Shale oil mobility and adsorbability prediction method and device based on machine learning

Abstract

The invention belongs to the technical field of petroleum geological exploration, and particularly relates to a shale oil mobility and adsorbability prediction method and device based on machine learning. The method comprises the following steps: S1, acquiring a continental facies shale data set, wherein the continental facies shale data set comprises a shale mineralogy type, basic geochemical parameters, porosity and multi-temperature-order pyrolysis data; s2, the continental facies shale data set is input into a machine learning algorithm model, predicted shale oil mobility parameters and adsorbability parameters are output, and the machine learning algorithm model comprises RF-PSO and RF-DE. According to the method, the optimal OMC and OSC prediction models are screened out based on mineralogy, basic geochemistry and porosity data, so that the OMC and OSC of the shale can be accurately calculated.

Description

Technical Field

[0001] The present invention belongs to the technical field of petroleum geological exploration, and specifically relates to a method and device for predicting the mobility and adsorption of shale oil based on machine learning. Background Art

[0002] Although the world is advocating energy structure adjustment to reduce CO2 emissions, oil will still play an important role in global energy supply before 2035. With the development of horizontal drilling and hydraulic fracturing technology, shale oil and gas have gradually occupied an important position in oil exploration and development. In recent years, my country has made remarkable achievements in continental shale oil and gas exploration. However, compared with North American marine shales, my country's continental shale reservoirs often show strong heterogeneity. Therefore, clarifying the key geological parameters in continental shale reservoirs has important guiding significance for optimizing exploration deployment.

[0003] Exploration practice shows that the actual production of shale oil reservoirs is largely controlled by oil mobility. Free oil in the pore and fracture system of shale has good fluidity and is considered to be movable oil. According to previous studies, the content of free oil per unit mass of organic matter can represent the oil mobility in shale (OMC, Wang et al., 2022). Adsorbed oil adsorbed on kerogen and mineral surfaces is often difficult to mobilize under realistic exploration and development conditions and is considered to be immobile oil. The content of adsorbed oil per unit mass of organic matter can represent the oil adsorption capacity (OSC) in shale. Since OMC and OSC in shale are controlled by many factors and the influencing mechanism is very complex, the current understanding of the controlling factors of oil mobility in shale is still unclear, which makes it difficult to predict oil mobility.

[0004] Among the current resource evaluation methods, the parameter S1 (conventional pyrolysis free hydrocarbon content) obtained by the conventional pyrolysis method can represent the free oil content in shale to a certain extent, and is widely used as an important geological parameter. However, with the advancement of technical means, the conventional pyrolysis method has been proven to have shortcomings. S1 contains both free oil and a portion of adsorbed oil, making it difficult to accurately reflect the mobility of oil. As a more effective testing technology, the multi-temperature stage pyrolysis method can more precisely characterize the different occurrence forms of shale oil. S 1-1 and S 1-2 Represent the contents of light oil and light medium components in the sample respectively. 1-1 and S 1-2 The parameters can more accurately characterize the content of movable oil in shale. However, since the multi-stage pyrolysis method is relatively new and the testing cost is high, there is a general lack of sufficient and accurate shale oil mobility data in the industry, which restricts shale oil and gas exploration and development as well as the understanding of shale mobility.

[0005] Machine learning (ML) has become a powerful tool for mining complex patterns in data. However, ML is essentially a black box model, and it is difficult to understand how the model makes predictions by understanding its internal mechanisms. To overcome this challenge, interpretable machine learning techniques have received widespread attention. These techniques can quantify the importance of each input feature to model predictions from different perspectives, and can also provide insights into the complex synergies between input features. Summary of the invention

[0006] In order to accurately and cost-effectively analyze the mobility and adsorption of shale oil, the present invention adopts a machine learning model and provides a method and device for predicting the mobility and adsorption of shale oil based on machine learning.

[0007] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:

[0008] The shale oil mobility and adsorption prediction method based on machine learning includes the following steps:

[0009] S1, obtaining a continental shale data set, wherein the continental shale data set includes shale mineralogy type, basic geochemical parameters, porosity, and multi-temperature stage pyrolysis data;

[0010] S2, inputting the continental shale data set into a machine learning algorithm model, and outputting predicted shale oil mobility parameters and adsorption parameters, wherein the machine learning algorithm model includes RF-PSO and RF-DE.

[0011] As a preferred solution, S2 also includes preprocessing the continental shale data set and then inputting it into the machine learning algorithm model, wherein the preprocessing includes detection and deletion of outliers and interpolation of missing values.

[0012] As a preferred embodiment, the shale mineralogy types include quartz, feldspar, plagioclase, potassium feldspar, calcite, dolomite, aragonite, siderite, gypsum, anhydrite, clay minerals and pyrite; the basic geochemical parameters include organic matter abundance and pyrolysis hydrocarbon peak temperature.

[0013] As a preferred embodiment, the multi-temperature stage pyrolysis data includes multi-temperature stage pyrolysis of light oil components; multi-temperature stage pyrolysis of light and medium oil components; multi-temperature stage pyrolysis of heavy hydrocarbons and asphalt components; and multi-temperature stage pyrolysis of kerogen pyrolysis cracking regenerated hydrocarbons.

[0014] As a preferred solution, the machine learning algorithm models include MLP-SGD, MLP-Lbfgs, MLP-Adam, RF-PSO, RF-GA, RF-DE, XGBoost-PSO, XGBoost-GA, XGBoost-DE, SVM-PSO, SVM-GA, and SVM-DE.

[0015] As a preferred solution, in the process of training the machine learning algorithm model, a regression model is applied to develop and evaluate the model, and the coefficient of determination and root mean square error are defined as follows:

[0016]

[0017] In the formula, is the predicted data, y i is the actual value, is the actual data mean, n is the number of data in each data set, R 2 is the coefficient of determination and RMSE is the root mean square error.

[0018] As a preferred solution, the best model for predicting shale oil mobility parameters and adsorption parameters is screened out from the machine learning algorithm model, and the screening method includes an average prediction performance evaluation based on five-fold cross validation.

[0019] As a preferred solution, the method further includes step S3: using a variety of explanatory algorithms to interpret the mobility and adsorption parameter prediction model to obtain the influencing factors of shale oil mobility and adsorption.

[0020] As a preferred solution, the explanatory algorithms include PHI, SHAP and PDP.

[0021] Based on the same concept, a shale oil mobility and adsorption prediction device based on machine learning is also proposed, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute any of the above-mentioned shale oil mobility and adsorption prediction methods based on machine learning.

[0022] Compared with the prior art, the present invention has the following beneficial effects: based on mineralogy, basic geochemistry and porosity data, the present invention constructs and compares a variety of machine learning-based models, and screens out the optimal OMC and OSC prediction models, so that the OMC and OSC of shale can be accurately calculated. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 is a flow chart of the method for predicting shale oil mobility and adsorption based on machine learning in Example 1;

[0024] Figure 2 It is a schematic diagram of the OMC and OSC prediction process based on interpretable machine learning in Example 2;

[0025] Figure 3 It is the correlation analysis between the features and the target data in Example 2;

[0026] Figure 4 is the performance evaluation result of the machine learning prediction model in Example 2;

[0027] Figure 5 It is the intersection diagram of the predicted value and the actual value of the optimal OMC and OSC prediction model in Example 2;

[0028] Figure 6 is the optimal OMC and OSC prediction model replacement importance ranking in Example 2;

[0029] Figure 7 It is the SHAP ranking of the optimal OMC and OSC prediction models in Example 2;

[0030] Figure 8 It is the result that the main characteristics of the optimal OMC and OSC prediction model in Example 2 partially depend on the results. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with test examples and specific implementation methods. However, this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the content of the present invention belong to the scope of the present invention.

[0032] Example 1

[0033] The shale oil mobility and adsorption prediction method based on machine learning is shown in the flow chart below. Figure 1 As shown, the following steps are included:

[0034] Step (1) collects shale mineralogy (quartz, feldspar, plagioclase, potassium feldspar, calcite, dolomite, aragonite, siderite, gypsum, anhydrite, clay minerals and pyrite), basic geochemistry (TOC, T max ), porosity and multi-temperature pyrolysis data (S 1-1 , S 1-2 , S 2-1 , S 2-2 ), as the original dataset:

[0035] Step (2), performing preprocessing operations on the raw data in step (1), performing outlier detection and missing value interpolation on the raw data, and calculating OMC and OSC based on TOC and multi-temperature pyrolysis data, the calculation formula is as follows:

[0036]

[0037]

[0038] Among them, OMC is the mobility of shale oil, OSC is the adsorption of shale oil, S 1-1It is a multi-stage pyrolysis light oil component, S 1-2 It is a multi-stage pyrolysis light medium oil component, S 2-1 It is the multi-temperature stage pyrolysis heavy hydrocarbons and colloidal asphalt components, and TOC is the abundance of organic matter.

[0039] Step (3), dividing the preprocessed data into two parts: features and targets. Features include mineralogy, basic geochemical parameters, and porosity, and targets include OMC and OSC. The original data are divided into two subsets according to different targets.

[0040] Step (4), before model development, the data subsets are randomly grouped into training sets and test sets. The training set accounts for 80% of the data subset and is used for ML model development; the test set accounts for 20% of the data subset and is used to evaluate the generalization ability of the optimal model. In order to reduce the impact of the dimensions of different features on model prediction, after dividing the training set and the test set, the training set is standardized, and the test set is standardized using the mean and standard deviation obtained from the training set to avoid data leakage caused by introducing test set data during the model development phase;

[0041] Step (5), in the training set data, four classic machine learning algorithms (SVM, RF, XGBoost, MLP) are used to build the model, and six optimization algorithms (PSO, GA, DE, SGD, Lbfgs, Adam) are used in combination, and finally 12 machine learning models (MLP-SGD, MLP-Lbfgs, MLP-Adam, RF-PSO, RF-GA, RF-DE, XGBoost-PSO, XGBoost-GA, XGBoost-DE, SVM-PSO, SVM-GA, SVM-DE) are developed. It should be emphasized that SGD, Lbfgs and Adam are used to optimize weights and biases in the training phase of the MLP model. DE, GA and PSO algorithms are responsible for optimizing the hyperparameters in the RF, XGBoost and SVM models;

[0042] Step (6) applies common regression model indicators such as the coefficient of determination (R 2 ) and root mean square error (RMSE), defined as follows:

[0043]

[0044] In the formula, is the predicted data, y i is the actual value, is the actual data mean, and n is the size of each data set. The average prediction performance (R 2 CV and RMSECV ) was used to evaluate 12 machine learning models and then select the optimal model;

[0045] Step (7), use PFI, SHAP, and PDP techniques to interpret the optimal OMC and OSC prediction models obtained in step (6).

[0046] The abbreviations and key terms are defined as follows: S1: conventional pyrolysis free hydrocarbon content; S2: conventional pyrolysis cracked hydrocarbon content; S 1-1 : Multi-stage pyrolysis of light oil components; S 1-2 : Multi-stage pyrolysis of light and medium oil components; S 2-1 : Multi-stage pyrolysis of heavy hydrocarbons and asphalt colloid components; S 2-2 :Multi-temperature pyrolysis kerogen pyrolysis cracking regeneration hydrocarbons; Tmax: pyrolysis hydrocarbon peak temperature; TOC: organic matter abundance; OMC: shale oil mobility; OSC: shale oil adsorption; SVM: support vector machine algorithm; RF: random forest algorithm; XGBoost: extreme gradient boosting tree; MLP: multi-layer perceptron; PSO: particle swarm algorithm; GA: genetic optimization algorithm; DE: differential optimization algorithm; PFI: permutation feature importance; SHAP: Shapley additive interpretation; PDP: partial dependence plot.

[0047] Example 2

[0048] In order to solve the problem of missing research data on shale oil mobility and improve the understanding of the factors controlling shale oil mobility, the present invention proposes a prediction method for shale oil OMC and OSC based on interpretable machine learning using mineral composition, geochemical parameters and porosity.

[0049] Figure 2 A flow chart of a shale oil OMC and OSC prediction method according to an embodiment of the present invention is shown. Example 2 of the present invention studies a certain area, and the study conducts a series of analytical tests on black shales from the Qianfoya Formation and Da'anzhai Section. The research data includes TOC test, multi-temperature pyrolysis test, X-ray diffraction test, helium porosity, etc.

[0050] According to the above steps, the mineralogy (quartz, feldspar, plagioclase, potassium feldspar, calcite, dolomite, aragonite, siderite, gypsum, anhydrite, clay minerals and pyrite) and basic geochemistry (TOC, T max ), porosity and multi-temperature pyrolysis data (S 1-1 , S 1-2 , S 2-1 , S 2-2 ), as the original data set.

[0051] Subsequently, preprocessing operations were performed on the raw data, including outlier detection and missing value interpolation. OMC and OSC were calculated based on TOC and distribution pyrolysis data, and the calculation formula is as follows:

[0052]

[0053] Among them, OMC is the mobility of shale oil, OSC is the adsorption of shale oil, S 1-1 It is a multi-stage pyrolysis light oil component, S 1-2 It is a multi-stage pyrolysis light medium oil component, S 2-1 It is the multi-temperature stage pyrolysis heavy hydrocarbons and colloidal asphalt components, and TOC is the abundance of organic matter.

[0054] The preprocessed data is divided into two parts: features and targets. Features include mineralogy, basic geochemical parameters, and porosity, and targets include OMC and OSC. The raw data is divided into two subsets according to different targets. Finally, the preprocessed raw data is divided into two subsets: (1) Data subset 1 contains 13 features and corresponding OMCs; (2) Data subset 2 contains 13 features and corresponding OSCs.

[0055] Figure 3 The correlation characteristics between the features and targets in the examples are shown. Geochemical parameters (TOC, T max ) provided the strongest correlation trend with the target (R>0.6), TOC was positively correlated with OMC and OSC, T max There is a negative correlation between the content of quartz and calcite and OMC and OSC.

[0056] Before model development, the data subsets were randomly divided into training sets and test sets. The training set accounted for 80% of the data subset and was used for ML model development; the test set accounted for 20% of the data subset and was used to evaluate the generalization ability of the optimal model. In order to reduce the impact of the dimensions of different features on model prediction, after dividing the training set and the test set, the training set was standardized, and the mean and standard deviation obtained from the training set were used to standardize the test set to avoid data leakage caused by introducing test set data during the model development stage.

[0057] In the training set data, four classic machine learning algorithms (SVM, RF, XGBoost, MLP) were used to build the model, and six optimization algorithms (PSO, GA, DE, SGD, Lbfgs, Adam) were used in combination. Finally, 12 machine learning models (MLP-SGD, MLP-Lbfgs, MLP-Adam, RF-PSO, RF-GA, RF-DE, XGBoost-PSO, XGBoost-GA, XGBoost-DE, SVM-PSO, SVM-GA, SVM-DE) were developed. It should be emphasized that SGD, Lbfgs, and Adam are used to optimize weights and biases in the training phase of the MLP model. DE, GA, and PSO algorithms are responsible for optimizing the hyperparameters in the RF, XGBoost, and SVM models.

[0058] The commonly used evaluation index of regression model, namely the coefficient of determination (R 2 ) and root mean square error (RMSE), defined as follows:

[0059]

[0060] In the formula, is the predicted data, y i is the actual value, is the actual data mean, and n is the size of each data set. The average prediction performance (R 2 CV and RMSE CV ) was used to evaluate 12 machine learning models and then select the optimal model.

[0061] Figure 4 The best results obtained by 12 machine learning models are shown. Among them, RF-PSO achieved the best OMC prediction results, with R 2 RF-DE achieved the best OMC prediction results, with R 2 They are 0.96 and 0.80 respectively, and the RMSE are 0.08 and 0.19 respectively.

[0062] Figure 5 The cross plots of the predicted values ​​and actual values ​​of the optimal OMC and OSC prediction models are shown. It can be observed that the model has a good prediction effect on OMC and OSC, and the predicted OMC and OSC are very similar to the actual values.

[0063] After the optimal model is determined in the model development process, PFI, SHAP, and PDP technologies are used to explain the optimal OMC and OSC prediction models. PFI, SHAP, and PDP are three different interpretable technologies to analyze how the optimal prediction model is output and what factors are more important to the results. Machine learning is a black box model, so it is difficult to really derive its operating mechanism. Therefore, as a preferred solution, multiple interpretable technologies are also used to explain the optimal OMC and OSC prediction models, and then their results are comprehensively analyzed to find out which factors are more important to the results.

[0064] The permutation feature importance graph obtained by PFI, the SHAP global dependency graph obtained by SHAP, and the partial dependency graph obtained by PDP can be combined to provide insights into the controlling factors of OMC and OSC. Figure 6 The permutation feature importance ranking of the main features of the optimal OMC and OSC models is shown respectively. max Two geochemical parameters have the greatest contribution to the accurate prediction of OMC and OSC. In addition, calcite, porosity, siderite and pyrite are more important for the prediction of OMC; porosity, pyrite, quartz and clay minerals are more important for the prediction of OSC.

[0065] The SHAP value can be used to study the impact of each feature on model prediction, thereby explaining the optimal machine learning model. Figure 7 The SHAP value ranking of each feature of the optimal OMC and OSC models is shown respectively. The features along the y-axis from top to bottom reflect the difference in the SHAP importance of the model. For the OMC model, the first six features of the SHAP summary graph are TOC, T max , calcite, porosity, pyrite and siderite; the first six features of the SHAP summary diagram of the OSC model are TOC, T max , pyrite, clay minerals, porosity, calcite.

[0066] A partial dependence plot (PDP) can represent the impact of two features on the model output under interaction. Figure 8 The partial dependence results of each feature of the optimal OMC and OSC models are shown respectively.

[0067] TOC and T max The TOC values ​​are positively and negatively correlated with the overall model output, respectively. This trend can be observed in both the OMC and OSC models. At the same time, when TOC is relatively low, the size of TOC has an important influence on the values ​​of OMC and OSC, and this influence is more obvious in OMC. When TOC is high, the increase in TOC is no longer accompanied by an increase in OMC and OSC, while the increase in TOC is not accompanied by an increase in OMC and OSC. maxThe negative impact on OMC and OSC is enhanced. In addition, pyrite shows a positive contribution to both OMC and OSC. This positive contribution is mainly reflected in the fact that lower pyrite content tends to reduce the output of the model. The interaction between clay minerals and quartz shows the influence of mineral composition and lithofacies type on OMC and OSC predictions. Among them, clay minerals and quartz contribute very little to OMC, and the overall impact is not significant. However, in the OSC model, the clay mineral content shows a positive contribution to the model, and the increase in calcite content has a negative impact on OSC.

[0068] Beneficial effects brought by the technical solution of the present invention:

[0069] (1) Based on mineralogy, basic geochemistry and porosity data, the present invention constructs and compares a variety of machine learning-based models to screen out the optimal OMC and OSC prediction models, thereby accurately calculating the OMC and OSC of shale.

[0070] (2) The present invention can make full use of a large amount of mineralogical and basic geochemical test data accumulated in the previous shale exploration and development process, and can quantitatively evaluate the OMC and OSC of shale without conducting additional large-scale analytical tests, which helps to provide theoretical guidance for the exploration and development of shale oil and gas.

[0071] (3) The present invention uses a variety of interpretability technologies to quantitatively analyze the effects of mineral composition, basic geochemical parameters and porosity on shale mobility, which helps to deepen the understanding of the factors controlling shale oil mobility.

[0072] The specific implementation examples described in this article are merely examples of the spirit of the present invention. Those skilled in the art of the present invention may make various modifications or supplements to the specific implementation examples described or take corresponding measures, but they will not deviate from the spirit of the present invention or exceed the scope defined by the attached claims.

Claims

1. A method for predicting shale oil mobility and adsorption based on machine learning, characterized in that: The following steps are involved: S1, obtaining a continental shale data set, wherein the continental shale data set includes shale mineralogy type, basic geochemical parameters, porosity, and multi-temperature stage pyrolysis data; S2, inputting the continental shale data set into a machine learning algorithm model, and outputting predicted shale oil mobility parameters and adsorption parameters, wherein the machine learning algorithm model includes RF-PSO and RF-DE.

2. The method for predicting shale oil mobility and adsorption based on machine learning according to claim 1, characterized in that: S2 also includes preprocessing the continental shale data set and inputting it into the machine learning algorithm model, wherein the preprocessing includes detection and deletion of outliers and interpolation of missing values.

3. The method for predicting shale oil mobility and adsorption based on machine learning according to claim 1, characterized in that: The shale mineralogy types include quartz, feldspar, plagioclase, potassium feldspar, calcite, dolomite, aragonite, siderite, gypsum, anhydrite, clay minerals and pyrite; the basic geochemical parameters include organic matter abundance and pyrolysis hydrocarbon peak temperature.

4. The method for predicting shale oil mobility and adsorption based on machine learning according to claim 1, characterized in that: Multi-temperature pyrolysis data include multi-temperature pyrolysis light oil components; Multi-temperature stage pyrolysis of light and medium oil components; Multi-stage pyrolysis of heavy hydrocarbons and asphalt colloid components; Multi-stage pyrolysis of kerogen to regenerate hydrocarbons by pyrolysis and cracking.

5. The method for predicting shale oil mobility and adsorption based on machine learning according to claim 1, characterized in that: The machine learning algorithm models include MLP-SGD, MLP-Lbfgs, MLP-Adam, RF-PSO, RF-GA, RF-DE, XGBoost-PSO, XGBoost-GA, XGBoost-DE, SVM-PSO, SVM-GA, and SVM-DE.

6. The method for predicting shale oil mobility and adsorption based on machine learning according to claim 5, characterized in that: In the process of training the machine learning algorithm model, a regression model is applied for model development and evaluation, and the coefficient of determination and root mean square error are defined as follows: In the formula, is the forecast data, is the actual value, is the actual data mean, n is the number of data in each data set, R 2 is the coefficient of determination, is the root mean square error.

7. The method for predicting shale oil mobility and adsorption based on machine learning according to claim 5, characterized in that: The optimal model for predicting shale oil mobility parameters and adsorption parameters is screened out from the machine learning algorithm model, and the screening method includes an average prediction performance evaluation based on five-fold cross validation.

8. The method for predicting shale oil mobility and adsorption based on machine learning according to any one of claims 1 to 7, characterized in that: The method also includes step S3: using a variety of explanatory algorithms to interpret the mobility and adsorption parameter prediction model to obtain the influencing factors of shale oil mobility and adsorption.

9. The method for predicting shale oil mobility and adsorption based on machine learning according to claim 8, characterized in that: The explanatory algorithms include PHI, SHAP and PDP.

10. A device for predicting shale oil mobility and adsorption based on machine learning, characterized in that: It includes at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the shale oil mobility and adsorption prediction method based on machine learning as described in any one of claims 1 to 9.

Images