Weak clouded thin reservoir and cloud-ash interbed identification method, system and device

Abstract

The invention discloses a weak clouded thin reservoir and cloud-ash interbed identification method, system and equipment, and the method comprises the steps: firstly, carrying out the deep mining of geological features sensitive to the weak clouded thin reservoir and the cloud-ash interbed, and constructing a training data set through the geological features and a logging interpretation lithofacies curve; a democratic neural network association method is introduced to construct a machine learning model, and a training data set is utilized to train the machine learning model, so that the machine learning model can automatically learn and capture geologic features of a weak clouded thin reservoir and a cloud-ash interbed, the data processing efficiency and recognition precision are improved, the complex and time-consuming geologic interpretation process in a traditional means is reduced, and the geologic interpretation efficiency of the weak clouded thin reservoir and the cloud-ash interbed is improved. Therefore, exploration work can be carried out more quickly, geological differences of different regions can be better adapted, and optimization of oil and gas resource development decisions is facilitated.

Description

Technical Field

[0001] This application belongs to the field of oil and gas field development technology, specifically relating to a method, system and equipment for identifying weakly clouded thin reservoirs and cloud-gray interlayers. Background Technology

[0002] In the field of oil and gas exploration, especially for the identification of thin reservoirs and interbedded cloud and ash formations, traditional methods face a series of challenges. Thin reservoirs typically have low oil and gas saturation, making them difficult to identify accurately using traditional exploration techniques. Interbedded cloud and ash formations often exhibit stratigraphic complexity, making it difficult for traditional methods to distinguish and describe them.

[0003] Weakly dolomitized thin reservoirs and interbedded dolomitic-grayishite formations possess unique geological characteristics, which are weakly represented in seismic and electromagnetic exploration data, increasing the difficulty of accurate identification. For example, the Maokou and Qixia Formations of the Permian in central Sichuan have significant exploration and development potential, but the heterogeneity of geological structures and stratigraphic parameters has not been fully considered, resulting in significant differences in production enhancement and stimulation effects. The preliminary results obtained are insufficient to guide the comprehensive development of gas reservoirs in this block. In particular, the Maokou Formation has been eroded, with interbedded high-impedance dolomite reservoirs and relatively low-impedance gray dolomite reservoirs, making identification using conventional seismic amplitude analysis impossible. The dolomite reservoirs are mostly thin or even ultra-thin, with a thickness of less than 15m, and the impedance difference between the thin dolomite layers and the surrounding rocks is small, making identification using simple impedance inversion very difficult. For stratigraphic oil and gas exploration and development, the presence of interbedded thin dolomite and gray dolomite significantly increases the difficulty of exploration and development.

[0004] It is evident that traditional methods are ineffective in handling these special geological conditions. Therefore, there is an urgent need to develop a faster, more accurate, and more adaptable identification technology for thin reservoirs with weak cloud formation and interbedded cloud and ash layers. Summary of the Invention

[0005] To overcome the shortcomings of traditional exploration techniques in identifying thin, weakly clouded reservoirs and interbedded cloud-gray layers, this application proposes a method, system, and device for identifying thin, weakly clouded reservoirs and interbedded cloud-gray layers. Based on large-scale exploration data and incorporating a democratic neural network association method, this application can automatically learn and capture the geological characteristics of thin, weakly clouded reservoirs and interbedded cloud-gray layers, greatly improving data processing efficiency and identification accuracy.

[0006] This application is achieved through the following technical solution:

[0007] A method for identifying weakly clouded thin reservoirs and cloud-gray interlayers, the method comprising:

[0008] Based on the various seismic attributes extracted from the 3D seismic data volume, common features are sought from different types of seismic attributes to obtain multiple types of seismic attributes that are sensitive to weakly clouded thin reservoirs and interbedded cloud and ash layers; the common features refer to the characteristics that can indicate the different responses of the same type of reservoirs to the surrounding rock under different scenarios or well locations.

[0009] A training dataset is constructed by using various seismic attributes and well logging interpretation lithofacies curves as feature inputs and target outputs, respectively.

[0010] A machine learning model is constructed using the democratic neural network association method, and the constructed machine learning model is trained using the training dataset to obtain a lithofacies classification model; the constructed machine learning model is a parallel learning model that integrates multiple different types of neural networks.

[0011] The probability of reservoir facies distribution is predicted using the aforementioned lithofacies classification model, and the lithofacies with the highest probability is taken as the final predicted lithofacies.

[0012] In some implementations, the multiple seismic attributes extracted from the three-dimensional seismic data volume are used to find common features among different types of seismic attributes, thereby obtaining multiple types of seismic attributes sensitive to weakly clouded thin reservoirs and interbedded cloud and ash layers. Specifically, these include:

[0013] Multiple types of seismic attributes are extracted from the three-dimensional seismic data volume, and each type of seismic attribute includes multiple features;

[0014] To improve the effectiveness and reliability of identification, common characteristics are sought among different types of earthquake attributes.

[0015] By using the inline profile of well-through seismic attributes, seismic attributes that can clearly distinguish between the upper and lower surrounding rocks are extracted from the common features.

[0016] By using horizontal slices of the dolomite reservoir obtained through well drilling, the planar distribution of the reservoir along the horizontal direction is clearly defined, and seismic attributes that can characterize the horizontal distribution of the reservoir are extracted from the common features.

[0017] In some implementations, the construction of a training dataset using multiple seismic attributes and well-logging interpretation lithofacies curves as feature inputs and target outputs, respectively, specifically includes:

[0018] Multiple seismic attributes are used as input features, and well logging interpretation lithofacies curves are used as the target output.

[0019] Well-seismic calibration was performed, and the well-logged interpretation lithofacies curves in the depth domain were converted to time-depth to obtain the well-logged interpretation lithofacies curves in the time domain.

[0020] Upsample the time-domain well logging interpretation lithofacies curves based on the sampling rate of the seismic attribute data to match the sampling interval of the seismic attribute data, forming input-output sample pairs, and construct a training dataset using the input-output sample pairs.

[0021] In some implementations, training the constructed machine learning model using the training dataset to obtain a lithofacies classification model specifically includes:

[0022] The training dataset is used to perform initial training on multiple neural networks in the constructed machine learning model until the training error meets the preset target value, thus completing the initial training of the model.

[0023] From the earthquake attributes, other earthquake attributes that can help improve the model's predictive ability are extracted as soft data, and the soft data is predicted using a machine learning model that has already completed its initial training. The consistency and reliability of these prediction results are evaluated through a democratic voting system.

[0024] Soft data that has a higher approval rate than a threshold in the democratic voting system is selected and added to the original training dataset to expand the training dataset;

[0025] The expanded training dataset is used to retrain multiple neural networks in the machine learning model, adjusting the network parameters until the iteration requirements are met.

[0026] In some implementations, the original training dataset includes only hard data, while the expanded training dataset includes both hard and soft data.

[0027] The hard data is highly reliable data, including seismic attributes that can clearly distinguish between the upper and lower surrounding rocks and seismic attributes that can characterize the horizontal distribution of the reservoir, and has a large weight.

[0028] The soft data includes seismic attributes that are not very obvious but still contain information about the reservoir, and have a lower weight.

[0029] In some embodiments, the identification method further includes:

[0030] The trained machine learning model is validated to ensure its prediction accuracy and generalization ability. The validated model is then used as a lithofacies classification model.

[0031] Secondly, this application proposes a system for identifying weakly clouded thin reservoirs and cloud-gray interlayers, the system comprising:

[0032] The feature extraction module is configured to: based on multiple seismic attributes extracted from the three-dimensional seismic data volume, search for common features from different types of seismic attributes to obtain multiple types of seismic attributes that are sensitive to weakly clouded thin reservoirs and interbedded cloud and ash layers; the common features refer to features that can indicate the different responses of the same type of reservoirs to the surrounding rock under different scenarios or well locations;

[0033] A training set construction module is configured to construct a training dataset by using multiple seismic attributes and well-logged interpretation lithofacies curves as feature inputs and target outputs, respectively.

[0034] The model training module is configured to: construct a machine learning model using the democratic neural network association method, and train the constructed machine learning model using the training dataset to obtain a lithofacies classification model; the constructed machine learning model is a parallel learning model integrating multiple different types of neural networks.

[0035] And a prediction module, which is configured to: use the lithofacies classification model to predict the probability of reservoir lithofacies distribution, and use the lithofacies with the highest probability as the final predicted lithofacies.

[0036] In some implementations, the feature extraction module further includes:

[0037] The first extraction unit is configured to extract multiple types of seismic attributes from the three-dimensional seismic data volume, each type of seismic attribute including multiple features;

[0038] The second extraction unit is configured to: search for common features from different types of earthquake attributes to improve the effectiveness and reliability of identification;

[0039] The third extraction unit is configured to extract seismic attributes that can clearly distinguish between the upper and lower surrounding rocks from the common features by using the inline profile of the well seismic attributes.

[0040] And, the fourth extraction unit is configured to: determine the horizontal distribution of the reservoir by passing through a horizontal slice of the dolomite reservoir, and extract seismic attributes that can characterize the horizontal distribution of the reservoir from the common features.

[0041] In some implementations, the training set construction module further includes:

[0042] The data acquisition unit is configured to: take multiple seismic attributes as input features and well logging interpretation lithofacies curves as target outputs;

[0043] The time-depth conversion unit is configured to: perform well-seismic calibration, convert the well-logging interpretation lithofacies curves in the depth domain to time-depth, and obtain well-logging interpretation lithofacies curves in the time domain;

[0044] And an upsampling unit, which is configured to: upsample the time-domain well-logging interpretation lithofacies curves according to the sampling rate of the seismic attribute data to match the sampling interval of the seismic attribute data, form input-output sample pairs, and construct a training dataset using the input-output sample pairs.

[0045] In some implementations, the model training module further includes:

[0046] The initial training unit is configured to: use the training dataset to perform initial training on multiple neural networks in the constructed machine learning model until the training error meets a preset target value, thereby completing the initial training of the model;

[0047] The soft data extraction unit is configured to: extract other earthquake attributes from earthquake attributes that can help improve the model's predictive ability as soft data, use a machine learning model that has been initially trained to predict the soft data, and evaluate the consistency and reliability of these prediction results through a democratic voting system.

[0048] An expansion unit is configured to: filter out soft data that has a higher approval rate than a threshold in the democratic voting system, add it to the original training dataset, and expand the training dataset;

[0049] And a retraining unit, configured to: retrain multiple neural networks in the machine learning model using the expanded training dataset, adjusting network parameters until the iteration requirements are met.

[0050] In some implementations, the model training module further includes:

[0051] The model validation unit is configured to validate the trained machine learning model to ensure its prediction accuracy and generalization ability, and the validated model serves as a lithofacies classification model.

[0052] Thirdly, this application proposes a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described method.

[0053] This application proposes a method, system, and device for identifying thin, weakly clouded reservoirs and interbedded cloud-gray layers. First, it deeply mines geological features sensitive to these features and uses well logging interpretation lithofacies curves to construct a training dataset. Then, it introduces a democratic neural network association method to build a machine learning model. The model is trained using the training dataset, enabling it to automatically learn and capture the geological features of thin, weakly clouded reservoirs and interbedded cloud-gray layers. This improves data processing efficiency and identification accuracy, reduces the complex and time-consuming geological interpretation process of traditional methods, allows exploration work to proceed more rapidly, better adapts to geological differences in different regions, and helps optimize oil and gas resource development decisions.

[0054] This application proposes a method, system, and device for identifying weakly clouded thin reservoirs and cloud-gray interlayers. During the training process, soft data and democratic voting are introduced to expand the training dataset, and the expanded training dataset is used to retrain the machine learning model, thereby improving the model's performance. Attached Figure Description

[0055] The accompanying drawings, which are included to provide a further understanding of the embodiments of this application and form part of this application, do not constitute a limitation on the embodiments of this application. In the drawings:

[0056] Figure 1 This is a flowchart of the identification method according to an embodiment of this application;

[0057] Figure 2 This is a schematic diagram of the machine learning model architecture in an embodiment of this application;

[0058] Figure 3 This is a block diagram illustrating the principle of the identification system according to an embodiment of this application;

[0059] Figure 4 The six types of seismic attributes selected for well A are shown in the inline profile; where (a) instantaneous frequency; (b) thin layer factor; (c) RMS frequency; (d) weighted average frequency; (e) RMS amplitude; and (f) integral absolute amplitude.

[0060] Figure 5 The six types of seismic attributes selected for well B are shown in the inline profile; where (a) instantaneous frequency; (b) thin layer factor; (c) RMS frequency; (d) weighted average frequency; (e) RMS amplitude; and (f) integral absolute amplitude.

[0061] Figure 6 The six preferred seismic attributes are shown in the horizontal slices along the Maokou Formation dolomite reservoir of Well A; where (a) instantaneous frequency; (b) thin-layer factor; (c) RMS frequency; (d) weighted average frequency; (e) RMS amplitude; and (f) integral absolute amplitude.

[0062] Figure 7 The six preferred seismic attributes are shown in the horizontal slices along the Maokou Formation dolomite reservoir of Well B; where (a) instantaneous frequency; (b) thin-layer factor; (c) RMS frequency; (d) weighted average frequency; (e) RMS amplitude; and (f) integral absolute amplitude.

[0063] Figure 8 This is a schematic diagram of well seismic calibration for the synthetic seismic record of well B.

[0064] Figure 9 The confusion matrix for all well classification results in the training dataset;

[0065] Figure 10 The following are predicted lithofacies profiles of wells A, B, and C; where (a) is well A; (b) is well B; and (c) is well C.

[0066] Figure 11 The spatial distribution probability prediction results of the Maokou Formation limestone;

[0067] Figure 12 The spatial distribution probability prediction results of the Maokou Formation dolomite;

[0068] Figure 13 The spatial distribution probability prediction results of the Maokou Formation dolomitic limestone;

[0069] Figure 14 The spatial distribution probability prediction results of the Maokou Formation calcareous dolomite;

[0070] Figure 15 The spatial distribution probability prediction results for argillaceous limestone;

[0071] Figure 16 The results show the predicted spatial distribution of the three-dimensional lithofacies of the Maokou Formation. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the embodiments and accompanying drawings. The illustrative embodiments and descriptions of this application are only for explaining this application and are not intended to limit this application.

[0073] Example 1

[0074] To quickly overcome the shortcomings of traditional technologies in identifying thin, weakly clouded reservoirs and interbedded cloud-grey layers, and to improve the accuracy and efficiency of identification, this embodiment proposes a method for identifying thin, weakly clouded reservoirs and interbedded cloud-grey layers. First, seismic data is analyzed to fully extract seismic attributes that clearly distinguish the upper and lower surrounding rocks, as well as identification factors that characterize the horizontal distribution of the reservoir, as training data. Then, the Democratic Neural Networks Association (DNNA) method is used to model a machine learning system, train the neural network, and apply the trained model to identify thin, weakly clouded reservoirs and interbedded cloud-grey layers. This improves the reliability of identification, provides more reliable technical support for exploration work, helps to accurately assess the hydrocarbon-bearing performance of the reservoir, and provides a scientific basis for reservoir development.

[0075] like Figure 1 As shown, the identification method proposed in this embodiment specifically includes the following steps:

[0076] Step 100: Based on the various seismic attributes extracted from the 3D seismic data volume, common features are sought from different types of seismic attributes to obtain multiple types of seismic attributes sensitive to weakly clouded thin reservoirs and interbedded cloud and ash layers. The common features refer to characteristics that can indicate different responses of the same type of reservoir compared to the surrounding rock under different scenarios or well locations.

[0077] Step 200: Construct a training dataset using multiple seismic attributes and well-logged interpreted lithofacies curves as feature inputs and target outputs, respectively. The well-logged interpreted lithofacies curves include those for limestone, dolomite, dolomitic limestone, calcareous dolomite, and argillaceous limestone.

[0078] Step 300: A machine learning model is constructed using the democratic neural network association method, and the constructed machine learning model is trained using the training dataset to obtain a lithofacies classification model. The constructed machine learning model includes multiple neural networks of different types.

[0079] Step 400: Use the lithofacies classification model to predict the probability of reservoir lithofacies distribution, and use the lithofacies with the highest probability as the final predicted lithofacies.

[0080] In one optional implementation, step 100 is specifically implemented as follows:

[0081] Step 101: Extract various types of seismic attributes from the 3D seismic data volume, including amplitude statistics, instantaneous parameters, frequency statistics, and thin-layer properties. Each type of seismic attribute includes multiple features.

[0082] Step 102: Find common features from different types of earthquake attributes to improve the effectiveness and reliability of identification.

[0083] Step 103: Extract seismic attributes that can clearly distinguish the upper and lower surrounding rocks from common features using the inline profile of well seismic attributes.

[0084] Step 104: By slicing the dolomite reservoir horizontally through the well, the planar distribution of the reservoir along the horizontal direction is clarified, and seismic attributes that can characterize the horizontal distribution of the reservoir are extracted from common features.

[0085] It should be noted that seismic attributes refer to the geometric, kinematic, dynamic, or statistical characteristics of seismic waves derived from pre-stack and post-stack seismic data through mathematical transformations. Based on the kinematic and dynamic characteristics of waves, seismic attributes can be divided into nine major categories: amplitude, frequency, phase, energy, waveform, wave impedance, wave velocity, correlation, and ratio. Each major category contains several to over twenty types of characteristics. From the basic definition of seismic attributes, they are physical quantities characterizing the morphology, kinematic features, and statistical features of seismic waves, possessing clear physical meaning. Due to the significant heterogeneity of dolomite reservoirs and the inability to form strong impedance difference interfaces with the surrounding rocks, wave impedance inversion results are insufficient to determine the distribution characteristics of dolomite reservoirs. The comprehensive seismic attribute analysis method is less constrained by stratigraphy and geological models, fully extracts detailed information from the seismic data volume, and amplifies information for thin reservoir identification. Therefore, it is suitable for identifying thin dolomite / gray dolomite reservoirs with irregular shapes and strong lateral heterogeneity. Based on this, this embodiment extracts seismic attributes (including seismic attributes that can clearly distinguish the upper and lower surrounding rocks and seismic attributes that can characterize the horizontal distribution of the reservoir) that are sensitive to weakly clouded thin reservoirs and interbedded cloud and ash layers by analyzing well logging and seismic response data. Subsequent lithofacies identification is then performed based on these seismic attributes, which improves the accuracy and reliability of the identification.

[0086] In one optional implementation, the process of generating the training dataset in step 200 is as follows:

[0087] Step 201: Use multiple seismic attributes as input features and well logging interpretation lithofacies curves as the target output.

[0088] Step 202: Perform well-seismic calibration by converting the depth-domain well-logged interpretation lithofacies curves to time-depth to obtain the time-domain well-logged interpretation lithofacies curves.

[0089] Seismic attribute data is generally time-domain data. Therefore, in order to maintain consistency with seismic attribute data, this embodiment converts the well logging interpretation lithofacies curves in the depth domain into lithofacies interpretation curves in the time domain, i.e., performs time-depth conversion. This is done by establishing a velocity model (formation acoustic velocity model), calculating the time point corresponding to the depth point through the velocity model, and performing effective interpolation and matching, thereby converting the well logging interpretation lithofacies curves in the depth domain into the corresponding time-domain data.

[0090] Step 203: Upsample the time-domain well logging interpretation lithofacies curves according to the sampling rate of the seismic attribute data to match the sampling interval of the seismic attribute data, forming input-output sample pairs, thereby constructing a training dataset.

[0091] The training dataset consists of a large number of input-output sample pairs, which can be represented as T={(xw,C)} m In the formula}, xw is defined by a d-dimensional vector x and weights w, where x is the input feature, C is the target output, and m is the number of sample pairs. For hard data, the weights w are higher; for soft data, the weights w are lower. Hard data refers to the seismic attributes extracted in step 100 used for training. These attributes have higher weights and are considered highly reliable, including features that clearly distinguish between upper and lower surrounding rocks and features that characterize the horizontal distribution of the reservoir. Soft data refers to other seismic attributes that may help improve the model's predictive ability. This soft data may include seismic attributes that are not very obvious but still contain information about the reservoir, and can be used to expand the training dataset to improve model performance.

[0092] In one optional implementation, step 300 employs the democratic neural network association method to construct a machine learning model. The constructed machine learning model is a parallel learning model integrating multiple types of neural networks, specifically as follows: Figure 2 As shown, the constructed machine learning model includes n neural networks of different types, such as convolutional neural networks, deep neural networks, and recurrent neural networks, where n is an integer greater than or equal to 3. All n neural networks are trained using the same training dataset, and the final prediction result is obtained by fusing the prediction results of the n neural networks.

[0093] It should be noted that before training, the data in the training dataset needs to be normalized. This is a common technique in machine learning and will not be elaborated on here.

[0094] To improve the reliability and stability of the prediction results, this embodiment employs a two-stage training method to train the constructed machine learning model. First, hard data is used for initial training of n neural networks. Then, soft data is introduced to stabilize the prediction results, reduce the bootstrap error rate, and enrich the training dataset. Finally, the expanded training dataset is used to retrain the n neural networks. The specific training process is as follows:

[0095] Step 301: Initially train multiple neural networks in the constructed machine learning model using the training dataset until the training error meets the preset target value, thus completing the initial training of the model. The training dataset used in this step is the same as the training dataset generated in step 200 above, which contains only hard data.

[0096] Step 302: Extract other earthquake attributes from the earthquake attributes that may help improve the model's predictive ability, i.e., soft data, and use the machine learning model that has completed its initial training to make predictions on the soft data, and use a democratic voting system to evaluate the consistency and reliability of these prediction results.

[0097] Step 303: Soft data that has a higher approval rate than the threshold in the democratic voting system is selected, added to the original training dataset to expand the training dataset, and assigned a lower weight to the soft data to reflect its relative uncertainty.

[0098] Step 304: Retrain multiple neural networks in the machine learning model using the expanded training dataset (containing hard data and soft data), adjust the network parameters until the iteration requirements are met, so that the model can comprehensively learn information from hard data and soft data.

[0099] After training is completed, the trained model can be validated to ensure its prediction accuracy and generalization ability. The final model is used as a lithofacies classification model. This lithofacies classification model is applied to the seismic data of the entire work area to predict the probability of reservoir lithofacies distribution, and the lithofacies with the highest probability is used as the final predicted lithofacies.

[0100] Compared to traditional technologies, this embodiment uses DNNA technology to integrate the prediction results of multiple neural network models. This not only effectively reduces the bias caused by a single model and improves the overall prediction accuracy, but also allows the integration of multiple neural network models to better cope with the uncertainty and complexity of the data. Even if a certain model performs poorly in some situations, other models can still provide supplementary information, thereby improving the robustness of the entire system.

[0101] Traditional machine learning algorithms typically assume that input features are independent, which is not always true in practice. This embodiment employs DNNA technology to allow the use of non-independent or correlated data features as input, thereby capturing the potential relationships within the data more comprehensively. For example, in seismic data interpretation, there may be complex correlations between different seismic attributes, and DNNA technology can utilize this correlation information to make more accurate predictions.

[0102] Traditional machine learning algorithms typically rely solely on hard data for training, while the DNNA technology used in this embodiment can effectively combine soft data. Through the collaborative work of different networks, it comprehensively utilizes these multi-source data for prediction and optimizes model performance through supervised methods.

[0103] The identification method proposed in this embodiment first deeply mines geological features sensitive to weakly clouded thin reservoirs and interbedded cloud and ash layers. Using these geological features and well logging interpretation lithofacies curves, a training dataset is constructed. A democratic neural network association method is introduced to build a machine learning model. This model is then trained using the training dataset, enabling it to automatically learn and capture the geological features of weakly clouded thin reservoirs and interbedded cloud and ash layers. This improves data processing efficiency and identification accuracy, reduces the complex and time-consuming geological interpretation process of traditional methods, allows for faster exploration, and better adapts to geological differences in different regions, thus helping to optimize oil and gas resource development decisions. Furthermore, the identification method proposed in this embodiment introduces soft data and democratic voting during the training process to expand the training dataset. The expanded training dataset is then used to retrain the machine learning model, further improving its performance.

[0104] Based on the same technical concept described above, this embodiment also proposes a system for identifying weakly clouded thin reservoirs and cloud-gray interlayers, such as... Figure 3 As shown, the identification system proposed in this embodiment includes:

[0105] The feature extraction module is configured to: extract various seismic attributes from the 3D seismic data volume, and search for common features among different types of seismic attributes to obtain multiple types of seismic attributes sensitive to weakly clouded thin reservoirs and interbedded cloud and ash layers. The common features refer to characteristics that can indicate different responses of the same type of reservoir compared to the surrounding rock under different scenarios or well locations.

[0106] The training set construction module is configured to construct a training dataset using multiple seismic attributes and well-logging interpretation lithofacies curves as feature inputs and target outputs, respectively. The well-logging interpretation lithofacies curves include lithofacies curves for limestone, dolomite, dolomitic limestone, calcareous dolomite, and argillaceous limestone.

[0107] The model training module is configured to construct a machine learning model using the democratic neural network association method and train the constructed machine learning model using a training dataset to obtain a lithofacies classification model. The constructed machine learning model includes multiple neural networks of different types.

[0108] And a prediction module, which is configured to: predict the probability of reservoir facies distribution using a lithofacies classification model, and use the lithofacies with the highest probability as the final predicted lithofacies.

[0109] In one optional implementation, the feature extraction module further includes:

[0110] The first extraction unit is configured to extract various types of seismic attributes from the 3D seismic data volume, including amplitude statistics, instantaneous parameters, frequency statistics, and thin-layer properties. Each type of seismic attribute includes multiple features.

[0111] The second extraction unit is configured to search for common features from different types of seismic attributes to improve the effectiveness and reliability of identification.

[0112] The third extraction unit is configured to extract seismic attributes that can clearly distinguish between the upper and lower surrounding rocks from common features using inline profiles of well-pass seismic attributes.

[0113] The fourth extraction unit is configured to: clarify the horizontal distribution of the reservoir by passing through horizontal slices of the dolomite reservoir, and extract seismic attributes that can characterize the horizontal distribution of the reservoir from common features.

[0114] In one optional implementation, the training set construction module further includes:

[0115] The data acquisition unit is configured to take seismic attributes as input features and well logging interpretation lithofacies curves as target outputs.

[0116] The time-depth conversion unit is configured to perform well-seismic calibration, convert the well-logged interpretation lithofacies curves in the depth domain to time-depth, and obtain the well-logged interpretation lithofacies curves in the time domain.

[0117] In addition, there is an upsampling unit, which upsamples the time-domain well logging interpretation lithofacies curves according to the sampling rate of the seismic attribute data to match the sampling interval of the seismic attribute data, forming input-output sample pairs to construct a training dataset.

[0118] In one optional implementation, the model training module further includes:

[0119] The initial training unit is configured to perform initial training on multiple neural networks in the constructed machine learning model using the training dataset until the training error meets a preset target value, thus completing the initial training of the model. The training dataset used in this process is the training dataset generated by the training set module, which contains only hard data.

[0120] The soft data extraction unit is configured to extract other earthquake attributes from earthquake attributes that may help improve the model's predictive ability, i.e., soft data, and use a machine learning model that has completed its initial training to predict the soft data, and to evaluate the consistency and reliability of these prediction results through a democratic voting system.

[0121] An expansion unit is configured to: filter out soft data that has a higher approval rate than a threshold in the democratic voting system, add it to the original training dataset to expand the training dataset, and assign lower weights to the soft data to reflect its relative uncertainty.

[0122] In addition, a retraining unit is configured to retrain multiple neural networks in the machine learning model using an expanded training dataset (containing hard and soft data), adjusting network parameters so that the model can comprehensively learn information from both hard and soft data.

[0123] Furthermore, the model training module also includes:

[0124] The model validation unit is configured to validate the trained model to ensure its prediction accuracy and generalization ability, and the final model is used as a lithofacies classification model.

[0125] Example 2:

[0126] The target reservoir is the Maokou Formation of the Permian system in central Sichuan. Reservoir lithofacies prediction was conducted, dividing it into three sections from bottom to top: Maokou Formation 1, 2, and 3. The structure is gentle, exhibiting a monocline characteristic with a south-high, north-low elevation. East-west trending faults are present, with a thickness of 185–254 m. The strata are predominantly limestone, with dolomite approximately 5 m thick, representing an ultrathin layer, making identification extremely difficult. The Maokou Formation dolomite reservoir underwent multiple phases of alteration after limestone deposition, resulting in a general mixture of limestone and dolomite, with complex lithology and physical properties. The reservoir thickness is generally thin, with insignificant seismic response, leading to multiple interpretations in seismic prediction and making accurate reservoir identification challenging. Multiple sets of thin dolomite reservoirs are developed and superimposed on each other. Within the target well area, the physical properties of the Maokou Formation dolomite / lime-lime thin layer are not significantly different from those of the upper and lower limestone. On conventional logging curves, the dolomite / lime-bearing dolomite section generally exhibits slightly lower P-wave velocity and natural gamma, and slightly higher neutron porosity and density. Within the well area, the impedance difference between the thin dolomite layers of the Maokou Formation and the upper and lower surrounding rocks is relatively small, making it difficult to identify weakly dolomitized thin reservoirs in the study area using simple impedance inversion. Conventional seismic amplitude analysis cannot identify the relatively high impedance dolomite reservoirs and the relatively low impedance lime-bearing dolomite reservoirs within the Maokou Formation.

[0127] This embodiment uses the identification method or system proposed in the above embodiments to predict the reservoir lithofacies of the Maokou Formation of the Permian System in central Sichuan. The specific process is as follows:

[0128] Step S1:

[0129] To obtain seismic attributes sensitive to weakly clouded thin reservoirs, multiple types of seismic attributes, including amplitude statistics, instantaneous parameters, frequency statistics, and thin-layer characteristics, were extracted from the 3D seismic data volume. Common features were sought from different seismic attributes to assess the effectiveness and reliability of the identification. Seismic attributes that can clearly distinguish between the upper and lower surrounding rocks and those that can characterize the horizontal distribution of the reservoir were extracted from the common features, resulting in six types of seismic attributes: instantaneous frequency, thin-layer factor, RMS frequency, weighted average frequency, RMS amplitude, and integral absolute amplitude. Figure 4 and Figure 5 The inline profiles of the six optimized seismic attribute types for wells A and B are shown respectively. It can be seen that, compared to the conventional seismic amplitude profile, the optimized six seismic attributes exhibit a significantly differentiating response from the surrounding rock in the dolomite / gray dolomite reservoir section at the well location. On the integral absolute amplitude attribute profile, the thin dolomite reservoir shows a positive amplitude response. Figure 4 (f)), while the thin layer of argillaceous dolomite shows a negative amplitude response ( Figure 5 (f)). This is consistent with the previous analysis results on the relative acoustic impedance of these two types of reservoirs. Frequency statistics properties show higher resolution for thin reservoirs, especially the RMS frequency and weighted average frequency properties, which clearly characterize the lateral distribution of the reservoir. In addition, thin-layer indicator factors were specifically added to characterize the spatial distribution of the reservoir. Figure 4 (b) and Figure 5 (b) Clearly demonstrates the lateral distribution of the thin-layer factor corresponding to the reservoir segment.

[0130] Figure 6 and Figure 7 Time slices extracted horizontally from dolomite reservoirs along well A and argillaceous dolomite reservoirs along well B, representing six types of seismic attributes, are shown. Horizontal slices reveal the planar distribution of reservoirs along the horizontal direction. The figures demonstrate that thin-layer indicator factors, RMS frequency, weighted average frequency, and integral absolute amplitude are effective reservoir identification factors, characterizing reservoir distribution in the horizontal direction. However, it is also noteworthy that the reservoirs exhibit strong lateral heterogeneity, likely distributed more along the local geological horizons in the horizontal direction.

[0131] Step S2:

[0132] Based on the analysis of geophysical logging and seismic response presented earlier, it is known that the weakly dolomitized reservoir of the Maokou Formation in the work area is extremely thin, exhibits strong lateral heterogeneity, and its elastic parameters are not significantly different from those of the surrounding rock. At the logging scale and in conventional seismic responses, the reservoir is difficult to distinguish from the upper and lower limestone strata. Conventional impedance inversion also struggles to effectively characterize the spatial distribution of the thin reservoir due to the limitations of vertical resolution in seismic data and the weak impedance difference between the reservoir and the surrounding rock. A comprehensive analysis of multi-seismic attribute responses reveals that the six selected seismic attributes can effectively identify dolomite and calcareous dolomite reservoirs within the Maokou Formation both vertically and horizontally. Therefore, a novel approach is proposed: using multiple seismic attributes and well-logged interpreted lithofacies as feature inputs and target outputs, respectively, to establish a nonlinear mapping relationship within a neural network. Once the neural network is trained, the multi-seismic attributes of the target layer can be input into the trained network to quickly obtain the three-dimensional spatial distribution of multiple lithofacies.

[0133] The DNNA method employs an "ensemble model" composed of multiple neural network models, integrating various data sample sets to predict formation lithology far from the wellbore. While running these neural networks in parallel, it learns from multi-resolution seismic data using different strategies and correlations, thereby minimizing bias. Furthermore, in the secondary training phase, "soft data" is introduced, and its inclusion in the training set is determined through a "voting" process, preventing overlearning and stabilizing the neural network training process.

[0134] (1) Generation of training data

[0135] Well-seismic calibration is the most fundamental and crucial step in the entire neural network modeling process. Well-seismic matching unifies the input and target output in the vertical space, and accurate sample pairs are an important guarantee for the final performance of data-driven methods. Figure 8 The diagram illustrates well-seismic calibration using a composite seismic record synthesized from sonic transit time and density logging, taking well B as an example. The correlation coefficient between the composite seismic record and the well-side seismic trace is 0.49. Simultaneously with well-seismic calibration, the check shot curve is corrected to obtain a more accurate time-depth conversion relationship, ensuring the accuracy of the lithofacies interpretation curve when converted to the time domain.

[0136] Next, upsampling of the well logging lithofacies curves was performed to coarsen the target output of the dataset. Simultaneously, after multiple trials and verifications, experiments showed that using the six selected seismic attributes mentioned above (instantaneous frequency, thin-layer factor, RMS frequency, weighted average frequency, RMS amplitude, and integral absolute amplitude) for spatial distribution prediction of lithofacies in the target layer was the most effective method. Input-output sample pairs from six wells in the study area were used for training the neural network model, and the lithofacies curves before and after coarsening were compared. It can be seen that the lithofacies curves before and after upsampling are basically consistent, ensuring the effectiveness of the training process.

[0137] (2) Model training results

[0138] The hyperparameters of the neural network model and the initial weight model parameters were configured. A standard normalization method was used to normalize the input multi-class seismic attributes to a training dataset with a mean of 0 and a variance of 1. This operation normalizes input features from different numerical scales to the same scale, avoiding the influence of numerical scale differences on the relative importance of features and facilitating network convergence. Furthermore, data points from the well perimeter were extracted as supplementary data to expand the training dataset, resulting in a final training sample size of 2844. The training learning rate was gradually decreased from 0.01 to 0.001, the tolerance of adjacent neurons was set to 0.5, and the effective neuron movement ratio was 0.1. The final classification method was the Gaussian probability distribution method, using the probability of the target point being determined by the 20 nearest neighbors.

[0139] Once the training error is reduced to a certain target value, the training converges, and the model is established. Figure 9 The confusion matrix of all well classification results in the training set is shown. Classes 1 to 5 correspond to lithofacies of limestone, dolomite, dolomitic limestone, calcareous dolomite, and argillaceous limestone, respectively, with an overall training accuracy of 0.85. The neural network model correctly predicted the vast majority of lithofacies sampling points, and achieved 100% prediction accuracy on the lithofacies of focus (dolomite and calcareous dolomite). This fully demonstrates the effectiveness and reliability of the method in establishing a nonlinear mapping between multiple seismic attributes and well logging lithofacies curves.

[0140] (3) Well-passing lithofacies prediction profile

[0141] To further demonstrate the prediction results of the neural network model at the well location, a cross-sectional view of the lithofacies prediction results was plotted, as shown below. Figure 10 As shown, the upper and lower boundaries of the lithofacies logging curves mapped on the profile extend throughout the entire Maokou Formation. It is evident that the predicted lithofacies match the logging lithofacies well at the well location. The cross-section reveals that the overall lithology of the Maokou Formation is dominated by limestone and dolomitic limestone, interbedded with dolomite, calcareous dolomite, and argillaceous limestone, and vertically superimposed. Multiple sets of argillaceous limestone strata develop in the overlying and underlying strata, interbedded with the limestone strata, with the argillaceous content gradually decreasing from top to bottom, with the Maokou Formation having the lowest argillaceous content. Within the Maokou Formation, two distinct sets of calcareous dolomite are visible, distributed nearly horizontally and exhibiting good lateral distribution. Dolomite is less common within the Maokou Formation, with only a few scattered thin layers and limited lateral distribution.

[0142] Step S3:

[0143] Based on the established neural network lithofacies classification model, we can apply the model to the seismic attribute data of the Maokou Formation throughout the entire work area to predict the probability of lithofacies distribution, and use the lithofacies with the highest probability as the final predicted lithofacies. Based on the neural network prediction results, we output the three-dimensional spatial distribution probabilities of five lithofacies: limestone, dolomite, dolomitic limestone, calcareous dolomite, and argillaceous limestone, respectively. (See...) Figures 11 to 15 As shown.

[0144] from Figures 11 to 15 The underlying mechanism by which the model makes classification predictions is understandable; the lithofacies with the highest probability are ultimately assigned to the sampling points. The area enclosed by the red and green three-dimensional interfaces in the figure represents the Maokou Formation, the focus of this study. It can be seen that limestone and dolomitic limestone are the main lithofacies of the Maokou Formation. Dolomite is sparsely distributed, with only a small, non-uniform distribution in the upper part of the Maokou 2 Member and the lower part of the Maokou 3 Member. Limestone dolomite is more abundant than dolomite within the Maokou Formation, exhibiting better lateral distribution and weaker heterogeneity. Within the Maokou Formation, limestone dolomite is mainly composed of multiple thin layers, interspersed with limestone. Meanwhile, argillaceous limestone is mainly distributed in the overlying and underlying layers, occurring in thick layers.

[0145] Figure 16 This paper presents the three-dimensional lithofacies spatial distribution of the entire Maokou Formation as predicted by a neural network model. It can be seen that the Maokou Formation roughly consists of two sets of laterally continuous dolomite reservoirs, alternating with limestone and dolomitic limestone. The dolomite reservoirs within the target interval are extremely thin and highly heterogeneous.

[0146] Therefore, the identification method or system proposed in this embodiment can better address the difficulties in identifying weakly clouded thin reservoirs and interbedded cloud and gray layers, improving the reliability of reservoir identification methods, providing more reliable technical support for exploration work, helping to accurately assess the oil and gas performance of reservoirs, and providing a scientific basis for the formulation of development plans. Through more intelligent and automated technical means, the input of human resources is reduced, and the economic benefits of oil and gas exploration are improved.

[0147] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0148] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0149] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0150] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0151] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of this application. It should be understood that the above description is only a specific embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for identifying weakly clouded thin reservoirs and cloud-gray interlayers, characterized in that, The identification method includes: Based on the various seismic attributes extracted from the 3D seismic data volume, common features are sought from different types of seismic attributes to obtain multiple types of seismic attributes that are sensitive to weakly clouded thin reservoirs and interbedded cloud and ash layers; the common features refer to the characteristics that can indicate the different responses of the same type of reservoirs to the surrounding rock under different scenarios or well locations. A training dataset is constructed by using various seismic attributes and well logging interpretation lithofacies curves as feature inputs and target outputs, respectively. A machine learning model is constructed using the democratic neural network association method, and the constructed machine learning model is trained using the training dataset to obtain a lithofacies classification model; the constructed machine learning model is a parallel learning model that integrates multiple different types of neural networks. The probability of reservoir facies distribution is predicted using the aforementioned lithofacies classification model, and the lithofacies with the highest probability is taken as the final predicted lithofacies.

2. The method for identifying weakly clouded thin reservoirs and cloud-gray interlayers according to claim 1, characterized in that, The aforementioned method, based on multiple seismic attributes extracted from 3D seismic data volumes, seeks common characteristics from different types of seismic attributes to obtain multiple seismic attributes sensitive to weakly clouded thin reservoirs and interbedded cloud and ash layers, specifically including: Multiple types of seismic attributes are extracted from the three-dimensional seismic data volume, and each type of seismic attribute includes multiple features; To improve the effectiveness and reliability of identification, common characteristics are sought among different types of earthquake attributes. By using the inline profile of well-through seismic attributes, seismic attributes that can clearly distinguish between the upper and lower surrounding rocks are extracted from the common features. By using horizontal slices of the dolomite reservoir obtained through well drilling, the planar distribution of the reservoir along the horizontal direction is clearly defined, and seismic attributes that can characterize the horizontal distribution of the reservoir are extracted from the common features.

3. The method for identifying weakly clouded thin reservoirs and cloud-gray interlayers according to claim 1, characterized in that, The construction of a training dataset, which uses multiple seismic attributes and well-logging interpretation lithofacies curves as feature inputs and target outputs respectively, specifically includes: Multiple seismic attributes are used as input features, and well logging interpretation lithofacies curves are used as the target output. Well-seismic calibration was performed, and the well-logged interpretation lithofacies curves in the depth domain were converted to time-depth to obtain the well-logged interpretation lithofacies curves in the time domain. Upsample the time-domain well logging interpretation lithofacies curves based on the sampling rate of the seismic attribute data to match the sampling interval of the seismic attribute data, forming input-output sample pairs, and construct a training dataset using the input-output sample pairs.

4. The method for identifying weakly clouded thin reservoirs and cloud-gray interlayers according to claim 1, characterized in that, The process of training the constructed machine learning model using the training dataset to obtain a lithofacies classification model specifically includes: The training dataset is used to perform initial training on multiple neural networks in the constructed machine learning model until the training error meets the preset target value, thus completing the initial training of the model. From the earthquake attributes, other earthquake attributes that can help improve the model's predictive ability are extracted as soft data, and the soft data is predicted using a machine learning model that has already completed its initial training. The consistency and reliability of these prediction results are evaluated through a democratic voting system. Soft data that has a higher approval rate than a threshold in the democratic voting system is selected and added to the original training dataset to expand the training dataset; The expanded training dataset is used to retrain multiple neural networks in the machine learning model, adjusting the network parameters until the iteration requirements are met.

5. The method for identifying weakly clouded thin reservoirs and cloud-gray interlayers according to claim 4, characterized in that, The original training dataset only includes hard data, while the expanded training dataset includes both hard and soft data. The hard data is highly reliable data, including seismic attributes that can clearly distinguish between the upper and lower surrounding rocks and seismic attributes that can characterize the horizontal distribution of the reservoir, and has a large weight. The soft data includes seismic attributes that are not very obvious but still contain information about the reservoir, and have a lower weight.

6. The method for identifying weakly clouded thin reservoirs and cloud-gray interlayers according to any one of claims 1-5, characterized in that, The identification method further includes: The trained machine learning model is validated to ensure its prediction accuracy and generalization ability. The validated model is then used as a lithofacies classification model.

7. A system for identifying weakly clouded thin reservoirs and cloud-gray interlayers, characterized in that, The identification system includes: The feature extraction module is configured to: based on multiple seismic attributes extracted from the three-dimensional seismic data volume, search for common features from different types of seismic attributes to obtain multiple types of seismic attributes that are sensitive to weakly clouded thin reservoirs and interbedded cloud and ash layers; the common features refer to features that can indicate the different responses of the same type of reservoirs to the surrounding rock under different scenarios or well locations; A training set construction module is configured to construct a training dataset by using multiple seismic attributes and well-logged interpretation lithofacies curves as feature inputs and target outputs, respectively. The model training module is configured to: construct a machine learning model using the democratic neural network association method, and train the constructed machine learning model using the training dataset to obtain a lithofacies classification model; the constructed machine learning model is a parallel learning model integrating multiple different types of neural networks. And a prediction module, which is configured to: use the lithofacies classification model to predict the probability of reservoir lithofacies distribution, and use the lithofacies with the highest probability as the final predicted lithofacies.

8. The identification system for weakly clouded thin reservoirs and cloud-gray interlayers according to claim 7, characterized in that, The feature extraction module further includes: The first extraction unit is configured to extract multiple types of seismic attributes from the three-dimensional seismic data volume, each type of seismic attribute including multiple features; The second extraction unit is configured to: search for common features from different types of earthquake attributes to improve the effectiveness and reliability of identification; The third extraction unit is configured to extract seismic attributes that can clearly distinguish between the upper and lower surrounding rocks from the common features by using the inline profile of the well seismic attributes. And, the fourth extraction unit is configured to: determine the horizontal distribution of the reservoir by passing through a horizontal slice of the dolomite reservoir, and extract seismic attributes that can characterize the horizontal distribution of the reservoir from the common features.

9. The identification system for weakly clouded thin reservoirs and cloud-gray interlayers according to claim 7, characterized in that, The training set construction module also includes: The data acquisition unit is configured to: take multiple seismic attributes as input features and well logging interpretation lithofacies curves as target outputs; The time-depth conversion unit is configured to: perform well-seismic calibration, convert the well-logging interpretation lithofacies curves in the depth domain to time-depth, and obtain well-logging interpretation lithofacies curves in the time domain; And an upsampling unit, which is configured to: upsample the time-domain well-logging interpretation lithofacies curves according to the sampling rate of the seismic attribute data to match the sampling interval of the seismic attribute data, form input-output sample pairs, and construct a training dataset using the input-output sample pairs.

10. The identification system for weakly clouded thin reservoirs and cloud-gray interlayers according to claim 7, characterized in that, The model training module also includes: The initial training unit is configured to: use the training dataset to perform initial training on multiple neural networks in the constructed machine learning model until the training error meets a preset target value, thereby completing the initial training of the model; The soft data extraction unit is configured to: extract other earthquake attributes from earthquake attributes that can help improve the model's predictive ability as soft data, use a machine learning model that has been initially trained to predict the soft data, and evaluate the consistency and reliability of these prediction results through a democratic voting system. An expansion unit is configured to: filter out soft data that has a higher approval rate than a threshold in the democratic voting system, add it to the original training dataset, and expand the training dataset; And a retraining unit, configured to: retrain multiple neural networks in the machine learning model using the expanded training dataset, adjusting network parameters until the iteration requirements are met.

11. A system for identifying weakly clouded thin reservoirs and cloud-gray interlayers according to any one of claims 7-10, characterized in that, The model training module also includes: The model validation unit is configured to validate the trained machine learning model to ensure its prediction accuracy and generalization ability, and the validated model serves as a lithofacies classification model.

12. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-6.

Images