A wellbore multiphase flow digital twin monitoring method and system

Abstract

This invention discloses a digital twin monitoring method and system for multiphase flow in wellbore, relating to the field of oil and gas field monitoring technology. The invention includes: establishing a coupled forward mechanism model of multiphase flow dynamics and fiber optic elastic-optical effect in wellbore; constructing a deep neural network with spatiotemporal coordinates as input and fluid state parameters as output; constructing a composite loss function including DAS observation data residuals, fluid dynamics control equation residuals, and boundary condition residuals; and using real-time DAS data stream to drive the network for iterative solution. By introducing fluid dynamics laws into the neural network, this invention solves the problems of difficult quantitative inversion and poor physical interpretability in traditional DAS monitoring. Even under pressurized conditions, complex flow patterns, or high-noise environments, it maintains extremely high prediction accuracy and physical consistency, achieving low-cost, full-well-section, and visualized dynamic monitoring of oil and gas well production. This significantly reduces reliance on expensive downhole logging instruments and improves the level of digital management in oilfields.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field monitoring technology, and more specifically, to a digital twin monitoring method and system for multiphase flow in wellbore. Background Technology

[0002] With the in-depth development of unconventional oil and gas resources, wellbore flow conditions are becoming increasingly complex. Real-time monitoring of multiphase flow parameters (such as phase flow rate, gas holdup, and flow pattern characteristics) is of decisive significance for optimizing production systems, adjusting lifting processes, and diagnosing downhole faults.

[0003] Currently, wellbore fluid profile monitoring primarily relies on Production Logging Tools (PLTs) or purely data-driven AI models. However, PLT operations are interventional, carrying the risk of downhole debris and incurring high costs. They only provide instantaneous fluid profiles and cannot capture transient flow changes. While surface multiphase flow meters can achieve continuous measurement, they cannot determine the dynamic distribution of fluid along the wellbore or the location of accumulated fluid, limiting their guidance for deep well lift processes. Furthermore, pure AI models often predict results that violate physical laws (such as mass non-conservation) when faced with unfamiliar formation conditions.

[0004] In recent years, distributed fiber optic acoustic sensing (DAS) technology has gradually become a research hotspot in wellbore monitoring due to its advantages such as high temperature and high pressure resistance, full wellbore coverage, and non-invasive monitoring. However, existing DAS multiphase flow interpretation techniques mainly have two limitations:

[0005] Physical methods based on feature engineering primarily rely on sound velocity tracking or energy extraction in specific frequency bands. These methods are typically based on simplified empirical formulas, neglecting the complex nonlinear characteristics of fluid dynamics. Their interpretation accuracy significantly decreases when dealing with slug flows, emulsion flows, or low-velocity conditions.

[0006] Purely data-driven deep learning methods directly establish the mapping relationship between the DAS spatiotemporal matrix and flow labels through architectures such as Convolutional Neural Networks (CNNs). These "black box" models lack physical interpretability, do not satisfy the laws of mass and momentum conservation, and heavily rely on massive amounts of labeled data. When actual downhole conditions (such as pressure, temperature, and fluid properties) exceed the coverage of the training set, the model's generalization ability is extremely poor, often resulting in predictions that contradict physical realities.

[0007] Therefore, there is an urgent need to explore a wellbore multiphase flow monitoring method that can make full use of DAS high spatiotemporal resolution data, strictly follow the physical conservation laws of fluid mechanics, and achieve high-precision and robust inversion under sparse labeled samples. Summary of the Invention

[0008] To address the shortcomings of existing technologies, this invention aims to provide a digital twin monitoring method for multiphase flow in wellbore systems. It seeks to solve the problems of poor physical interpretability, high dependence on large amounts of tag data, and low quantitative inversion accuracy under complex unsteady conditions in existing DAS multiphase flow monitoring technologies. This is achieved by constructing a "optical-fluid coupling" digital twin model, embedding the partial differential equations of multiphase flow dynamics and the fiber optic-elastic-optical effect mechanism as strong physical constraints into the loss function of the neural network. During data assimilation, the network output is forced to satisfy physical conservation laws. To achieve the above objectives, this invention provides the following technical solution:

[0009] A digital twin monitoring method for multiphase flow in wellbore, comprising the following steps:

[0010] S1. Establish a coupling forward mechanism model of the multiphase flow dynamics model of the wellbore and the fiber optic elastic-optic effect. The multiphase flow dynamics model is used to characterize the evolution of fluid state parameters (including phase flow velocity, pressure, and phase holding ratio) in the wellbore with time and well depth; the fiber optic elastic-optic effect model is used to characterize the mapping relationship between the dynamic pressure field of the pipe wall induced by fluid flow and the optical signal measured by the distributed fiber acoustic wave sensor (DAS).

[0011] The construction logic of the fiber response part of the coupling forward mechanism model is as follows: the turbulent pulsating pressure generated by the fluid in the wellbore during the flow process acts on the inner surface of the pipe wall, causing radial elastic deformation of the pipe string. This elastic deformation is transmitted to the fiber core through the mechanical coupling between the wellbore and the fiber sheath, causing a change in the axial strain rate of the fiber.

[0012] S2. Construct the Physical Information Neural Network (PINN) architecture. The network input layers use time coordinate t and well depth coordinate z, while the multiphase flow state variables serve as the output layers. Automatic differentiation techniques are used to calculate the partial derivatives of the output variables with respect to the input coordinates.

[0013] S3. Define a composite loss function L that includes multiple physical constraints. total To ensure that PINN's predictions both conform to the observed data and obey physical laws, a composite loss function L is defined. total Composite loss function L total L total L from DAS observation data data The residual L of the governing equations in fluid dynamics pde and the sparse point sensor boundary constraint term L bc Weighted composition.

[0014] S4. Pre-train the Physical Information Neural Network (PINN) based on historical production data, and then connect it to the real-time DAS data stream for inversion iteration, adjusting the composite loss function L. totalThe solution is minimized, the weights and bias parameters of the physical information neural network are updated, and the multiphase flow parameters of the entire well section are output.

[0015] A digital twin monitoring system for multiphase flow in a well is provided to implement the aforementioned digital twin monitoring method for multiphase flow in a well. The system includes: a distributed optical fiber sensing cable pre-installed in the well, an optical fiber demodulator on the ground, an edge computing server, and a visualization terminal. The edge computing server is equipped with a trained PINN architecture model.

[0016] Compared with the prior art, the present invention has the following beneficial effects:

[0017] This invention innovatively introduces a physical information neural network, transforming traditional "black box" deep learning into a "gray box" solution process with clear physical meaning. As long as the physical equations are correctly constructed, the network prediction results satisfy mass and momentum conservation, avoiding logical fallacies such as "water without a source" that may arise from purely data-driven models. By introducing an optical-fluid coupling mechanism, it is equivalent to deploying a virtual multiphase flow meter along every meter of the wellbore. This method significantly reduces the dependence on massive amounts of labeled training data, requiring only a small amount of wellhead or bottomhole pressure / temperature data as boundary constraints to achieve parameter inversion across the entire well section. It realizes low-cost, high-precision, and visualized digital twins of wellbore flow, solving the technical bottleneck of traditional methods in dealing with complex unsteady flows (such as slug flow and severe fluid accumulation). Compared to traditional deep learning, this invention ensures that the inversion results strictly adhere to mass and momentum conservation, eliminating non-physical prediction errors.

[0018] High-precision quantification: By constructing an acoustic-flow coupling mechanism, the qualitative acoustic waterfall diagram is transformed into a quantitative flow velocity and holdup curve, with the relative error controlled within 5%.

[0019] Strong generalization ability: Because it incorporates general fluid dynamics equations, the model has a strong adaptability to flow patterns or operating conditions not appearing in the training set.

[0020] Low-cost deployment: It utilizes only existing fiber optic resources and conventional wellhead sensors, eliminating the need for expensive downhole electronic instruments, and features virtual metering capabilities, which can replace some multiphase flow meters.

[0021] Visual and intuitive: It can generate fluid phase distribution maps of the entire well section in real time, intuitively displaying the location of accumulated fluid and the migration trajectory of slugs, assisting engineers in making quick decisions. Attached Figure Description

[0022] Figure 1 This is a schematic diagram illustrating the steps of a digital twin monitoring method for multiphase flow in a wellbore, as provided in an embodiment of the present invention.

[0023] Figure 2This is a schematic diagram illustrating the workflow of a digital twin monitoring method for multiphase flow in wellbore, provided in an embodiment of the present invention.

[0024] Figure 3 A schematic diagram of a digital twin monitoring method for multiphase flow in wellbore provided in an embodiment of the present invention;

[0025] Figure 4 This invention provides a schematic diagram of the coupling forward mechanism model association in a digital twin monitoring method for multiphase flow in wellbore.

[0026] Figure 5 This invention provides a curve comparing the inversion results with the actual values ​​in a digital twin monitoring method for multiphase flow in wellbore. Detailed Implementation

[0027] To make the above-mentioned objectives, features and advantages of the present invention more apparent and understandable, the technical solution of the present invention will be described in conjunction with the accompanying drawings, which realizes the real-time inversion of gas-liquid two-phase flow parameters of the entire well section without changing the downhole tubing structure.

[0028] Figure 1 and Figure 2 As shown, the present invention provides a digital twin monitoring method for multiphase flow in wellbore, comprising the following steps:

[0029] S1. Establish a forward coupling mechanism model between the multiphase flow dynamics model of the wellbore and the fiber optic elastic-optic effect.

[0030] Figure 4 The model presents the logical connections of the coupled forward mechanism, clearly demonstrating the transmission path from fluid dynamics to DAS signals. Fluid flow within the wellbore follows fluid dynamics laws, and the turbulent pulsating pressure generated by the fluid flow induces tubing deformation. This deformation is transmitted to the optical fiber through mechanical coupling, triggering the fiber optic elastic-optic effect, which causes a change in the axial strain rate of the optical fiber. This change in strain rate is converted into a detectable optical signal by the distributed fiber acoustic wave sensor (DAS), and this process is directly related to the rate of change of turbulent pulsating pressure over time.

[0031] S11. Constructing a multiphase flow dynamics model

[0032] This model is used to characterize the evolution of fluid state parameters within the wellbore over time and at well depth. This embodiment uses the drift-fluid model, which is suitable for gas-liquid two-phase flow. It considers the slippage effect between gas and liquid while avoiding the computational instability caused by too many closed relationships in the two-fluid model.

[0033] The core of the model includes a mixed mass conservation equation and a mixed momentum conservation equation, the simplified form of which is as follows:

[0034]

[0035]

[0036] Where t is the time coordinate; z is the well depth coordinate; For mixed density; For mixing flow rates; The fluid pressure inside the wellbore; It is the acceleration due to gravity; This is the friction term.

[0037] S12. Constructing a fiber optic-elastic effect model

[0038] This model is used to characterize the mapping relationship between the dynamic pressure field of the pipe wall induced by fluid flow and the optical signal measured by the distributed optical fiber acoustic sensor (DAS).

[0039] Turbulent pulsating pressure generated by fluid flow inside the wellbore Acting on the inner surface of the pipe wall, it causes radial elastic deformation of the tubing. Due to It is the fluid pressure inside the wellbore. The high-frequency fluctuation component in the fluid and the fact that the DAS system primarily responds to dynamic strain rate necessitate establishing a relationship between the time derivative of fluid pressure and fiber strain rate in the model. Based on the theory of thick-walled cylinders and the elastic-optical effect, its simplified coupling relationship is expressed as:

[0040]

[0041] in, This is the axial strain rate of the optical fiber, i.e., the direct observation value of DAS; The comprehensive elastic coupling coefficient is related to the Young's modulus, Poisson's ratio, tube diameter-to-wall thickness ratio, and shear modulus of the fiber coating. It refers to the turbulent pulsating pressure generated by the fluid flowing inside the pipe.

[0042] This relationship establishes a direct physical bridge between the invisible "pressure field changes inside the fluid" and the visible "DAS optical signal." This allows neural networks to inversely deduce the pressure wave propagation characteristics and flow pattern structure inside the fluid by observing the intensity and phase changes of the DAS signal.

[0043] S2. Constructing the PINN architecture for physical information neural networks.

[0044] Figure 2 and Figure 3As shown, a Physical Information Neural Network (PINN) architecture is constructed to replace the traditional black-box neural network. The network input layers use time coordinate t and well depth coordinate z, while the multiphase flow state variables serve as the output layers. Automatic differentiation techniques are used to calculate the partial derivatives of the output variables with respect to the input coordinates.

[0045] S21. Define network inputs and outputs.

[0046] like Figure 3 As shown, the input layer of the Physical Information Neural Network (PINN) architecture consists of normalized time coordinates t and well depth coordinates z; the output layer contains four neurons, corresponding to the dimensionless fluid state variables to be inverted: gas holdup α, gas phase velocity v. g Liquid phase velocity v l And the fluid pressure P inside the wellbore.

[0047] S22, Network Structure Initialization

[0048] Construct a fully connected feedforward neural network (FNN) with 5 hidden layers, each containing 50 neurons. The activation function chosen is either the Tanh or Swish function, as these functions are smooth and infinitely differentiable. This ensures that the network output will not exhibit vanishing or exploding gradients when automatically differentiating and calculating the second derivative, thus satisfying the mathematical requirements for solving partial differential equations in fluid mechanics.

[0049] S23. Configure the automatic differentiation module

[0050] Utilize the Automatic Differentiation module of deep learning frameworks (such as PyTorch or TensorFlow). In the computational graph, directly construct the partial derivative operators of the output variables with respect to the input variables (e.g., ...). , This step eliminates the need for manual derivation of the difference equations and allows for the accurate calculation of the derivative values ​​used for physical residual verification.

[0051] S3. Define a composite loss function that includes multiple physical constraints.

[0052] To ensure that PINN's predictions both conform to the observed data and obey physical laws, a composite loss function L is defined. total The composite loss function is a weighted sum of the residual terms from the DAS observation data, the residual terms from the fluid dynamics control equations, and the boundary constraint terms from the sparse point sensors.

[0053] S31. Define the residual term of DAS observation data.

[0054]

[0055] in, This is the measured DAS strain rate waterfall plot data; It utilizes the coupling model in step S1, based on the network output. The calculated predicted acoustic data must be such that the network output can reconstruct the observed acoustic signal.

[0056] S32. Define the residual terms of the fluid dynamics governing equations.

[0057] The network output of a, v g v l Substituting P and its automatic derivative into the mass and momentum conservation equations of the drift flow model, the residuals are calculated:

[0058]

[0059] Where, N col R represents the number of configuration points, i.e., the total number of data points randomly sampled in the spatiotemporal domain for calculating the residuals of partial differential equations; mass and R mom These are the residuals of the mass and momentum equations, respectively. Ideally, if the predictions perfectly conform to physical laws, this term should be 0.

[0060] S33. Define boundary condition constraints.

[0061]

[0062] Where, N bc P represents the number of boundary condition data points. gauge The measured value is the pressure gauge reading at the wellhead or bottom of the well; P pred This represents the network's output value at the corresponding location. This item utilizes high-precision point sensors to anchor the absolute pressure values ​​across the entire well section.

[0063] The final composite loss function is the weighted sum of the three: .

[0064] Where, λ data , λ pde , λ bc These are the weighting coefficients for each corresponding part. The term forces the model to fit the observed data. This term forces the model to obey the laws of physics. The method utilizes existing high-precision point sensor data at the wellhead / bottom to anchor absolute values. This multi-task constraint learning strategy greatly improves the robustness and uniqueness of the inversion results under complex operating conditions.

[0065] S4. Online inversion based on historical production data pre-training and real-time DAS data stream

[0066] This step is divided into three stages to ensure the convergence speed and accuracy of the model in practical applications.

[0067] S41, Historical Data Pre-training

[0068] First, PINN was pre-trained based on historical production data. Historical production records from the wells over the past three months (including wellhead / bottomhole pressure, temperature, and production volume data) were selected, and simulated data generated by a multiphase flow simulator (such as OLGA) were used as prior knowledge. Time and well depth were used as inputs, and fluid parameters were used as labels for initial training of PINN.

[0069] During training, the Adam optimizer (an adaptive moment estimation optimization algorithm suitable for rapid descent) was used for the first 5000 iterations, followed by the L-BFGS optimizer (a quasi-Newton method suitable for high-precision solutions) for the next 2000 iterations, until the loss function converged. This process enabled the network to initially grasp the basic physical laws of fluid distribution in the well.

[0070] S42, Real-time DAS Data Stream Online Inversion

[0071] The pre-trained PINN is connected to the real-time data stream: ① Data access: Real-time reception of DAS acoustic matrix data transmitted back from the downhole fiber optic cable (generating a new time slice every second); ② Iterative update: Every 10 seconds serves as a calculation window, based on the new DAS data D... DAS Calculate L data And combined with the residual L of the fluid dynamics control equation pde ③ Adaptive weight adjustment: For complex operating conditions, the system introduces an adaptive mechanism. For example, when slug flow (manifested as high-energy sloping stripes on an acoustic waterfall plot) is detected using the short-time energy characteristics of the DAS signal, the weight λ of the momentum equation residual is automatically increased. pde For severe transient flow conditions such as slug flow, the penalty force of physical constraints is enhanced to prevent overfitting of the neural network when processing high-frequency oscillating data, thus ensuring accurate capture of the position and velocity of the slug front.

[0072] S43. Results Output and Verification

[0073] When the composite loss function L total Converging to a preset threshold (e.g., 1×10) -5 When the inversion is complete, the spatiotemporal coordinates of the entire well section are input into PINN to output the multiphase flow parameter profile of the entire well section.

[0074] In addition, the present invention proposes a wellbore multiphase flow digital twin monitoring system to implement the above-mentioned wellbore multiphase flow digital twin monitoring method. The system hardware and software include: a distributed optical fiber sensing cable pre-installed in the wellbore, an optical fiber demodulator on the ground, an edge computing server and a visualization terminal. The edge computing server is equipped with the above-mentioned trained PINN inference model.

[0075] To verify the effectiveness of this embodiment, a production well in an oilfield is used as Example 1. The well has a vertical depth H = 2000 meters, an inner diameter of tubing D = 0.1 meters, and high-precision pressure and temperature sensors are installed at the wellhead and bottom. The bottom boundary condition is set as sinusoidal fluctuation in gas production to simulate periodic slug flow conditions. Figure 5 As shown in the figure, the gas holdup at a well depth of 1000 meters changes over time. The solid red line represents the gas holdup curve output by the PINN inversion method of this invention; the dashed black line represents the ground truth value from the numerical simulation. The results show that the two methods agree very well, with a mean absolute percentage error (MAPE) of less than 4.5%. In particular, during the period from t=200s to t=300s, significant slug bubbles were generated in the wellbore, and the gas holdup suddenly increased to over 0.8. Traditional pure data-driven models often cannot predict such extreme values ​​not found in the training set, while this invention, due to its superior performance, is more effective. The strong physical constraints (especially the mass conservation equation) "forced" the network to deduce the bubble aggregation process when it detects a change in pressure gradient, thus accurately capturing the slug characteristics.

[0076] Furthermore, the error between the virtual pressure profile obtained by inversion and the measured value of the bottom hole pressure gauge is only 0.2 MPa, proving that the system has the ability to replace the downhole permanent pressure gauge (PDG).

[0077] Example 2: Robustness test under different noise levels. Gaussian white noise of 10%, 20%, and 30% was artificially added to the original DAS data to test the model's noise resistance. Experiments show that, due to the "regularization" effect of the physical equations, even at a 30% noise level, the velocity profile retrieved by PINN remains smooth without significant oscillations, demonstrating the method's good adaptability to harsh noise environments in the field.

[0078] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A digital twin monitoring method for multiphase flow in wellbore, characterized in that, The method includes the following steps: S1. Establish a coupling forward mechanism model of multiphase flow dynamics model of wellbore and fiber optic elastic-optic effect. The multiphase flow dynamics model is used to characterize the evolution of fluid state parameters in the wellbore with time and well depth. The fiber optic elastic-optic effect model is used to characterize the mapping relationship between the dynamic pressure field of the pipe wall induced by fluid flow and the optical signal measured by distributed fiber acoustic wave sensor (DAS). S2. Construct the Physical Information Neural Network (PINN) architecture, using time coordinate t and well depth coordinate z as the network input layer, and multiphase flow state variables as the network output layer. Utilize automatic differentiation techniques to calculate the partial derivatives of the network output variables with respect to the input coordinates. S3. Define a composite loss function L that includes multiple physical constraints. total Composite loss function L total L from DAS observation data data The residual L of the governing equations in fluid dynamics pde and the sparse point sensor boundary constraint term L bc Weighted composition; S4. Pre-train the Physical Information Neural Network (PINN) based on historical production data, and then connect it to the real-time DAS data stream for inversion iteration, adjusting the composite loss function L. total Minimize the solution, update the weights and bias parameters of the physical information neural network, and output the multiphase flow parameters of the entire well section; Step S4 is divided into three stages: S41. Pre-train PINN based on historical production data. Select the historical production records of oil wells over the past 3 months, combine them with the simulation data generated by the multiphase flow simulator as prior knowledge, take time and well depth as input, and fluid parameters as labels to perform initial training on PINN. During training, the Adam optimizer is used for the first 5000 iterations, and then the L-BFGS optimizer is switched to the last 2000 iterations until the composite loss function converges. S42. Real-time DAS data stream online inversion: The pre-trained PINN is connected to the real-time data stream: ① Data access: Real-time reception of DAS acoustic matrix data transmitted back from the downhole fiber optic cable; ② Iterative update: Every 10 seconds serves as a calculation window, and L is calculated based on the new DAS data. data And combined with the residual L of the fluid dynamics control equation pde ③ Adaptive weight adjustment: For complex working conditions, the system introduces an adaptive mechanism. S43, when the composite loss function L total When the convergence reaches the preset threshold, the inversion is complete. At this point, the spatiotemporal coordinates of the entire well section are input into PINN, which outputs the multiphase flow parameter profile of the entire well section. The construction process of the PINN physical information neural network architecture described in S2 is as follows: S21. Define the network input and output. The input layer of the Physical Information Neural Network (PINN) architecture consists of normalized time coordinates t and well depth coordinates z. The output layer contains four neurons, corresponding to the dimensionless fluid state variables to be inverted: gas holdup α, gas phase velocity v. g Liquid phase velocity v l And the fluid pressure P inside the wellbore; S22. Network structure initialization: Construct a fully connected feedforward neural network (FNN) with 5 hidden layers and 50 neurons in each hidden layer; the activation function can be either Tanh or Swish. S23. Configure the automatic differentiation module. Using the automatic differentiation module of the deep learning framework, the partial derivative operator of the output variable with respect to the input variable is directly constructed in the computation graph.

2. The method for monitoring multiphase flow in a wellbore according to claim 1, characterized in that, The fluid state parameters in the wellbore described in S1 include phase flow velocity, pressure, and phase retention rate.

3. The digital twin monitoring method for multiphase flow in wellbore according to claim 1, characterized in that, The core of the multiphase flow dynamics model described in S1 includes the mixed mass conservation equation and the mixed momentum conservation equation, which are simplified as follows: ； ； Where t is the time coordinate; z is the well depth coordinate; For mixed density; For mixing flow rates; The fluid pressure inside the wellbore; It is the acceleration due to gravity; This is the friction term.

4. The method for monitoring multiphase flow in a wellbore according to claim 1, characterized in that, The simplified coupling relationship of the fiber elasto-optic effect model described in S1 is as follows: ； in, This is the axial strain rate of the optical fiber, i.e., the direct observation value of DAS; The comprehensive elastic coupling coefficient is related to the Young's modulus, Poisson's ratio, tube diameter-to-wall thickness ratio, and shear modulus of the fiber coating. It refers to the turbulent pulsating pressure generated by the fluid flowing inside the pipe.

5. The digital twin monitoring method for multiphase flow in a wellbore according to claim 1, characterized in that, The residual term L of the DAS observation data described in S3 data Defined as: ； in, This is the measured DAS strain rate waterfall plot data; It utilizes the coupling forward mechanism model in step S1, based on the network output. The calculated predicted acoustic wave data; The residual L of the fluid dynamics governing equation pde Defined as: ； Where, N col R represents the number of configuration points, i.e., the total number of data points randomly sampled in the spatiotemporal domain for calculating the residuals of partial differential equations; mass and R mom These are the residuals of the mass and momentum equations, respectively. Ideally, if the prediction perfectly conforms to the laws of physics, this term is 0. The sparse point sensor boundary constraint term L bc Defined as: ； Where, N bc P represents the number of boundary condition data points. gauge The measured value is the pressure gauge reading at the wellhead or bottom of the well; P pred This represents the network's output value at the corresponding location; The composite loss function is a weighted sum of the three factors: ； Where, λ data , λ pde , λ bc These are the weighting coefficients for each corresponding part; The term enables the model to fit the observed data. The term makes the model obey the laws of physics. The project utilizes existing high-precision point sensor data at the wellhead or bottom to anchor absolute values.

6. A digital twin monitoring system for multiphase flow in wellbore, characterized in that, The wellbore multiphase flow digital twin monitoring system is used to implement the wellbore multiphase flow digital twin monitoring method according to any one of claims 1-5.

7. A digital twin monitoring system for multiphase flow in a wellbore according to claim 6, characterized in that, The wellbore multiphase flow digital twin monitoring system includes: a distributed optical fiber sensing cable pre-installed in the wellbore, an optical fiber demodulator on the ground, an edge computing server and a visualization terminal, and a trained PINN architecture model deployed in the edge computing server.

Images