A method for predicting total organic carbon in formations based on parameter fusion

Abstract

The present invention discloses a method for predicting formation total organic carbon (TOC) based on parameter fusion, which relates to the technical field of petroleum exploration and development. The method comprises: obtaining the measured TOC value and multiple logging parameters of an exploration well, and using the logging parameters whose correlation coefficients with the measured TOC value are greater than a set threshold value as characteristic logging parameters; constructing an improved deep neural network, inputting the characteristic logging parameters into the improved deep neural network to obtain a predicted TOC value; training the improved deep neural network based on the error between the predicted TOC value and the measured TOC value; and, during the TOC prediction process, selecting characteristic logging parameters from the multiple logging parameters of the exploration well to be tested and inputting them into the trained improved deep neural network to obtain a TOC prediction result. The present invention enhances the decoupling capability of multi-factor coupling effects, effectively removes outliers, and improves data quality.

Description

Technical Field

[0001] The present invention relates to the technical field of petroleum exploration and development, and in particular to a method for predicting total organic carbon in formations based on parameter fusion. Background Art

[0002] The prediction of total organic carbon (TOC) in formations is necessary in many fields. It can improve oil and gas exploration efficiency, optimize development plans, assist in geological research and resource assessment, enhance corporate operational efficiency and management level, promote scientific research and technological progress, and provide strong support for the development of related fields.

[0003] Traditionally, TOC prediction relies on complex pyrolysis experiments on core or cuttings samples, a process that is not only time-consuming and labor-intensive, but also costly. Among existing studies, Patent CN117192613A discloses a TOC prediction method and device based on lithologic control. Patent C116859467A discloses a TOC prediction method for source rocks. Patent CN15877463A discloses a TOC prediction method for shale gas reservoirs, and Patent CN114943060A discloses a shale gas total organic carbon prediction method based on deep learning and interpolation regression. Patent CN13835138A discloses a method for predicting the total organic carbon content of shale based on a deep encoding and decoding network. Patent CN13050191A discloses a shale oil TOC prediction method and device based on dual parameters. Patent CN12147713A discloses a segmented prediction method for the total organic carbon content of mudstone. In the Journal of China University of Petroleum, Vol. 41, No. 4, 2017, Chen Haifeng et al. conducted a detailed TOC evaluation of source rocks based on the variable coefficient logarithmic resistivity method. Using the variable coefficient logarithmic resistivity method to evaluate TOC in source rocks in this area, they considered the empirical parameters in the logarithmic resistivity method as undetermined coefficients. Based on the geological significance of the model parameters, they conducted TOC logging evaluations for 120 wells in the entire area, targeting undetermined parameters in both single wells and the "zoning and phase" prediction model. They used logging to calculate TOC for hierarchical source rock evaluation. Using the "drilling-determined frequency, sequence-determined distribution" method, they determined the thickness of source rocks of different abundance levels in the first member of the Nan Formation. In the Journal of Petroleum Geology and Recovery, Vol. 25, No. 4, 2018, Bian Leibo et al. conducted an optimized logarithmic resistivity method and its application to predicting total organic carbon content in medium- to deep-seated source rocks. They optimized the logarithmic resistivity method using a normalization method and introduced the natural gamma parameter to establish an optimized total organic carbon content prediction model. In Volume 34, Issue 3, 2019, Progress in Geophysics, Chen Hao et al. optimized and applied well logging prediction methods for TOC content in source rocks of the Juyanhai Depression in the Yin'e Basin. Combining the geological characteristics of the Juyanhai Depression with relevant logging parameters, they conducted a comparative error analysis of the prediction results from different methods and concluded that the multivariate regression method was the best choice for predicting TOC content in source rocks in the study area. They also proposed a multivariate regression prediction model suitable for this study area, using the ∆logR parameter (a relevant variable and coefficient in the method for identifying and calculating total organic carbon in formations using logging data), deep lateral resistivity, and natural gamma ray as basic parameters. In Volume 36, Issue 1, 2021, Progress in Geophysics, Lu Pengyu et al. predicted organic carbon content in the Lunpola Basin using a neural network approach.

[0004] In the existing technology, the implementation process of the TOC prediction method based on logging data includes: using logging instruments to measure in the target well section to obtain various logging data including natural gamma, resistivity, and acoustic time difference; correcting and filtering the collected raw logging data to remove noise and outliers; based on existing geological knowledge and sample data, using statistical analysis, machine learning and other methods to establish a relationship model between logging data and TOC; inputting the pre-processed logging data into the established model to calculate the predicted value of TOC.

[0005] Formations are characterized by diverse lithologies and strong heterogeneity, which influence logging responses to varying degrees. This complicates the relationship between logging data and TOC, making it difficult to accurately model. For example, in shale formations, the mineral composition and organic matter types and contents vary widely, affecting logging response characteristics and leading to prediction errors. Summary of the Invention

[0006] Based on this, it is necessary to provide a method for predicting total organic carbon in formations based on parameter fusion to address the above technical issues.

[0007] The embodiment of the present invention provides a method for predicting total organic carbon in a formation based on parameter fusion, comprising:

[0008] Obtaining a measured value of formation total organic carbon and multiple logging parameters of an exploration well, and using a logging parameter whose correlation coefficient with the measured value of formation total organic carbon is greater than a set threshold as a characteristic logging parameter of formation total organic carbon;

[0009] Connecting a feature mapping layer and a fusion layer between the input layer and the first hidden layer of the deep neural network to obtain an improved deep neural network;

[0010] The characteristic logging parameters are input into the improved deep neural network, and the ∆logR parameter is calculated through the input layer. The ∆logR parameter is nonlinearly transformed through the feature mapping layer and mapped into the logging feature. The logging feature and the characteristic logging parameters are weighted and fused through the fusion layer to obtain the fused data. The fused data is used to make a prediction to obtain the predicted value of total organic carbon in the formation.

[0011] According to the error between the predicted value of total organic carbon in the formation and the measured value of total organic carbon in the formation, the improved deep neural network is trained to obtain the trained improved deep neural network;

[0012] In the process of predicting total organic carbon in the formation, multiple logging parameters of the exploration well to be tested are obtained, and characteristic logging parameters are selected from the multiple logging parameters and input into the trained improved deep neural network to obtain the total organic carbon prediction results of the exploration well to be tested.

[0013] Optionally, the ∆logR parameter is calculated by improving the input layer of the deep neural network, specifically by:

[0014] The ∆logR parameter is calculated by improving the input layer of the deep neural network based on the following formula:

[0015]

[0016] Where RT is the resistivity, is the resistivity baseline, AC is the acoustic time difference, It is the acoustic time difference baseline.

[0017] Optionally, the ∆logR parameter is nonlinearly transformed through the feature mapping layer to map the ∆logR parameter to the well logging feature, specifically including

[0018] The feature mapping layer adopts a multi-layer perceptron structure and processes the ∆logR parameter using nonlinear transformation;

[0019] The hidden layer neurons of the multi-layer perceptron structure perform weighted summation and activation function operation on the ∆logR parameters after nonlinear transformation, and map the ∆logR parameters to well logging features.

[0020] Optionally, the improved deep neural network includes multiple hidden layers, and the output layer is connected in series after the last hidden layer; the activation function of the output layer of the improved deep neural network is a scaled exponential linear unit;

[0021] By improving the multiple hidden layers of the deep neural network, the fused data are processed layer by layer, and the extracted features are combined in the last hidden layer to obtain the global features; by improving the output layer of the deep neural network, the global features are operated by scaled exponential linear units for prediction, and the predicted value of total organic carbon in the formation is obtained.

[0022] Optionally, the improved deep neural network is trained based on the error between the predicted value of the total organic carbon in the formation and the measured value of the total organic carbon in the formation, specifically including:

[0023] Based on the predicted and measured values ​​of total organic carbon in the formation, an adaptive moment estimation optimization function was constructed to train the improved deep neural network.

[0024] Optionally, the method further includes: determining, for each characteristic well logging parameter, an SVM parameter of a support vector machine algorithm according to parameter distribution characteristics of the characteristic well logging parameter, so as to train a classifier of the support vector machine algorithm;

[0025] Outliers that deviate from the normal data distribution pattern are identified and eliminated through the trained support vector machine algorithm classifier.

[0026] Optionally, before inputting the characteristic logging parameters into the improved deep neural network, the characteristic logging parameters are smoothed using an exponentially weighted moving average, and the formula is expressed as:

[0027]

[0028] Among them, y[t] represents the smoothed value at the tth moment, x[t] represents the observed value at the tth moment, is the smoothing coefficient.

[0029] The above-mentioned method for predicting total organic carbon in formations based on parameter fusion provided by the embodiment of the present invention has the following beneficial effects compared with the prior art:

[0030] Based on the correlation coefficients between various logging parameters and the measured values ​​of formation total organic carbon, the present invention uses logging parameters with correlation coefficients greater than a set threshold as characteristic logging parameters of formation total organic carbon. The ∆logR parameter is then fused with the characteristic logging parameters through parameter mapping and weighted summation. This enhances the ability to decouple the effects of multiple factors and can remove logging parameters that are weakly correlated with or irrelevant to formation total organic carbon prediction, thus resolving the problem in the prior art where different lithologies have varying degrees of influence on logging responses. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A method flow chart of a method for predicting total organic carbon in a formation based on parameter fusion provided in one embodiment

[0032] Figure 2 A flowchart of an implementation method for predicting total organic carbon in a formation based on parameter fusion provided in one embodiment;

[0033] Figure 3 A DNN model framework diagram of a method for predicting total organic carbon in formations based on parameter fusion provided in one embodiment;

[0034] Figure 4 A TOC prediction scatter plot of a DNN model of a method for predicting total organic carbon in formations based on parameter fusion provided in one embodiment;

[0035] Figure 5 This is a TOC prediction scatter plot of the DNN fusion ∆logR parameter of a formation total organic carbon prediction method based on parameter fusion provided in one embodiment. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0037] Currently, there are the following problems in the field of TOC prediction:

[0038] (1) Traditional methods are complex and costly: Traditional pyrolysis experimental methods that rely on core or cuttings samples are not only time-consuming and labor-intensive, but also come with high costs, which greatly limits their application in a wide range of exploration activities.

[0039] (2) Limited prediction accuracy: Although a number of patents and research results have proposed TOC prediction methods based on well logging data, these methods still cannot meet the needs of high-precision exploration in terms of prediction accuracy, especially under complex geological conditions.

[0040] (3) Parameter singleness and strong subjectivity: Existing prediction methods mostly rely on single or limited logging parameters, such as ∆logR parameter, deep lateral resistivity and natural gamma, which may lead to the prediction results being greatly affected by parameter selection. There is also a lack of systematic standards for handling outliers, and human subjectivity is strong, which affects the stability and reliability of the prediction results.

[0041] In one embodiment, a method for predicting total organic carbon in a formation based on parameter fusion is provided, such as Figure 1 As shown, the method includes:

[0042] 1. Obtain the measured formation total organic carbon value and multiple logging parameters of the exploration well. The logging parameter whose correlation coefficient with the measured formation total organic carbon value is greater than a set threshold is used as the characteristic logging parameter of the formation total organic carbon.

[0043] For each characteristic logging parameter, the SVM parameters of the support vector machine algorithm are determined based on the parameter distribution characteristics of the characteristic logging parameter to train the support vector machine algorithm classifier. The trained support vector machine algorithm classifier identifies and eliminates outliers that deviate from the normal data distribution pattern.

[0044] 2. If Figure 3 As shown in FIG, the feature mapping layer and the fusion layer are connected between the input layer and the first hidden layer of the deep neural network to obtain an improved deep neural network.

[0045] The feature mapping layer uses a multi-layer perceptron structure to process the ∆logR parameters using nonlinear transformations. The hidden layer neurons in the multi-layer perceptron structure perform weighted summation and activation function operations on the ∆logR parameters after nonlinear transformation, mapping the ∆logR parameters into well logging features.

[0046] 3. Input the characteristic logging parameters into the improved deep neural network. The ∆logR parameter is calculated through the input layer. The ∆logR parameter is nonlinearly transformed in the feature mapping layer, mapping it to logging features. The fusion layer then performs a weighted summation of the logging features and characteristic logging parameters to generate fused data. Based on this fused data, a prediction is performed to obtain the predicted total organic carbon value for the formation.

[0047] The improved deep neural network consists of multiple hidden layers, with an output layer connected in series after the last hidden layer. The activation function of the output layer of the improved deep neural network is a scaled exponential linear unit. The fused data is processed layer by layer through the multiple hidden layers of the improved deep neural network, and the extracted features are combined in the last hidden layer to obtain a global feature. The output layer of the improved deep neural network performs a scaled exponential linear unit operation on the global feature to obtain a predicted value of total organic carbon in the formation.

[0048] 4. According to the error between the predicted value of total organic carbon in the formation and the measured value of total organic carbon in the formation, the improved deep neural network is trained to obtain the trained improved deep neural network.

[0049] Based on the predicted and measured values ​​of total organic carbon in the formation, an adaptive moment estimation optimization function was constructed to train the improved deep neural network.

[0050] 5. In the process of total organic carbon prediction of the formation, multiple logging parameters of the exploration well to be tested are obtained, and characteristic logging parameters are selected from the multiple logging parameters and input into the trained improved deep neural network to obtain the total organic carbon prediction results of the exploration well to be tested.

[0051] like Figure 2 As shown, a specific embodiment of the present invention is provided:

[0052] Step 1: Data selection and preprocessing

[0053] Data Collection: First, we collected detailed logging data from multiple exploration wells, including but not limited to acoustic transit time (AC), gamma ray (GR), resistivity (RT), density (DEN), and neutron lifetime (CNL). We also ensured that each well had a corresponding measured TOC value as tag data.

[0054] Correlation Analysis: Statistical analysis methods (such as the Pearson correlation coefficient) were used to calculate the correlation between each logging parameter and TOC. Based on the results of this correlation analysis, logging parameters with significant correlations with TOC (such as AC, GR, RT, DEN, and CNL in this example) were selected as core input data for building the machine learning model. Specifically, the correlation coefficients between AC and GR both exceeded 0.5, indicating a strong correlation; RT followed closely at 0.47, showing a relatively strong association; DEN's 0.33 indicated a moderate correlation; and CNL's -0.34 revealed a negative correlation with TOC. This unique characteristic was also taken into account to fully capture the multi-dimensional relationships between the data.

[0055] Data preprocessing: A support vector machine (SVM) algorithm is used to detect and process outliers in the filtered well logging data. Specifically, an SVM classifier is trained to identify and remove outliers that significantly deviate from the normal data distribution pattern. This process can be achieved by setting appropriate SVM parameters (such as the kernel function and penalty coefficient) to ensure accurate and efficient outlier detection.

[0056] After using the SVM algorithm to remove outliers, the performance of the XGBoost model (Extreme Gradient Boosting) was significantly improved. MAE decreased from 1.6911 to 1.4271, RMSE decreased from 3.4876 to 2.4503, and R 2 It increased from 0.5484 to 0.7609. These indicators show that the prediction accuracy and stability of the model have been significantly improved.

[0057] Similarly, the SVM algorithm also has a significant effect in removing outliers in the Deep Neural Network (DNN) model. The MAE of the DNN model slightly decreased from 1.0551 to 0.9806, and the RMSE slightly decreased from 1.8857 to 1.81, while the R 2 The data was smoothed using an exponentially weighted moving average (EWMA). The EWMA formula is as follows:

[0058]

[0059] Where y[t] represents the smoothed value at the tth moment, and x[t] represents the observed value at the tth moment. is the smoothing coefficient, which determines the weight of the observation in the calculation. The value range is [0,1], and the value in this example is 0.3.

[0060] Step 2: DNN model construction and training

[0061] Building a DNN model: This carefully designed model consists of an input layer, four hidden layers, and an output layer. Each hidden layer contains 64 and 128 neurons, respectively, providing the model with ample processing power. The input layer receives preprocessed well logging data, which serves as the foundation for the model's learning and predictions.

[0062] Hidden layers are the core of DNN models. They process input data layer by layer through nonlinear transformations, extracting features from the data. These features begin with low-level, local information, such as the most basic and direct elements in the data. Layer by layer, through continuous learning and processing, they are gradually combined and refined into high-level, global features. This process of feature extraction and combination is one of the unique advantages and characteristics of DNN models.

[0063] At the model's output layer, high-level features processed by the hidden layers are used to predict TOC values. To enhance the model's nonlinear representation capabilities, a scaled exponential linear unit (SELU) activation function was chosen, which helps the model better capture complex patterns in the data.

[0064] In terms of optimization, Adam (Adaptive Moment Estimation) was selected as the optimization function. It combines the advantages of the momentum method and the RMSProp method, can efficiently update the model weights, and enable the model to converge quickly during training.

[0065] The multi-layered architecture of DNNs not only enables the model to systematically extract and combine features layer by layer, but also greatly enhances the model's expressiveness and generalization capabilities. This shift from local to global, from low-level to high-level features, enables DNNs to demonstrate exceptional performance in handling complex tasks. Whether for deep data mining or accurate prediction of unknown data, DNNs have demonstrated tremendous strength and potential.

[0066] Comparative Model Construction: Four machine learning models were used for comparison: K-Nearest Neighbors (KNN), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and XGBoost. Each model was trained using preprocessed well logging data as input and measured TOC values ​​as output. The number of neighbors was set in the KNN model; the number of trees and the tree depth were set to 8 in the RF model; and the learning rate was set to 0.01 and the tree depth was set to 10 in the GBDT and XGBoost models. These parameters can be optimized using methods such as cross-validation or grid search.

[0067] Cross-validation: A ten-fold cross-validation strategy was used to train and validate the above models. Specifically, the preprocessed data was randomly divided into ten subsets. Nine of these subsets were selected as training sets, and the remaining subset was used as the validation set for model evaluation. This process was repeated ten times, using a different validation set each time, to comprehensively evaluate the generalization ability and predictive performance of each model.

[0068] Step 3: ∆logR parameter fusion and prediction

[0069] ∆logR parameter calculation: The traditional ∆logR parameter is calculated based on well logging data. This parameter reflects the abundance of organic matter in the formation. The ∆logR parameter calculation formula is as follows:

[0070]

[0071] Where RT is the resistivity, is the resistivity baseline, AC is the acoustic time difference, The acoustic time difference baseline

[0072] Feature fusion and model improvement:

[0073] Constructing a ∆logR-based feature mapping layer: Design a dedicated feature mapping layer to input the calculated ∆logR parameters. Within this layer, nonlinear transformations are applied to the ∆logR parameters. Using a multilayer perceptron (MLP) architecture, the ∆logR parameters are mapped into a new set of features through weighted summation of hidden layer neurons and activation function operations. This enhances the nonlinear relationship between the ∆logR parameters and other logging parameters, unlocking deeper characteristic information, rather than simply using them as additional input.

[0074] Improved DNN model structure: The aforementioned ∆logR-based feature mapping layer is added after the DNN model's input layer. The output of the feature mapping layer is then fused with other preprocessed logging parameters (acoustic travel time, gamma ray radiation, resistivity, density, and neutron lifetime) in a new fusion layer. This fusion layer uses a weighted summation approach to enable more efficient information interaction and integration between the mapped ∆logR parameter and the other logging parameters within the model, which then serve as the input to subsequent hidden layers.

[0075] Joint training and optimization: The improved DNN model is jointly trained, and the parameters of all layers in the model are optimized simultaneously (including the parameters of the feature mapping layer and the fusion layer, as well as the parameters of the hidden layer, input layer, and output layer of the original DNN model). During training, the backpropagation algorithm adjusts the values ​​of various model parameters based on the error between the predicted and measured values ​​to minimize the error. A suitable optimizer (such as the Adam optimizer) is used to accelerate model convergence and ensure that the model fully learns the effective information from the fused features, thereby improving the model's predictive performance.

[0076] Through the improved fusion process described above, the ∆logR parameter can be better integrated with the DNN model, giving full play to its role in improving the model's prediction accuracy and providing more accurate and reliable support for the prediction of total organic carbon content in the formation.

[0077] Figure 4 is the TOC prediction scatter plot of the DNN model, and Figure 5 This is the TOC prediction scatter plot of DNN fusion ∆logR parameters. It can be seen that Figure 5 The scatter distribution is higher than Figure 4 Therefore, after integrating the ∆logR parameter, the prediction performance of the model has been significantly improved. Taking the DNN model as an example, its mean absolute error MAE is reduced to 0.4193, and the root mean square error RMSE is also reduced to 1.01196. At the same time, the coefficient of determination R 2 The changes in these indicators fully demonstrate the significant improvement of the fusion strategy on the prediction accuracy of the DNN model.

[0078] The KNN model's predictive performance is also impressive. After incorporating the ∆logR parameter, the KNN model's MAE significantly decreased from 1.6772 to 0.9762, the RMSE also decreased from 3.1395 to 1.7875, and the R² increased from 0.6427 to 0.8307. This series of improvements fully demonstrates the significant effect of incorporating the traditional ∆logR parameter on improving model performance.

[0079] Through the above-mentioned specific implementation methods, the TOC content prediction method based on multi-parameter logging data and machine learning algorithms described in the present invention can be effectively implemented, achieving high-precision prediction of the total organic carbon content in the formation, and providing reliable and efficient geological parameter support for oil exploration and development. By constructing multiple machine learning models and performing cross-validation, the model with the best performance is selected for prediction, significantly improving the prediction accuracy of TOC content. The ten-fold cross-validation strategy is adopted in the cross-validation process, which not only comprehensively evaluates the generalization ability of each model, but also improves the adaptability and universality of the model under different geological conditions through model optimization and fusion, providing strong support for high-precision exploration under complex geological conditions.

[0080] The above-described embodiments merely illustrate several implementations of the present invention. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, and all such variations and improvements fall within the scope of protection of the present invention.

Claims

1. A method for predicting total organic carbon in formations based on parameter fusion, characterized in that: include: Obtaining a measured value of formation total organic carbon and multiple logging parameters of an exploration well, and using a logging parameter whose correlation coefficient with the measured value of formation total organic carbon is greater than a set threshold as a characteristic logging parameter of formation total organic carbon; Connecting a feature mapping layer and a fusion layer between the input layer and the first hidden layer of the deep neural network to obtain an improved deep neural network; The characteristic logging parameters are input into the improved deep neural network, and the ∆logR parameter is calculated through the input layer. The ∆logR parameter is nonlinearly transformed through the feature mapping layer and mapped into the logging feature. The logging feature and the characteristic logging parameters are weighted and fused through the fusion layer to obtain the fused data. The fused data is used to make a prediction to obtain the predicted value of total organic carbon in the formation. According to the error between the predicted value of total organic carbon in the formation and the measured value of total organic carbon in the formation, the improved deep neural network is trained to obtain the trained improved deep neural network; In the process of predicting total organic carbon in the formation, multiple logging parameters of the exploration well to be tested are obtained, and characteristic logging parameters are selected from the multiple logging parameters and input into the trained improved deep neural network to obtain the total organic carbon prediction results of the exploration well to be tested.

2. The method for predicting total organic carbon in a formation based on parameter fusion according to claim 1, wherein: The calculation of the ∆logR parameter by improving the input layer of the deep neural network specifically includes: The ∆logR parameter is calculated by improving the input layer of the deep neural network based on the following formula: Where RT is the resistivity, is the resistivity baseline, AC is the acoustic time difference, It is the acoustic time difference baseline.

3. The method for predicting total organic carbon in a formation based on parameter fusion according to claim 1, wherein: The ∆logR parameter is nonlinearly transformed through the feature mapping layer to map the ∆logR parameter to the logging feature, specifically including: The feature mapping layer adopts a multi-layer perceptron structure and processes the ∆logR parameter using nonlinear transformation; The hidden layer neurons of the multi-layer perceptron structure perform weighted summation and activation function operation on the ∆logR parameters after nonlinear transformation, and map the ∆logR parameters to well logging features.

4. The method for predicting total organic carbon in a formation based on parameter fusion according to claim 1, wherein: The improved deep neural network includes multiple hidden layers, and the output layer is connected in series after the last hidden layer; the activation function of the output layer of the improved deep neural network is a scaled exponential linear unit; By improving the multiple hidden layers of the deep neural network, the fusion data is processed layer by layer, and the extracted features are combined in the last hidden layer to obtain the global features; By improving the output layer of the deep neural network, the global features are scaled exponential linear unit operations are performed for prediction to obtain the predicted value of total organic carbon in the formation.

5. The method for predicting total organic carbon in formations based on parameter fusion according to claim 1, wherein: According to the error between the predicted value of total organic carbon in the formation and the measured value of total organic carbon in the formation, the improved deep neural network is trained, specifically including: Based on the predicted and measured values ​​of total organic carbon in the formation, an adaptive moment estimation optimization function was constructed to train the improved deep neural network.

6. The method for predicting total organic carbon in formations based on parameter fusion according to claim 1, wherein: The method further includes: for each characteristic well logging parameter, determining an SVM parameter of a support vector machine algorithm according to parameter distribution characteristics of the characteristic well logging parameter, so as to train a classifier of the support vector machine algorithm; Outliers that deviate from the normal data distribution pattern are identified and eliminated through the trained support vector machine algorithm classifier.

7. The method for predicting total organic carbon in a formation based on parameter fusion according to claim 1, wherein: It also includes using exponentially weighted moving average to smooth the characteristic logging parameters before inputting them into the improved deep neural network. The formula is expressed as: Among them, y[t] represents the smoothed value at the tth moment, x[t] represents the observed value at the tth moment, is the smoothing coefficient.

Images