Method for determining viscous friction force model of drilling well drainage liquid after marine riser dissociation by fusing mechanism and data

Abstract

The invention belongs to the technical field of offshore oil and gas deepwater exploration, and provides a method for determining a viscous friction force model of drilling well drainage liquid after marine riser dissociation, and the method comprises the steps: S1, obtaining original data, comprise the pressure difference between the top and the bottom of the riser, the liquid column height, the cross section area of the inner wall of the riser, the flow velocity of liquid in the riser, the viscous friction force of discharged liquid of the inner wall of the riser and the like; s2, analyzing data features, and analyzing and screening Top-N key features of the data features; s3, processing and classifying the data, and executing a data feature standardization operation by using a Min-Max scaling normalization method; s4, determining and optimizing a data driving model structure, and selecting a random configuration neural network as a core tool; s5, training the model, and alternately executing network node growth and a Dropout regularization algorithm; and S6, testing the model, and verifying whether the Top-N key features on the test data are matched or not.

Description

Technical Field

[0001] The present invention belongs to the technical field of deepwater marine oil and gas exploration, and in particular relates to a method for determining a viscous friction force model of downhole fluid discharge after riser separation that integrates mechanism and data. Background Art

[0002] When disconnecting a deepwater riser after completion of a drilling operation or in an emergency during drilling, the viscous frictional resistance of the drilling fluid against the inner wall of the riser must be considered, as it directly impacts pump pressure requirements, circulation system design, and well control safety. Factors influencing friction include the height of the drilling fluid column, where friction loss is proportional to the fluid column volume, and the flow rate, where the frictional pressure drop is proportional to the square of the flow rate. Research on the mechanics of viscous frictional forces in drilling fluid drainage will enrich and advance deepwater drilling riser theory, improve deepwater drilling safety, and provide new insights into the multidisciplinary field of petroleum engineering and marine engineering in my country.

[0003] Methods for measuring or calculating friction mainly include experimental measurement, on-site monitoring, empirical formulas, and numerical simulation. Friction calculation methods that collect a large amount of data and analyze based on empirical formulas include:

[0004] (1) The Darcy-Weisbach equation calculation method is applicable to Newtonian fluids (such as water-based drilling fluids). The key parameters include friction coefficient, hydraulic diameter (the riser and the drill pipe annulus), drilling fluid density, and average flow rate, which are defined as follows:

[0005]

[0006] in:

[0007] ΔP is the pressure difference;

[0008] f is the Darcy friction coefficient, which is related to the Reynolds number and the pipe roughness;

[0009] L is the length of the pipeline;

[0010] D is the inner diameter of the pipe;

[0011] ρ is the liquid density;

[0012] v is the liquid flow rate.

[0013] (2) Non-Newtonian fluid correction (Herschel-Bulkley model) calculation method, applicable to high-viscosity drilling fluids (such as oil-based muds), key parameters include consistency coefficient, rheological index and yield stress.

[0014] (3) Computational Fluid Dynamics (CFD) simulation: This method uses CFD software (such as ANSYS Fluent and OpenFOAM) to simulate the flow of drilling fluid within the riser and calculate the friction pressure drop. Its advantages include the ability to analyze complex flow patterns (turbulence, secondary flow) and actual operating conditions such as drill pipe rotation and eccentric annuli. Its limitations include high computational cost and the need for accurate input of rheological parameters.

[0015] Numerical models come in various forms. The following uses the full liquid column model as an example. The full liquid column model is a simplified model used to analyze the dynamic behavior of liquid columns in vertical or inclined pipelines. It is primarily used in fields such as petroleum engineering, drilling fluid mechanics, and wellbore flow analysis. Its core assumption is to treat the entire liquid column as a continuous, incompressible fluid column, ignoring the complex interactions between the gas and liquid phases. It is suitable for single-phase flow or low-gas-holding liquid column systems.

[0016] The core of the full liquid column model is the hydrostatic equilibrium and momentum conservation. It is usually calculated by combining the total pressure balance relationship between the static liquid column pressure and the flow pressure drop. Its basic form is as follows:

[0017]

[0018] in:

[0019] f w and f m are the friction coefficients of seawater and mud, respectively;

[0020] ρ w and ρ m are their respective densities;

[0021] v f is the velocity of the fluid relative to the tensioner and its peripheral auxiliary devices;

[0022] a f is the acceleration of the fluid relative to the tensioner and its peripheral auxiliary devices;

[0023] l w is the height of the liquid column of the seawater portion;

[0024] l r is the total riser length;

[0025] A e is its hydraulic cross-sectional area, i.e. the difference between the outer ring area and the inner ring area;

[0026] d r It is its hydraulic diameter, which is the sum of the circumference of the outer ring and the circumference of the inner ring.

[0027] Different numerical models differ significantly in their applicable conditions, complexity, and considerations. The full liquid column model is suitable for single-phase liquid columns and low gas fraction scenarios, with relatively low complexity, primarily considering hydrostatic pressure and friction losses. The slug flow model, targeting intermittent gas-liquid two-phase flows, has moderate to high complexity, requiring a focus on analyzing gas-liquid interactions and the dynamic characteristics of the slug unit. The two-fluid model, suitable for high gas fractions and complex multiphase flows, is the most complex and requires detailed simulation of complex mechanisms such as interphase momentum exchange and interfacial effects. Each model should be selected based on the actual flow characteristics and accuracy requirements.

[0028] It is important to note that traditional modeling methods for the viscous friction of fluid drainage during riser drilling rely on physical laws or empirical formulas, and the large sample size of data leads to difficulties in parameter calibration and poor adaptability. Combined with existing mechanism analysis, data-driven modeling can significantly improve model generalization by exploiting inherent patterns in the data. However, the limited available public data on fluid drainage during riser drilling and the small sample size make large-scale model learning difficult. Furthermore, traditional neural network learning models are prone to overfitting in small sample environments. Existing modeling methods based on random configuration networks can be applied to dynamic modeling problems with limited time series data. Dropout technology, combined with its incremental learning properties, can alleviate the overfitting problem. It is noted that while the dynamic time series modeling method based on random configuration networks and the incremental learning properties of dropout technology offer significant advantages in their respective fields, there has been no systematic study on their combined application to modeling the viscous friction of fluid drainage during riser drilling. Summary of the Invention

[0029] In order to overcome the reliance of traditional modeling methods for the viscous friction resistance of drainage fluid during riser drilling on prior knowledge and to address the accuracy issues of friction force data analysis models under mechanism fusion data and small sample conditions, the present invention proposes a data-driven modeling method for the viscous friction force of drainage fluid during riser decomposition after the riser is decomposed, which integrates the mechanism and data. The technical solution adopted to achieve the above purpose includes the following steps:

[0030] Step S1: Acquire raw data, including sample data of the mechanism model and sensor measurement data, including the pressure difference between the top and bottom of the riser, the height of the liquid column, the cross-sectional area of ​​the inner wall of the pipe, the flow rate of the liquid in the pipe, and the viscous friction of the discharged liquid on the inner wall of the pipe;

[0031] Step S2: Analyze data features. Based on the riser drilling drain pressure differential, liquid column height, flow rate, and viscous friction data, analyze and select the top-N key features. The key features include the peak frequency and time of the friction between the liquid in the pipe and the pipe wall, and the specific time required for the entire pipe liquid to be completely discharged from the bottom.

[0032] Step S3: Process and classify the data, using the Min-Max scaling normalization method to perform data feature standardization operations, which serve as the input and output data of the riser drilling drainage fluid viscous friction data-driven model that integrates the mechanism and data. The data includes the pressure difference between the top and bottom of the riser, the liquid column height, the flow rate of the liquid in the pipe, and the drainage fluid viscous friction on the inner wall of the riser. The input data selection range is the three types of data: liquid column height, flow rate, and friction, and the output data is friction type data;

[0033] Step S4: Determine and optimize the data-driven model structure, and establish a data-driven model that describes the viscous friction force of the fluid draining down the inner wall of the riser. The method for determining the data-driven model that describes the viscous friction force of the fluid draining down the inner wall of the riser includes: fusing multiple mechanism model data and measurement data, selecting a randomly configured neural network as a core tool, and introducing the Dropout algorithm in deep learning methods on the fully connected layer structure of a traditional neural network to determine and optimize the model structure;

[0034] The steps of establishing a data-driven model to describe the viscous friction force of the drainage fluid on the inner wall of the riser and determining and optimizing the model structure include:

[0035] Step S401: Encode data. Encode the sample data and measurement data set of each model as a training set, and reserve a portion as a test data set. The data includes the pressure difference between the top and bottom of the riser, the height of the liquid column, the flow rate of the liquid in the riser, and the viscous friction of the drainage liquid on the inner wall of the riser.

[0036] Step S402: Determine the model architecture, select a randomly configured network, gradually add hidden nodes using an incremental construction method, and optimize the number of randomly configured nodes through a supervision mechanism to determine the model structure;

[0037] Step S403: Optimizing the model structure: introducing the Dropout algorithm in the process of determining the random configuration network model structure, randomly discarding hidden layer nodes, and achieving dynamic sparsification of the network structure for optimizing the model structure;

[0038] Step S5: Train the model. A hybrid training strategy is used to train the data-driven model describing the viscous friction of the drainage fluid on the inner wall of the riser. The network node growth and Dropout regularization algorithms are performed alternately. The settings are: training data x, y, Dropout probability p, randomly configured network constraint parameter λ, and the maximum number of hidden layer nodes L. max ;

[0039] The random configuration network model using the Dropout algorithm outputs data y, and the mathematical expression of its L-node network structure is: Where x is the input data, y is the output data, L is the number of hidden layer nodes, and w l and b lare the randomly generated weights and biases for the lth node, β l is the output weight, usually obtained by the least squares method, g() is the activation function, usually the Sigmoid function can be selected, p is the Dropout probability, p l is the Dropout mask;

[0040] The training steps of the data-driven model of viscous friction force of the drainage fluid on the inner wall of the riser by alternately executing the network node growth and Dropout regularization algorithms include:

[0041] Step S501: Initialize the number of nodes L = 0, f0(x) = 0, e0 = y, and define the residual error: e L =yf L (x);

[0042] Step S502: incrementally adding nodes and randomly generating candidate node parameters, discarding some nodes until the stopping condition is met;

[0043] For each newly added L+1 node, a random weight and bias w is generated L+1 ~U(-a,a),b L+1 ~U(-b,b), satisfies uniform sampling or Gaussian distribution, where a and b are given constants;

[0044] Calculate candidate hidden layer node output Dynamically generate a random Dropout mask for each node, uniformly sample to generate p L+1 ~U(0,1), after the comparison operation, it is assigned 1 or 0, which is True or False respectively, that is, p L+1 =p L+1 >p;

[0045] Hidden layer node output after updating Dropout:

[0046] Step S503: Calculate the random configuration network output weights where β=[β1…β l ] is the hidden layer output matrix, h L The pseudo-inverse of (x), h L+1 =[h1…h L+1 ], y is the target output;

[0047] Step S504: Update the network and residual, add new nodes, and update the network output: f L+1 (x) = f L (x)+β L+ 1h L+1 (x), calculate the new residual: e L+1=yf L+1 (X);

[0048] Step S505: Check the stopping condition: If the training meets the error condition Where λ∈(0,1) is the preset random configuration network constraint parameter, then the training is stopped; or the network structure reaches the maximum number of nodes L during the training process max , then stop training; otherwise continue to add nodes;

[0049] Step S6: Test the model to test the trained data-driven model describing the viscous friction of the drainage fluid on the inner wall of the riser, and verify whether the Top-N key features on the test data are consistent.

[0050] The present invention improves the accuracy of modeling by combining a random configuration network with a dynamic Dropout mechanism, and can be applied to the data modeling process of viscous friction force of drainage fluid in riser drilling. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 : Random configuration network diagram

[0052] Figure 2 Flowchart of data-driven modeling of viscous friction of downhole fluid drainage integrating mechanism and data

[0053] Figure 3 :Schematic diagram of the determination and optimization process of the data-driven model of the viscous friction force of the drainage fluid

[0054] Figure 4 : Schematic diagram of the training process of the data-driven model of viscous friction of drainage fluid

[0055] Figure 5 : Model performance comparison curve

[0056] Figure 6 : Accuracy of the drainage fluid viscous friction data driven model of Example 1

[0057] Figure 7 : Accuracy of the drainage fluid viscous friction data driven model of Example 2 DETAILED DESCRIPTION

[0058] In order to clearly and completely describe the technical solutions and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Note that the specific technical solutions described in the embodiments herein are only used to explain the present invention and are not intended to limit the present invention. Based on the embodiments of the present invention, embodiments obtained by ordinary technicians in this field using the same inventive concept of the present invention should be included in the scope of protection of the present invention.

[0059] In the context of existing engineering practice and related research, combined with the Darcy-Weisbach equation,

[0060]

[0061] and the full liquid column model equation,

[0062]

[0063] According to the public data set of the paper, the model environment parameters and initial conditions are set, including water depth and total height of liquid column l r =1000m, riser hydrodynamic diameter d r =0.4826m, the friction coefficients of seawater and mud are f w =0.002,f m =0.004, the initial conditions of velocity and acceleration are 0, and the simulation results of the simulation data model on the length of the liquid column and the friction force are as follows: the peak time is about 14s, and the peak friction force is about 1.89×10 5 N, the time required for the equivalent whole-pipe drilling fluid to flow down from the bottom is about 51 seconds.

[0064] Note that after the riser is dissociated, seawater rushes into the riser body from the valve at the top of the riser, and the mixed mud flows down and into the seawater from the bottom of the riser. In order to fully analyze the mechanistic characteristics of the viscous friction force of the drilling fluid during this process, the existing mechanism model and data analysis model have the following problems:

[0065] Analysis of the viscous friction data for drilling fluid drainage obtained using the full-fluid-column model shows that after the mud drainage process is complete, the mud column height, velocity, and acceleration are forced to zero, meaning that the friction data after the 51-second drainage is disregarded. However, given the presence of seawater within the riser, the velocity and height of the fluid column are clearly not zero. Therefore, the full-fluid-column model for the viscous friction of drilling fluid drainage has some unmodeled mechanisms after drainage, which require urgent analysis and improvement of the model's accuracy.

[0066] Analysis of the viscous friction exponential function model for drilling fluid drainage, derived through model identification using a publicly available dataset, indicates that the external structural system can address the practical issue of non-zero friction data after drainage. However, this model is based entirely on structural variables and fails to capture drainage fluid characteristics, such as the changes in the length and velocity of the fluid column of a mixed mud and seawater drainage system. This suggests that this type of friction time series model contains some unmodeled mechanisms, and its accuracy needs to be improved.

[0067] To address these issues and address the limited availability of publicly available data on riser drainage fluids, resulting in small sample sizes and the resulting difficulty in implementing large-scale model learning, this paper proposes a method for determining a viscous friction model for drainage fluids after riser decoupling by integrating mechanisms and data. This method describes the relationship between three types of drainage fluid data: liquid column height, flow rate, and viscous friction. The following examples illustrate this method; practical applications are not limited to these data.

[0068] Figure 1 It is a randomly configured network diagram, in which the data required by the network include the pressure difference between the top and bottom of the watertight pipe, the liquid column height, the flow rate of the liquid in the pipe, and the viscous friction of the discharged liquid on the inner wall of the watertight pipe. The input data selection range is three types of data: liquid column height, flow rate, and friction force, and the output data is friction force type data.

[0069] Figure 2 The data-driven modeling flow chart of the viscous friction force of downhole drainage that integrates the mechanism and data is shown in the figure. Figure 2 To verify the effects of the present invention, the present invention provides the following examples.

[0070] Example 1

[0071] To verify the effect of the present invention, this embodiment uses full liquid column model simulation data for testing, and the specific steps are as follows:

[0072] Step S1: Obtain sample data for multiple mechanism models, including the pressure difference between the top and bottom of the riser, the height of the liquid column, the cross-sectional area of ​​the inner wall of the riser, the flow rate of the liquid in the riser, and the viscous friction of the discharged liquid on the inner wall of the riser;

[0073] Step S2: Analyze data features. Based on the riser drilling drain pressure differential, liquid column height, flow rate, and viscous friction data, analyze and select the top-N key features. The key features include the peak frequency and time of the friction between the liquid in the pipe and the pipe wall, and the specific time required for the entire pipe liquid to be completely discharged from the bottom.

[0074] Step S3: Process and classify the data, using the Min-Max scaling normalization method to perform data feature standardization operations, and use the data as the input and output data of the riser drilling drainage fluid viscous friction data-driven model that integrates the mechanism and data. The data includes the liquid column height, the flow rate of the liquid in the pipe, and the drainage fluid viscous friction force on the inner wall of the riser obtained by combining the pressure difference between the top and bottom of the riser. The input data is selected as two types of data: liquid column height and flow rate, and the output data is friction type data;

[0075] Step S4: Determine and optimize the data-driven model structure, and establish a data-driven model that describes the viscous friction force of the fluid draining down the inner wall of the riser. The method for determining the data-driven model that describes the viscous friction force of the fluid draining down the inner wall of the riser includes: fusing multiple mechanism model data, selecting a randomly configured neural network as a core tool, and introducing the Dropout algorithm in deep learning methods on the fully connected layer structure of a traditional neural network to determine and optimize the model structure.

[0076] Step S5: Train the model. A hybrid training strategy is used to train the data-driven model describing the viscous friction of the drainage fluid on the inner wall of the riser. The network node growth and Dropout regularization algorithms are performed alternately. The following settings are used:

[0077] In the training data, the input data x is the liquid column height and flow rate, the output data y is the viscous friction data of the discharge liquid, the Dropout probability p is 0.2, the random configuration network constraint parameter λ is 0.1, and the maximum number of hidden layer nodes L is 0. max is 50;

[0078] Alternately execute the network node growth and Dropout regularization algorithms to train a data-driven model of the viscous friction force of the drainage fluid on the inner wall of the riser;

[0079] Step S6: Test the model, test the trained data-driven model describing the viscous friction of the liquid discharged from the inner wall of the riser, and verify whether the Top-N key features on the test data are consistent. The key features include the peak frequency and time of the friction between the liquid in the pipe and the pipe wall, and the specific time required for the entire pipe liquid to be discharged from the bottom.

[0080] like Figure 3 As shown, the steps of establishing a data-driven model to describe the viscous friction force of the drainage fluid on the inner wall of the riser and determining and optimizing the model structure include:

[0081] Step S401: Encode data. Encode the sample data and measurement data set of each model as a training set, and reserve a portion as a test data set. The data includes the pressure difference between the top and bottom of the riser, the height of the liquid column, the flow rate of the liquid in the riser, and the viscous friction of the drainage liquid on the inner wall of the riser.

[0082] Step S402: Determine the model architecture, select a randomly configured network, gradually add hidden nodes using an incremental construction method, and optimize the number of randomly configured nodes through a supervision mechanism to determine the model structure;

[0083] Step S403: Optimizing the model structure. In the process of determining the random configuration network model structure, the Dropout algorithm is introduced to randomly discard hidden layer nodes to achieve dynamic sparsification of the network structure for optimizing the model structure.

[0084] like Figure 4 As shown, the training steps of the data-driven model of viscous friction force of the drainage fluid on the inner wall of the riser by alternately executing the network node growth and Dropout regularization algorithms include:

[0085] Step S501: Initialize the number of nodes L = 0, f0(x) = 0, e0 = y, and define the residual error: e L =yf L (x);

[0086] Step S502: incrementally adding nodes and randomly generating candidate node parameters, discarding some nodes until the stopping condition is met;

[0087] For each newly added L+1 node, a random weight and bias w is generated L+1 ~U(-a,a),b L+1 ~U(-b,b), satisfies uniform sampling or Gaussian distribution, where a and b are given constants;

[0088] Calculate candidate hidden layer node output Dynamically generate a random Dropout mask for each node, uniformly sample to generate p L+1 ~U(0,1), after the comparison operation, it is assigned 1 or 0, which is True or False respectively, that is, p L+1 =p L+1 >p;

[0089] Hidden layer node output after updating Dropout:

[0090] Step S503: Calculate the random configuration network output weights where β=[β1…β l ] is the hidden layer output matrix, h L The pseudo-inverse of (x), h L+1 =[h1…h L+1 ], y is the target output;

[0091] Step S504: Update the network and residual, add new nodes, and update the network output: f L+1 (x) = f L (x)+β L+ 1h L+1 (x), calculate the new residual: e L+1 =yf L+1 (X);

[0092] Step S505: Check the stopping condition: If the training meets the error condition Then stop training; or the network structure reaches the maximum number of nodes L during training max, then stop training; otherwise continue to add nodes;

[0093] Figure 5 The model performance comparison curve shows the data of three mechanism models. The training data comes only from model 1, and the test data comes from models 2 and 3. Figure 5 As shown in the figure, compared with the training model, the output results of the data-driven model have a difference of about 0.1s in the time to reach the peak frequency and the time to complete the drilling fluid leakage. The modeling results of the Top-N key features meet the requirements, have good generalization effect, and effectively reduce the amount of data.

[0094] Figure 6 The mean square error (MSE) results of the data-driven model of the viscous friction force of the drainage fluid in this embodiment are shown. Figure 6 As shown in the figure, after adding the Dropout algorithm, the model accuracy is improved, and the accuracy of modeling is increased.

[0095] Example 2

[0096] To verify the effect of the present invention, this embodiment uses the time series data of the viscous friction force of the drainage fluid to conduct a test. The specific steps are as follows:

[0097] Step S1: Obtain sample data, time series data of the viscous friction force of the drainage fluid on the inner wall of the riser;

[0098] Step S2: Analyze the data features and select the top-N key features, including the peak frequency and time of the friction between the liquid in the tube and the tube wall, and the specific time required for the entire tube liquid to be discharged from the bottom;

[0099] Step S3: Process and classify the data, perform data feature normalization using the Min-Max scaling normalization method, and use the data as input and output data for a data-driven model of viscous friction force of drainage fluid in riser drilling that integrates the mechanism and data. The input and output data are both friction time series data.

[0100] Step S4: Determine and optimize the data-driven model structure to establish a data-driven model describing the viscous friction of the fluid draining down the inner wall of the riser. The method for determining the data-driven model describing the viscous friction of the fluid draining down the inner wall of the riser includes: selecting a randomly configured neural network as a core tool, and introducing a Dropout algorithm from a deep learning method into the fully connected layer structure of a traditional neural network to determine and optimize the model structure.

[0101] Step S5: Train the model. A hybrid training strategy is used to train the data-driven model describing the viscous friction of the drainage fluid on the inner wall of the riser. The network node growth and Dropout regularization algorithms are performed alternately. The following settings are used:

[0102] In the training data, the input data x is the liquid column height and flow rate, the output data y is the viscous friction data of the discharge liquid, the Dropout probability p is selected as 0.2 and 0.5 respectively, the random configuration network constraint parameter λ is 0.1, and the maximum number of hidden layer nodes L is max is 50;

[0103] Alternately execute the network node growth and Dropout regularization algorithms to train a data-driven model of the viscous friction force of the drainage fluid on the inner wall of the riser;

[0104] Step S6: Test the model, test the trained data-driven model describing the viscous friction of the liquid discharged from the inner wall of the riser, and verify whether the Top-N key features on the test data are consistent. The key features include the peak frequency and time of the friction between the liquid in the pipe and the pipe wall, and the specific time required for the entire pipe liquid to be discharged from the bottom.

[0105] like Figure 7 As shown in FIG, after adding the Dropout algorithm, the accuracy of the drainage fluid viscous friction data-driven model of Example 2 is improved, which effectively improves the accuracy of modeling on time series data.

[0106] Although embodiments of the present invention have been described, it will be apparent to those skilled in the art that modifications and substitutions may be made to these embodiments without departing from the principles of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for determining a viscous friction model for downhole fluid discharge after riser separation that integrates mechanism and data, characterized in that: The following steps are involved: Step S1: Acquire raw data, including sample data of the mechanism model and sensor measurement data, including the pressure difference between the top and bottom of the riser, the height of the liquid column, the cross-sectional area of ​​the inner wall of the pipe, the flow rate of the liquid in the pipe, and the viscous friction of the discharged liquid on the inner wall of the pipe; Step S2: Analyze data features. Based on the riser drilling drain pressure differential, liquid column height, flow rate, and viscous friction data, analyze and select the top-N key features. The key features include the peak frequency and time of the friction between the liquid in the pipe and the pipe wall, and the specific time required for the entire pipe liquid to be completely discharged from the bottom. Step S3: Process and classify the data, using the Min-Max scaling normalization method to perform data feature standardization operations, and use the data as the input and output data of the riser drilling drainage fluid viscous friction data-driven model that integrates the mechanism and data. The data includes the pressure difference between the top and bottom of the riser, the liquid column height, the flow rate of the liquid in the pipe, and the drainage fluid viscous friction of the riser inner wall. The input data selection range is the three types of data: liquid column height, flow rate, and friction, and the output data is friction type data; Step S4: Determine and optimize the data-driven model structure, and establish a data-driven model that describes the viscous friction force of the fluid draining down the inner wall of the riser. The method for determining the data-driven model that describes the viscous friction force of the fluid draining down the inner wall of the riser includes: fusing multiple mechanism model data and measurement data, selecting a randomly configured neural network as a core tool, and introducing the Dropout algorithm in deep learning methods on the fully connected layer structure of a traditional neural network to determine and optimize the model structure; Step S5: training the model by adopting a hybrid training strategy to train a data-driven model describing the viscous friction of the drainage fluid on the inner wall of the riser, alternatingly executing the network node growth and Dropout regularization algorithms; Step S6: Test the model to test the trained data-driven model describing the viscous friction of the drainage fluid on the inner wall of the riser, and verify whether the Top-N key features on the test data are consistent.

2. The method for determining the viscous friction force model of downhole drainage after riser separation according to the fusion mechanism and data of claim 1 is characterized in that: The steps of establishing a data-driven model to describe the viscous friction force of the drainage fluid on the inner wall of the riser in step S4 and determining and optimizing the model structure include: Step S401: Encode data. Encode the sample data and measurement data set of each model as a training set, and reserve a portion as a test data set. The data includes the pressure difference between the top and bottom of the riser, the height of the liquid column, the flow rate of the liquid in the riser, and the viscous friction of the drainage liquid on the inner wall of the riser. Step S402: Determine the model architecture, select a randomly configured network, gradually add hidden nodes using an incremental construction method, and optimize the number of randomly configured nodes through a supervision mechanism to determine the model structure; Step S403: Optimizing the model structure. In the process of determining the random configuration network model structure, the Dropout algorithm is introduced to randomly discard hidden layer nodes to achieve dynamic sparsification of the network structure for optimizing the model structure.

3. The method for determining the viscous friction force model of downhole drainage after riser separation according to the fusion mechanism and data of claim 1 is characterized in that: Step S5: Train the model. A hybrid training strategy is used to train the data-driven model describing the viscous friction of the drainage fluid on the inner wall of the riser. The network node growth and Dropout regularization algorithms are performed alternately. The settings are: training data x, y, Dropout probability p, randomly configured network constraint parameter λ, and the maximum number of hidden layer nodes L. max ; The random configuration network model using the Dropout algorithm outputs data y, and the mathematical expression of its L-node network structure is: Where x is the input data, y is the output data, L is the number of hidden layer nodes, and w l and b l are the randomly generated weights and biases for the lth node, β l is the output weight, usually obtained by the least squares method, g() is the activation function, usually the Sigmoid function can be selected, p is the Dropout probability, p l is the Dropout mask.

4. The method for determining the viscous friction force model of downhole drainage after riser separation according to claim 3, characterized in that: The training steps of the data-driven model of viscous friction force of the drainage fluid on the inner wall of the riser by alternately executing the network node growth and Dropout regularization algorithms include: Step S501: Initialize the number of nodes L = 0, f0(x) = 0, e0 = y, and define the residual error: e L =yf L (x); Step S502: incrementally adding nodes and randomly generating candidate node parameters, discarding some nodes until the stopping condition is met; Step S503: Calculate the random configuration network output weights where β=[β1…β l ] is the hidden layer output matrix, h L The pseudo-inverse of (x), h L+1 =[h1 … h L+1 ], y is the target output; Step S504: Update the network and residual, add new nodes, and update the network output: f L+1 (x) = f L (x)+β L+1 h L+1 (x), calculate the new residual: e L+1 =yf L+1 (X); Step S505: Check the stopping condition: If the training meets the error condition Where λ∈(0,1) is the preset random configuration network constraint parameter, then the training is stopped; or the network structure reaches the maximum number of nodes L during the training process max , then stop training; otherwise, continue adding nodes.

5. The method for determining the viscous friction force model of downhole drainage after riser separation according to the fusion mechanism and data of claim 4 is characterized in that: Step S502: incrementally adding nodes and randomly generating candidate node parameters, discarding some nodes until the stopping condition is met; For each newly added L+1 node, a random weight and bias w is generated L+1 ~U(-a,a),b L+1 ~U(-b,b), satisfies uniform sampling or Gaussian distribution, where a and b are given constants; Calculate candidate hidden layer node output Dynamically generate a random Dropout mask for each node, uniformly sample to generate p L+1 ~U(0,1), after the comparison operation, it is assigned 1 or 0, which is True or False respectively, that is, p L+1 =p L+1 >p; Hidden layer node output after updating Dropout:

Images