A method for describing gas drive front based on multi-modal fusion deep learning model

Abstract

The present application relates to the technical field of oil and gas field development, and particularly relates to a gas drive front edge description method based on a multi-modal fusion deep learning model, which comprises the following steps: collecting multi-modal data, constructing a multi-modal comprehensive database; performing data preprocessing and cleaning on the multi-modal comprehensive database; constructing a multi-modal fusion deep learning model; training and hyperparameter optimizing the multi-modal fusion deep learning model; inputting real-time data into the trained model to output a front edge prediction result; adding new data meeting the requirements to the multi-modal comprehensive database according to the output prediction result, triggering model incremental training, and updating parameters to adapt to reservoir dynamic changes; and through integration of static geological data, dynamic production data, real-time monitoring data and historical front edge data, rapid, accurate and full-domain dynamic prediction of the gas drive front edge is realized.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field development technology, specifically to a gas drive front description method based on a multimodal fusion deep learning model. Background Technology

[0002] In the field of oil and gas field development, gas drive technology is a core means of enhancing oil recovery in low-permeability and tight reservoirs. Its application effect is closely related to the migration patterns of the gas drive front. Accurately capturing the location, velocity, and dynamic changes of the gas drive front is crucial for optimizing injection and production parameters, preventing gas channeling, and improving oil recovery, directly impacting the large-scale promotion and maximization of benefits of gas drive technology. However, due to the heterogeneity of reservoir geological conditions, the complexity of fluid seepage, and the dynamic nature of the displacement process, traditional gas drive front description methods have consistently struggled to achieve efficient and accurate predictions, becoming a significant bottleneck in the industry's development.

[0003] Currently, the mainstream gas-driven front characterization technologies are mainly divided into three categories, all of which have inherent defects that are difficult to overcome. The first category is physical detection methods, such as the synthetic interferometric electromagnetic wave detection method proposed in patent CN118669115A and the ultrasonic detection method disclosed in CN107515246B. Although these methods have high accuracy at the laboratory core scale, in actual reservoir applications, they can only achieve local measurements at the "point" and "line" level, and cannot complete the full-area "surface" and "volume" characterization of the entire reservoir. Moreover, the equipment is expensive, the deployment is complex, the data acquisition and inversion interpretation cycle is long, the real-time performance is extremely poor, and it is difficult to support rapid on-site decision-making.

[0004] The second category is numerical simulation. Related technologies, such as the schemes disclosed in patents CN111305801B and CN110863806A, solve the seepage equation and characterize the leading edge through methods such as finite difference and finite element. The mechanism is rigorous, but the computational efficiency is extremely low. A single simulation takes several hours to several days, which cannot meet the needs of rapid scheme comparison in production sites. At the same time, its accuracy is highly dependent on the accuracy of geological models and physical parameters, which are often difficult to obtain accurately, resulting in high uncertainty and multiple solutions in the simulation results under complex reservoir conditions.

[0005] The third category is the manual experience method, such as the calculation method based on empirical formulas or simplified models mentioned in patent CN110863806A. Although the calculation is simple and fast, it relies heavily on the personal experience of engineers, lacks a unified quantitative standard, has poor reproducibility and generalizability, and is oversimplified, failing to consider key factors such as complex geological structures, faults and heterogeneity. The prediction results are rough and have large errors, and hysteresis analysis can only be performed after the production well has produced a significant response.

[0006] The inherent defects of the aforementioned traditional technologies severely restrict the refined management of gas drive technology. Therefore, there is an urgent need for a technical solution that can break through the limitations of existing technologies and achieve rapid, accurate, and full-domain description of the gas drive front, so as to provide support for the efficient implementation of gas drive oil projects. Summary of the Invention

[0007] To address the problems existing in the prior art, this invention provides a gas drive front description method based on a multimodal fusion deep learning model, which solves the problem that traditional methods are difficult to accurately obtain many key parameters and have significant uncertainties in actual reservoir development.

[0008] To address the above problems, the technical solution of this invention is as follows: A method for describing the gas-driven leading edge based on a multimodal fusion deep learning model, comprising the following steps:

[0009] S1. Collect multimodal data and construct a multimodal integrated database. The multimodal data includes static geological data, dynamic production data, real-time monitoring data, and historical frontier data of the target reservoir.

[0010] S2. Perform data preprocessing and cleaning on the multimodal integrated database;

[0011] S3. Construct a multimodal fusion deep learning model based on Transformer-Ctrl-LSTM-FCN;

[0012] S4. Train and optimize the hyperparameters of the multimodal fusion deep learning model;

[0013] S5. Input the real-time data into the trained model and output the leading edge prediction result;

[0014] S6. Based on the output prediction results, add the new data that meets the requirements to the multimodal integrated database, trigger incremental training of the model, and update the parameters to adapt to the dynamic changes of the reservoir.

[0015] Furthermore, S3 includes:

[0016] S31. Construct a Transformer encoder to process static geological data, and add a local attention mechanism to enhance the extraction of geological features;

[0017] S32. An improved Ctrl-LSTM network is used to process dynamic production data and extract time-series features;

[0018] S33. For one-dimensional scalar data in real-time monitoring data and historical leading edge data, a fully connected neural network (FCN) feature extraction module is constructed. The deep network structure is used to maximize feature extraction of scalar data, providing high-quality feature input for multimodal fusion.

[0019] S34, Cross-modal fusion layer: Based on a dynamic attention weight allocation mechanism, it fuses features from the three modalities of S31-S33;

[0020] S35. A leading edge prediction output layer is constructed using a fully connected layer and a Softmax activation function to output the air-driven leading edge image.

[0021] Furthermore, in S32, the improved Ctrl-LSTM network includes:

[0022] S321. Construct a time memory transfer path;

[0023] S322. Construct a spatial memory transfer path;

[0024] S323. Based on the attention mechanism, the outputs of the temporal memory transfer path and the spatial memory transfer path are deeply integrated to generate comprehensive temporal features.

[0025] Furthermore, in S323, the expression for deep fusion is:

[0026] ,

[0027] in, For dynamic feature vectors, , , , is the network weight parameter, is the input data at the current time step, is the control signal, is the Sigmoid activation function, is the hidden state at the previous time step, is the cell state at the current time step, and is the spatial memory state at the current time step.

[0028] Furthermore, in S34, the fusion formula for the attention weight dynamic allocation mechanism is:

[0029] ,

[0030] in, This is a global multimodal feature vector. For static feature vectors, For dynamic feature vectors, For scalar eigenvectors, This represents the attention weight.

[0031] Furthermore, S4 includes:

[0032] S41. Divide the preprocessed multimodal integrated database into a training set and a test set according to the proportions.

[0033] S42. Use a hybrid loss function for model training;

[0034] S43. Combine the Adam optimizer, early stopping strategy and Bayesian optimization to adjust hyperparameters;

[0035] S44. Calibrate the model using on-site measured data.

[0036] Furthermore, in S42, the expression for the hybrid loss function is:

[0037] ,

[0038] in, , , These are the weighting coefficients. For mean square error loss, For structural similarity loss, Marginal loss.

[0039] Furthermore, in S42,

[0040] The mean square error loss ,in, is the number of samples, is the original pixel value matrix of the actual gas saturation field at the gas-driven leading edge, and is the original pixel value matrix of the gas saturation field at the gas-driven leading edge predicted by the model.

[0041] The structural similarity loss ,in, This represents a real image at the level of structural features. To predict the representation of an image at the level of structural features;

[0042] The edge loss ,in, This is a gradient magnitude map of a real image after processing it using the Sobel operator. This is a map showing the gradient magnitude of the predicted image after processing with the Sobel operator.

[0043] Furthermore, S2 includes the following processing of the multimodal integrated database:

[0044] Outliers are removed, and missing values ​​are filled using interpolation or generative models.

[0045] Normalization: Mapping data of different dimensions to the [0,1] interval to eliminate the influence of dimension differences. The expression is: ,in, For normalized values, These are the initial data values. This is the minimum value of the parameter. This is the maximum value of the parameter;

[0046] Data fusion: Dynamic production data is mapped to well location grids through Gaussian transformation to construct a spatiotemporal fusion sample of "well location + injection and production" parameters.

[0047] Furthermore, in S31, the local attention mechanism uses the following formula to calculate the attention weights:

[0048] ;

[0049] in, The query vector representing the location of the i-th geological grid is obtained by linear projection from the input static geological data; and Let these represent the key vector and value vector of the j-th geological grid location, respectively. represents the scaling factor for the dimension of the key vector; softmax is the normalization exponential function.

[0050] Compared with existing technologies, this invention has the following advantages: it uses Gaussian transform to map one-dimensional time-series injection and production parameters to a three-dimensional well location grid, constructing a spatiotemporal fusion sample of "well location + injection and production", which allows time-series data to be accurately aligned with spatial location; based on the attention mechanism fusion formula, the model adaptively learns the contribution weights of each mode, realizing deep coupling of static spatial features, dynamic time-series features and scalar potential features. This multi-modal fusion strategy solves the industry pain point of incomplete information in a single mode, enabling the frontal description to fully reflect the complex seepage law and gas-driven evolution characteristics of the reservoir.

[0051] For static geological data, a Transformer encoder with a local attention mechanism is used. This encoder can capture the global spatial distribution of porosity and permeability, and accurately extract key local features such as fractures, high-permeability strips, and faults, thus solving the problem that traditional models are insufficient in characterizing geological heterogeneity.

[0052] For dynamic production data, the improved Ctrl-LSTM network, through unique time memory transfer paths, spatial memory transfer paths, and spatiotemporal fusion output formulas, can quickly adjust the forget gate according to control signals, efficiently capture the temporal change trend of injection and sampling parameters, adapt to jump values ​​in production dynamics, and solve the problem of lag in response to dynamic mutation features of traditional LSTM.

[0053] The FCN network mines latent features from scalar data and outputs global features after attention fusion. It does not rely on the derivation of complex physical equations, reducing the computation time of traditional numerical simulations from hours to days to seconds or minutes. At the same time, the hybrid loss function optimizes numerical accuracy, image structure similarity, and leading edge accuracy respectively. The edge loss extracts the gradient map and calculates MSE through the Sobel operator, forcing the model to focus on the accuracy of the steep gradient region at the leading edge, keeping the prediction error within 5%, and achieving the dual goals of "high accuracy" and "high efficiency".

[0054] Employing a data-driven deep learning model, the model is trained using massive amounts of multimodal historical data to automatically learn the nonlinear laws governing the evolution of the gas drive front, completely eliminating subjective interference from human experience. The model exhibits strong robustness to geological heterogeneity in complex reservoirs (such as fractures, faults, and high-permeability zones), and can achieve quantitative output without relying on precise and scarce physical property parameters, using only conventionally available geological, production, and monitoring data. Attached Figure Description

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 This is the Transformer-Ctrl-LSTM-FCN model architecture of the present invention;

[0057] Figure 3 This is the five-point well pattern of the present invention;

[0058] Figure 4 The present invention uses SGS to randomly generate a heterogeneous permeability field;

[0059] Figure 5 The Latin square design of dynamic injection and extraction parameters for this invention;

[0060] Figure 6 This is a spatiotemporal fusion alignment sample for the "well location + injection and production" of this invention;

[0061] Figure 7 This invention relates to the Transformer-Ctrl-LSTM-FCN multimodal deep learning fusion model;

[0062] Figure 8 This is the training loss for the multimodal model in this invention;

[0063] Figure 9 To improve the training accuracy of the multimodal model in this invention;

[0064] Figure 10 This paper compares the prediction error of the CO2 displacement leading edge using the Transformer-CTRL-LSTM-FCN multimodal deep learning fusion model of this invention. Detailed Implementation

[0065] This application presents a gas-driven front description method based on a multimodal fusion deep learning model. By integrating static geological data, dynamic production data, real-time monitoring data, and historical front data, a Transformer-Ctrl-LSTM-FCN hybrid network architecture is constructed, enabling fast, accurate, and dynamic prediction of the gas-driven front across the entire domain. The steps are as follows:

[0066] S1. Collect multimodal data and construct a multimodal integrated database. The multimodal data includes static geological data, dynamic production data, real-time monitoring data, and historical frontier data of the target reservoir.

[0067] S2. Perform data preprocessing and cleaning on the multimodal integrated database;

[0068] S3. Construct a multimodal fusion deep learning model based on Transformer-Ctrl-LSTM-FCN;

[0069] S4. Train and optimize the hyperparameters of the multimodal fusion deep learning model;

[0070] S5. Input the real-time data into the trained model and output the leading edge prediction result;

[0071] S6. Based on the output prediction results, add the new data that meets the requirements to the multimodal integrated database, trigger incremental training of the model, and update the parameters to adapt to the dynamic changes of the reservoir.

[0072] Specifically, it includes the following steps:

[0073] S1. Multimodal Data Collection and Preprocessing: Collect multi-dimensional modal data of the target reservoir and construct a comprehensive database, including:

[0074] ① Static geological data: porosity, permeability, saturation, formation thickness, lithological distribution, fracture development characteristics, etc.;

[0075] ② Dynamic production data: gas injection rate, injection and production pressure, oil / gas / water production, gas-oil ratio, cumulative gas injection volume, etc.;

[0076] ③ Real-time monitoring data: gas-oil ratio, CO2 component mole fraction, gas phase saturation, etc.;

[0077] ④ Historical leading edge data: existing labeled data such as leading edge position, moving velocity, and mixing state obtained from physical monitoring, numerical simulation or field testing.

[0078] S2. Data preprocessing and cleaning are performed. Through normalization and spatiotemporal fusion, data from different sources and with different structures are unified into a format that the model can process, improving data quality. Cleaning and imputation operations reduce the interference of noise and erroneous data on the model, and spatiotemporal fusion samples ensure that the model can simultaneously learn the spatial characteristics of the reservoir and the temporal evolution of production dynamics, which is a prerequisite for achieving accurate prediction.

[0079] ① Use reservoir engineering knowledge or statistical methods to identify and remove abnormal data caused by equipment failure, etc. For missing values, use interpolation or generative models to fill them in reasonably to ensure data continuity;

[0080] ② Normalization: Data with different dimensions (such as pressure in MPa, yield in m³ / d) are mapped to the [0,1] interval, eliminating dimensional differences, accelerating model convergence, and preventing certain features from dominating model training due to excessively large values. The expression is as follows: ,in, For normalized values, These are the initial data values. This is the minimum value of the parameter. This is the maximum value of the parameter;

[0081] ③ Data Fusion: Well network data and dynamic production data are aligned according to time series and spatial location. To enhance the local feature information of single-well data, spatiotemporal fusion samples of "well location + injection and production" parameters are generated. One-dimensional time-series injection and production parameters are mapped to the well location grid through Gaussian transformation. The fused sample set represents the injection and production parameters of each well at different time steps. The darker the color at the well location, the greater the injection and production volume, thus accurately linking time-series data with spatial location. This processing specifically uses a two-dimensional Gaussian kernel convolution algorithm. The specific parameters and calculation process are as follows:

[0082] Constructing the Gaussian kernel function: A two-dimensional Gaussian convolution kernel G of size k×k is constructed, where the kernel center coordinates are (0, 0). The weight value at any point (u, v) in the kernel is calculated using the following formula:

[0083] ,

[0084] The specific parameters are defined as follows: (Standard deviation): controls the width (smoothing level) of the Gaussian function; in this embodiment, the preferred value range is [0.5, 1.5], to preserve the sampling features while smoothing high-frequency noise; k (kernel size): the size of the Gaussian kernel, usually an odd number; in this embodiment, k=3 or k=5; (u, v): represents the distance coordinates of a point within the convolution kernel relative to the kernel center, with a value range of... .

[0085] S3. Construct a multimodal fusion deep learning model architecture (Transformer-Ctrl-LSTM-FCN), designing a three-level network architecture based on "feature extraction + modality fusion + temporal modeling," such as... Figure 2 As shown, the details are as follows:

[0086] S31. For static geological data (structured data): Construct a Transformer encoder to extract three-dimensional global geological features, capture the spatial distribution patterns of parameters such as porosity and permeability, and add a local attention mechanism to enhance the extraction of geological features and improve the model's ability to characterize geological heterogeneity, especially local fractures, high-permeability strips, faults, etc.

[0087] S32. For dynamic production data (time series data): Improve the Ctrl-LSTM network to extract time series features and capture the changing trends of parameters such as injection and production parameters and output over time. Ctrl-LSTM can quickly adjust the forget gate according to the control signal to speed up the removal of old information and adapt to the jump values ​​of real production dynamics. The changes in production dynamics are driven by operating condition decisions (control). After improving the control gate, by making the gate control visible to the operation signal, a direct correlation is established between the gas injection strategy and the output response, which can more accurately grasp the characteristics of drastic changes in production dynamics.

[0088] The specific steps are as follows:

[0089] S321, Ctrl-LSTM time memory transfer path:

[0090] ,

[0091] ,

[0092] ,

[0093] ,

[0094] in, Candidate cell states for temporal memory pathways. This serves as the input gate for the time-memory path. Forgetting gates for the time memory path, This represents the current state of the cell. , , This is the trainable weight matrix in the time memory path. , , In the time-memory path, the trainable weight matrix connecting the hidden states of the previous time step is... Input data for the current time step. For control signals, It is the Sigmoid activation function. This is the hidden state from the previous time step.

[0095] S322, Ctrl-LSTM Spatial Memory Transfer Path:

[0096] ,

[0097] ,

[0098] ,

[0099] ,

[0100] ,

[0101] in, These are candidate spatial states for spatial memory paths. It is a control gate for spatial memory paths. This serves as the input gate for the spatial memory path. Forgetting gates for spatial memory paths The spatial memory state at the current time step. , , , Trainable weight matrices in spatial memory paths In the spatial memory path, a control gate is innovatively introduced to connect the trainable weight matrix of the spatial memory M_{t-1} from the previous time step. and spatial memory state The control gate explicitly receives injection / import operation signals (control signals). (For example, instructions to change the gas injection rate) enable the model to respond quickly to sudden changes in production strategies and adapt to rapid changes in real production.

[0102] The spatiotemporal information fusion output of S323 and Ctrl-LSTM is calculated as follows:

[0103] ,

[0104] in, For dynamic feature vectors, , , , These are the network weight parameters. This represents the cell state at the current time step.

[0105] S33. For one-dimensional scalar data: Construct a fully connected neural network (FCN) for extraction. For zero-dimensional scalar data such as reservoir temperature and crude oil density, extract the features of the scalar data to the greatest extent.

[0106] S34, Cross-modal fusion layer: Based on an attention mechanism, this layer fuses features from the three modalities mentioned above. The attention mechanism enables the model to intelligently weigh the importance of different data sources, addressing the pain point of incomplete information from a single modality and enhancing the model's robustness and generalization ability. By calculating the contribution weights of different modal features to the leading-edge description, the proportion of each modal feature is dynamically adjusted to solve the problem of incomplete information from a single modality, thus integrating the static feature vectors... Dynamic feature vectors , and scalar eigenvectors The features are fused to form a global multimodal feature vector Hfusion. The fusion formula is as follows:

[0107] ,

[0108] in, These are attention weights, obtained through adaptive learning by the network, and .

[0109] S35, Leading Edge Prediction Output Layer: Employs a fully connected layer and the Softmax activation function to output the air-driven leading edge image.

[0110] S4. Train and optimize the hyperparameters of the multimodal fusion deep learning model. The specific steps are as follows:

[0111] S41. Divide the preprocessed multimodal integrated database into a training set and a test set in an 8:2 ratio;

[0112] S42. Use a hybrid loss function for model training, the expression of which is:

[0113] ,

[0114] in, These are the weighting coefficients.

[0115] Mean squared error loss ensures the numerical accuracy of the overall pixel values ​​in the predicted image.

[0116] ;

[0117] Structural similarity loss measures the similarity between the predicted image and the real image in terms of structure, brightness, and contrast, improving the visual quality of the image and preventing blurry results. ,

[0118] Edge loss is crucial for ensuring the accuracy of leading edge prediction. First, the Sobel operator is used to extract the edges (gradient maps) of the real and predicted leading edge images. Then, the MSE between these two gradient maps is calculated. This forces the model to accurately reconstruct the sharp boundaries of the leading edge; otherwise, it will be significantly penalized. The expression is:

[0119] .

[0120] S43. The Adam optimizer with an adaptive learning rate is used for parameter updates, combined with an early stopping strategy. Training is stopped when the validation set accuracy no longer improves to prevent overfitting. Automated methods such as Bayesian optimization are used to search for network hyperparameters (such as the number of Transformer layers and the number of LSTM hidden units) to obtain optimal model performance.

[0121] S44. Fine-tune the trained model using a small amount of field measurement data to make the prediction results more consistent with the actual reservoir conditions.

[0122] S5. Dynamic Prediction and Real-time Adjustment of the Leading Edge: Real-time Input Data: Input the current geological parameters, production dynamics data, and monitoring signals of the reservoir into the trained model; Leading Edge Prediction Output: The model outputs the position and shape of the leading edge in real time.

[0123] S6. Dynamic adjustment mechanism: If there is a deviation between the prediction results and the field monitoring feedback (error exceeding 5%), new data will be automatically added to the sample set, triggering incremental training of the model and updating parameters to adapt to dynamic changes in the reservoir (such as changes in formation pressure, the generation of new fractures, etc.).

[0124] Taking CO2 flooding in a certain oil field as an example, the method of this application is illustrated as follows:

[0125] S1: Collect and organize basic parameters of geological and development data for a specific oilfield's G block: porosity, permeability, saturation, reservoir thickness, pressure, viscosity, relative permeability curves, reservoir well locations, and historical operating procedures of injection and production wells; construct a CO2 flooding conceptual model based on the actual reservoir information, using a basic five-point well network, such as... Figure 3 As shown, the study involves sample data from different modalities, such as... Figure 2 As shown, the data specifically includes 0-dimensional numerical data: crude oil density, initial formation pressure, initial oil saturation, reservoir temperature, fracture length, and fracture angle; 1-dimensional time-series data mainly consists of dynamically changing injection and production parameters; and 2-dimensional / 3-dimensional spatial data includes: permeability (e.g., ...). Figure 4 As shown), well location (such as...) Figure 3 As shown in the figure), porosity (Table 1), where the 0-dimensional numerical data were generated using Latin hypercube; the permeability field was realized through sequential Gaussian simulation (SGS), as shown in the figure. Figure 4As shown, SGS has been widely used in permeability field construction; 0-dimensional data and 1-dimensional dynamic well control parameter time series data are obtained by using Latin hypercube design under given constraints, such as... Figure 5 As shown, the generalization ability of this patent is verified by randomly constructing a sample set.

[0126] Table 1. Dataset Parameter Range Table (Referencing CO2 Flooding in Block G of a Certain Oilfield)

[0127]

[0128] S2: Data preprocessing and cleaning: ① Remove outliers using reservoir engineering methods (e.g., simulation data from abnormal schemes caused by randomly generated unreasonable parameters); ② Normalization: Map data of different dimensions (e.g., pressure MPa, injection / production parameters m³ / d) to the [0,1] interval to eliminate the influence of dimensional differences; ③ Data fusion: Align well network data and dynamic production data according to time series and spatial location to construct multimodal data samples. First, normalize the injection / production parameters, then assign values ​​to the injection / production parameters at each well location according to Gaussian transformation. The fused sample set represents the injection / production parameters of each well at different time steps. The darker the color at the well location, the greater the injection / production fluid volume, such as... Figure 6 As shown.

[0129] S3: Extract features from multimodal sample data based on the constructed multimodal fusion deep learning model (Transformer-Ctrl-LSTM-FCN), such as... Figure 7 As shown.

[0130] S4: Model Training and Hyperparameter Optimization ① Dataset Partitioning: The preprocessed multimodal sample set was divided into training and test sets in an 8:2 ratio; ② Loss Function Design: A hybrid loss function (Mean Squared Error (MSE) + Cross-Entropy Loss) was used to optimize the leading edge position error; ③ Training Strategy: An adaptive learning rate (Adam optimizer) was used, combined with an early stopping strategy to prevent overfitting. Bayesian optimization of network hyperparameters was performed based on the validation set accuracy. The training effect on the test set was as follows: Figure 8 , Figure 9 As shown.

[0131] S5-S6: Dynamic prediction and real-time adjustment of the oil front. Current geological parameters, production dynamics data, and monitoring signals of the reservoir are input into the trained model. The model outputs the position and shape of the oil front in real time, such as... Figure 10 As shown.

[0132] This application constructs a complete, intelligent, and adaptive gas drive front description system through the above six closely linked steps. It fundamentally overcomes the three major bottlenecks of traditional methods in terms of efficiency (computation speed), accuracy (adaptability to heterogeneity), and cost (dependence on scarce parameters). Its successful implementation will provide strong technical support for gas drive development in oil and gas fields, realize refined and intelligent reservoir management, and ultimately achieve the goal of improving recovery rate and reducing development costs.

[0133] The above specific embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to examples, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for describing the air-driven leading edge based on a multimodal fusion deep learning model, characterized in that, Includes the following steps: S1. Collect multimodal data and construct a multimodal integrated database. The multimodal data includes static geological data, dynamic production data, real-time monitoring data, and historical frontier data of the target reservoir. S2. Perform data preprocessing and cleaning on the multimodal integrated database; S3. Construct a multimodal fusion deep learning model based on Transformer-Ctrl-LSTM-FCN; S4. Train and optimize the hyperparameters of the multimodal fusion deep learning model; S5. Input the real-time data into the trained model and output the leading edge prediction result; S6. Based on the output prediction results, add the new data that meets the requirements to the multimodal integrated database, trigger incremental training of the model, and update the parameters to adapt to the dynamic changes of the reservoir. S3 includes: S31. Construct a Transformer encoder to process static geological data, and add a local attention mechanism to enhance the extraction of geological features; S32. An improved Ctrl-LSTM network is used to process dynamic production data and extract time-series features; S33. For one-dimensional scalar data in real-time monitoring data and historical leading edge data, a fully connected neural network (FCN) feature extraction module is constructed. The deep network structure is used to maximize feature extraction of scalar data, providing high-quality feature input for multimodal fusion. S34, Cross-modal fusion layer: Based on a dynamic attention weight allocation mechanism, it fuses features from the three modalities of S31-S33; S35. A leading edge prediction output layer is constructed using a fully connected layer and a Softmax activation function to output the air-driven leading edge image; In S32, the improved Ctrl-LSTM network includes: S321. Construct a time memory transfer path; S322. Construct a spatial memory transfer path; S323. Based on the attention mechanism, the outputs of the temporal memory transfer path and the spatial memory transfer path are deeply integrated to generate comprehensive temporal features; In S323, the expression for deep fusion is: ， in, For dynamic feature vectors, , , , These are the network weight parameters. Input data for the current time step. For control signals, It is the Sigmoid activation function. This is the hidden state from the previous time step. This represents the cell state at the current time step. This represents the spatial memory state at the current time step. In S34, the fusion formula for the dynamic allocation mechanism of attention weights is: ， in, This is a global multimodal feature vector. For static feature vectors, For dynamic feature vectors, For scalar eigenvectors, This represents the attention weight.

2. The description method according to claim 1, characterized in that: S4 includes: S41. Divide the preprocessed multimodal integrated database into a training set and a test set according to the proportions. S42. Use a hybrid loss function for model training; S43. Combine the Adam optimizer, early stopping strategy and Bayesian optimization to adjust hyperparameters; S44. Calibrate the model using on-site measured data.

3. The description method according to claim 2, characterized in that: In S42, the expression for the hybrid loss function is: ， in, , , These are the weighting coefficients. For mean square error loss, For structural similarity loss, This is the edge loss.

4. The description method according to claim 3, characterized in that: In S42, The mean square error loss ,in, For the sample size, This represents the original pixel value matrix of the actual gas saturation field at the gas-driven leading edge. This is the original pixel value matrix of the gas saturation field at the gas-driven leading edge predicted by the model. The structural similarity loss ,in, This represents a real image at the level of structural features. Predict the representation of an image at the level of structural features; The edge loss ,in, This is a gradient magnitude map of a real image after processing it using the Sobel operator. This is a map showing the gradient magnitude of the predicted image after processing with the Sobel operator.

5. The description method according to claim 1, characterized in that, S2 includes the following processing of the multimodal integrated database: Outliers are removed, and missing values ​​are filled using interpolation or generative models. Normalization: Mapping data of different dimensions to the [0,1] interval to eliminate the influence of dimension differences. The expression is: ,in, For normalized values, These are the initial data values. This is the minimum value of the parameter. This is the maximum value of the parameter; Data fusion: Dynamic production data is mapped to well location grids through Gaussian transformation to construct a spatiotemporal fusion sample of "well location + injection and production" parameters.

6. The description method according to claim 1, characterized in that: In S31, the local attention mechanism has the following formula for calculating attention weights: ， in, Q i Indicates the first i The query vector for each geological grid location is obtained by linear projection from static geological data input. K j and V j They represent the first j Key vectors and value vectors for each geological grid location; D k represents the scaling factor for the dimension of the key vector; softmax is the normalization exponential function.

Images