of the embodiments above are only for the purpose of helping to understand the method and core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A well logging method for identifying fluid types in tight sandstone gas reservoirs, characterized in that, include: Sensitive logging curves of blind wells in the tight sandstone gas reservoir are obtained; wherein the fluid type of the blind well is unknown. The fluid type of the blind well is identified based on the sensitive logging curve of the blind well and a pre-generated fluid determination model. The fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model. The convolutional neural network is used to extract the first feature data of the sensitive logging curve of the blind well. The ensemble algorithm model consists of multiple machine learning models and is used to identify the fluid type of the blind well based on the first feature data. The convolutional neural network includes a fully connected layer.

2. The well logging identification method according to claim 1, characterized in that, The acquisition of sensitive logging curves for blind wells in the tight sandstone gas reservoir includes: The fluid type of the known wells is determined based on fluid test data and production data from known wells within the tight sandstone gas reservoir; Calculate the mutual information between each logging curve of the known well and its corresponding fluid type; Among the logging curves, the sensitive logging curve for the blind well is selected based on the mutual information.

3. The well logging identification method according to claim 2, characterized in that, The steps for generating the convolutional neural network include: The second feature data of the sensitive logging curve of the known well and the third feature data of the corresponding fluid type are extracted through the convolutional layer of the initial model of the convolutional neural network. The initial model is forward-propagated and backward-propagated using the second feature data and the third feature data to generate the convolutional neural network.

4. The well logging identification method according to claim 3, characterized in that, The second feature data of the sensitive logging curves of the known well and the third feature data of the corresponding fluid type are extracted through the convolutional layers of the initial model of the convolutional neural network, including: The convolutional layer performs a convolution operation on the sensitive logging curves of the known well and the corresponding fluid type to extract the second feature data and the third feature data.

5. The well logging identification method according to claim 3, characterized in that, Identifying the fluid type of the blind well based on its sensitive logging curves and a pre-generated fluid determination model includes: The first feature data of the sensitive logging curve of the blind well is extracted by the convolutional neural network; The fluid type of the blind well is identified using the first feature data and the integrated algorithm model.

6. The well logging identification method according to claim 3, characterized in that, The steps for generating the ensemble algorithm model include: Training data is generated based on the second feature data and the third feature data; The training data is input into the multiple machine learning models to generate multiple fluid type recognition results; The results of the multiple fluid type identifications are input into a weak classifier to calculate the errors of each of the multiple machine learning models. The classifier weights of the multiple machine learning models are modified based on the error. The ensemble algorithm model is generated based on the multiple machine learning models and their respective classifier weights.

7. The well logging identification method according to any one of claims 1 to 6, characterized in that, The sensitive logging curves include at least one of the following: sonic transit time curve, deep lateral resistivity curve, shallow lateral resistivity curve, compensated neutron logging curve, and natural gamma logging curve.

8. A logging identification device for fluid types in tight sandstone gas reservoirs, characterized in that, include: A sensitive logging curve acquisition module is used to acquire sensitive logging curves of blind wells in the tight sandstone gas reservoir; wherein the fluid type of the blind well is unknown. A fluid type identification module is used to identify the fluid type of the blind well based on the sensitive logging curve of the blind well and a pre-generated fluid determination model. The fluid determination model is generated by a pre-generated convolutional neural network and an ensemble algorithm model. The convolutional neural network is used to extract the first feature data of the sensitive logging curve of the blind well. The ensemble algorithm model consists of multiple machine learning models and is used to identify the fluid type of the blind well based on the first feature data. The convolutional neural network includes a fully connected layer.

9. An electronic device, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the logging identification method for fluid types in tight sandstone gas reservoirs as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the logging identification method for fluid types in tight sandstone gas reservoirs as described in any one of claims 1 to 7.

Images