Residual oil distribution prediction method and system, control device and storage medium

Abstract

The invention discloses a remaining oil distribution prediction method and system, a control device and a storage medium, and the method comprises the steps: obtaining a multi-scale fusion feature vector according to wide-area apparent resistivity data, seismic attribute data, optical fiber time sequence signal data, nano tracer tracking data and pore network point cloud data, inputting a physical constraint neural network to obtain a residual oil saturation initial value and a pressure gradient initial value of each region of the target layer, wherein a loss function comprises data fitting item loss, Darcy law physical constraint loss and mass conservation constraint loss; the real-time water content and the bottomhole flowing pressure of the production well serve as observation data, and a static parameter correction value, a remaining oil saturation correction value and a pressure gradient correction value are obtained through correction based on an ensemble Kalman filtering algorithm; the static parameter correction value and the pressure gradient correction value are fed back to a physical constraint neural network for loss function updating; and determining a remaining oil distribution result of the target layer. According to the invention, the accuracy and timeliness of residual oil distribution prediction can be improved.

Description

Technical Field

[0001] The present application relates to the technical field related to oil and gas exploration and development, and in particular to a method, system, control device and storage medium for predicting residual oil distribution. Background Art

[0002] Predicting the distribution of remaining oil is a key step in improving oil and gas recovery during exploration and development, but it faces numerous challenges. Conventional seismic methods, for example, have limited spatial resolution, making it difficult to identify thin layers or small-scale remaining oil. Four-dimensional seismic analysis is expensive and lacks sensitivity to low-saturation remaining oil. Conventional well logging techniques can only detect areas near the wellbore.

[0003] Traditional remaining oil distribution prediction models have the following disadvantages: (1) Relying on static geological models, it is unable to dynamically respond to real-time changes in reservoir properties and fluid distribution during the injection and production process.

[0004] (2) There is a lack of effective integration of multi-source data such as seismic, well logging, and production dynamics data, making cross-scale modeling difficult.

[0005] (3) Pure data-driven models may not conform to the physical laws of seepage, resulting in unreliable prediction results.

[0006] (4) The existing monitoring system has a long update cycle, making it difficult to make immediate development decisions. Summary of the Invention

[0007] The present application aims to solve at least one of the technical problems existing in the prior art. To this end, the present application proposes a method for predicting residual oil distribution, which can improve the accuracy and timeliness of residual oil distribution prediction.

[0008] The present application also provides a remaining oil distribution prediction system, a control device for executing the above-mentioned remaining oil distribution prediction method, and a computer-readable storage medium.

[0009] According to the remaining oil distribution prediction method of the first embodiment of the present application, the method includes: Acquire wide-area apparent resistivity data, seismic attribute data, fiber optic time series signal data, nano-tracer tracking data, pore network point cloud data, and real-time production well water cut and bottomhole pressure of the target layer; performing spatiotemporal alignment processing and feature fusion processing on the wide-area apparent resistivity data, the seismic attribute data, the optical fiber time series signal data, the nanotracer tracking data, and the pore network point cloud data to obtain a multi-scale fusion feature vector; Inputting the multi-scale fusion feature vector into a physical constraint neural network to obtain initial values ​​of remaining oil saturation and pressure gradient in each region of the target layer, wherein the loss function of the physical constraint neural network includes data fitting term loss, Darcy's law physical constraint loss, and mass conservation constraint loss; Using the real-time water cut and bottom hole flowing pressure of the production well as observation data, the static parameters in the Darcy's law physical constraint loss and the mass conservation constraint loss, the initial value of the remaining oil saturation, and the initial value of the pressure gradient as state vectors, and performing parameter correction on the static parameters, the initial value of the remaining oil saturation, and the initial value of the pressure gradient based on an ensemble Kalman filter algorithm to obtain static parameter correction values, remaining oil saturation correction values, and pressure gradient correction values; Feeding back the static parameter correction value and the pressure gradient correction value to the physical constraint neural network to update the loss function; The remaining oil distribution result of the target layer is determined according to the remaining oil saturation correction value of each area of ​​the target layer.

[0010] The remaining oil distribution prediction method according to the embodiment of the present application has at least the following beneficial effects: This application uses multi-method, cross-scale data fusion of wide-area apparent resistivity data with commonly used remaining oil monitoring data. This approach addresses the limited resolution, dynamic monitoring difficulties, and "blind spots" in heterogeneous reservoirs associated with traditional methods, while also addressing the limited timeliness of wide-area electromagnetic methods. The multi-scale fused feature vectors are then fed into a physical constraint neural network to obtain initial values ​​for remaining oil saturation and pressure gradients in each region of the target layer. This physical constraint neural network embeds physical laws (Darcy's law and the law of conservation of mass), combining hard constraints (physical laws directly embedded in the network) with soft constraints (loss function regularization) to balance data-driven and physical laws, improving model reliability. Finally, using real-time production well water cut and bottomhole flowing pressure as observation data, and the static parameters from the Darcy's law physical constraint loss and mass conservation constraint loss, as well as initial values ​​for remaining oil saturation and pressure gradient, as state vectors, these static parameters, initial values ​​for remaining oil saturation, and initial values ​​for pressure gradient are modified using an ensemble Kalman filter algorithm, improving the accuracy and timeliness of remaining oil distribution prediction.

[0011] According to some embodiments of the present application, performing spatiotemporal alignment processing and feature fusion processing on the wide-area apparent resistivity data, the seismic attribute data, the optical fiber time series signal data, the nanotracer tracking data, and the pore network point cloud data to obtain a multi-scale fusion feature vector includes: performing feature extraction on the wide-area apparent resistivity data, the seismic attribute data, the optical fiber time series signal data, the nanotracer tracking data, and the pore network point cloud data to obtain wide-area apparent resistivity features, seismic attribute features, optical fiber time series signal features, nanotracer tracking features, and pore network point cloud features; Based on the feature projection method, a joint loss function is added to the feature projection layer to force different modal features to satisfy the same preset physical relationship. The wide-area apparent resistivity feature, the seismic attribute feature, the fiber time series signal feature, the nanotracer tracking feature, and the pore network point cloud feature are mapped into a unified feature space to obtain a unified feature of wide-area apparent resistivity, a unified feature of seismic attributes, a unified feature of fiber time series signals, a unified feature of nanotracer tracking, and a unified feature of pore network point cloud. The wide-area apparent resistivity unified feature, the seismic attribute unified feature, the optical fiber timing signal unified feature, the nanotracer tracking unified feature, and the pore network point cloud unified feature are weighted and fused to obtain the multi-scale fusion feature vector.

[0012] According to some embodiments of the present application, the seismic attribute data is feature extracted using a convolutional neural network, the optical fiber timing signal data is feature extracted using a long short-term memory network, and the pore network point cloud data is feature extracted using a graph neural network embedded with a seepage equation.

[0013] According to some embodiments of the present application, the Darcy's law physical constraint loss is obtained by the following steps: Obtaining a plurality of samples of observed remaining oil flow rates and a pressure gradient correction value predicted by the physical constraint neural network each time, wherein the observed remaining oil flow rates are determined by the optical fiber timing signal data or the nanotracer tracking data; Based on Darcy's law, the current remaining oil predicted flow rate is calculated according to the pressure gradient correction value obtained by the previous physical constraint neural network prediction, so as to obtain the remaining oil predicted flow rate of multiple samples; The Darcy's law physical constraint loss is determined based on the observed remaining oil flow rate and the predicted remaining oil flow rate of all samples.

[0014] According to some embodiments of the present application, the mass conservation constraint loss is obtained by the following steps: Obtaining porosity at multiple time steps, predicted remaining oil flow rate, and predicted remaining oil saturation value predicted by the physical constraint neural network, wherein the porosity is determined by the pore network point cloud data; Calculate the time change rate of the product of the porosity and the predicted value of the remaining oil saturation at each time step, and the divergence of the product of the predicted remaining oil flow rate and the predicted value of the remaining oil saturation at each time step; calculating the residuals of the time rate of change and the divergence in the time series; The residual of the time change rate and the divergence in the time series is used as the mass conservation constraint loss.

[0015] According to some embodiments of the present application, the constraint formula of the loss function of the physical constraint neural network is: ; ; ; ; in, is the loss function of the physical constraint neural network, is the data fitting loss, is the physical constraint loss of Darcy's law, is the mass conservation constraint loss, 、 and are the weights of the data fitting term loss, the Darcy's law physical constraint loss, and the mass conservation constraint loss, respectively. is the total number of samples, The physical constraint neural network is i The predicted value of remaining oil saturation of samples, For the i The true label value of the remaining oil saturation of samples, is the total number of samples, The physical constraint neural network is j The predicted flow rate of remaining oil for each sample, For the j The remaining oil observed flow rate of each sample is determined by the optical fiber timing signal data or the nanotracer tracking data. is the total number of time steps, is the porosity, The physical constraint neural network is k The predicted value of remaining oil saturation in time steps, The physical constraint neural network is k The remaining oil predicted flow rate for each time step.

[0016] According to some embodiments of the present application, the static parameter includes a permeability field; and the method further includes: Divide each area of ​​the target layer into grids to obtain multiple grids; extracting a permeability field of a target grid, the residual oil saturation correction value, and a difference between the residual oil saturation correction values ​​of the target grid and an adjacent grid, wherein the target grid is one of the plurality of grids; If the residual oil saturation correction value of the target grid is greater than a preset first threshold, and the difference between the residual oil saturation correction values ​​of the target grid and an adjacent grid is greater than a preset second threshold, the target grid is encrypted and divided into multiple subgrids, and each subgrid is interpolated according to the permeability field of the target grid and the residual oil saturation correction value to recalibrate the physical constraint neural network model.

[0017] According to the residual oil distribution prediction system of the second embodiment of the present application, the system includes: Data acquisition unit, used to obtain wide-area apparent resistivity data, seismic attribute data, fiber time series signal data, nano-tracer tracking data, pore network point cloud data, and real-time production well water cut and bottom hole pressure of the target layer; a multi-scale fusion feature vector determination unit, configured to perform spatiotemporal alignment processing and feature fusion processing on the wide-area apparent resistivity data, the seismic attribute data, the optical fiber time series signal data, the nanotracer tracking data, and the pore network point cloud data to obtain a multi-scale fusion feature vector; an initial prediction unit, configured to input the multi-scale fusion feature vector into a physical constraint neural network to obtain initial values ​​of remaining oil saturation and pressure gradient in each region of the target layer, wherein the loss function of the physical constraint neural network includes a data fitting term loss, a Darcy's law physical constraint loss, and a mass conservation constraint loss; a parameter correction unit, configured to use the real-time water cut and bottom hole flowing pressure of the production well as observation data, and the static parameters in the Darcy's law physical constraint loss and the mass conservation constraint loss, the initial value of the remaining oil saturation, and the initial value of the pressure gradient as state vectors, and perform parameter correction on the static parameters, the initial value of the remaining oil saturation, and the initial value of the pressure gradient based on an ensemble Kalman filter algorithm to obtain a static parameter correction value, a remaining oil saturation correction value, and a pressure gradient correction value; A correction parameter feedback unit, configured to feed back the static parameter correction value and the pressure gradient correction value to the physical constraint neural network for loss function update; The remaining oil distribution result determining unit is used to determine the remaining oil distribution result of the target layer according to the remaining oil saturation correction value of each area of ​​the target layer.

[0018] The remaining oil distribution prediction system according to the embodiment of the present application has at least the following beneficial effects: This application uses multi-method, cross-scale data fusion of wide-area apparent resistivity data with commonly used remaining oil monitoring data. This approach addresses the limited resolution, dynamic monitoring difficulties, and "blind spots" in heterogeneous reservoirs associated with traditional methods, while also addressing the limited timeliness of wide-area electromagnetic methods. The multi-scale fused feature vectors are then fed into a physical constraint neural network to obtain initial values ​​for remaining oil saturation and pressure gradients in each region of the target layer. This physical constraint neural network embeds physical laws (Darcy's law and the law of conservation of mass), combining hard constraints (physical laws directly embedded in the network) with soft constraints (loss function regularization) to balance data-driven and physical laws, improving model reliability. Finally, using real-time production well water cut and bottomhole flowing pressure as observation data, and the static parameters from the Darcy's law physical constraint loss and mass conservation constraint loss, as well as initial values ​​for remaining oil saturation and pressure gradient, as state vectors, these static parameters, initial values ​​for remaining oil saturation, and initial values ​​for pressure gradient are modified using an ensemble Kalman filter algorithm, improving the accuracy and timeliness of remaining oil distribution prediction.

[0019] According to a third embodiment of the present application, a control device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the remaining oil distribution prediction method described in the first embodiment. Because the control device utilizes all the technical solutions of the remaining oil distribution prediction method of the above embodiment, it at least has all the beneficial effects brought about by the technical solutions of the above embodiment.

[0020] According to a fourth embodiment of the present application, a computer-readable storage medium stores computer-executable instructions for executing the remaining oil distribution prediction method described in the first embodiment. Because the computer-readable storage medium employs all of the technical solutions of the remaining oil distribution prediction method described in the aforementioned embodiment, it at least has all of the beneficial effects provided by the technical solutions of the aforementioned embodiment.

[0021] Other features and advantages of the present application will be set forth in the following description, and in part will be apparent from the description, or may be learned by practicing the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] The above and / or additional aspects and advantages of the present application will become apparent and easily understood from the description of the embodiments in conjunction with the following drawings, in which: Figure 1 This is a flowchart of a method for predicting residual oil distribution according to an embodiment of the present application. DETAILED DESCRIPTION

[0023] The following describes in detail embodiments of the present application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application and are not to be construed as limiting the present application.

[0024] In the description of this application, if there is a description of first, second, etc., it is only for the purpose of distinguishing technical features, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features.

[0025] In the description of this application, it should be understood that descriptions involving orientation, such as the orientation or positional relationship indicated by up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing this application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on this application.

[0026] In the description of this application, it should be noted that, unless otherwise clearly defined, terms such as setting, installing, and connecting should be understood in a broad sense, and technical personnel in the relevant technical field can reasonably determine the specific meaning of the above terms in this application based on the specific content of the technical solution.

[0027] The following will be combined Figure 1 The remaining oil distribution prediction method of the embodiment of the present application is described clearly and completely. Obviously, the embodiment described below is only a part of the embodiment of the present application, not all the embodiments.

[0028] refer to Figure 1 , Figure 1 This is a flowchart of a method for predicting residual oil distribution according to an embodiment of the present application.

[0029] According to the remaining oil distribution prediction method of the first embodiment of the present application, the method includes: Acquire wide-area apparent resistivity data, seismic attribute data, fiber optic time series signal data, nano-tracer tracking data, pore network point cloud data, and real-time production well water cut and bottomhole pressure of the target layer; Perform spatiotemporal alignment and feature fusion processing on wide-area apparent resistivity data, seismic attribute data, fiber time series signal data, nanotracer tracking data, and pore network point cloud data to obtain multi-scale fusion feature vectors. The multi-scale fused feature vectors are input into a physical constraint neural network to obtain the initial values ​​of remaining oil saturation and pressure gradient in each area of ​​the target layer. The loss function of the physical constraint neural network includes data fitting loss, Darcy's law physical constraint loss, and mass conservation constraint loss. The real-time water cut and bottom hole flowing pressure of the production wells are used as observation data. The static parameters, initial values ​​of remaining oil saturation, and initial values ​​of pressure gradient in the physical constraint loss and mass conservation constraint loss of Darcy's law are used as state vectors. The static parameters, initial values ​​of remaining oil saturation, and initial values ​​of pressure gradient are corrected based on the ensemble Kalman filter algorithm to obtain corrected values ​​of static parameters, remaining oil saturation, and pressure gradient. Feeding back the static parameter correction value and the pressure gradient correction value to the physical constraint neural network to update the loss function; The remaining oil distribution result of the target layer is determined according to the remaining oil saturation correction value of each area of ​​the target layer.

[0030] In some embodiments, wide-area apparent resistivity data is obtained by: Acquire well logging data, seismic data, and wide-area electromagnetic data of the target layer; The wide-area electromagnetic data are denoised and then inverted using well logging and seismic data as constraints to obtain wide-area apparent resistivity data.

[0031] As you can understand, seismic attribute data includes wave impedance and velocity difference data, which are processed through time-lapse seismic (4D seismic) to extract dynamic characteristics. Fiber optic time-series signal data is used to analyze temperature and acoustic signal anomalies and invert fluid velocity around the wellbore. Nanotracer tracking data is combined with mass spectrometry to determine seepage paths and swept volumes. Pore network point cloud data is extracted from core CT scan images using image segmentation algorithms.

[0032] After obtaining the wide-area apparent resistivity data, seismic attribute data, fiber timing signal data, nanotracer tracking data, and pore network point cloud data of the target layer, it is also necessary to set the data credibility threshold (such as the seismic wave impedance change error is less than 5%, and the fiber temperature drift is less than 0.5°C) to automatically eliminate outliers.

[0033] In some embodiments of the present application, wide-area apparent resistivity data, seismic attribute data, optical fiber time series signal data, nanotracer tracking data, and pore network point cloud data are subjected to spatiotemporal alignment and feature fusion processing to obtain a multi-scale fusion feature vector, including: Feature extraction is performed on wide-area apparent resistivity data, seismic attribute data, fiber time series signal data, nanotracer tracking data, and pore network point cloud data to obtain wide-area apparent resistivity features, seismic attribute features, fiber time series signal features, nanotracer tracking features, and pore network point cloud features; Based on the feature projection method, a joint loss function is added to the feature projection layer to force different modal features to satisfy the same preset physical relationship. The wide-area apparent resistivity features, seismic attribute features, fiber timing signal features, nanotracer tracking features, and pore network point cloud features are mapped into a unified feature space. The unified features of wide-area apparent resistivity, seismic attribute features, fiber timing signal features, nanotracer tracking features, and pore network point cloud features are obtained. Weight allocation and feature fusion are performed on the unified features of wide-area apparent resistivity, seismic attributes, fiber timing signals, nanotracer tracking and pore network point cloud to obtain a multi-scale fusion feature vector.

[0034] In some embodiments of the present application, convolutional neural networks are used for feature extraction of seismic attribute data, long short-term memory networks are used for feature extraction of optical fiber time series signal data, and graph neural networks embedded with seepage equations are used for feature extraction of pore network point cloud data.

[0035] The embodiment of the present application adopts a spatiotemporal alignment algorithm to unify the spatial grids and timestamps of multi-source data, constructs a heterogeneous data encoder through different feature extraction methods, and performs feature extraction on wide-area apparent resistivity data, seismic attribute data, fiber timing signal data, nanotracer tracking data, and pore network point cloud data.

[0036] It's understandable that wide-area apparent resistivity data directly reflects differences in the electrical properties of reservoir fluids (such as the conductivity differences between oil, water, and gas). Traditional inversion methods (such as regularized least squares and Bayesian inversion) can efficiently resolve resistivity distributions without relying on deep learning models. Nanotracer tracking data, which changes over time, directly reflects the migration paths and velocities of fluids. These data are typically processed using material balance equations or analytical models (such as convection-diffusion equations) with low computational complexity. These two types of data, due to their clear physical meaning, established interpretation methods, and efficient traditional algorithms, typically do not require complex neural network processing.

[0037] Seismic attribute data is typically three-dimensional and exhibits local spatial correlation. The stratigraphic features it reflects (such as faults and river channels) may recur at different locations. Furthermore, it exhibits multi-scale characteristics, meaning that thin layers (<5m thick) and thick layers require different-scale convolution kernels for extraction. The advantages of convolutional neural networks include using convolution kernels to extract local features (such as impedance anomalies) using a sliding window, adapting to the spatial continuity of seismic attribute data. For example, 3D convolutional neural networks can capture the characteristics of three-dimensional geological volumes.

[0038] Fiber optic time series signal data is a time series (e.g., minute-by-minute temperature / acoustic sampling) with dynamic evolution characteristics. Processes such as fluid flow and pressure propagation require capturing hourly or even daily trends, and time series modeling is required to filter out wellsite vibration and instrument noise. Long-short-term memory networks offer the advantage of gating mechanisms: they can control information flow through forget gates, input gates, and output gates, resolving long-term dependencies. They also possess sequence modeling capabilities, enabling modeling of both gradual and sudden changes in temperature / acoustic signals during waterflooding (e.g., oil-water front breakthroughs).

[0039] The pore network point cloud data consists of nodes (pores) and edges (throats), with complex topological relationships, which are suitable for graph structure representation. The permeability between pores is determined by the connectivity of the throat, and it is necessary to capture the neighborhood interactions between nodes, that is, their data has local connectivity. Each node contains attributes such as pore volume and throat radius, and edges may carry flow resistance information. Graph neural networks have good adaptability to graph structures: through the message passing mechanism, the characteristics of neighboring nodes are aggregated to directly model the pore-throat topological relationship. The seepage equation is embedded in the graph neural network, and Darcy's law constraints are added in the message passing process. For example, the edge weight is calculated by the throat radius and fluid viscosity: ;Formula (1) in, represents the edge weight, i.e., the throat ij The conductivity (or flow conducting capacity) of Indicates the throat ij The radius, represents the fluid viscosity, Indicates the throat ij length.

[0040] Formula (1) is used to calculate the conductivity of the throat, which is used to simulate the flow of fluid in the pores. ij The conductivity of a fluid is used to describe the flow of fluid through the pores connected in the pore network. i and j The flow capacity of the throat is related to the throat geometry parameters and fluid properties. ij The radius is the cross-sectional radius of the throat, which directly affects the flow capacity (the conductivity is proportional to the fourth power of the radius). The viscosity of the fluid characterizes the internal friction resistance of the fluid during flow. The greater the viscosity, the lower the conductivity. ij The length of the throat is the extension distance. The longer the length, the greater the flow resistance and the lower the conductivity.

[0041] Convolutional neural networks provide macroscopic geological structural constraints, long-short-term memory networks capture dynamic development responses, and graph neural networks embedded with flow equations reveal the influence of microscopic pore structure on flow. These three networks, targeting gridded, serialized, and unstructured data, jointly construct a multiscale "pore-wellbore-reservoir" model. Through multimodal fusion, they achieve a closed-loop "static characterization-dynamic update" of remaining oil distribution, improving accuracy by 20% to 30% compared to a single network.

[0042] Based on the feature projection method, a joint loss function is added to the feature projection layer to force different modal features to satisfy the same preset physical relationship. For example, seismic attribute features and pore network point cloud features must satisfy: ;Formula (2) in, is the resistivity-saturation conversion function calibrated by core experiments, is the earthquake attribute characteristic, is the pore network point cloud feature, is the square of the Euclidean norm, which is used to quantify the difference between two eigenvectors in the unified space. is the preset difference threshold.

[0043] It should be noted that the specific joint loss function can be selected according to actual conditions. The physical relationship mentioned above is only a specific embodiment and cannot be regarded as a limitation of this application.

[0044] In some embodiments, in order to comprehensively utilize the complementary information of multimodal data, overcome the limitations of a single data source, and improve the accuracy and robustness of the remaining oil prediction model, the feature fusion weight is automatically adjusted according to the data confidence (such as the fiber signal-to-noise ratio), and the weights of the unified feature of wide-area apparent resistivity, the unified feature of seismic attributes, the unified feature of fiber time series signals, the unified feature of nanotracer tracking, and the unified feature of pore network point cloud are determined.

[0045] In some embodiments of the present application, the structure of the physical constraint neural network is designed as follows: Input layer: multi-scale fused feature vector (dimension 256).

[0046] Hidden layer: 5 fully connected layers, with 128, 64, 32, 16, and 8 neurons in each layer respectively. The activation function for the first 4 layers uses LeakyReLU ( = 0.01), and the last layer uses Sigmoid (limiting the saturation output to 0-1).

[0047] The five fully connected layers can gradually fuse information of different scales. Too deep networks (such as those with more than 10 layers) are prone to overfitting on small data sets, while the five layers can achieve a balance between expressiveness and generalization.

[0048] LeakyReLU introduces a small slope in the negative region ( ), to avoid neuron "death". LeakyReLU has the following advantages: (1) Alleviate gradient disappearance: the gradient in the negative area is , avoiding the zero gradient problem of ordinary ReLU. (2) Sparse activation: only some neurons are activated, improving computational efficiency. (3) Suitable for intermediate layers: stable training in deep networks, suitable for feature abstraction stage.

[0049] Sigmoid compresses the output to the interval [0, 1] and is suitable for probability or proportion prediction. The value of residual oil saturation is in the interval [0, 1]. Sigmoid can directly output a normalized value, and the output can be regarded as the "oil content probability", which facilitates subsequent decision-making (such as target area optimization).

[0050] Physical constraint layer: Darcy's law and mass conservation equation are hard-wired into the loss function.

[0051] Output layer: Main output: Remaining oil saturation distribution ( x , y , z ) (resolution 0.1m); auxiliary output: confidence index C ( x , y , z ) (0-1 probability value, confidence level >90%). The output results can be compared with wide-area apparent resistivity data (wide-area apparent resistivity profile) to ensure model accuracy.

[0052] Specifically, the wide-area apparent resistivity profile shows the distribution of wide-area apparent resistivity across different regions. Generally speaking, water-bearing areas have low apparent resistivity, while oil and gas reservoirs have high apparent resistivity. Within a reservoir, higher oil saturation generally leads to an increasing apparent resistivity. Therefore, based on this characteristic, the remaining oil saturation distribution calculated by the physical constraint neural network can be compared with the wide-area apparent resistivity profile to determine the reliability of the calculated results and whether they conform to the actual geological conditions of the study area.

[0053] In the confidence calculation, the uncertainty quantification method uses the Monte Carlo Dropout technique. In the inference phase, Dropout is randomly applied to the network activation layer (retention rate p =0.8), and T = 50 forward propagations. Calculate the standard deviation of the remaining oil saturation prediction value and normalize it to the confidence level. C= 1 - Standard deviation.

[0054] Dropout itself is a regularization technique that prevents overfitting by randomly "turning off" a subset of neurons during training. Monte Carlo Dropout, on the other hand, continues to use Dropout during the inference (prediction) phase, performing multiple forward propagations and estimating the model's uncertainty based on the variance of the predicted results. When using Monte Carlo Dropout during inference, the activation function determines the range of variation in the output after turning off neurons: if Sigmoid is used, the output remains between 0 and 1, resulting in a limited range of fluctuation across multiple predictions and a small standard deviation. If ReLU is used, negative values ​​are reset to zero, potentially causing some predicted values ​​to fluctuate wildly and have a large standard deviation. Therefore, the choice of activation function affects the model's sensitivity to Dropout, and thus the accuracy of the confidence calculation.

[0055] It should be noted that the principle and function of Monte Carlo Dropout are prior arts known to those skilled in the art and will not be described in detail here.

[0056] Next, the physical constraint embedding method is described in detail.

[0057] Darcy's law constraints: Darcy's equation: The Darcy's law for single-phase flow is extended to oil-water two-phase flow. The constraint formula for the oil phase flow rate is: ;Formula (3) in, is the oil phase flow rate, is the oil-water relative permeability, is the permeability (absolute permeability), is the viscosity of oil and water, is the pressure gradient, is the residual oil saturation.

[0058] By converting the Darcy speed As an intermediate variable, it is constrained to meet the observed remaining oil flow rate (the observed remaining oil flow rate is determined by fiber optic timing signal data or nanotracer tracking data). After the physical constraint neural network outputs the remaining oil saturation, the pressure gradient is calculated through automatic differentiation. The pressure gradient at the previous moment is used to calculate the new oil phase flow rate at the next moment.

[0059] Constraints on the mass conservation equation: Continuity equation: ;Formula (4) in, is the porosity, is the residual oil saturation, is the oil phase flow rate.

[0060] In time series data (such as minute-level production dynamics), computing and The residual is added to the loss function as a regularization term to force the network prediction results to meet the quality conservation.

[0061] In some embodiments of the present application, the constraint formula of the loss function of the physical constraint neural network is: ;Formula (5) ;Formula (6) ;Formula (7) ;Formula (8) in, is the loss function of the physical constraint neural network, is the data fitting loss, is the physical constraint loss of Darcy's law, is the mass conservation constraint loss, 、 and are the weights of data fitting term loss, Darcy’s law physical constraint loss, and mass conservation constraint loss, respectively. is the total number of samples, The physical constraint neural network is i The predicted value of remaining oil saturation of samples, For the i The true label value of the remaining oil saturation of samples, is the total number of samples, The physical constraint neural network is j The predicted flow rate of remaining oil for each sample, For the j The remaining oil observation flow rate of each sample is determined by the optical fiber timing signal data or nano-tracer tracking data. is the total number of time steps, is the porosity, The physical constraint neural network is k The predicted value of remaining oil saturation in time steps, The physical constraint neural network is k The remaining oil predicted flow rate for each time step.

[0062] In some embodiments of the present application, Darcy's law physical constraint loss is obtained by the following steps: Obtaining the remaining oil observed flow rate of multiple samples and the pressure gradient correction value obtained by each physical constraint neural network prediction, wherein the remaining oil observed flow rate is determined by optical fiber time series signal data or nanotracer tracking data; Based on Darcy's law, the current remaining oil predicted flow rate is calculated according to the pressure gradient correction value obtained by the previous physical constraint neural network prediction, so as to obtain the remaining oil predicted flow rate of multiple samples; The Darcy's law physical constraint loss is determined based on the observed remaining oil flow rate and the predicted remaining oil flow rate for all samples.

[0063] In some embodiments of the present application, the mass conservation constraint loss is obtained by the following steps: Obtain porosity at multiple time steps, predicted remaining oil velocity, and predicted remaining oil saturation using a physical constraint neural network. Porosity is determined using pore network point cloud data. Calculate the time change rate of the product of the porosity and the predicted value of the remaining oil saturation at each time step, and the divergence of the product of the predicted remaining oil flow rate and the predicted value of the remaining oil saturation at each time step; Compute residuals of temporal rate of change and divergence in time series; The residuals of the temporal rate of change and divergence in the time series are used as mass conservation constraint losses.

[0064] It is understood that the time change rate of the product of porosity and the predicted value of residual oil saturation represents the accumulation of residual oil per unit time. The divergence of the product of the predicted residual oil flow rate and the predicted value of residual oil saturation represents the outflow rate of residual oil.

[0065] The data fitting loss represents the mean squared error between the model's predictions and the true labels (a common loss function in supervised learning). The Darcy's law physics constraint loss describes fluid flow in porous media and enforces the Darcy equation in physics-constrained neural networks. The mass conservation constraint loss ensures that model predictions adhere to the law of mass conservation.

[0066] Dynamic adjustment based on data reliability 、 and : ;Formula (9) in, 、 、 They are the predicted standard deviations of the data fitting term loss, Darcy's law physical constraint loss, and mass conservation constraint loss, estimated by Monte Carlo Dropout.

[0067] Next, the training process of the physical constraint neural network is explained.

[0068] The pre-training dataset includes synthetic data from public reservoir models (such as SPE10 and UNISIM-II) and historical oil field data (which needs to be desensitized).

[0069] Pre-training task: only use Perform supervised learning and initialize network weights.

[0070] Physical constraint fine-tuning: Two-stage training: Stage 1: Fix the encoding layer weights and train only the decoding layer, focusing on physical constraints ( and Phase 2: Unfreeze all network layers and jointly optimize the data fitting term and the physical constraint term.

[0071] The two-stage training is mainly due to: (1) Stable training and gradient balance. When optimizing data fitting terms (such as mean square error) and physical constraint terms (such as Darcy's law residual) simultaneously, the gradient directions and magnitudes of different loss terms may conflict, resulting in unstable optimization or falling into local optimality. Therefore, a phased strategy is adopted: Phase 1: Prevent the random initial weights of the encoding layer from interfering with the learning of physical laws, ensuring that the decoder has a preliminary understanding of how to generate physical understanding. Phase 2: Based on physical consistency, the encoding layer is allowed to adjust the feature extraction method to further improve data fitting accuracy.

[0072] (2) Gradually introduce complexity. Phase 1: Simplification: After fixing the encoding layer, the optimization objective is simplified to “satisfying physical laws in a known feature space,” reducing model complexity and accelerating convergence. Phase 2: Comprehensive optimization: After the decoding layer initially satisfies physical constraints, the encoding layer parameters are released, allowing the model to more flexibly capture complex patterns in the data.

[0073] The advantages of training in two stages can be summarized as follows: (1) Prevent overfitting. Stage 1: Optimize only the decoding layer, reducing the number of adjustable parameters and reducing sensitivity to noisy data. Stage 2: Adjust the encoding layer under the "protection" of physical constraints to prevent the model from deviating from the physical solution due to complexity.

[0074] (2) Improve convergence efficiency. Phase 1: Fast convergence: Simplify the optimization objective (only physical constraints) to enable the decoding layer to quickly find a feasible solution. Phase 2: Fine tuning: Based on good initialization, joint optimization accelerates global convergence.

[0075] (3) Domain adaptability. Low-data scenarios: When labeled data is scarce, the physical constraints in stage 1 can serve as strong regularization, reducing reliance on data. Multimodal data fusion: The encoding layer in stage 2 can learn the correlation between cross-modal features (such as the mapping of seismic attributes to pore structure).

[0076] Optimizer Settings: AdamW optimizer (learning rate 10−4, weight decay 10−5). Cosine annealing with a learning rate of 50 epochs was used. Regularization and anti-overfitting: Dropout was used in all hidden layers with a dropout rate of 0.3. Early stopping: Training was terminated after the validation set loss did not decrease for 10 consecutive epochs.

[0077] The dynamic correction of the model in this application is achieved using the digital twin dynamic update module, which is based on the ensemble Kalman filter algorithm, combined with reservoir numerical simulation and real-time data assimilation technology to dynamically correct the permeability field and the initial value of the remaining oil saturation.

[0078] It is understandable that the state vector x Including all parameters and variables that need to be dynamically updated. In the embodiment of the present application, the state vector includes but is not limited to neural network weights, static parameters in Darcy's law physical constraint loss and mass conservation constraint loss, initial value of residual oil saturation and initial value of pressure gradient, wherein the static parameters include but are not limited to permeability field and porosity.

[0079] In some embodiments, the state vector x The constraint formula is: ;Formula (10) in, represents the neural network weights, K represents the permeability field, Indicates the initial value of residual oil saturation, N represents the number of neural network weights, M Indicates the number of reservoir grids (e.g. 10,000 grids).

[0080] It should be noted that the parameters and variables shown in formula (10) are only examples, and more parameters and variables may be included.

[0081] First, initialize the set and sample the initial state from the prior distribution: (1) Neural network weights . represents a normal distribution, Represents the neural network weights From a mean , the covariance matrix is Sampling from a multidimensional normal distribution. is a diagonal covariance matrix, where It is the identity matrix, with diagonal elements of 1 and off-diagonal elements of 0. 0.1 is the variance value, indicating that the variance of each weight dimension is 0.1 and that different weights are independent of each other (covariance is 0). Variance 0.1: The random perturbation amplitude of weight initialization is small to avoid gradient explosion or vanishing. Diagonal matrix: Each weight dimension is initialized independently, without correlation assumption. Neural network weights The initialization method is: with pre-trained weights is the mean (such as the basic model parameters in transfer learning), from the mean , random sampling from an independent normal distribution with a variance of 0.1, generating initial weights, for the i Weight , whose distribution is , all weights form a diagonal covariance matrix .

[0082] (2) Permeability field , mean Derived from geological models, it reflects the logarithmic mean of permeability, which is related to the rock pore structure. represents the permeability field Obeys lognormal distribution, its natural logarithm Normal distribution: , yes The mean of The scale (median), yes The variance of The width of the distribution (uncertainty).

[0083] (3) Initial value of residual oil saturation . Indicates the initial value of residual oil saturation It obeys uniform distribution in the range of 0.2-0.4.

[0084] In the prediction step, each ensemble member Run the simplified reservoir model (alternative model): ;Formula (11) in, Indicates the i The state vector of each ensemble member at the prediction step represents the model's estimate of the reservoir state at the next moment. Indicates the i The state vector of each ensemble member after updating in the analysis step is the reservoir parameters and dynamic variables corrected by the observed data. Represents an alternative model, used to replace the traditional numerical simulator (such as finite difference method), which is implemented by: Based on the physical constraint neural network of this application, input the current state , output the reservoir status at the future moment (remaining oil saturation and pressure field ). is process noise, indicating uncertainty in the model, has a mean of 0, and the covariance matrix is The normal distribution of The effects of include the simplification errors of the surrogate model M, the spatial variability of reservoir parameters (such as permeability field), and unmodeled physical processes (such as complex phase changes). is the covariance matrix of the process noise, which is used to quantify the uncertainty of the model prediction. For example, the diagonal elements represent the state variables (such as 、 ), and the off-diagonal elements represent the correlation between variables (such as the dynamic coupling of residual oil saturation and pressure gradient).

[0085] It should be noted that the residual oil saturation and pressure field It is the core input of the prediction step and the basis of the update step. and pressure field Supports reservoir dynamic optimization and ensemble Kalman filter algorithm data assimilation to achieve efficient and high-precision development decision-making.

[0086] In addition, it should be noted that the pressure gradient is the pressure change per unit length in the direction of fluid flow, which can be understood as the pressure difference between two positions. The pressure field refers to the pressure distribution at each point in the fluid. Generally speaking, the pressure field is a scalar field, meaning that each point in space has a corresponding pressure value. In this application, both the pressure gradient and the pressure field are intermediate results and variables used to update data.

[0087] Calculate the predicted mean and covariance : ;Formula (12) ;Formula (13) in, Indicates the number of ensemble members and the number of model samples used to characterize uncertainty. For the i The state vector of the prediction step of each ensemble member. is the ensemble mean of the prediction step, that is, the prediction mean, which is the average value of the state vectors of all ensemble members. It represents the best estimate of the reservoir state by the model and serves as the benchmark for subsequent analysis steps. is the collective covariance matrix of the prediction step, which is used to quantify the uncertainty and correlation of the state variables. Its diagonal elements are the values ​​of each state variable (such as 、 ), reflecting the uncertainty of its prediction, and the off-diagonal elements are the covariances between different variables (such as the correlation between the permeability field and the pressure gradient).

[0088] Calculate the predicted mean and covariance The uses of include: (1) supporting the calculation of subsequent Kalman gain. (2) quantifying model uncertainty. (3) supporting dynamic data assimilation. In subsequent update steps, the observed data are used to correct the predicted mean and covariance. (4) supporting real-time decision making. and Adjust the injection and production rate to maximize the recovery rate. (5) Risk warning: A sudden increase in the variance of high pressure in the covariance matrix may indicate the risk of fracture opening or water breakthrough, triggering regulatory measures.

[0089] The observation data uses real-time production well water content and bottom hole pressure.

[0090] In the update step, define the observation operator H , mapping the state vector to the observation space: ;Formula (14) in, Indicates the i The predicted value of the observation vector of each set member in the prediction step, that is, the predicted value of the model for the actual observation data, is used to compare with the real observation data to calibrate the model. H is the observation operator, which is used to map the state vector to the observation space. Indicates the i The state vector of each ensemble member at the prediction step. The water cut of the production well indicates the volume percentage of water in the production fluid of the production well, reflecting the effect of water injection development. A high water cut of the production well (such as greater than 80%) may indicate water flooding breakthrough or residual oil dispersion. It is the bottom hole flowing pressure, that is, the flowing pressure at the bottom of the production well, which is used to monitor reservoir energy. An abnormal decrease may indicate reservoir depletion or fluid breakthrough.

[0091] It should be noted that the observation operator is the core tool that connects model predictions with actual observations. Its functions include: (1) Spatial and physical quantity conversion: Its core function is to map the model's state vector to the observation space, enabling direct comparison between model predictions and actual observations, thereby calibrating model parameters and reducing prediction errors. (2) Data assimilation bridge: It supports the calculation of Kalman gain and calibrates model parameters. (3) Uncertainty management: It quantifies the uncertainty of predictions and observations through covariance transmission.

[0092] The Kalman gain is calculated. A high Kalman gain indicates greater confidence in the observed data and requires a significant revision of the model prediction. A low Kalman gain indicates greater confidence in the model prediction and requires a minor adjustment. It should be noted that the specific calculation process and principles of the Kalman gain are well known to those skilled in the art and will not be elaborated here.

[0093] After calculating the Kalman gain, each parameter and dynamic variable is dynamically corrected according to the Kalman gain.

[0094] Permeability field update: ;Formula (15) in, For grid cells j The updated permeability field, For grid cells j The permeability field before updating, is the adjustment of the permeability field, which is calculated from the Kalman gain and the observation residual, or generated by random perturbations, The variance of the adjustment amount of the permeability field is used to control the amplitude of the random disturbance. In the fracture area (permeability field>100mD), it can be increased , to enhance the flexibility of parameter adjustment.

[0095] Correction of the initial value of residual oil saturation: ;Formula (16) in, The corrected oil phase is at position j The residual oil saturation correction value, represents the number of set members, For the i The set members are at positions j The initial value of residual oil saturation.

[0096] For high confidence areas (greater than 0.9), the corrected value of remaining oil saturation is directly adopted, and for low confidence areas, the initial value of remaining oil saturation is retained.

[0097] It should be noted that the static parameters and dynamic variables mentioned above are only examples, and more may be included. In addition, the specific correction process and principle will not be described in detail here, and may be changed according to actual conditions.

[0098] In addition, it should be noted that the principle and function of the ensemble Kalman filter algorithm are existing technologies known to those skilled in the art, and the specific parameter update process and corresponding principles will not be explained in detail here.

[0099] In some embodiments, in the process of performing parameter correction on static parameters, initial values ​​of residual oil saturation, and initial values ​​of pressure gradient based on the ensemble Kalman filter algorithm, multi-GPU parallelization is adopted, and 100 state vectors are input in a single batch for processing, which can improve the model update speed.

[0100] In this embodiment, the trigger mechanism for parameter correction anomaly detection is: the deviation between observed and predicted data is greater than 5% for water cut in the production well or greater than 2% for bottomhole flowing pressure, and the confidence threshold is less than 0.85. The adaptive update frequency is: in normal mode, assimilation is performed every 5 minutes. In abnormal mode, the update is immediately triggered and switches to assimilation every minute.

[0101] In the embodiments of this application, the updated neural network weights based on the Ensemble Kalman Filter algorithm are synchronized in real time to the physical constraint neural network to ensure consistency in online learning. The updated permeability field serves as the input to the physical constraint neural network, affecting the calculation of flow velocity in Darcy's law.

[0102] In some embodiments of the present application, the static parameter includes a permeability field; and the method further includes: Divide each area of ​​the target layer into grids to obtain multiple grids; extracting a permeability field of a target grid, a residual oil saturation correction value, and a difference between the residual oil saturation correction values ​​of the target grid and an adjacent grid, where the target grid is one of the multiple grids; If the remaining oil saturation correction value of the target grid is greater than a preset first threshold, and the difference between the remaining oil saturation correction values ​​of the target grid and the adjacent grid is greater than a preset second threshold, the target grid is encrypted and divided into multiple subgrids, and each subgrid is interpolated according to the permeability field and the remaining oil saturation correction value of the target grid to recalibrate the physical constraint neural network model.

[0103] In some embodiments, the first threshold is 40%, and the second threshold is 15%. 1m 0.1m) is divided into 8 sub-grids (0.5m) 0.5m The permeability field is interpolated using log-linear interpolation, and the remaining oil saturation correction value is based on inverse distance weighted interpolation.

[0104] The dynamic adjustment process is: (1) Identify the infill area: Mark the high saturation area based on the real-time predicted distribution of the remaining oil saturation correction value.

[0105] (2) Grid splitting: Call the AMR (Adaptive Mesh Refinement) library (such as p4est) to generate sub-grids.

[0106] (3) Parameter transfer: interpolate the parent grid (target grid) parameters to the child grid and initialize the new set members.

[0107] (4) Model rebalancing: Increase the number of set members only in the encrypted area (e.g., from 100 to 120).

[0108] In some embodiments of the present application, a real-time visualization interface can also be used to output a dynamic model, annotating "permeability field, initial value of remaining oil saturation, and neural network weight."

[0109] This application utilizes cross-scale data temporal and spatial alignment and heterogeneous encoding to integrate macroscopic wide-area apparent resistivity, seismic, mesoscopic fiber optic, nanotracer, and microscopic CT data to address the "blind spot" problem in heterogeneous reservoirs. Wide-area apparent resistivity data offers high resolution, wide coverage, and low cost, significantly improving the accuracy and efficiency of remaining oil distribution monitoring while reducing costs.

[0110] This application adds a joint loss function in the feature projection layer to force different modal features to satisfy the same physical relationship and perform cross-modal physical consistency constraints.

[0111] This application combines hard constraints (Darcy's law and the law of conservation of mass) with soft constraints (loss function) to achieve both prediction accuracy and physical rationality.

[0112] Based on digital twins and ensemble Kalman filter algorithms, this application realizes real-time assimilation of oilfield parameters and second-level model correction, and realizes real-time closed-loop optimization of oilfield development strategies.

[0113] This application sets up an adaptive grid encryption and anomaly triggering mechanism, triggers model recalibration according to the confidence threshold, and dynamically refines the grid in high-saturation areas, thereby improving local prediction accuracy.

[0114] According to the remaining oil distribution prediction method of the present invention, a multi-method, cross-scale data fusion is performed using wide-area apparent resistivity data and commonly used remaining oil monitoring data. This method addresses the limited resolution, dynamic monitoring difficulties, and "blind spots" in heterogeneous reservoirs associated with traditional methods, while also addressing the timeliness issues of wide-area electromagnetic methods. The multi-scale fused feature vectors are then input into a physical constraint neural network to obtain initial remaining oil saturation and pressure gradient values ​​for each region of the target layer. The physical constraint neural network embeds physical laws (Darcy's law and the law of conservation of mass), combining hard constraints (physical laws directly embedded in the network) with soft constraints (loss function regularization) to balance data-driven and physical laws, improving model reliability. Finally, the real-time water cut and bottomhole flowing pressure of production wells are used as observation data, and the static parameters, initial remaining oil saturation, and initial pressure gradient values ​​from the Darcy's law physical constraint loss and mass conservation constraint loss are used as state vectors. These static parameters, initial remaining oil saturation, and initial pressure gradient values ​​are corrected using the ensemble Kalman filter algorithm, improving the accuracy and timeliness of remaining oil distribution prediction.

[0115] According to the remaining oil distribution prediction system of the second embodiment of the present application, the system includes a data acquisition unit, a multi-scale fusion feature vector determination unit, an initial prediction unit, a parameter correction unit, a correction parameter feedback unit and a remaining oil distribution result determination unit.

[0116] Data acquisition unit, used to obtain wide-area apparent resistivity data, seismic attribute data, fiber time series signal data, nano-tracer tracking data, pore network point cloud data, and real-time production well water cut and bottom hole pressure of the target layer; A multi-scale fusion feature vector determination unit is used to perform spatiotemporal alignment and feature fusion processing on wide-area apparent resistivity data, seismic attribute data, optical fiber time series signal data, nano-tracer tracking data, and pore network point cloud data to obtain a multi-scale fusion feature vector; The initial prediction unit is used to input the multi-scale fusion feature vector into the physical constraint neural network to obtain the initial values ​​of the remaining oil saturation and pressure gradient in each area of ​​the target layer. The loss function of the physical constraint neural network includes the data fitting term loss, Darcy's law physical constraint loss, and mass conservation constraint loss. A parameter correction unit is used to use the real-time water cut and bottom hole flowing pressure of the production well as observation data, and the static parameters in the physical constraint loss of Darcy's law and the mass conservation constraint loss, the initial value of the remaining oil saturation, and the initial value of the pressure gradient as state vectors, and perform parameter correction on the static parameters, the initial value of the remaining oil saturation, and the initial value of the pressure gradient based on the ensemble Kalman filter algorithm to obtain static parameter correction values, remaining oil saturation correction values, and pressure gradient correction values; A correction parameter feedback unit is used to feed back the static parameter correction value and the pressure gradient correction value to the physical constraint neural network to update the loss function; The remaining oil distribution result determination unit is used to determine the remaining oil distribution result of the target layer according to the remaining oil saturation correction value of each area of ​​the target layer.

[0117] According to an embodiment of the present invention, the remaining oil distribution prediction system utilizes multi-method, cross-scale data fusion of wide-area apparent resistivity data with commonly used remaining oil monitoring data. This solves the problems of traditional methods, such as insufficient resolution, difficulty in dynamic monitoring, and "blind spots" in heterogeneous reservoirs, while also addressing the timeliness issues of wide-area electromagnetic methods. The multi-scale fused feature vectors are then input into a physical constraint neural network to obtain initial remaining oil saturation and pressure gradient values ​​for each region of the target layer. The physical constraint neural network embeds physical laws (Darcy's law and the law of conservation of mass), combining hard constraints (physical laws directly embedded in the network) with soft constraints (loss function regularization), balancing data-driven and physical laws and improving model reliability. Finally, the real-time water cut and bottomhole flowing pressure of production wells are used as observation data, and the static parameters, initial remaining oil saturation, and initial pressure gradient values ​​from the Darcy's law physical constraint loss and mass conservation constraint loss are used as state vectors. These static parameters, initial remaining oil saturation, and initial pressure gradient values ​​are corrected using the ensemble Kalman filter algorithm, improving the accuracy and timeliness of remaining oil distribution prediction.

[0118] Since the remaining oil distribution prediction system adopts all the technical solutions of the remaining oil distribution prediction method of the above embodiment, it has at least all the beneficial effects brought by the technical solutions of the above embodiment, which will not be described in detail here.

[0119] In addition, an embodiment of the present application further provides a control device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor. The processor and the memory may be connected via a bus or other means.

[0120] The memory, as a non-transient computer-readable storage medium, can be used to store non-transient software programs and non-transient computer executable programs. In addition, the memory may include a high-speed random access memory and may also include a non-transient memory, such as at least one disk storage device, a flash memory device, or other non-transient solid-state storage device. In some embodiments, the memory may optionally include a memory remotely arranged relative to the processor, and these remote memories may be connected to the processor via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0121] The non-transient software program and instructions required to implement the remaining oil distribution prediction method of the above embodiment are stored in the memory, and when executed by the processor, the remaining oil distribution prediction method of the above embodiment is executed.

[0122] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, i.e., they may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment.

[0123] In addition, an embodiment of the present application also provides a computer-readable storage medium, which stores computer-executable instructions, and the computer-executable instructions are executed by a processor or controller, for example, by the processor of the above embodiment, so that the above processor can execute the remaining oil distribution prediction method in the above embodiment.

[0124] Those skilled in the art will appreciate that all or some of the steps and systems disclosed above can be implemented as software, firmware, hardware, or any suitable combination thereof. Some or all of the physical components may be implemented as software executed by a processor, such as a central processing unit, digital signal processor, or microprocessor, or as hardware, or as an integrated circuit, such as an application-specific integrated circuit. Such software may be distributed on computer-readable media, which may include computer storage media (or non-transitory media) and communication media (or transient media). As is well known to those skilled in the art, the term computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storing information, such as computer-readable instructions, data structures, program modules, or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technologies, CD-ROM, digital versatile disks (DVDs) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by a computer. Furthermore, as is well known to those skilled in the art, communication media typically embodies computer-readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave or other transport mechanism, and may include any information delivery media.

[0125] The embodiments of the present application are described in detail above in conjunction with the accompanying drawings, but the present application is not limited to the above embodiments. Various changes can be made within the scope of knowledge possessed by ordinary technicians in the relevant technical field without departing from the purpose of the present application.

Claims

1. A method for predicting residual oil distribution, characterized in that: The method comprises: Acquire wide-area apparent resistivity data, seismic attribute data, fiber optic time series signal data, nano-tracer tracking data, pore network point cloud data, and real-time production well water cut and bottomhole pressure of the target layer; performing spatiotemporal alignment processing and feature fusion processing on the wide-area apparent resistivity data, the seismic attribute data, the optical fiber time series signal data, the nanotracer tracking data, and the pore network point cloud data to obtain a multi-scale fusion feature vector; Inputting the multi-scale fusion feature vector into a physical constraint neural network to obtain initial values ​​of remaining oil saturation and pressure gradient in each region of the target layer, wherein the loss function of the physical constraint neural network includes data fitting term loss, Darcy's law physical constraint loss, and mass conservation constraint loss; Using the real-time water cut and bottom hole flowing pressure of the production well as observation data, the static parameters in the Darcy's law physical constraint loss and the mass conservation constraint loss, the initial value of the remaining oil saturation, and the initial value of the pressure gradient as state vectors, and performing parameter correction on the static parameters, the initial value of the remaining oil saturation, and the initial value of the pressure gradient based on an ensemble Kalman filter algorithm to obtain static parameter correction values, remaining oil saturation correction values, and pressure gradient correction values; Feeding back the static parameter correction value and the pressure gradient correction value to the physical constraint neural network to update the loss function; The remaining oil distribution result of the target layer is determined according to the remaining oil saturation correction value of each area of ​​the target layer.

2. The method for predicting residual oil distribution according to claim 1, wherein: The performing spatiotemporal alignment processing and feature fusion processing on the wide-area apparent resistivity data, the seismic attribute data, the optical fiber time series signal data, the nanotracer tracking data, and the pore network point cloud data to obtain a multi-scale fusion feature vector includes: performing feature extraction on the wide-area apparent resistivity data, the seismic attribute data, the optical fiber time series signal data, the nanotracer tracking data, and the pore network point cloud data to obtain wide-area apparent resistivity features, seismic attribute features, optical fiber time series signal features, nanotracer tracking features, and pore network point cloud features; Based on the feature projection method, a joint loss function is added to the feature projection layer to force different modal features to satisfy the same preset physical relationship. The wide-area apparent resistivity feature, the seismic attribute feature, the fiber time series signal feature, the nanotracer tracking feature, and the pore network point cloud feature are mapped into a unified feature space to obtain a unified feature of wide-area apparent resistivity, a unified feature of seismic attributes, a unified feature of fiber time series signals, a unified feature of nanotracer tracking, and a unified feature of pore network point cloud. The wide-area apparent resistivity unified feature, the seismic attribute unified feature, the optical fiber timing signal unified feature, the nanotracer tracking unified feature, and the pore network point cloud unified feature are weighted and fused to obtain the multi-scale fusion feature vector.

3. The method for predicting residual oil distribution according to claim 2, wherein: The seismic attribute data is feature extracted using a convolutional neural network, the optical fiber time series signal data is feature extracted using a long short-term memory network, and the pore network point cloud data is feature extracted using a graph neural network embedded with a seepage equation.

4. The method for predicting residual oil distribution according to claim 1, wherein: The Darcy's law physical constraint loss is obtained by the following steps: Obtaining a plurality of samples of observed remaining oil flow rates and a pressure gradient correction value predicted by the physical constraint neural network each time, wherein the observed remaining oil flow rates are determined by the optical fiber timing signal data or the nanotracer tracking data; Based on Darcy's law, the current remaining oil predicted flow rate is calculated according to the pressure gradient correction value obtained by the previous physical constraint neural network prediction, so as to obtain the remaining oil predicted flow rate of multiple samples; The Darcy's law physical constraint loss is determined based on the observed remaining oil flow rate and the predicted remaining oil flow rate of all samples.

5. The method for predicting residual oil distribution according to claim 4, wherein: The mass conservation constraint loss is obtained by the following steps: Obtaining porosity at multiple time steps, predicted remaining oil flow rate, and predicted remaining oil saturation value predicted by the physical constraint neural network, wherein the porosity is determined by the pore network point cloud data; Calculate the time change rate of the product of the porosity and the predicted value of the remaining oil saturation at each time step, and the divergence of the product of the predicted remaining oil flow rate and the predicted value of the remaining oil saturation at each time step; calculating the residuals of the time rate of change and the divergence in the time series; The residual of the time change rate and the divergence in the time series is used as the mass conservation constraint loss.

6. The method for predicting residual oil distribution according to claim 1, wherein: The constraint formula of the loss function of the physical constraint neural network is: ； ； ； ； in, is the loss function of the physical constraint neural network, is the data fitting loss, is the physical constraint loss of Darcy's law, is the mass conservation constraint loss, 、 and are the weights of the data fitting term loss, the Darcy's law physical constraint loss, and the mass conservation constraint loss, respectively. is the total number of samples, The physical constraint neural network is i The predicted value of remaining oil saturation of samples, For the i The true label value of the remaining oil saturation of samples, is the total number of samples, The physical constraint neural network is j The predicted flow rate of remaining oil for each sample, For the j The remaining oil observed flow rate of each sample is determined by the optical fiber timing signal data or the nanotracer tracking data. is the total number of time steps, is the porosity, The physical constraint neural network is k The predicted value of remaining oil saturation in time steps, The physical constraint neural network is k The remaining oil predicted flow rate for each time step.

7. The method for predicting residual oil distribution according to claim 1, wherein: The static parameters include a permeability field; the method further includes: Divide each area of ​​the target layer into grids to obtain multiple grids; extracting a permeability field of a target grid, the residual oil saturation correction value, and a difference between the residual oil saturation correction values ​​of the target grid and an adjacent grid, wherein the target grid is one of the plurality of grids; If the residual oil saturation correction value of the target grid is greater than a preset first threshold, and the difference between the residual oil saturation correction values ​​of the target grid and an adjacent grid is greater than a preset second threshold, the target grid is encrypted and divided into multiple subgrids, and each subgrid is interpolated according to the permeability field of the target grid and the residual oil saturation correction value to recalibrate the physical constraint neural network model.

8. A residual oil distribution prediction system, characterized in that: The system comprises: Data acquisition unit, used to obtain wide-area apparent resistivity data, seismic attribute data, fiber time series signal data, nano-tracer tracking data, pore network point cloud data, and real-time production well water cut and bottom hole pressure of the target layer; a multi-scale fusion feature vector determination unit, configured to perform spatiotemporal alignment processing and feature fusion processing on the wide-area apparent resistivity data, the seismic attribute data, the optical fiber time series signal data, the nanotracer tracking data, and the pore network point cloud data to obtain a multi-scale fusion feature vector; an initial prediction unit, configured to input the multi-scale fusion feature vector into a physical constraint neural network to obtain initial values ​​of remaining oil saturation and pressure gradient in each region of the target layer, wherein the loss function of the physical constraint neural network includes a data fitting term loss, a Darcy's law physical constraint loss, and a mass conservation constraint loss; a parameter correction unit, configured to use the real-time water cut and bottom hole flowing pressure of the production well as observation data, and the static parameters in the Darcy's law physical constraint loss and the mass conservation constraint loss, the initial value of the remaining oil saturation, and the initial value of the pressure gradient as state vectors, and perform parameter correction on the static parameters, the initial value of the remaining oil saturation, and the initial value of the pressure gradient based on an ensemble Kalman filter algorithm to obtain a static parameter correction value, a remaining oil saturation correction value, and a pressure gradient correction value; A correction parameter feedback unit, configured to feed back the static parameter correction value and the pressure gradient correction value to the physical constraint neural network for loss function update; The remaining oil distribution result determining unit is used to determine the remaining oil distribution result of the target layer according to the remaining oil saturation correction value of each area of ​​the target layer.

9. A control device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the remaining oil distribution prediction method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing computer-executable instructions, characterized in that: The computer-executable instructions are used to execute the remaining oil distribution prediction method according to any one of claims 1 to 7.

Images