Practical application of artificial intelligence model optimization based on multi-source data in oil and gas reservoir evaluation(https://doi.org/10.63386/620073)

Authors: Yu Ji¹, Lirong Dou^1*, Kunye Xiao¹

 Affiliations:

¹Research Institute of Petroleum Exploration and Development, PetroChina, Beijing 100083, China

Emails of Authors:

First author: 15833767617@163.com

Second author: Dlirong@petrochina.com.cn

Third author: xiaokunye@petrochina.com.cn

Corresponding author: Dlirong@petrochina.com.cn

Project Funding:

This research was financially supported by grants from the National Natural Science Foundation (92255302)

Authors’ bio:

Ji Yu, born in 1993 in Hebei, completed his undergraduate studies at Yangtze University and earned a Master’s degree in Petroleum Engineering from the University of Southern California (USC). He is currently pursuing a Doctoral degree in Mineral Exploration and Surveying at the Research Institute of Petroleum Exploration and Development. His primary research interest is the application of artificial intelligence in the oil and gas industry.

Li Rong Dou, born in August 1965 in Yangzhou, Jiangsu, is a Professor and Senior Engineer with a Doctoral degree. He currently serves as the Director of the National Oil and Gas Strategic Research Center and the Executive Director and President of the China Petroleum Exploration and Development Research Institute. He is also a doctoral supervisor at multiple institutions, including Yangtze University and China University of Petroleum (Beijing). Additionally, he holds leadership roles at the National Energy Oil and Gas R&D Center and the Key Laboratory of Efficient Exploration and Development of XAI and Natural Hydrogen Resources.

Xiao Kunye graduated from Wuhan University of Geosciences. His main research areas include geology and oil and gas exploration and development. He has conducted multiple studies in the Doseo Basin of Chad and the Termit Basin of Niger in West Africa, addressing challenges related to structural complexity caused by multiple phases of strike-slip movements and unclear reservoir formation patterns.

Practical application of artificial intelligence model optimization based on multi-source data in oil and gas reservoir evaluation

Abstract: This paper analyzes the multi-source characteristics of reservoir data based on the artificial intelligence model of multi-source data, covering exploration data, development data, oil and water well production data and other related reservoir data. Based on the concept of transfer learning, different network branches of source and target domains are constructed, and the Transformer network is reconfigured and optimized to achieve unified processing of different data. Combining the advantages of CNN and Transformer, a hybrid artificial intelligence model is established to learn the spatial features of reservoirs in the image, accurately evaluate the reservoir features, and provide an innovative and effective solution for reservoir identification. The experimental results show that the AI model based on multi-source data has an extremely low error rate for depth prediction of each well, with the depth prediction error rate ranging from 0.001% to 0.0026%, and most of the permeability, maximum permeability, and minimum permeability prediction error rates within 0.5%, with a recall HR of 0.81, an average Pearson’s correlation coefficient of 0.99957, and a coefficient of determination of (R²) 0.99899. Through the optimized and comprehensive application of artificial intelligence model for multi-source data, reservoir characteristics can be identified and evaluated more accurately, providing powerful support for oil and gas exploration and development, with a view to promoting the progress of oil and gas exploration and development technology and guaranteeing stable energy supply.

Keywords: multi-source data; artificial intelligence model; transfer learning; transformer network; oil and gas exploration and development

1. Introduction

In the field of energy, the efficient exploration and development of oil and gas resources has always been the focus of global attention, and the evaluation of oil and gas reservoirs, as a key link in the process of oil and gas exploration and development, plays a crucial role in accurately assessing the value of reservoir resources and formulating scientific and reasonable exploitation plans [1]. The traditional means of oil and gas reservoir evaluation mainly relies on a single data source or simple analysis methods, however, the reservoir geological structure is complex and variable, involving rock physical properties, fluid distribution, tectonic features and other levels, a single data often can not comprehensively and accurately portray the whole picture of the reservoir, resulting in the limitations of the evaluation results, and it is difficult to meet the needs of the current fine exploration and efficient development [2-3]. Multi-source data acquisition technology has been constantly innovated, covering data acquisition in geology, geophysics, geochemistry and other multidisciplinary fields, which brings massive and abundant information for oil and gas reservoir evaluation. After learning multi-source data, the model can enhance the adaptability to complex reservoir situations, improve the generalization ability, and reduce the risk of erroneous evaluation due to single data bias [4]. By integrating data from different sources and formats, using advanced algorithms and models, and optimizing the modeling, the reservoir feature information can be extracted more effectively, and the multi-source data can be mined and analyzed in depth, revealing key information such as reservoir features and oil and gas transport patterns [5].

Therefore, it is of great significance to carry out the research on data algorithm modeling optimization of multi-source data in oil and gas reservoir evaluation, in order to promote the progress of oil and gas exploration and development technology, and improve the development and utilization efficiency of oil and gas resources. The dataset in this paper covers exploration data, development data, oil and water well production data and other related reservoir data, which provides a rich and accurate data base for subsequent model construction. Based on the concept of transfer learning, we construct an efficient intelligent model of Transformer transfer learning network, use TransformerBlock stacking to design the migration architecture, construct different network branches in the source and target domains, and design a loss function that includes the mean-square loss and the loss of the maximum difference in the mean, which can make the Transformer transfer learning network better adapt to the multi-source data and improve the model’s performance in the oil and gas industry. The lifting model is used in oil and gas. Aiming at the characteristics of time-frequency images of logging signals, a hybrid model combining the advantages of CNN and Transformer is constructed. CNN can effectively extract the local details of the image and capture the local patterns and structures in the spectral image, while Transformer extracts the more advanced abstract features by virtue of the self-attention mechanism to model different location features such as porosity, permeability, etc., and obtains the final prediction output to complete the classification and classification of reservoirs and the final prediction output. Output, complete the reservoir for classification and evaluation evaluation.

2. Relevant Works

In an application of complex reservoir characterization and intelligent model building, Akinwumiju A A research focused on the Kimmeridge Clay Formation shale reservoir in the North Sea, using machine learning techniques combined with established rock property equations to generate a neural network model of geochemical logging profiles. Numerical simulation techniques were used to model and analyze the key attributes of porosity and permeability of the shale reservoir, and the region was found to be as high as 9 wt% of the original total organic carbon (TOCo), 48/g of the original hydrocarbon production (S2o), and the KCF attribute map generated by the study of the KCF attributes can help the geologists to assess the potential of shale oil and gas development of the district, and to identify the areas with the prospect of exploitation [6]. Fernandes, F. B. In order to solve the mechanical diagenetic damage caused by pore collapse, a new perturbation solution method was proposed for modeling the hysteresis effect of pore transient collapse in production curve-dependent reservoirs. In the model construction, the effects of various aspects such as the mechanical properties of reservoir rocks, fluid pressure changes, and time factors are considered. The analysis shows that well shut-in pressure has an effect on permeability loss, which represents less than 5% of the permeability value in the previous decline cycle. Through asymptotic analysis and convolutional operations, the degree of compaction damage in the reservoir at different stages of mining is more accurately predicted, and the current status of the reservoir is more accurately assessed, which provides a scientific basis for the adjustment of the subsequent mining program [7]. Chen, C Taking the Dongying Formation as an example for the low-resistivity reservoirs formed in this area due to the high irreducible water saturation and the extra clay electrical conductivity, the quantitative assessment of the low-resistivity reservoirs in the study area A new water saturation model was applied to determine the main controlling factors of resistivity reservoirs, and specific identification criteria for these reservoirs were proposed so that the intelligent model can accurately deal with the evaluation of such reservoirs [8]. Zhao, G established an integrated approach to study the Gulong Shale oil reservoir. This reservoir is characterized by porosity, high maturity, etc., and the production performance shows the characteristic of “gas first and then oil”, and it is difficult to analyze the initial state of fluid existence and fluidity by conventional methods. The initial state of existence is quantitatively evaluated by equation of state (EOS) calculations, taking into account the nano-restriction effect, and combined with two-dimensional nuclear magnetic resonance (NMR) and molecular dynamics simulations. The saturation of the moveable fluid is quantified by centrifugal experiments to provide experimental data support for the smart model in fluid evaluation in shale oil reservoirs [9].

In the practice of intelligent model-related technology in reservoir development and pressure assessment, Wang, X et al. proposed a new method for predicting the development pattern of submarine reservoirs in response to the complex spatial distribution problem caused by the extremely non-homogeneous nature of submarine reservoirs. Using the box method and normalization formula to process and normalize the anomaly data from elemental logging and engineering logging, and then optimizing the sensitive parameters, a new method for predicting the development pattern of subduction reservoirs was established by using the optical gradient enhancer algorithm, deep neural network (DNN) and support vector machine (SVM). And through the comprehensive evaluation of F1 score, it promotes the optimal application of intelligent models in the dynamic prediction of reservoir development [10]. Lin, H et al. proposed an intelligent fusion model to predict the horizontal principal stress for the difficult problem of predicting the mechanical parameters of the rocks in complex lithological reservoirs, adopting the rocks of transitional shale reservoirs as the object of study. Based on the laboratory test data, machine learning algorithms such as nearest neighbor regression, support vector and random forest were selected to construct the intelligent fusion model for different rock mechanical parameters, which provides new model construction ideas and constraints for the intelligent model in the evaluation of reservoir stress characteristics [11]. Production capacity is limited due to excessive exposure time of drilling fluid equilibrium.Zhao, X et al. selected the main reservoir core of a block well H-1 in the East China Sea to conduct experiments to assess the dynamic damage of drilling fluid. Determining the range of reservoir permeability damage rate, combined with the experimental parameters, it was found that the reservoir permeability damage rate caused by drilling fluid intrusion ranged from 58.25% to 87.25%, and the overall dynamic damage could be categorized as medium to high, which provided experimental data for the intelligent modeling in the assessment of damage in low permeability reservoirs [12].

In recent years, with the rapid development of artificial intelligence technology, especially the successful application of big data algorithmic models in the field of natural language processing, machine translation, etc., it provides new ideas for the intelligent processing and analysis of oil and gas reservoir data.

3. Reservoir data multi-source analysis

As a complex system with multidisciplinary and multidisciplinary collaboration, the core business of an oilfield enterprise involves oil and gas exploration, development and production, and is highly dependent on the accurate management of underground reservoirs. However, reservoir management still faces many challenges in oilfield practice, especially the problem of multiple sources of data. That is, the amount of oil and gas production is the result of collaboration between many departments and various disciplines, not a department or a discipline or field to do it can be realized, the more important thing, in fact, is how much capacity of the reservoir, how much oil and gas resources [13]. As a large number of reservoir data have different attributes, formats and accuracies, the purpose of multiplicity analysis is to reveal the intrinsic connection and differences between these data, and to provide a basis for subsequent data processing and model training. At the same time, data standards are not uniform, and different data models are constructed by different standards, and these different sources and classes of data bring great challenges to reservoir management. At present, the multiple sources of reservoir data are mainly reflected in the following aspects:

(1) Exploration data comes from the data generated by the exploration process of oil and gas resources. It can be divided into two categories of basic data and graphic data from the major categories, and the classification model of basic database of exploration data is shown in Figure 1, and only the basic database construction is divided into 14 major categories.

Figure 1 Classification Model of basic database of exploration data

(2) Development data comes from the data generated in the process of oil and gas development, mainly including drilling geological information, additional information on drilling geology, drilling geological design, target point data of inclined wells and horizontal wells, well location data of straight wells, well coordinates data, data on small layers of single wells, geological stratification data, and stratification data of oil and gas layers. Common data of coring design, data of scrapped wells, basic data of oilfield, data of tectonic elements, data of reservoir properties, data of reservoir fluid properties, data of water layer evaluation, etc.

(3) Oil and water well production data comes from the data generated by the production process of oil wells, mainly the data of single well production and water injection volume of oil and gas wells. It mainly includes daily production data of oil and gas wells. Newly commissioned daily production data, water injection well data, measure well data, variation well data, check production monthly data and sales data, etc.

(4) Other reservoir data are more from the data of oil and gas resources census, exploration, evaluation, etc., in which there are more seismic and non-seismic data in physical and chemical exploration, especially the seismic reservoir prediction data is very important for reservoir management.

Reservoir management and research for the purpose of oilfield multi-source data, will involve a very wide range of data needs, so that both the establishment of a very powerful database, including the basic database, graphic database, professional database and so on. At the same time, it is also necessary to establish data visualization platform, research application platform and other multidisciplinary integrated collaborative work environment. Thus, it can change the way of reservoir management and research, realize the unity of data flow, information flow and business flow, achieve the efficient organization of data, and improve the efficiency and level of scientific research as a whole. The multiple sources of data make the degree of change of reservoir management larger and more complex [14].

4. Transformer Migration Learning Network Intelligent Modeling for Oil and Gas Reservoir Evaluation

4.1 General Network Architecture

4.1.1 Feature extraction network

Migration learning network architectures can be applied in oil and gas reservoir evaluation to solve the problems of data scarcity and model generalization capability. In oil and gas reservoir evaluation, high-quality data may be very scarce. Migration learning can utilize models that have been trained on related domains or tasks and migrate their knowledge to new tasks, thus reducing the dependence on large amounts of new data. The feature extraction layer in a feature migration learning network can be useful for processing reservoir data with multivariate nature, learning the common features of the reservoir data, thus enabling uniform processing of different data. This helps to recognize and evaluate the reservoir more accurately. According to the concept of migration learning to construct the source and target domains, the source domain refers to the logging and core data of the old workover, and the target domain refers to the logging data of the new workover, which is needed to predict the reservoir parameters, this paper constructs the Tiansformer Migration Learning Network which contains the following three core parts:

(1) Construct Transformer feature extraction network reservoir is affected by many geological factors such as tectonic movement, diagenesis and depositional environment, and there is a complex nonlinear mapping relationship between reservoir parameters and logging parameters [15].

(2) In this paper, the Transformer network is reconstructed and optimized that is, the Transformer Block introduces a multi-attention mechanism, which is able to explore the association between logging data and reservoir parameters in different subspaces, and then build a feed-forward neural network to carry out forward computation based on this association, to achieve feature extraction at different levels.

(3) At the same time, residual connectivity and layer normalization are added in each part to avoid model training failure due to gradient disappearance. This feature extraction network can accurately capture the key information in multi-source data, which lays the foundation for subsequent reservoir evaluation.

4.1.2 Model Architecture Design

After completing the construction of the feature extraction network, the next key step is the design of the Transformer migration learning model architecture, which is a key prerequisite for reservoir prediction and the basis for realizing the migration function. Different network branches in the source and target domains are constructed, and the logging data are forward-operated with the model weight coefficient matrix and bias vector to obtain the prediction output. Based on the feature data obtained from the middle layer of the source and target domains, the maximum mean difference is used to calculate the difference in the distribution of feature data between the source and target domains, showing that the source domain is the logging and core data of the old workover and the target domain is the logging data of the new workover. The logging data are forward-operated with the model weight coefficient matrix and bias vector to obtain the predicted output [16]. The whole network architecture is new and can accurately realize the reservoir parameter prediction to provide a reliable basis for the reservoir development decision, which further improves the overall framework of the migration learning network.

4.1.3 Network optimization guidelines

After having a reasonable network architecture, in order for the model to be better trained and achieve accurate prediction, the criterion that guides the model training, i.e., the loss function, needs to be designed to optimize the parameters of the Transformer migration learning network. The loss function designed in this paper includes the mean-square loss and the maximum mean difference loss, which is the mean-square difference between the predicted values of the source domain and the true values of the reservoir parameters, and the maximum mean difference loss is the difference in the data distribution of the features of the source domain and the target domain in the high-dimensional space. Using stochastic gradient descent algorithm, the migration function is achieved by minimizing the loss value and optimizing the model weight coefficient matrix and bias vector [17]. Using stochastic gradient descent algorithm, the model weight coefficient matrix and bias vector are optimized by minimizing the loss value to achieve the migration function. Figure 2 shows the Transformer migration learning network, through this optimization criterion, the Transformer migration learning network can better adapt to multi-source data, improve the accuracy and generalization ability of the model in oil and gas reservoir evaluation, and provide reliable technical support for actual oil and gas exploration and development.

Figure 2 Transformer migration learning network

4.2 CNN-Transformer reservoir identification and evaluation

4.2.1 Normalization

While Transformer excels in capturing global information, CNN is equally indispensable in extracting local features. Combining the advantages of both, a more powerful evaluation model for reservoir identification can be constructed. The Transformer model used in this paper employs scaled dot product attention as the component unit of multi-head attention [18]. Firstly, the dot product of query (0) and key () is computed, and then divided by the scaling factor  scaled to get the attention score, which is then normalized with the value to compute the dot product to get the final output:

(1)

where  denotes the matrices of enquiry, key and value, respectively,  is the -dimensional size of the -region, and  is the scaling factor. The effect is to turn the attention matrix into a standard normal distribution to make the matrix more stable after Softmax normalization process [19].

4.2.2 Multiple Attention Structures

The structure of multi-head attention is shown in Fig. 3, which means integrating multiple scaled dot product attentions, each scaled dot product attention has a query, key and value, and feeds into another linear mapping by connecting the separately computed attention values. Multiple attention heads are computed in parallel, which helps to capture information from different subspaces at different locations and extract more features. In the oil and gas reservoir identification task, it provides richer feature information for reservoir identification [20].

Figure 3 Multi head attention structure

4.2.3 Construction of hybrid models

For the characteristics of time-frequency images of logging signals, a hybrid model based on Convolutional Neural Network (CNN) and Transformer is proposed, which takes full advantage of the respective strengths of CNN and Transformer.CNN is able to efficiently extract the local detailed features of the image and capture the local patterns and structures in the spectral image. Transformer, on the other hand, is able to capture the global information of the image and model the features at different locations through the self-attention mechanism to understand the overall context of the image [21-22]. The main structure of the model can be divided into 3 parts, and Figure 4 shows the structure of the CNN-Transformer model.

The first part is the feature extraction module which is used to process the input data and extract the low-level features of the image. The module first employs two consecutive 3×3 convolutional and activation function layers designed to extract the underlying detailed features of the image. Subsequently, a downsampling operation is performed through a convolutional layer with a step size of 2 to reduce the size of the feature map by half and increase the number of channels to twice the original size to enhance the representation of the features [23].

The second part of the model is the self-attention module, which adopts the Transformer model encoder structure, where each encoder layer consists of a self-attention mechanism and a fully-connected feed-forward network, interspersed with downsampling convolutional layers with a step size of 2, to lessen the model parameters and computation, and to improve the computational efficiency of the model. It also helps to extract higher level abstract features and improve the model’s ability to perceive the overall features.

The third part is the classification module, which firstly performs average pooling on each channel data of the advanced features output from the second part through an average pooling layer, and further compresses the spatial dimension of the feature map. The final predicted output is obtained by mapping the features to the classification results through a fully connected layer [24]. The final prediction output is obtained by mapping the features to the classification results through the fully connected layer. The hybrid model can identify reservoir features more comprehensively and accurately, improve the precision and reliability of reservoir identification and evaluation, and provide an innovative and effective solution for reservoir identification and evaluation.

Figure 4 CNN transformer model structure

5. Results of practical application in oil and gas reservoir evaluation

5.1 Oil and gas reservoir data acquisition

The experimental data come from the Liaohe Oilfield, one of the important oil and gas production bases in China, which is rich in thick oil resources. Its complex geological conditions, including many types of reservoirs and oil and gas deposits, provide a rich data source for multivariate data analysis and oil and gas reservoir evaluation. The Liaohe Oilfield’s annual crude oil extraction capacity exceeds 10 million tons, and its annual natural gas extraction capacity reaches 800 million cubic meters.In 2024, the Liaohe Oilfield’s oil and gas production equivalent amounted to 10,157,000 tons, maintaining a steady production of 10 million tons for 39 consecutive years. 100 horizontal wells in Liaohe Oilfield. Each horizontal well contains 10 logging curves, reservoir parameter data, formation parameter data, logging interpretation and analysis program, well data, reservoir data, and conclusion data. Logging curve is to measure various physical properties of the formation or rock along the depth of the well, which can reflect important information such as porosity, permeability, water saturation and so on. Five common logging curves, resistivity (RT), density (DEN), neutron (CNL), acoustic time difference (HAC) and borehole diameter (CAL), were selected for the experiment. The data segments of logging curves 550-850m with sampling interval of 0.124m are selected respectively, and corrected using reservoir parameter data and formation parameter data. The effective thickness, oil saturation, crude oil reserves and other important indexes of each horizontal profile can be calculated.

According to the above data set, the basic statistics of the data set is shown in Table 1. And the density of dataset-(number of logging operator-reservoir parameter interactions)/(number of logging operators-number of reservoirs), so the dilution of dataset=1-dataset density.

Table 1 Basic statistics of dataset

Table 2 shows the oilfield oil and gas reservoir core kerogen data from target wells in different regions, which were analyzed and processed using three commonly used logging interpretation models, porosity analysis interpretation model, complex lithology analysis interpretation model and clay mineral analysis interpretation model. Remove surface impurities and contaminants to ensure that the core surface is clean and tidy for subsequent experimental operations. Unique numbers were assigned to each core sample, such as K5 core from Liaohe Slant 38-22 well, N8 core from Liaohe Shallow 17-9 well, etc., and detailed records were made of their corresponding well numbers, core numbers and depths in the subsurface. These numbers and record information are used throughout the experimental process, and the experimental data are traceable.

Table 2 Kerosene data of oil and gas reservoir cores in the oilfield

5.2 Experimental evaluation of oil and gas reservoirs in oil fields

Table 3 shows the comparison between the oil and gas reservoir model prediction and the actual data, with the depth prediction values ranging from 1,251.53 to 3,013.47, and the error rate of the depth prediction values ranging from 0.001% to 0.0026%, and with the Liaohe Slanting 38-22 well, for example, the error is only 0.0024%. It shows that the depth characteristics in the data can be accurately captured in the measurement. The prediction error rates of gas measurement permeability, maximum permeability and minimum permeability are also controlled at a low level, mostly within 0.5%, indicating that the multi-attention mechanism in this paper is able to capture data information more comprehensively by working in parallel with several different attention heads, which can characterize and pay attention to complex features and concerns related to porosity and permeability in the input data from different perspectives. Injury rate is an important parameter to measure the characteristics of a specific material or geological structure, which affects the storage and transmission performance of the material, for the injury rate, the model prediction error rates are all less than 0.065%-0.142%, which can accurately predict the reservoir injury. The prediction results of each parameter fit well with the actual experimental data, and the predicted values show unique distribution characteristics and change rules, showing strong performance advantages, further verifying the accuracy and reliability of the prediction data, and providing a powerful tool for the accurate assessment of oil and gas reservoirs.

Table 3 Comparison between oil and gas reservoir model prediction and actual data

5.3 Analysis of experimental accuracy

When the logging operator carries out the operation of the logging interpretation model, the porosity analysis and processing interpretation model, the complex lithology analysis and processing interpretation model, and the clay mineral analysis and interpretation model, the parameters required for the operation are recommended based on the reservoir characteristics of the two Liaohe wells, A and B, respectively, as well as the oil-bearing status of the logging data, and the oil-bearing status of the unknown well (C) in Xinjiang is estimated. With the above model default parameter settings on the logging dataset Welldatales, the hybrid intelligent model combines the advantages of CNN and Transformer to effectively extract local features as well as capture global dependencies. Comparison with LSTM method, plain Bayesian classifier, SVM similar algorithms, respectively, in the recall metrics (HR), normalized discount cumulative gain (NDG) of the data for correlation test, to analyze the model’s ability to deal with the complex nonlinear relationships, and the gap with deep learning models in feature interaction mining.

Figure 5 shows a comparison of the NDGC metrics, where the performance of the models tends to converge as the dimensionality increases. Larger hidden dimensions do not necessarily lead to better model performance. It can be seen that the proposed recommendation algorithm in this paper outperforms all other algorithms in the evaluation metric HR on the well logging domain dataset Welldatales. Long Short-Term Memory Network (LSTM), Plain Bayes, and Support Vector Machine (SVM) on the 256-segment dataset, with an average accuracy MRR of 0.65, 0.69, and 0.52, the hybrid model consistently outperforms the other baselines on the well logging dataset even if the hidden dimension is relatively small.

Figure 5 Comparison of ndgc indicators

Fig. 6 shows the comparison of HR metrics, LSTM, plain Bayes, and SVM in the 256-segment dataset, with recall HRs of 0.69, 0.67, and 0.55, and 0.81 for the hybrid model.On the task of logging reservoir parameter recommendation, this paper’s algorithm has high accuracy in the well-logging domain dataset for the problem of logging reservoir parameter prediction, especially when the self-attention mechanism is used, the Compared with similar algorithms, it has better performance performance when dealing with long sequence data. The higher accuracy of the hybrid model indicates a better prediction performance.

Figure 6 HR index comparison

5.4 Reservoir Observation Prediction Impedance

In this paper, the model is used to test the validity of the proposed method by calculating the reflectivity of the model using the impedance difference between the two layers.The oil and gas reservoir response tracts are obtained by convolving the reflectivity with the Ricker subwave at 30 Hz.100 data are selected from 15,000 reservoir characteristic response trace data. The 100 channels of data were selected as the labeled dataset, which is less than 1% of the total data. For the remaining 14,900 channels of data, 10% are randomly selected as the validation dataset, and the rest of the data are used as unlabeled data, where both labeled and unlabeled data are used for the training of the network. The impedance results of reservoir observation prediction are shown in Fig. 7, which shows that the results between the hybrid model prediction results and the real impedance are close to each other, and the absolute error does not exceed 4 impedance values, and the overall error of the predicted impedance phase is small and close to the real impedance.

Figure 7 Reservoir observation and prediction impedance results

Randomly selected 39 groups of observation data, the comparison of experimental results recorded by reservoir observation is shown in Fig. 8, where the scatter points indicate the results of each group of data, the horizontal line is the mean value, and the upper and lower lines represent the maximum and minimum values. The average Pearson correlation coefficients (PCCs) of the results predicted by the Transformer-based model are 0.99112, and the coefficient of determination (R²) is 0.98034, whereas the average Pearson correlation coefficients (PCCs) of the results predicted by the CNN-Transformer are 0.99957, and the coefficient of determination (R²) is 0.99899, and the average Pearson correlation coefficient (PCCs) of the results predicted by the CNN- Transformer outperforms the Transformer model in both metrics.

Figure 8 Comparison of experimental results of reservoir observation records

6. Conclusion

In this paper, the application of intelligent model optimization of multi-source data in the evaluation of oil and gas reservoirs is studied in depth, and the attributes, formats and accuracies of multi-source data are analyzed in detail to reveal the intrinsic connections and differences between the data. Combined with CNN to extract the characteristic parameters of the reservoir, capture the local patterns and structures in the spectral image, complete the reservoir for classification and evaluation, and the results are as follows:

(1) The error rate of the depth prediction value is between 0.001% and 0.0026%, for example, the error of Liaohe Slope 38-22 well is only 0.0024%. 15000 data containing the characteristic response trajectory of the reservoir in the region of anomalous gas-bearing layer and complex structure, the absolute error between the predicted impedance value and the real impedance value is not more than 4 impedance value. It shows that the multi-head attention mechanism is introduced into the Transformer Block to realize the feature improvement at different levels, and the different network branches of the source domain and the target domain can be constructed to accurately realize the prediction of reservoir parameters, and the stochastic gradient descent algorithm is used to optimize the weight coefficient matrix and bias vector of the model, so as to achieve the migration function and solve the problems of gradient disappearance and gradient explosion faced by long sequences.

(2) The hybrid intelligent model outperforms the algorithms such as LSTM, plain Bayesian classifier and SVM in the evaluation index HR and NDGC. The hybrid intelligent model has an evaluation index HR of 0.92 and an average accuracy MRR of 0.81. The results verify that the migration learning network can better adapt to multi-source data, improve the accuracy of reservoir parameter prediction, and promote the intelligent development in the field of oil and gas exploration and development.

References

[1] Zhang, W., Li, H., Fan, Y., Lv, M., Zhou, Y., Wang, L., … & Yang, X. (2024). Adsorption mechanism study of viscoelastic surfactants with various molecular structures on oil and gas reservoir rocks. Journal of Molecular Liquids, 415, 126408.

[2] Yang, H., Xu, Z., Zhao, Y., Zhang, B., Cui, W., Kang, W., … & Xu, P. (2024). A strong stability gel foam for water shutoff during oil and gas reservoir development. Physics of Fluids, 36(2).

[3] Zhang, C., Yin, T., Wu, S., Yin, Y., Feng, W., Li, Y., … & Hu, G. (2023). Architectural element analysis of nonmarine oil and gas reservoir in China, the research history, progress, and future trend: A review. Interpretation, 11(1), SA127-SA154.

[4] Zhang, K., Yu, H. Q., Ma, X. P., Zhang, J. D., Wang, J., Yao, C. J., … & Yao, J. (2022). Multi-source information fused generative adversarial network model and data assimilation based history matching for reservoir with complex geologies. Petroleum Science, 19(2), 707-719.

[5] Fan, D., Lai, S., Sun, H., Yang, Y., Yang, C., Fan, N., & Wang, M. (2025). Review of Machine Learning Methods for Steady State Capacity and Transient Production Forecasting in Oil and Gas Reservoir. Energies, 18(4), 842.

[6] Akinwumiju, A. A., Satterfield, D., & Phethean, J. J. (2024). Evaluation of shale oil and gas plays-Part I: Shale reservoir property modelling of the North Sea Kimmeridge Clay Formation. Marine and Petroleum Geology, 164, 106824.

[7] Fernandes, F. B., Braga, A. M., Gildin, E., & Soares, A. C. (2024). Asymptotic-Convolution Model for Compaction Damage Evaluation in Depletion-Dependent Oil Reservoirs During Alternating Drawdown/Buildup Cycles. Transport in Porous Media, 151(7), 1539-1578.

[8] Chen, C., Pan, B., Lin, X., Zhang, P., Zhang, L., Ren, S., … & Li, Y. (2025). Evaluation of Low-Resistivity Reservoirs Based on the New Three-Water Model and Comprehensive Identification of Fluid Properties Using a Committee Machine-A Case Study of the Dongying Formation Member 1 Reservoir in the Gangdong Fault Zone. Journal of Geophysics and Engineering, gxaf022.

[9] Zhao, G., Cheng, L., Jia, P., Liu, Y., Feng, H., Kuang, T., & Wang, Q. (2024). Initial Occurrence State and Movability Evaluation of the Gulong Shale Oil Reservoir, Songliao Basin. Energies, 17(6), 1358.

[10] Wang, X., Mao, M., Yang, Y., Yuan, S., Guo, M., Li, H., … & Ye, X. (2024). Prediction Technology of a Reservoir Development Model While Drilling Based on Machine Learning and Its Application. Processes, 12(5), 975.

[11] Lin, H., Liu, X., Xiong, J., Liang, L., Wu, J., & Li, B. (2024). Reservoir horizontal principle stress prediction using intelligent fusion model based on physical model constraints: a case study of Daji Block, Eastern Ordos Basin, North China. Acta Geophysica, 72(4), 2565-2579.

[12] Zhao, X., Jiang, Y., Xu, P., Yu, J., & Xie, L. (2024). Method for the Quantitative Evaluation of Low-Permeability Reservoir Damage in the East China Sea Based on Experimental Evaluation and Modeling Calculation. Processes, 12(7), 1406.

[13] Zhang, Y., Meng, F., Xu, J., Liu, B., Wu, X., Shi, K., & Wang, R. (2023). Oil and gas exploration potential in the exclusive economic zone of the Cuba’s sector of the Gulf of Mexico. Geological Journal, 58(11), 4133-4145.

[14] Benson, A. L., Clarkson, C. R., & Zeinabady, D. (2025). Analysis of enhanced geothermal system flowback and circulation test data for fracture and reservoir characterization. Geoenergy Science and Engineering, 245, 213532.

[15] Zhang, S., Zhu, C., & Guo, X. (2024). Wind-speed multi-step forecasting based on variational mode decomposition, temporal convolutional network, and transformer model. Energies, 17(9), 1996.

[16] Xu, Q., Li, Y., & Xie, M. (2024). Oil Species Identification Based on the Fluorescence Spectroscopic Analysis Using the Excitation-Emission Matrix and Transfer Learning. Water, Air, & Soil Pollution, 235(10), 642.

[17] Xin, Y., Li, W., Zhang, H., Tian, H., Fu, X., & Xie, Z. (2024). Mechanisms by Which an Evaporated Lagoon Sedimentation System Controls Source–Reservoir Preservation in Lei32 Sub-Member Unconventional Oil and Gas. Energies, 17(4), 964.

[18] Gu, D., Zheng, R., Cheng, P., Zhou, S., Yan, G., Liu, H., … & Liao, M. (2024). Single Well Production Prediction Model of Gas Reservoir Based on CNN-BILSTM-AM. Energies, 17(22), 5674.

[19] Zhao, C., Jia, Y., Qu, Y., Zheng, W., Hou, S., & Wang, B. (2024). Forecasting Gas Well Classification Based on a Two-Dimensional Convolutional Neural Network Deep Learning Model. Processes, 12(5), 878.

[20] Deng, X., Kang, X., Yang, D., Fu, W., & Luo, T. (2024). Advancing Quantitative Seismic Characterization of Physical and Anisotropic Properties in Shale Gas Reservoirs with an FCNN Framework Based on Dynamic Adaptive Rock Physics Modeling. Applied Sciences, 14(4), 1469.

[21] Hou, L., Ma, C., Tang, W., Zhou, Y., Ye, S., Chen, X., … & Hou, M. (2024). DDViT: Advancing lithology identification on FMI image logs through a dual modal transformer model with less information drop. Geoenergy Science and Engineering, 234, 212662.

[22] Gao, M., Wei, C., Zhao, X., Huang, R., Li, B., Yang, J., … & Xiong, L. (2023). Intelligent optimization of gas flooding based on multi-objective approach for efficient reservoir management. Processes, 11(7), 2226.

[23] Li, F., Guo, X., Qi, X., Feng, B., Liu, J., Xie, Y., & Gu, Y. (2025). A Surrogate Model-Based Optimization Approach for Geothermal Well-Doublet Placement Using a Regularized LSTM-CNN Model and Grey Wolf Optimizer. Sustainability, 17(1), 266.

[24] Yang, X., Zhang, C., Zhao, S., Zhou, T., Zhang, D., Shi, Z., … & Zhang, Y. (2023). CLAP: Gas Saturation Prediction in Shale Gas Reservoir Using a Cascaded Convolutional Neural Network–Long Short-Term Memory Model with Attention Mechanism. Processes, 11(9), 2645.