A reservoir fluid identification method combining LIBS and random forest algorithms

Abstract

This invention proposes a reservoir fluid identification method combining LIBS and random forest algorithms, comprising four steps: LIBS spectral data collection and processing, construction of a random forest algorithm model, training of the random forest algorithm model, and identification of unknown reservoir fluids using the model. This invention first inputs features based on LIBS spectral data from well logging samples, and studies the impact of the number of input features and different feature combinations on the algorithm's prediction results. Then, it uses the algorithm to learn the nonlinear relationship between the input features and reservoir information in the well. Finally, it comprehensively identifies the reservoir based on the learning results, achieving a comprehensive analysis of multiple fluid identification factors. This reduces the ambiguity caused by a single fluid identification factor, improves the accuracy and reliability of oil and gas reservoir fluid identification, and solves the problem of accuracy disclosure in existing oil and gas reservoir fluid identification technologies.

Description

Technical Field

[0001] This invention relates to the field of oil and gas reservoir fluid detection technology, and in particular to a reservoir fluid identification method combining LIBS and random forest algorithms. Background Technology

[0002] Reservoir fluid property identification is a crucial step in the exploration phase, thus it is of great significance for oilfield exploration and development. Currently, oil and gas reservoir types are primarily determined using gas logging triangulation charts and Pixler charts. However, these methods have strong regional limitations and significant errors, failing to effectively guide oil and gas field exploration and development. These issues necessitate improvements to traditional techniques. Therefore, inventing a method for identifying reservoir fluid properties using gas logging data is a problem we need to solve.

[0003] Since Smith proposed the concept of fluid identification factors in 1987, numerous scholars have constructed a large number of fluid identification factors based on rock physics models. Therefore, various reservoir fluid identification methods centered on fluid identification factors have emerged, such as elastic parameters like Poisson's ratio, Young's modulus, and Lamé constant. While each of these fluid identification factors has its advantages, using them individually inevitably leads to uncertainties and multiple solutions. Traditional core observation methods, although limited by the amount of core material, can only reflect the fluid distribution in the vicinity of the wellbore, but this method offers relatively accurate identification results. Therefore, how to comprehensively utilize the nonlinear relationship between multiple fluid identification factors and the fluid distribution in the well for three-dimensional fluid identification is a challenging problem in comprehensive reservoir prediction.

[0004] Many large international logging companies and research institutions, leveraging their advantages in logging technology and instruments, focus on employing new methods to identify oil and gas reservoirs. For example, Doveton J et al. utilized wireline logging technology and employed dynamic programming for logging comparison, selecting the optimal match from a wide range of possibilities to accurately identify oil and gas reservoirs. Additionally, some researchers have achieved certain results using conventional mathematical statistical pattern recognition techniques to identify oil and gas reservoirs based on logging curves.

[0005] Traditional methods for identifying oil and gas reservoirs using well logging data require the establishment of rock volume models, well logging response equations, statistical empirical formulas, and a large amount of expert experience. However, as the injected water content and water saturation of oil fields increase, the reservoir resistivity does not decrease monotonically, resulting in a very poor match between the interpretation results and the actual test results. Therefore, this invention proposes a reservoir fluid identification method combining LIBS and random forest algorithms to solve the problems existing in the prior art. Summary of the Invention

[0006] To address the aforementioned problems, the present invention aims to propose a reservoir fluid identification method combining LIBS and random forest algorithms. This method offers the advantage of multi-method detection and solves the problems in the prior art.

[0007] To achieve the objectives of this invention, the invention is implemented through the following technical solution: a reservoir fluid identification method combining LIBS and random forest algorithms, comprising the following steps:

[0008] Step 1: LIBS Spectral Data Collection and Processing

[0009] Select an experimental area containing several sets of well logs. Extract several sets of LIBS spectra from the well log data of the gas-bearing and water-bearing reservoirs in each set of well logs. Obtain LIBS spectral data of several sets of well log samples. Then, perform preprocessing to obtain processed LIBS spectral data.

[0010] Step 2: Construction of the Random Forest Algorithm Model

[0011] Multiple random forest decision trees are constructed by using two randomness factors: random repeated sampling and random feature selection. The specific steps are as follows:

[0012] S1: Bootstrap resampling method;

[0013] S2: Random feature selection index;

[0014] S3: Determine the optimal split node;

[0015] S4: Decision Classification;

[0016] Step 3: Random Forest Algorithm Model Training

[0017] The single-class fluid identification factor is introduced into the random forest algorithm model constructed in step two. The validation set samples are used to identify whether they contain gas or water. Based on the identification results, fluid identification factors are selected. These fluid identification factors are then arranged into 1, 2, ..., N classes to form a new training dataset. The random forest algorithm model is then applied to the new validation set samples to make a judgment. From the judgment results, the categories and number of fluid identification factors suitable for identifying gas and water in this area are found.

[0018] Step 4: Identify unknown reservoir fluids using models

[0019] Based on the random forest algorithm model trained in step three, fluid identification is performed on the unknown sample. First, the LIBS spectral data of the unknown sample is detected, and then the LIBS spectral data of the unknown sample is input. The model is used to obtain the identification results of gas-bearing reservoirs and water-bearing reservoirs of the unknown sample.

[0020] A further improvement is made in the following: In step one, the LIBS spectral data is collected in the range of 180–790 nm. The LIBS spectral data is collected from three sets of spectral channels, which cover 180–350 nm, 350–580 nm and 580–790 nm respectively. Each spectral channel has 2048 pixels.

[0021] A further improvement is made in step one, where the preprocessing method involves normalizing the total area of ​​the LIBS spectral data. The preprocessing formula is shown below:

[0022]

[0023] In the formula, I normalization It is the spectral intensity after preprocessing, I origin I represents the original spectral intensity. total It is the sum of the intensities of all data in the spectrum.

[0024] A further improvement is made in step one, where the processed LIBS spectral data is divided into a training set, a validation set, and a test set.

[0025] A further improvement lies in the following: In step two, the Bootstrap resampling method specifically involves: denoting the original training dataset of reservoir fluid LIBS spectra generated based on well logging data as D = ([x... i ]1, [x i ]2,…,[x i ] n y i The Bootstrap sampling method is used to randomly sample K times from D, forming K independently distributed training data subsets {D}. k , k = 1, 2, ..., K}.

[0026] A further improvement is made in step two, where the random feature selection metric is specifically: a decision tree with maximum growth is built for each of the K datasets without any pruning, and then the information gain ratio of a subset of features randomly selected from the training subset is calculated using the information gain ratio formula.

[0027] A further improvement is made in the following step: In step two, determining the optimal splitting node specifically involves splitting the candidate fluid identification factor with the largest information gain ratio, recalculating the information gain ratio according to the information gain ratio formula, repeating the splitting step until the information gain ratio is less than a set threshold, and finally generating a forest model containing K decision trees.

[0028] A further improvement is made in the following step: In step two, the decision classification is specifically as follows: a random forest classifier is constructed by integrating numerous decision trees for fluid identification based on well logging data, and then the random forest classifier is used to classify the obtained fluid identification factors. The prediction results of all decision trees in the forest are statistically analyzed, and the result with the highest frequency among the classification results of each decision tree is selected as the final classification result.

[0029] A further improvement is made in step three, when verifying the effectiveness of the algorithm in identifying gas and water content, firstly, several training subsets are generated from the original training dataset using the Bootstrap repeated sampling method. Then, a portion of feature factors are randomly selected from each training subset, and the internal nodes are split according to the principle of maximizing information gain ratio to generate several sets of freely growing decision trees. These decision trees are then used to sequentially identify the gas and water content of the validation set samples and record the identification results. The result with the highest frequency is selected from several sets of identification results as the final identification result of that set of validation set samples.

[0030] The beneficial effects of this invention are as follows: This reservoir fluid identification method combining LIBS and random forest algorithms first inputs features (fluid identification factors) based on LIBS spectral data of well logging samples, and studies the influence of the number of input features and different feature combinations on the algorithm prediction results. Then, the algorithm is used to learn the nonlinear relationship between the input features and the reservoir information in the well. Finally, the reservoir is comprehensively identified based on the learning results, realizing the comprehensive analysis of multiple fluid identification factors. This can reduce the ambiguity caused by a single fluid identification factor, improve the accuracy and reliability of oil and gas reservoir fluid identification, and solve the problem of accuracy disclosure in oil and gas reservoir fluid identification in the prior art. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the overall steps of the method of the present invention.

[0032] Figure 2 This is a schematic diagram of the overall structural framework of the method of the present invention.

[0033] Figure 3 This is a schematic diagram of the decision tree construction process of the random forest algorithm of this invention.

[0034] Figure 4 This is a schematic diagram of the random forest algorithm for fluid identification in this invention. Detailed Implementation

[0035] To enhance understanding of the present invention, the present invention will be further described in detail below with reference to embodiments. These embodiments are only used to explain the present invention and do not constitute a limitation on the scope of protection of the present invention.

[0036] Example 1

[0037] according to Figure 1 As shown, this embodiment proposes a reservoir fluid identification method combining LIBS and random forest algorithms, including the following steps:

[0038] Step 1: LIBS Spectral Data Collection and Processing

[0039] LIBS spectral data of well logging samples were collected, preprocessed, and then input back into the well logging samples. The LIBS spectral data acquisition range was 180–790 nm, and the LIBS spectral data was acquired from three spectral channels, covering 180–350 nm, 350–580 nm, and 580–790 nm respectively. Each spectral channel has 2048 pixels. The specific preprocessing method was to perform total area normalization on the LIBS spectral data. The preprocessing formula is shown below:

[0040]

[0041] In the formula, I normalization It is the spectral intensity after preprocessing, I origin I represents the original spectral intensity. total The sum of all data intensities in the spectrum is used to further divide the processed LIBS spectral data into training, validation, and test sets for subsequent model training and validation.

[0042] Step 2: Construction of the Random Forest Algorithm Model

[0043] Multiple random forest decision trees are constructed by using two randomness factors: random repeated sampling and random feature selection. The specific steps are as follows:

[0044] S1: The Bootstrap resampling method denotes the original training dataset of reservoir fluid LIBS spectra generated based on well logging data as D=([x i ]1, [x i ]2,…,[x i ] n y i The Bootstrap sampling method is used to randomly sample K times from D, forming K independently distributed training data subsets {D}. k ,k=1,2,…,K};

[0045] S2: Random feature selection index; Build decision trees with maximum growth without any pruning based on K datasets, and then calculate the information gain ratio of a subset of features randomly selected from the training subset according to the information gain ratio formula;

[0046] S3: Determine the optimal splitting node, split the candidate fluid identification factor with the largest information gain ratio, and recalculate the information gain ratio according to the information gain ratio formula. Then repeat the splitting step until the information gain ratio is less than the set threshold. Finally, generate a forest model containing K decision trees.

[0047] S4: Decision classification integrates numerous decision trees for fluid identification based on well logging data to construct a random forest classifier. Then, the random forest classifier is used to classify the obtained fluid identification factors, and the prediction results of all decision trees in the forest are statistically analyzed. The result with the highest frequency among the classification results of each decision tree is selected as the final classification result.

[0048] Step 3: Random Forest Algorithm Model Training

[0049] The single-class fluid identification factor is introduced into the random forest algorithm model constructed in step two. The validation set samples are used to identify whether they contain air or water. Based on the identification results, fluid identification factors are selected, and these fluid identification factors are arranged into 1, 2, ..., N classes to form a new training dataset. The random forest algorithm model is then applied to the new validation set samples to make judgments. From the judgment results, the categories and quantities of fluid identification factors suitable for identifying air and water in this area are found. When verifying the effectiveness of the algorithm in identifying air and water, several training subsets are first generated from the original training dataset according to the Bootstrap repeated sampling method. Then, a portion of feature factors are randomly selected from each training subset, and the internal nodes are split according to the principle of maximizing information gain ratio to generate several sets of freely growing decision trees. These decision trees are then used to identify air and water in the validation set samples in turn and the identification results are recorded. The highest frequency result is selected from several sets of identification results as the final identification result of the validation set samples.

[0050] Step 4: Identify unknown reservoir fluids using models

[0051] Based on the random forest algorithm model trained in step three, fluid identification is performed on the unknown sample. First, the LIBS spectral data of the unknown sample is detected, and then the LIBS spectral data of the unknown sample is input. The model is used to obtain the identification results of gas-bearing reservoirs and water-bearing reservoirs of the unknown sample.

[0052] Example 2

[0053] according to Figures 2-4 As shown, this embodiment proposes a reservoir fluid identification method combining LIBS and random forest algorithms, including the following steps:

[0054] Step 1: LIBS Spectral Data Collection and Processing

[0055] The experimental area was selected, containing several sets of well logs. In this embodiment, the X area of ​​the South China Sea offshore oil and gas field was selected as the research object. The area includes three wells: A, B, and C. Twenty different locations were extracted from the well log data of the gas-bearing and water-bearing reservoirs in each well group. LIBS spectra were collected, and after accumulating four pulse lasers at each sampling location, one spectrum was obtained by averaging four spectra. Therefore, this embodiment obtained a total of 1940 effective LIBS spectra from 97 samples. Then, preprocessing was performed to obtain the processed LIBS spectral data. The preprocessing method was to normalize the total area of ​​the LIBS spectral data. The preprocessing formula is shown below:

[0056]

[0057] In the formula, I normalization It is the spectral intensity after preprocessing, I origin I represents the original spectral intensity. total The sum of all data intensities in the spectrum is used. The processed LIBS spectral data is divided into training set, validation set and test set. The training set and validation set are used to train the network, and the test set is used to test the network. Then, the training network is built and trained using the training set and validation set. The resulting model can identify the LIBS spectral data of the well logging sample. The identification results are substituted into the evaluation index system, and the training network is optimized and improved based on the evaluation results.

[0058] Step 2: Construction of the Random Forest Algorithm Model

[0059] To improve the accuracy and precision of fluid identification, this embodiment constructs multiple random forest decision trees by employing two randomnesses: random repeated sampling and random feature selection. The specific steps are as follows:

[0060] S1: The Bootstrap resampling method denotes the original training dataset of reservoir fluid LIBS spectra generated based on well logging data as D=([x i ]1, [x i ]2,…,[x i ] n y i The Bootstrap sampling method is used to randomly sample K times from D, forming K independently distributed training data subsets {D}. k ,k=1,2,…,K};

[0061] Because each training data subset D k If each of these variables is a discrete random variable that can only take a finite number of values, then its probability distribution can be expressed as:

[0062] P(D k ) = pj j = 1, 2, ..., n

[0063] random variable D k Information entropy is defined as:

[0064]

[0065] Suppose that from the training data subset D k The selected feature factor A contains V values; therefore, the dataset D can be partitioned based on these V values. k And calculate the information entropy separately. Then, feature factor A for the training data subset D k The information gain Gain(D,A) is:

[0066]

[0067] Training data subset D k Given feature factor A, the corresponding information gain ratio GainRatio(D,A) is:

[0068]

[0069] in,

[0070] S2: Random feature selection index. Decision trees with maximum growth are built for each of the K datasets without any pruning. Then, the information gain ratio of a subset of features randomly selected from the training subset is calculated using the information gain ratio formula.

[0071] S3: Determine the optimal splitting node, split the candidate fluid identification factor with the largest information gain ratio, and recalculate the information gain ratio according to the information gain ratio formula. Then repeat the splitting step until the information gain ratio is less than the set threshold. Finally, generate a forest model containing K decision trees.

[0072] S4: Decision classification integrates numerous decision trees for fluid identification based on well logging data to construct a random forest classifier. Then, the random forest classifier is used to classify the obtained fluid identification factors, and the prediction results of all decision trees in the forest are statistically analyzed. The result with the highest frequency among the classification results of each decision tree is selected as the final classification result.

[0073] Step 3: Random Forest Algorithm Model Training

[0074] The random forest algorithm has the ability to optimize features. Therefore, by directly using this algorithm to optimize numerous fluid identification factors, we can obtain a combination of fluid identification factors suitable for comprehensive prediction by the random forest algorithm in the research area. Then, we first introduce the single-class fluid identification factors into the random forest algorithm model constructed in step two, and identify the gas and water content of the validation set samples. Based on the identification results, we select fluid identification factors and select 1, 2, ..., N classes of these fluid identification factors according to the permutation and combination method to form a new training dataset. Then, we apply the random forest algorithm model to judge the new validation set samples and find the category and number of fluid identification factors suitable for gas and water content identification in this area from the judgment results.

[0075] To verify the effectiveness of the algorithm in identifying gas and water content, 600 training subsets were first generated from the original training dataset using the Bootstrap repeated sampling method. Then, a subset of feature factors were randomly selected from each training subset, and the internal nodes were split according to the principle of maximizing information gain ratio to generate 600 freely growing decision trees. These decision trees were then used to identify gas and water content in the validation set samples and the identification results were recorded. The most frequent result among the 600 identification results was selected as the final identification result for that set of validation set samples.

[0076] Since dataset D is an imbalanced dataset, the F1-score method is used to evaluate the performance of the RF algorithm.

[0077]

[0078]

[0079]

[0080] The random forest algorithm is introduced to learn from the training dataset, and then the gas and water content of the validation set samples are identified based on the learning results.

[0081] Step 4: Identify unknown reservoir fluids using models

[0082] Based on the random forest algorithm model trained in step three, fluid identification is performed on the unknown sample. First, the LIBS spectral data of the unknown sample is detected. Then, the LIBS spectral data of the unknown sample is input, and the model obtains the identification results of gas-bearing and water-bearing reservoirs of the unknown sample.

[0083] In Examples 1 and 2, LIBS spectral data were obtained using a LIBS experimental setup, which is a prior art technique. This setup includes an Nd:YAG (yttrium aluminum garnet) laser beam with constant energy per pulse, focused on the surface of the target sample and positioned at the top of the XYZ translational sample stage. An optical probe is used to transfer the light emission from the laser plasma to a spectrometer. The spectrometer consists of three spectral channels, covering 180–350 nm (channel 1), 350–580 nm (channel 2), and 580–790 nm (channel 3), respectively. Each channel has 2048 pixels, thus each LIBS spectral data can be converted into a 3×2048 matrix.

[0084] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A reservoir fluid identification method combining LIBS and random forest algorithms, characterized in that: Includes the following steps: Step 1: LIBS Spectral Data Collection and Processing Select an experimental area containing several sets of well logs. Extract several sets of LIBS spectra from the well log data of the gas-bearing and water-bearing reservoirs in each set of well logs. Obtain LIBS spectral data of several sets of well log samples. Then, perform preprocessing to obtain processed LIBS spectral data. Step 2: Construction of the Random Forest Algorithm Model Multiple random forest decision trees are constructed by using two randomness factors: random repeated sampling and random feature selection. The specific steps are as follows: S1: Bootstrap resampling method; S2: Random feature selection index; S3: Determine the optimal split node; S4: Decision Classification; In step two, the Bootstrap resampling method specifically involves: denoting the original training dataset of reservoir fluid LIBS spectra generated based on well logging data as D = ([x... i ]1, [x i ]2,…,x i ] n y i The Bootstrap sampling method is used to randomly sample K times from D, forming K independently distributed training data subsets {D}. k In step two, the random feature selection index is as follows: Decision trees with maximum growth are built for each of the K datasets without any pruning. Then, the information gain ratio of a subset of randomly selected feature factors from the training subset is calculated using the information gain ratio formula. In step two, determining the optimal splitting node involves splitting the candidate fluid identification factor with the highest information gain ratio and recalculating the information gain ratio using the formula. This splitting step is repeated until the information gain ratio is less than a set threshold. Finally, a forest model containing K decision trees is generated. In step two, decision classification involves integrating numerous decision trees for fluid identification built based on well logging data to construct a random forest classifier. This random forest classifier is then used to classify the obtained fluid identification factors, and the prediction results of all decision trees in the forest are statistically analyzed. The decision tree with the highest frequency of occurrence among the classification results of each decision tree is selected as the final classification result. Step 3: Random Forest Algorithm Model Training The single-class fluid identification factor is introduced into the random forest algorithm model constructed in step two. The validation set samples are used to identify whether they contain gas or water. Based on the identification results, fluid identification factors are selected. These fluid identification factors are then arranged into 1, 2, ..., N classes to form a new training dataset. The random forest algorithm model is then applied to the new validation set samples to make a judgment. From the judgment results, the categories and number of fluid identification factors suitable for identifying gas and water in this area are found. Step 4: Identify unknown reservoir fluids using models Based on the random forest algorithm model trained in step three, fluid identification is performed on the unknown sample. First, the LIBS spectral data of the unknown sample is detected, and then the LIBS spectral data of the unknown sample is input. The model is used to obtain the identification results of gas-bearing reservoirs and water-bearing reservoirs of the unknown sample.

2. The reservoir fluid identification method combining LIBS and random forest algorithm according to claim 1, characterized in that: In step one, the LIBS spectral data is collected in the range of 180–790 nm. The LIBS spectral data is collected from three sets of spectral channels, which cover 180–350 nm, 350–580 nm and 580–790 nm respectively. Each spectral channel has 2048 pixels.

3. The reservoir fluid identification method combining LIBS and random forest algorithm according to claim 1, characterized in that: In step one, the preprocessing method involves normalizing the total area of ​​the LIBS spectral data. The preprocessing formula is shown below: In the formula, I normalization It is the spectral intensity after preprocessing, I origin I represents the original spectral intensity. total It is the sum of the intensities of all data in the spectrum.

4. The reservoir fluid identification method combining LIBS and random forest algorithm according to claim 1, characterized in that: In step one, the processed LIBS spectral data is divided into a training set, a validation set, and a test set.

5. The reservoir fluid identification method combining LIBS and random forest algorithm according to claim 1, characterized in that: In step three, when verifying the effectiveness of the algorithm in identifying gas and water content, several training subsets are first generated from the original training dataset using the Bootstrap repeated sampling method. Then, a portion of feature factors are randomly selected from each training subset, and the internal nodes are split according to the principle of maximizing information gain ratio to generate several sets of freely growing decision trees. These decision trees are then used to sequentially identify gas and water content in the validation set samples and record the identification results. The result with the highest frequency is selected from several sets of identification results as the final identification result of that set of validation set samples.

Images