Logging Identification Method for Reservoir Facies in Fractured-Vuggy Dolomite Reservoirs Based on AI: A Case Study of Ediacaran Dengying Formation, Sichuan Basin, China

Abstract

As an important target for deep to ultra-deep carbonate oil and gas exploration, Fractured-Vuggy dolomite reservoirs have strong heterogeneity. Accurate characterization of reservoir facies is crucial for their exploration and exploitation. Three methods, including the unsupervised intelligent clustering method of improved Fuzzy C-means clustering Algorithm Based on Density Sensitive Distance and Fuzzy Partrition (FCM-DSDFP), the fusion method of Principal Components Analysis (PCA) dimensionality reduction and noise reduction, and the principle of clustering feature analysis are applied to identify reservoir facies based on logging data. Based on the PCA method, the logging response characteristics of the reservoir facies are excavated, and the fusion characterization data of dimensionality reduction and noise reduction are extracted. The FCM-DSDFP unsupervised intelligent clustering method, a model that approximates the subsurface conditions is established, and the reliability of the model is tested according to the elbow rule and silhouette coefficient. Combining drilling core observation, Fractured-Vuggy type, partially cemented Fractured-Vuggy type, Pore-Vuggy type, Pore Type I, Pore Type II, Tight Type I, and Tight Type II are divided in the Dengying Formation 4th Member. Fractured-Vuggy type, partially cemented Fractured-Vuggy type, Pore-Vuggy Type I, Pore-Vuggy Type II, Pore Type I, Pore Type II, and Tight Type are divided in the Dengying Formation 2nd Member, respectively. Two methods were applied to verify the reservoir facies identification results based on intelligent algorithms. The first method is to compare the identification results with the reservoir facies types identified by core observations (Well PT103 and PS13). The second method is to verify the recognition results of intelligent algorithms by utilizing the relationship between reservoir facies types and bitumen. The test results show that the accuracy of the reservoir level identification is higher than 0.9, and the applicability is better than the commonly used algorithms such as FCM and K-means.

1. Introduction

With the deepening of oil and gas exploration and development, deep and ultra-deep paleo carbonate rocks have become an important global field exploration and development [1,2,3]. Due to multiple periods of tectonic movement and diagenetic alteration [4,5,6], the deep and ultra-deep ancient carbonate reservoirs type are mainly fractured-vuggy dolomite reservoirs [7,8,9]. Oil and gas exploration in such reservoirs, particularly those represented by the Ediacaran Dengying Formation in the Sichuan Basin, Refs. [10,11,12] has led to the discovery of large gas fields such as Anyue and Penglai in recent years. These discoveries mark the first economically profitable Precambrian exploration targets worldwide [13,14]. The proven reserves demonstrate a vast scale of accumulation and suggest broad exploration prospects [15,16].

However, Fractured-Vuggy dolomite reservoirs have the characteristics of complex fabric, strong heterogeneity, and fast spatial distribution [15,16]. It is difficult to identify reservoir facies, which restricts the expansion of exploration scale and efficient development [17]. Reservoir facies form the foundation of Fractured-Vuggy dolomite reservoirs and represent the macro-physical variations and micro-pore characteristics of rocks [16,17,18]. Different configurations of pore and fracture arrangements in various reservoir facies lead to variations in the storage and permeability capacities of mound structures, resulting in high heterogeneity [19,20]. Therefore, accurately identifying reservoir facies is crucial for subsequent research on Fractured-Vuggy dolomite reservoirs.

The vertical accuracy of logging curves is high, and their response characteristics are the comprehensive characterization of geological and geophysical information such as rock minerals, fluid properties, and reservoir space size. Identifying complex reservoir facies based on the inherent connections between various conventional logging curves has always been a research hotspot [21,22,23]. However, the diverse types of reservoirs in the Dengying Formation greatly increase the difficulty of data processing and analysis. With the widespread application of intelligent algorithms in the field of geology [24,25,26], the Fuzzy C-means clustering (FCM) method identifies reservoir clusters based on the inherent connections of the data itself [27,28]. Its results are not easily affected by human factors and can be used to identify reservoirs. However, the algorithm itself is prone to getting stuck in local optima, which affects the accuracy of clustering.

Therefore, this study is based on three methods: the unsupervised intelligent clustering method of improved Fuzzy C-means clustering Algorithm Based on Density Sensitive Distance and Fuzzy Partrition (FCM-DSDFP), the fusion method of Principal Components Analysis (PCA) dimensionality reduction and noise reduction, and the principle of clustering feature analysis. By using core, thin-section, logging, and full-diameter core physical property data for practical applications, accurate identification of reservoir phases in fractured dolomite gas reservoirs is achieved in order to provide a strong basis for subsequent geological research.

2. Geological Setting

The study area is located in the SW China Sichuan Basin (Figure 1a), where paleo-uplift structures formed during the Caledonian period were developed (Figure 1b). Subsequently, after undergoing multiple stages of tectonic activities, the paleo-uplift underwent overall inheritance and development and ultimately formed after the Himalayan Movement [29,30]. According to microbial content and sedimentary structure, the Ediacaran Dengying dolomite formation can be divided into four members [9]. Among them, the 2nd and 4th Members comprise microbial dolostromatolite, dololaminite, and dolograinstone [2], the 1st Member is mainly dolowackestone, and the 3rd Member comprises shale dolowackestone and pelitic dolomite. Due to the uplift movement during the Tongwan period of the Late Ediacaran, the upper part of the 2nd and 4th Member underwent erosion and extensive karstification, forming an unconformity surface (Figure 1c). Unlike the Anyue area, most of the Penglai area lacks the 3rd Member and 4th Member, and the 2nd Member is directly in unconformable contact with the overlying Cambrian strata.

Under this background, the widely distributed microbial mound–shoal complex within the 2nd and 4th Member of the Dengying dolomite have formed large-scale, high-quality mound–shoal reservoirs [31]. The potential hydrocarbon source rocks for the Dengying reservoir include the following three sets: Cambrian Qiongzhusi shale, the 3rd Member mudstone, and shale within the Doushantuo Formation [32].

3. Method and Process

3.1. Reservoir Facies Identification Method

3.1.1. Unsupervised Intelligent Clustering Method Based on FCM-DSDFP

FCM-DSDFP is an improved method of the FCM unsupervised clustering method. FCM adopts mechanisms such as membership degree and cluster center and uses alternating optimization algorithms for intelligent solution and clustering. It has fewer parameters and is more advantageous in complex mapping relationships. However, the random selection of initial clustering centers and the local optimality of the Euclidean distance algorithm make FCM susceptible to noise interference. The clustering effect of FCM is unstable and ignores global consistency, resulting in a decrease in clustering accuracy. FCM-DSDFP overcomes the shortcomings of FCM and achieves global optimal clustering by introducing density-sensitive distance and fuzzy entropy, which can be used to solve the prediction problem of complex reservoir facies.

Based on fuzzy mathematics theory, FCM can divide the logging dataset X (x₁, x₂, x₃…x_n) into k classes (k₁, k₂, k₃…k_k), with an objective function of:

$$Y={\sum }_i^n{\sum }_j^k{b}_{ij}^m{c}_{ij}^2$$

(1)

In the formula, b_ij is the membership degree of sample point x_i to the k_j of the cluster center, $c_{ij}$ is the Euclidean distance between the sample point $x_i$ and the $k_j$ of the cluster center, and m represents the fuzzy factor (m > 1).

On this basis, FCM-DSDFP uses density-sensitive distance to replace Euclidean distance for measurement, enabling the function to perform global optimal interpretation. At the same time, fuzzy entropy is introduced to compensate for function partitioning defects, resulting in a new objective function G, namely:

$$\begin{gathered} G={\displaystyle \frac{4}{e^p}}{\sum }_i^n{\sum }_j^k{b}_{ij}\left(1-b_{ij}\right)D_{ij}^{\rho }+\omega {\sum }_i^n{\sum }_j^k{b}_{ij}\mathit{ln}{b}_{ij} \\ \mathrm{s}.\mathrm{t}. p={\mathit{log}}_{e}4n(e-1)-1 \end{gathered}$$

(2)

In this formula, e is the number of classifications. The parameter p is used to compensate for the defects in the partition coefficient and clustering function (generally p > 0). $D_{ij}^{\rho }$ is density sensitive distance. ρ is the scale factor (ρ > 0). ω represents the adjustment factor representing fuzzy entropy.

The initial clustering center is no longer randomly selected, but is calculated through the average density sensitive distance, which can effectively solve the problem of noise interference and reduce the number of iterations. Its function is:

$$D_{avg}^{\rho }={\displaystyle \frac{1}{C_n^2}}{\sum }_{i=1}^n{\sum }_{j=1}^i{D}_{ij}^{\rho }$$

(3)

In the formula, $D_{avg}^{\rho }$ is the average density sensitive distance. $C_n^2$ is the number of combinations of any two sample points selected from n samples.

3.1.2. Fusion Method Based on PCA Dimensionality Reduction and Noise Reduction

Principal Components Analysis (PCA) is the most widely used data dimensionality reduction algorithm, which preserves the most important features of high-dimensional data. PCA can eliminate useless information, suppress noise, compress data, and achieve the goal of improving data processing speed by removing noise and unimportant features, making data simplified and easy to use. In the process of predicting the reservoir facies of Fractured-Vuggy dolomite gas reservoirs, multiple complex logging parameters such as GR, AC, DEN, CNL, RT, etc. are faced, which will increase the difficulty of distinguishing geological, geophysical, and other characteristics. Based on PCA, effective information in high-dimensional space can be extracted and replaced by the number of principal components when the cumulative contribution rate of information is greater than 90%, providing fused characterization data for reservoir facies recognition.

3.1.3. Principles of Cluster Feature Analysis

(1) The elbow rule is used to scientifically optimize logging data categories based on the distortion degree of SSE (Sum of Squared Error). That is, as the number of clusters increases, the data points in each cluster become closer, and their distortion degree decreases. When a certain inflection point occurs, and the degree of distortion decreases slowly, it is considered the optimal number of clusters. Its function is:

$$\mathrm{SSE}={\sum }_{i=1}^k{\sum }_{x\in k_i}^n|x-{\mu }_i|$$

(4)

In the formula, k_i is the i-th classification cluster. μ_i is the mean vector of the i-th class classification cluster.

(2) When evaluating the clustering effect based on the silhouette coefficient, the closer the silhouette coefficient is to 1, the more reasonable the cluster where the sample is located. The silhouette coefficient is close to −1, indicating that the data should be further divided into other clusters, resulting in poor classification performance.

3.2. The Recognition Process of Intelligent Clustering Based on PCA-FCM

Combining interdisciplinary theoretical knowledge such as geophysical logging, geology, and artificial intelligence, three methods were used: (1) unsupervised intelligent clustering method based on FCM-DSDFP, (2) fusion method based on PCA dimensionality reduction and noise reduction, and (3) clustering feature analysis to carry out intelligent learning-based reservoir facies identification.

The specific steps are (1) based on geological theory background, select standard wells for logging preprocessing and standardization, use PCA to extract favorable rock facies features from conventional logging data, and provide fused characterization data; (2) based on the FCM-DSDFP unsupervised intelligent clustering method, a model that approximates the underground reservoir is established using fused characterization data. The reliability of the model is tested using elbow rule and silhouette coefficient; (3) combining core, core testing, conventional logging, and imaging data, assign geological significance to each cluster.

4. Results and Testing

4.1. Reservoir Facies Identification Results

This study collected logging and core data from a total of 26 wells in Dengying Formation 4th Member (10 wells) and Dengying Formation 2nd Member (16 wells). Two cored wells were randomly selected for clustering effect testing, and the remaining well data were used for model clustering training. Among them, logging data are used as input for identifying reservoir facies, including GR, AC, CNL, DEN, RT, RXO, K, Th, U, and PE (

Table 1. Feature statistics of fusion methods based on PCA dimensionality reduction and noise reduction (DEN = Density logging, AC = Acoustic time difference logging, RT = True formation resistivity logging, RXO = Flushed zone formation resistivity logging, CNL = Neutron logging, GR = Natural gamma ray logging).

Dengying Formation 4th Member						Dengying Formation 2nd Member
	PC1	PC2	PC3	PC4	PC5		PC1	PC2	PC3	PC4	PC5
DEN	0.1	0.56	−0.76	0.29	−0.08	DEN	0.07	−0.72	0.01	−0.69	0.06
AC	0.03	−0.61	0.62	−0.49	0.01	AC	−0.05	0.01	0.95	−0.01	0.01
RT	−0.65	−0.15	−0.05	0.19	−0.72	RT	−0.7	0.01	−0.04	−0.02	0.71
RXO	−0.65	0.1	−0.13	0.26	0.69	RXO	−0.69	0.04	−0.03	−0.18	−0.69
CNL	0.37	−0.54	−0.05	0.76	−0.02	CNL	0.12	0.69	−0.01	−0.7	0.1
GR	0.23	0.14	−0.31	0.11	0.03	GR	0.1	0.01	0.02	0.15	0.22
K	0.16	0.01	0.04	0.26	−0.3	K	0.16	−0.01	0.01	0.14	0.1
Th	0.25	0.11	0.05	0.12	0.01	Th	0.25	−0.11	0.06	0.01	0.01
U	0.03	0.06	0.19	0.05	0.07	U	−0.03	0.06	−0.01	0.11	0.07
PE	0.26	0.16	0.21	−0.17	0.22	PE	−0.21	0.16	0.11	−0.17	0.12
Vector value	1.99	1.67	0.71	0.43	0.2	Vector value	1.77	1.35	1	0.65	0.24
Information	0.4	0.33	0.14	0.09	0.04	Information	0.35	0.27	0.2	0.13	0.05
Accumulated information	0.4	0.73	0.87	0.96	1	Accumulated information	0.35	0.62	0.82	0.95	1

). The relevant data are used for the qualitative classification of clusters, giving them geological significance, including core, thin-section, and full-diameter core physical property testing data. When using two cored wells as inspection standards, the differences in reservoir facies between the Dengying Formation 4th Member and Dengying Formation 2nd Member were taken into consideration. Based on the PCA fusion method, dimensionality reduction, and noise reduction were performed on the logging curves of the Dengying Formation 4th Member and Dengying Formation 2nd Member, with a cumulative information value greater than 0.9 to cover the complete information standard. Four principal component information curves were obtained for the Dengying Formation 4th Member and Dengying Formation 2nd Member, that is, PC1, PC2, PC3, and PC4 (Table 1).

In the principal component information parameters of Dengying Formation 4th Member, PC1 has a high negative correlation with RT and RXO, with values of −0.65 and −0.65, respectively. PC2 has a high correlation with DEN and AC, with a positive correlation with DEN and a negative correlation with AC of 0.56 and −0.61, respectively. PC3 has a high correlation with DEN and AC, with a negative correlation with DEN and a positive correlation with AC, with values of −0.76 and 0.62, respectively. PC4 has a high correlation with CNL and AC, with a positive correlation with CNL and a negative correlation with AC of 0.76 and −0.49, respectively.

In the principal component information parameters of Dengying Formation 2nd Member, PC1 has a high negative correlation with RT and RXO, with values of −0.69 and −0.70, respectively. PC2 has a high correlation with DEN and CNL, with a positive correlation with CNL and a negative correlation with DEN of 0.69 and −0.72, respectively. The positive correlation between PC3 and AC is high, at 0.95. PC4 has a high negative correlation with DEN and CNL, with values of −0.69 and −0.70, respectively.

Using the unsupervised intelligent method of FCM-DSDFP, cluster the principal component information parameters obtained from Dengying Formation 4th Member and Dengying Formation 2nd Member separately, with an iteration count of 10. Scientifically analyze the number of clusters and the clustering effect using the principle of clustering feature analysis. Among them, the distortion degree of Dengying Formation 4th Member and Dengying Formation 2nd Member shows a significant mutation at position 7, and the silhouette coefficients at position 7 are 0.84 and 0.77, respectively, which are closest to 1 (Figure 2a,b). Based on the two indicator coefficients, it is suggested that the Dengying Formation 4th Member and Dengying Formation 2nd Member are most suitable to be divided into seven categories, and the clustering effect is the best.

As shown in the cross-plot of PC1 and PC3 (Figure 3), the larger the PC1 in the Dengying Formation 4th Member, the lower the RT and RXO, the larger the PC3, the smaller the DEN, and the higher the AC, indicating a reservoir facies with well-developed reservoir space in the upper right corner and a relatively dense reservoir facies in the lower left corner. The larger the PC1 in the Dengying Formation 2nd Member, the lower the RT and RXO, the larger the PC3, and the higher the AC. The upper right corner indicates the reservoir facies with well-developed reservoir space, while the lower left corner indicates the relatively dense reservoir facies. In addition, each type of data has fuzzy transition areas, indicating a transition phenomenon between each type of reservoir facies and reservoir facies with little difference in characteristics, which is consistent with the observation results of the rock core.

4.2. Reservoir Facies Characteristics

4.2.1. Dengying Formation 4th Member

The core samples, characterized by the Fractured-Vuggy type, are brownish-gray. The lithology primarily consists of microbial dolostromatolite with a range of vug sizes from large to medium, with bitumen as the predominant fill material. Core porosity varies from 2.22% to 5.27%, with an average of 3.58%, and permeability ranges from 0.33 mD to 8.96 mD, averaging 0.63 mD. The Fractured-Vuggy type is characterized by conventional logging responses, including a significantly enlarged caliper, high AC values, elevated neutron porosity, and notably low resistivity with distinct invasion characteristics (Figure 4).

Partially cemented Fractured-Vuggy type cores are brownish-gray. The lithology primarily consists of microbial dolostromatolite with small vugs, often filled with bitumen. Core porosity varies from 2.20% to 5.36%, with an average of 3.25%, and permeability ranges from 0.28 mD to 6.78 mD, averaging 0.48 mD. Partially cemented Fractured-Vuggy type is characterized by conventional logging responses, including moderate to high AC value, significant fluctuations in neutron porosity, and notably low resistivity without distinct invasion characteristics (Figure 4).

Pore-Vuggy type samples are light and brownish-gray, with microbial dolograinstone as the predominant lithology. Porosity in the cores varies from 1.42% to 4.51%, with an average of 2.92%; permeability ranges from 0.17 mD to 4.87 mD, averaging 0.32 mD. The Pore-Vuggy type is characterized by conventional logging responses, including slightly enlarged caliper, medium to high resistivity, inconsistent invasion characteristics, and a wide range of densities (Figure 4).

Pore Type I samples are colored light and brownish-gray, predominantly composed of microbial dololaminite. Porosity varies from 0.26% to 3.57%, averaging 1.45%, while permeability ranges from 0.13 mD to 1.28 mD, with an average of 0.028 mD. Pore Type I is characterized by highly variable resistivity (Figure 4).

Pore Type II samples are light and brownish-gray, composed primarily of microbial dololaminite. Core porosity varies from 0.85% to 3.38%, with an average of 1.10%, while permeability ranges from 0.11 mD to 1.36 mD, with an average of 0.017 mD. Pore Type II is characterized by elevated natural gamma rays, high-density values, and no clear invasion characteristics (Figure 4).

Tight Type I samples are light gray and brownish-gray, primarily composed of dolowackestone. Porosity ranges from 0% to 2.15%, with an average of 1.02%, and permeability ranges from 0.01 mD to 0.47 mD, with an average of 0.018 mD. Tight Type I is characterized by a normal caliper, high-density values, extremely high resistivity, and indistinct invasion characteristics (Figure 4).

Tight Type II samples are light and brownish-gray, primarily composed of wackestone. Porosity varies from 0% to 1.90%, with an average of 0.25%, and permeability is below 0.01 mD. Tight Type II is characterized by a normal caliper, lower density than Tight Type I, extremely high resistivity, and no invasion characteristics (Figure 4).

4.2.2. Dengying Formation 2nd Member

Fractured-Vuggy types are colored light and brownish-gray in the core, predominantly comprising large to medium vug microbial dolostromatolite. The primary fill material is bitumen. Core porosity varies from 2.12% to 9.73%, averaging 4.65%; permeability ranges from 0.56 mD to 5.88 mD, averaging 1.15 mD. Fractured-Vuggy type is characterized by a significant enlargement of the caliper, extremely high AC, high neutron porosity, distinctly low resistivity, and invasion characteristics (Figure 5).

Partially cemented Fractured-Vuggy type core samples are colored light gray and brownish-gray, primarily comprising medium-vug small-fracture microbial dolostromatolites. The primary cement phase is bitumen. Core porosity varies from 1.89% to 12.39%, with an average of 4.24%; permeability ranges from 0.37 mD to 5.87 mD, with an average of 0.87 mD. Partially cemented Fractured-Vuggy type is characterized by a significant enlargement of the caliper, greatly increased AC, significantly lower density, a wide range of neutron porosity, extremely low resistivity, and obvious invasion characteristics (Figure 5).

Pore-Vuggy Type I samples are both light and brownish-gray, primarily composed of small-vug microbial dolograinstone. Core porosity ranges from 1.39% to 4.57%, averaging 2.18%, and permeability varies from 0.17 mD to 2.17 mD, averaging 0.20 mD. Pore-Vuggy Type I is characterized by a normal caliper, medium-to-low resistivity, subtler invasion characteristics compared to Fractured-Vuggy Type, and lower density (Figure 5).

Pore-Vuggy Type II samples are colored light gray and brownish-gray, primarily composed of small-vug breccia dolostone. Core porosity varies from 1.15% to 5.11%, averaging 2.10%, and permeability ranges from 0.13 mD to 1.22 mD, averaging 0.16 mD. Pore-Vuggy Type II is characterized by a normal caliper, medium to high resistivity, and slight invasion characteristics. Compared to Pore-Vuggy Type I, Pore-Vuggy Type II displays a significantly lower density (Figure 5).

Pore Type I samples are light and brownish-gray, primarily composed of fine-medium crystalline dolostone. Core porosity varies from 0.62% to 5.62%, averaging 1.66%, and permeability ranges from 0.08 mD to 4.22 mD, averaging 0.10 mD. Pore Type I is characterized by a normal caliper, elevated density values, and no invasion characteristics (Figure 5).

Pore Type II samples appear light gray and brownish-gray, primarily comprising microbial dolograinstone. Porosity varies from 0.44% to 5.11%, averaging 1.59%, and permeability ranges from 0.04 mD to 1.08 mD, with an average of 0.11 mD. Pore Type II is characterized by elevated gamma rays, enlarged caliper, low density, high resistivity, and no invasion characteristics (Figure 5).

The Tight Type in core samples is light gray, primarily composed of dolowackestone. Porosity varies from 0% to 2.88%, averaging 0.71%, and permeability ranges from 0 mD to 0.41 mD, with an average of 0.044 mD. Conventional logging responses show no enlargement or contraction caliper, high bulk density, high resistivity, and lack evident invasion characteristics (Figure 5).

4.3. Testing Methods

Two methods were applied to verify the reservoir facies identification results based on intelligent algorithms proposed in this study. The first method is to compare the identification results with the reservoir facies types identified by core observations (Well PT103 and PS13). As shown in Figure 6, the reservoir facies types divided by this method are highly consistent with the core observation results, and the boundary division of the types is accurate. As shown in Figure 6a,b, the boundary positions of the Fractured-Vuggy type or partially cemented Fractured-Vuggy type correspond well with the boundary positions divided by the intelligent algorithm.

The second method is to verify the recognition results of intelligent algorithms by utilizing the relationship between reservoir facies types and hydrocarbon accumulation (Figure 7). Previous studies have shown that the hydrocarbon accumulation process of the Ediacaran Dengying Formation can be divided into the following three stages (Figure 7). The first stage was the paleo-oil charge event of the Triassic, which formed the paleo-oil reservoir of the Dengying Formation [33]. During the 2nd stage, the Dengying paleo-oil reservoir was gradually cracked into gas and bitumen during the Middle Cretaceous period [34,35]. During the last stage, intense compression during the Himalayan period resulted in rapid uplift with the exposure of Dengying paleo-reservoirs, leaving only solid bitumen in the Dengying reservoir (Figure 8a).

Thus, the content of bitumen is controlled by the charging of paleo crude oil (Figure 8b). The migration and accumulation of crude oil require a higher lower porosity limit than natural gas. Li et al. (2016) proposed that the lower limit of porosity for the formation of paleo-oil reservoirs in carbonate reservoirs is 2.6% [36]. Taking the Dengying Formation 2nd Member as an example, the only reservoir types with porosity exceeding 2.5% are the Fractured-Vuggy type and partially cemented Fractured-Vuggy type (Figure 8c). Thus, only these two types of reservoir facies can participate in the accumulation and cracking process of paleo-oil reservoirs, thereby enriching bitumen (Figure 8d).

The content of bitumen can reflect the accuracy of identifying reservoir facies types (Figure 8e,f). This study used four principal component information curves as inputs and trained a model using SGBDT (Stochastic Gradient Boosting Decision Tree) to form a model that is close to the actual situation of subsurface oil reservoirs, achieving an accurate prediction of bitumen content. According to the parameter tuning of the grid search algorithm, the decision tree depth of the SGBDT algorithm is 12, the learning rate is 0.1, the number of iterations is determined as 100 based on the expected loss calculation formula, and the subsampling random factor of SGBDT is 70%.

According to the prediction results of the Dengying Formation 2nd Member in Well PT1, the correlation between the optimized predicted bitumen content and the experimentally measured bitumen content reached 0.87 (Figure 9), indicating that the prediction results based on the SGBDT model are reliable. As shown in Figure 10, the development positions of the Fractured-Vuggy type and partially cemented Fractured-Vuggy type are highly consistent with the high bitumen content interval, indicating a high level of reliability in reservoir facies prediction.

To assess the proposed algorithm’s applicability and the input data’s reliability, control experiments are carried out using four unsupervised learning algorithms (Figure 11): PCA-FCM-DSDFP, FCM-DSDFP, FCM, and K-means. The Cramer’s V and Crossplot coefficients served as standards for testing algorithm accuracy. Results showed that for denoised input data using the PCA algorithm, the Klem and convergence coefficients were 0.932 and 0.881, respectively, while for non-denoised data, they were 0.928 and 0.837. Non-denoised input data somewhat interfered with algorithm accuracy. In control experiments with FCM-DSDFP, FCM, and K-means algorithms, FCM-DSDFP accuracy was higher than FCM and K-means. This indirectly confirmed that PCA-FCM-DSDFP is more applicable than FCM and K-means, and it can be applied to similar reservoirs to further verify its suitability. It is noteworthy that for reservoirs with few categories, unsupervised learning algorithms are suitable. However, for reservoirs with complex categories, unsupervised learning classification remains challenging.

5. Conclusions

An unsupervised intelligent clustering method based on FCM-DSDFP, a fusion method based on PCA dimensionality reduction and noise reduction, and principles of cluster feature analysis were integrated to identify reservoir facies in Fractured-Vuggy dolomite reservoirs. Seven types of reservoir facies were identified in Dengying Formation 2nd Member and Dengying Formation 4th Member. Combining drilling core observation, conventional logging data, and identification results of reservoir facies, Fractured-Vuggy type, partially cemented Fractured-Vuggy type, Pore-Vuggy type, Pore Type I, Pore Type II, Tight Type I, and Tight Type II are divided in the Dengying Formation 4th Member. Fractured-Vuggy type, partially cemented Fractured-Vuggy type, Pore-Vuggy Type I, Pore-Vuggy Type II, Pore Type I, Pore Type II, and Tight Type are divided in the Dengying Formation 2nd Member. The observation results of drilling cores and the prediction of bitumen content are used to test the identification results; the algorithm’s accuracy was above 0.90, reflecting a high level of reliability in reservoir facies prediction. Compared with conventional algorithms, this method showed superiority in highly heterogeneous Fractured-Vuggy carbonate reservoirs, providing an approach for precise characterization. However, for reservoirs with extremely complex categories, unsupervised learning classification remains a significant challenge.