Physical-data cooperative driving hook load torque prediction method

Abstract

The invention discloses a physical-data cooperative driving hook load torque prediction method. The method comprises the following steps: S1, preparing physical constraint data and data constraint data; s2, constructing a PINN model, wherein the PINN model comprises a physical constraint item, a data constraint item and a loss function; the physical constraint item is a drill string dynamic model, and the data constraint item is a CNN-LSTM-Attention multi-dimensional sequential network model; s3, the physical constraint data and the data constraint data are adopted to train the PINN model, and a trained PINN model is obtained; and S4, predicting the load and torque of the hook by adopting the trained PINN model. According to the method, the theoretical basis of a physical model and the nonlinear fitting ability of machine learning can be organically combined, and besides data driving, the generalization ability and prediction accuracy of the model are enhanced by introducing physical constraint conditions.

Description

Technical Field

[0001] The present invention relates to the technical field of oil and gas field development, and in particular to a physics-data collaboratively driven hook load torque prediction method. Background Art

[0002] As oil and gas extraction continues to increase in depth, drilling technology faces increasingly challenging challenges. One of the main challenges facing traditional horizontal well drilling methods is excessive drill string friction and torque, which often leads to problems such as stuck drillstring and buckling, seriously impacting the transmission of weight-on-bit (WOB) and drilling safety. Accurately predicting drill string friction and torque can not only prevent complex downhole incidents such as stuck drillstring and buckling, but also optimize drilling parameters and improve drilling efficiency.

[0003] While drill string friction torque theoretical models can predict key parameters in oscillatory drilling, such as bottomhole weight on bit and tool face angle, they often face challenges in solving complex downhole conditions, including computational complexity, time-consuming calculations, and model parameter sensitivity. This is particularly true in complex conditions such as extended reach and deep wells, significantly reducing model solution efficiency. While machine learning models can efficiently process large amounts of data, purely data-driven models often have poor generalization capabilities when faced with unseen conditions, making it difficult to ensure stable and accurate predictions. Summary of the Invention

[0004] In view of the above problems, the present invention aims to provide a physics-data collaborative driven hook load torque prediction method.

[0005] The technical solutions of the present invention are as follows:

[0006] A physics-data collaboratively driven hook load torque prediction method includes the following steps:

[0007] S1: Prepare physical constraint data and data constraint data;

[0008] S2: Constructing a PINN model, wherein the PINN model includes a physical constraint item, a data constraint item, and a loss function; the physical constraint item is a drill string dynamics model, and the data constraint item is a CNN-LSTM-Attention multidimensional temporal network model;

[0009] S3: training the PINN model using the physical constraint data and the data constraint data to obtain a trained PINN model;

[0010] S4: Using the trained PINN model to predict the hook load and torque.

[0011] Preferably, in step S1, the physical constraint data include boundary conditions and initial conditions of the drill string dynamics model, and the data constraint data include actual drilling characteristic data and measured hook load and torque.

[0012] Preferably, the boundary conditions include the axial displacement and yaw angle of the wellhead, and the initial conditions include the initial axial force and initial torque at the bottom of the well; the actual drilling characteristic data includes time domain data, non-time series text data, and depth domain data;

[0013] The time domain data includes weight on bit, standpipe pressure, rotary table speed, total pump stroke, inlet temperature, outlet temperature, inlet conductance, outlet conductance, inlet density, outlet density, inlet flow rate, outlet flow rate, total pool volume, mechanical penetration rate, tool face angle, hook load and torque;

[0014] The non-time-series text data includes drilling fluid system and drilling tool assembly;

[0015] The depth domain data includes depth measurement, well inclination angle and azimuth angle.

[0016] Preferably, in step S2, the drill string dynamics model is:

[0017]

[0018] Where: R1 represents the equation established in the axial direction of the drill string based on Newton's second law; T t is the internal force component acting in the axial direction of the cross section, N; s is the distance from the drill string to the drill bit, m; μ a is the axial friction coefficient component; N is the normal pressure, MPa; c1 is the axial drilling fluid resistance coefficient; u is the axial displacement, m; t is the time, s; q s is the drill pipe line weight, N / m; is the average well inclination, rad; ρ is the drill string density, kg / m 3 ; A is the cross-sectional area of ​​the drill string, m 2 ; R2 represents the equation established in the circumferential direction of the drill string according to Newton's second law; M t is the torque, N·m; μ t is the tangential friction coefficient component; D j is the diameter of the drill string joint, m; c2 is the drilling fluid resistance coefficient in the torsional direction; θ is the angular displacement, rad; I p is the polar moment of inertia, m 4 .

[0019] Preferably, the axial friction coefficient component and the tangential friction coefficient component are respectively calculated by the following formulas:

[0020]

[0021] Where: μ is the friction coefficient between the drill string and the well wall; v a and v tare the velocity components in the axial and torsional directions, m / s, respectively.

[0022] Preferably, the friction coefficient between the drill string and the well wall is calculated by the following formula:

[0023]

[0024] Where: μ* is the steady-state reference value of the friction coefficient μ; a, b, and c are dimensionless model parameters; ε is a dimensionless parameter to eliminate model singularities; V * is the characteristic velocity, m / s.

[0025] Preferably, the positive pressure is calculated by the following formula:

[0026]

[0027]

[0028] M b =EI a k b (9)

[0029]

[0030] Where: N n and N b are the main direction and secondary normal pressure components, N; k b is the wellbore curvature, rad / m; k α is the rate of change of well inclination, rad / m; α is the well inclination, rad; k n is the natural curvature of the drill string centerline, rad / m; M b is the bending moment, N·m; k φ is the azimuth angle change rate, rad / m; E is the elastic modulus of the drill string, GPa; I a is the drill string section moment of inertia, m 4 ; G is the shear modulus, GPa.

[0031] Preferably, the axial drilling fluid resistance coefficient and the torsional drilling fluid resistance coefficient are respectively calculated by the following formulas:

[0032]

[0033] Where: D i 、D o and D b are the inner diameter, outer diameter and borehole diameter of the drill string, m; τ0 is the yield limit point, Pa; dx is the incremental length of the drill string, m; μ pv is the plastic viscosity, mPa·s; v ris the relative velocity between the drill string and the drilling fluid.

[0034] Preferably, in step S2, the CNN-LSTM-Attention multi-dimensional temporal network model includes an input layer, a CNN layer, an activation function layer, an LSTM layer, an Attention layer, a fully connected layer and a discard layer connected in sequence.

[0035] Preferably, in step S2, the loss function is:

[0036] L PINN (β)=λ·L phy (β)+(1-λ)·L data (β) (13)

[0037]

[0038] Where: L PINN (β) is the total loss function; λ is the weight coefficient; L phy (β) is the physical loss function; L data (β) is the data loss function; N' is the number of sample points; R1 represents the equation established in the axial direction of the drill string based on Newton's second law; R2 represents the equation established in the circumferential direction of the drill string based on Newton's second law; is the hook load predicted by the CNN-LSTM-Attention multi-dimensional temporal network model, N; HL is the measured hook load, N; is the torque predicted by the CNN-LSTM-Attention multi-dimensional temporal network model, N·m; TOR is the measured torque, N·m.

[0039] The beneficial effects of the present invention are:

[0040] The present invention can organically combine the theoretical basis of physical models and the nonlinear fitting ability of machine learning. In addition to data-driven, it also enhances the generalization ability and prediction accuracy of the model by introducing physical constraints. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1 Constructing schematics for physical constraint input data;

[0043] Figure 2Schematic diagram of forces and moments acting on a microelement segment of the drill string;

[0044] Figure 3 A schematic diagram showing a comparison between the hook load predicted by the drill string dynamics model and the measured hook load in a specific embodiment;

[0045] Figure 4 Schematic diagram showing a comparison between the torque predicted by the drill string dynamics model and the actually measured torque in a specific embodiment;

[0046] Figure 5 A query chart for the sliding friction coefficient under water-based drilling fluid conditions in a specific embodiment;

[0047] Figure 6 A query chart for the sliding friction coefficient under oil-based drilling fluid conditions in a specific embodiment;

[0048] Figure 7 A schematic diagram showing a comparison of two different friction coefficient calculation models for calculating bottom hole pressure in a specific embodiment;

[0049] Figure 8 A schematic diagram showing a comparison of tool face angles calculated using two different friction coefficient calculation models in a specific embodiment;

[0050] Figure 9 A schematic diagram of the structure of a CNN-LSTM-Attention multi-dimensional temporal network model in a specific embodiment;

[0051] Figure 10 This is a scatter cloud diagram of the hook load prediction of different models for testing the torsion swing well section in a specific embodiment;

[0052] Figure 11 A Taylor diagram showing the predicted hook load for different models in a torsion swing well section test in a specific embodiment;

[0053] Figure 12 This is a scatter cloud diagram of torque predictions of different models for a torsion swing well section in a specific embodiment;

[0054] Figure 13 A Taylor diagram showing torque predictions for different models in a torsion swing well section test in a specific embodiment;

[0055] Figure 14 A schematic diagram of a PINN model framework in a specific embodiment;

[0056] Figure 15 A graph showing the training loss curve of a PINN model in a specific embodiment;

[0057] Figure 16 A cloud diagram comparing the axial load and circumferential load prediction results of the PINN model and the theoretical model in a specific embodiment;

[0058] Figure 17 A schematic diagram of prediction results of different weight coefficient values ​​in a specific embodiment;

[0059] Figure 18 The hook load prediction results of different models on the test set in a specific embodiment are shown;

[0060] Figure 19 The torque prediction results of different models on the test set in a specific embodiment are shown;

[0061] Figure 20 Graph showing the hook load and torque error distribution for different models in a specific embodiment. DETAILED DESCRIPTION

[0062] The present invention is further described below with reference to the accompanying drawings and examples. It should be noted that, in the absence of conflict, the embodiments in this application and the technical features in the embodiments can be combined with each other. It should be noted that, unless otherwise specified, all technical and scientific terms used in this application have the same meanings as those commonly understood by those of ordinary skill in the art to which this application belongs. The use of similar words such as "include" or "comprising" in the present invention means that the elements or objects appearing before the word cover the elements or objects listed after the word and their equivalents, without excluding other elements or objects.

[0063] The present invention provides a physics-data collaborative driving hook load torque prediction method, comprising the following steps:

[0064] S1: Prepare physical constraint data and data constraint data.

[0065] In a specific embodiment, the physical constraint data include boundary conditions and initial conditions of a drill string dynamics model, and the data constraint data include actual drilling characteristic data and measured hook load and torque.

[0066] In a specific embodiment, the boundary conditions include the axial displacement and yaw angle of the wellhead, the initial conditions include the initial axial force and initial torque at the bottom of the well; the actual drilling characteristic data includes time domain data, non-time series text data, and depth domain data;

[0067] The time domain data includes drilling pressure, standpipe pressure, rotary table speed, total pump stroke, inlet temperature, outlet temperature, inlet conductance, outlet conductance, inlet density, outlet density, inlet flow rate, outlet flow rate, total pool volume, mechanical penetration rate, tool face angle, hook load and torque; the non-time series text data includes drilling fluid system and drill tool assembly; the depth domain data includes depth measurement, well inclination angle and azimuth.

[0068] In a specific embodiment, the construction process of the physical constraint data is as follows: Figure 1 As shown, the horizontal axis is the time grid and the vertical axis is the space grid. The blue area is the data point that needs to be constructed, the red dot represents the coordinate position of the bottom of the well at the initial moment, and the green dot represents the coordinate position adjacent to the bottom of the well at the initial moment. According to the force balance law and Hooke's law, the axial force T and torque M at the initial moment can be iterated. The specific iteration method is shown in the formula:

[0069] T(p,1)=T(p+1,1)-G s +μ·N (16)

[0070] u(p,1)=u(p+1,1)-(T(p+1,1)+T(p,1)) / 2·h x / (E·A s ) (17)

[0071] θ(p,1)=θ(p+1,1)-(M(p+1,1)+M(p,1)) / 2·h x / (G·I p ) (18)

[0072] Where: p is the spatial position point; h x is the spatial step length.

[0073] It is worth noting that since the model adopts angle control, the iteration direction of θ and M is from the wellhead to the bottom of the well.

[0074] For the construction of boundary conditions, the yaw angle control function is set according to the following trapezoidal wave function:

[0075]

[0076] Where: E is the elastic modulus of the drill string, GPa; A is the cross-sectional area of ​​the drill string, m 2 ; u is the axial displacement, m; s is the distance from the drill string to the drill bit, m; G t is the buoyancy of the drill string in the axial direction, N; W ob is the bit pressure on the indicator, N; F f is the total friction of the drill string, N.

[0077] The axial displacement of the bottom hole is set as follows:

[0078] u| p＝L =ROP·t (20)

[0079] Where: L is the well depth, m; ROP is the mechanical drilling rate, m / s; t is the time, s.

[0080] The actual drilling feature data comes from logging time library data, logging depth domain data, inclination measurement data, drilling log text data, drill tool assembly text data and mud performance data, etc. According to the time series and data type of these data, they are divided into time domain data, non-time series text data and depth domain data.

[0081] It should be noted that after obtaining the data constraint data, a pre-processing step is also included, which includes outlier processing, filtering and noise reduction, feature selection, feature data normalization, sample construction and data set division.

[0082] In a specific embodiment, the MAD method is used to detect outliers. After the outliers are detected, in order to eliminate the influence of the outliers, linear interpolation of adjacent non-outlier values ​​can be used to fill the outlier points; the Kalman filter method is used to filter and reduce noise on the logging data set; correlation analysis is used for feature selection; and maximum and minimum normalization is used to normalize the feature data.

[0083] In a specific embodiment, three correlation analysis methods, Pearson, Spearman, and Kendall, are used to analyze the linear and nonlinear correlations between hook load torque and characteristic parameters. Based on the analysis results, characteristic data obtained for predicting hook load include well depth, bit weight, standpipe pressure, rotary table speed, total pump stroke, outlet temperature, inlet density, inlet flow rate, total pool volume, drill string assembly, tool face angle, well inclination, azimuth, and mechanical penetration rate; characteristic data for predicting torque include well depth, bit weight, mechanical penetration rate, standpipe pressure, rotary table speed, total pump stroke, outlet temperature, inlet conductance, outlet flow rate, drill string assembly, and tool face angle.

[0084] S2: Construct a PINN model, which includes physical constraints, data constraints, and a loss function; the physical constraints are a drill string dynamics model, and the data constraints are a CNN-LSTM-Attention multidimensional temporal network model.

[0085] In a specific embodiment, the drill string dynamics model is:

[0086]

[0087] Where: R1 represents the equation established in the axial direction of the drill string based on Newton's second law; T t is the internal force component acting in the axial direction of the section, N; μ a is the axial friction coefficient component; N is the normal pressure, MPa; c1 is the axial drilling fluid resistance coefficient; q s is the drill pipe line weight, N / m; is the average well inclination, rad; ρ is the drill string density, kg / m3 ; R2 represents the equation established in the circumferential direction of the drill string according to Newton's second law; M t is the torque, N·m; μ t is the tangential friction coefficient component; D j is the diameter of the drill string joint, m; c2 is the drilling fluid resistance coefficient in the torsional direction; θ is the angular displacement, rad; I p is the polar moment of inertia, m 4 .

[0088] In the above embodiment, the kinetic model is derived by the following steps:

[0089] First, to study the friction and drag reduction of the drill string during sliding drilling and the optimization of the torsion-pendulum parameters in torsion-pendulum drilling, the following model assumptions are established:

[0090] (1) The drill string is regarded as a soft rope that can withstand axial force and torque. Its cross section always remains circular, and the shear force and bending moment of the cross section are ignored.

[0091] (2) The centerline of the drill string overlaps with the centerline of the borehole and is in uniform contact with the wellbore wall;

[0092] (3) Drill string buckling has nothing to do with the gap between the drill string and the wellbore wall;

[0093] (4) The damping force acting on the drill string only includes friction and drilling fluid resistance, and does not consider other mechanical resistance caused by the drill string penetrating the well wall and getting stuck.

[0094] The model is derived based on the actual drill string, and the natural coordinate system is established by taking the infinitesimal segment of the drill string. Its main normal direction, secondary normal direction and tangential direction coordinates are set as like Figure 2 As shown in the figure, the force of the drill string in this coordinate system is as follows, where s is the distance from the drill string to the drill bit, ds is the incremental length, t is the time, T, M, q s , N, f, and m are axial force, torque, linear weight, normal pressure, axial friction, and torsional friction torque, respectively.

[0095] According to Newton's second law, the equilibrium equation of the drill string in the axial and torsional directions can be obtained:

[0096]

[0097] Where: → represents a vector; is the drill string gravity vector;

[0098] A simplified analytical model is used to estimate the damping effect of the drilling fluid. The basic assumptions of the analytical model are as follows: (1) the drill string is always in the center; (2) the movement of the drill string has no effect on the axial flow of the drilling fluid; and (3) the drilling fluid is a Bingham fluid. The axial and torsional resistance coefficients of the drilling fluid can be expressed as:

[0099]

[0100] Where: D i 、D o and D b are the inner diameter, outer diameter and borehole diameter of the drill string, m; τ0 is the yield limit point, Pa; dx is the incremental length of the drill string, m; μ pv is the plastic viscosity, mPa·s; v r is the relative velocity between the drill string and the drilling fluid.

[0101] Ignoring the higher-order derivatives, the variables in (21) and (22) can be expanded using Taylor series at point (s, t) to obtain the first-order variables as follows:

[0102]

[0103] Decompose the force and torque, convert the vector expression into an equation, and omit the subscript. In the natural coordinate system, we can get:

[0104]

[0105] Where: T n and T b are the internal force components acting in the main normal direction and the secondary normal direction, N;

[0106] According to the Freinet formula, we can get the vector expression:

[0107]

[0108] Where: k α is the rate of change of well inclination, rad / m; k b is the wellbore curvature, rad / m; k φ is the azimuth angle change rate, rad / m.

[0109] Convert the vector expression into scalar form and substitute Equations (25)-(27) into Equations (23) and (24) to obtain the equations for the axial, main normal, secondary normal, and torsional directions:

[0110]

[0111] Where: N n and Nb are the main and secondary normal pressure components, N; α is the well inclination, rad; M b is the bending moment, N·m; Ia is the moment of inertia of the drill string section, m 4 .

[0112] The internal force components, bending moments, and torque acting in the axial direction of the cross section are expressed by the following formula:

[0113]

[0114] M b =EI a k b (9)

[0115]

[0116] Where: G is the shear modulus, GPa.

[0117] Combined with the positive pressure calculation method shown in formula (5):

[0118]

[0119] At the same time, the effects of high-order derivatives of the equation, high-order terms of curvature, and bending moment on the axial force are omitted, and the equations in the axial and circumferential directions are obtained respectively, and the drill string dynamic model shown in equation (1) is obtained.

[0120] It should be noted that the drill string dynamics model shown in the above embodiment has passed the accuracy verification. In a specific embodiment, taking a horizontal well drilled as an example, the comparison results of the hook load and torque predicted by the drill string dynamics model and the measured hook load and torque are as follows: Figure 3 and Figure 4 As shown. Figure 3 and Figure 4 It can be seen that the predicted results of the hook load and torque of the present invention are relatively close to the measured results. The relative error of the hook load is within 10%, and the relative error of the torque is basically kept within 12%, indicating that the drill string dynamics model has sufficient calculation accuracy.

[0121] In a specific embodiment, the axial friction coefficient component and the tangential friction coefficient component are respectively calculated by the following formulas:

[0122]

[0123]

[0124] Where: μ is the friction coefficient between the drill string and the well wall; v a and v t are the velocity components in the axial and torsional directions, m / s, respectively.

[0125] In a specific embodiment, the friction coefficient between the drill string and the well wall is calculated by the following formula:

[0126]

[0127] Where: μ* is the steady-state reference value of the friction coefficient μ; a, b, and c are dimensionless model parameters; ε is a dimensionless parameter to eliminate model singularities; V * is the characteristic velocity, m / s.

[0128] In the above embodiment, the calculation model for the friction coefficient between the drill string and the wellbore wall (DR friction model) is constructed using the widely accepted friction constitutive relationship (rate-state friction law). This rate-state friction law links changes in the friction coefficient with changes in the sliding rate and can describe the friction behavior between a variety of different materials. Furthermore, because the drill string is made of steel, a material with a certain yield point, its frictional resistance is independent of the contact area, and the effect of changes in normal stress on the friction coefficient is negligible. Once sliding friction stabilizes, the calculation model for the friction coefficient between the drill string and the wellbore wall, as shown in Equation (4), is obtained.

[0129] In the calculation model of the friction coefficient between the drill string and the wellbore shown in formula (4), the parameters involved are determined through sliding friction experiments. In a specific embodiment, a sliding friction experiment is conducted using a BRUKE UMT friction and wear tester. The positive pressure is set to 20-50N (corresponding to the actual downhole distributed positive pressure of 800N / m to 1500N / m) and the sliding speed is set to 0.3-2.1mm / s (corresponding to the actual drilling mechanical penetration rate of 1.08m / h to 7.56m / h). The friction coefficient changes of the drill string-wellbore specimen are simulated under different parameter conditions using oil-based drilling fluid and water-based drilling fluid as the fluid lubrication medium. The experimental scheme is shown in Table 1:

[0130] Table 1 Sliding friction experimental plan

[0131]

[0132]

[0133] The sliding friction coefficient under two fluid lubrication media is as follows Figure 5 and Figure 6 As shown, the model parameters of the calculation model of the friction coefficient between the drill string and the well wall shown in formula (4) under two drilling fluid environments are shown in Table 2:

[0134] Table 2 Model parameters determined by sliding friction experiments

[0135]

[0136] The friction coefficient between the drill string and the well wall can be directly and accurately calculated using the calculation model shown in formula (4) after the parameters are determined.

[0137] It should be noted that the friction calculation model of the above embodiment has passed the accuracy test. In a specific embodiment, the friction coefficient is calculated by using the model and the Coulomb friction model, and the friction coefficients calculated by the two models are used to determine the change in bottom hole pressure and tool face angle. The results are as follows: Figure 7 and Figure 8 As shown. Figure 7 and Figure 8 It can be seen that the calculation results of the friction calculation model of the above embodiment are more consistent with the actual drilling conditions.

[0138] In a specific embodiment, Figure 9 As shown in the figure, the CNN-LSTM-Attention multi-dimensional temporal network model includes an input layer, a CNN layer, an activation function layer, an LSTM layer, an Attention layer, a fully connected layer, and a dropout layer connected in sequence. Optionally, the activation function layer uses a Relu activation function. In this embodiment, in order to maintain the standard Attention calculation process without affecting the continuity and integrity of the feature expression, that is, after Attention(Q, K, V) = Softmax(), it directly enters the output part, and there is no need to perform additional Softmax on the attention matrix.

[0139] In the above embodiment, the CNN-LSTM-Attention multi-dimensional temporal network model is capable of processing multiple feature time series data. In drilling operations, data such as weight on bit, rotary table speed, and standpipe pressure are all multi-dimensional time series, and these data may have complex interdependencies. The present invention utilizes this CNN-LSTM-Attention multi-dimensional temporal network model to learn how to predict future values ​​of hook torque from the historical relationships between these multiple features.

[0140] In the CNN-LSTM-Attention multi-dimensional time series network model, the CNN layer is used to process the local features of time series data, the LSTM layer is used to capture the temporal dependencies in the data, and the Attention layer is responsible for dynamically assigning weights to different parts of the input, helping the model focus on key features. The fully connected layer can help the model better process complex time series data. The CNN-LSTM-Attention multi-dimensional time series network model described in the present invention combines the advantages of CNN and LSTM, can comprehensively consider the complexity of time series and the relationship between multiple features, and provide accurate prediction results. It is particularly suitable for prediction tasks involving multi-dimensional features and complex time series dependencies.

[0141] It should be noted that the CNN-LSTM-Attention multi-dimensional temporal network model in the above embodiment can accurately predict the load and torque of the hook. In a specific embodiment, the prediction accuracy of the hook load and torque of the CNN-LSTM-Attention multi-dimensional temporal network model of the present invention is compared with five single models including multi-layer perceptron (MLP), long short-term memory neural network (LSTM), convolutional neural network (CNN), gated recurrent unit (GRU), and Transformer neural network with self-attention mechanism, as well as four fusion models including CNN-MLP, CNN-LSTM, CNN-GRU and CNN-Transformer. The results are as follows: Figure 10-13 As shown. Figure 10-13 It can be seen that the CNN-LSTM-Attention multi-dimensional temporal network model adopted in the present invention can most accurately predict the hook load and torque.

[0142] In a specific embodiment, the loss function is:

[0143] L PINN (β)=λ·L phy (β)+(1-λ)·L data (β) (13)

[0144]

[0145] Where: L PINN (β) is the total loss function; λ is the weight coefficient; L phy (β) is the physical loss function; L data (β) is the data loss function; N' is the number of sample points; R1 represents the equation established in the axial direction of the drill string based on Newton's second law; R2 represents the equation established in the circumferential direction of the drill string based on Newton's second law; is the hook load predicted by the CNN-LSTM-Attention multi-dimensional temporal network model, N; HL is the measured hook load, N; is the torque predicted by the CNN-LSTM-Attention multi-dimensional temporal network model, N·m; TOR is the measured torque, N·m.

[0146] S3: Using the physical constraint data and the data constraint data to train the PINN model to obtain a trained PINN model.

[0147] In a specific embodiment, the training includes the following sub-steps:

[0148] (1) For the physical constraint part, the boundary conditions and the initial conditions are used as input data of the physical constraint items, and the hook load and torque calculated by the drill string dynamics model are used as labels of the physical constraint items; for the data constraint part, the pre-processed actual drilling features are used as input features of the CNN-LSTM-Attention multidimensional temporal network model, and the measured hook load and torque are used as labels of the CNN-LSTM-Attention multidimensional temporal network model;

[0149] It should be noted that the outputs calculated by the drill string dynamics model during the training process are two two-dimensional matrices of axial load (Tt) and circumferential load (Mt) that vary with time and depth. These two variables serve as intermediate variables of the loss function and need to be further calculated to obtain the wellhead hook load and torque as variables for constructing the physical constraint loss function. The method of calculating the hook load through the axial load is common knowledge, and the axial load is the torque.

[0150] (2) The results predicted by the PINN model are calculated according to Figure 14 The framework diagram shown is input into the physical constraint loss L phy (β) and data-constrained loss L data (β), and calculate the loss values ​​of the physical constraint term and the data constraint term respectively, and finally add them according to the weighted rule to get the final loss value L PINN (β);

[0151] (3) Through the back-propagation mechanism in the PINN model, the weights and bias terms of the network model are optimized to minimize the total loss value and obtain a trained PINN model.

[0152] S4: Using the trained PINN model to predict the hook load and torque.

[0153] In a specific embodiment, taking the Sichuan Z201H69-4 well as an example, the hook load and torque prediction method of the present invention is used to predict the hook load and torque. In this embodiment, since the training data required by the model consists of sample points based on measured data and physical model boundary conditions and initial conditions, in order to match the theoretical model and measured data, the output T obtained by the experimental results is converted into the output T of the model. t and M t The sample matrix size is fixed at 401×60000.

[0154] In this embodiment, the network model hyperparameters are set as follows: 100 hidden layer neurons, 3 convolution kernels, 0.5 dropout rate, 0.0015 learning rate, 500 training times, and 64 batch sizes; the weight coefficients of the physical constraint and data constraint are both set to 0.5. The iteration curve of the PINN model training process is shown in Figure 2. Figure 15 The PINN model's prediction results for axial load (Tt) and torsional displacement (Mt) and the theoretical model's (i.e., the drill string dynamics model of the present invention) calculation results for axial load (Tt) and torsional displacement (Mt) are shown in the cloud diagram, as well as the relative error between the theoretical model's calculated values ​​and the PINN model's predicted values. Figure 16 shown.

[0155] from Figure 16 It can be seen that: (1) The change in the axial load of the drill string from the wellhead to the bottom of the well at each moment is positive for tension and negative for compression. When drilling begins and the well depth is almost 2000m, the friction begins to accumulate slowly, so the axial load at the corresponding moment decreases in the tensile direction, and then continues to drill deeper until the axial load is reduced to 0. At this time, the hook load in the axial direction is equal to the friction and the buoyancy of the drill string. Then the drill string is subjected to greater friction in the bending section, so the axial load increases in the compression direction again, and decreases again when it is close to the bottom of the horizontal section. The axial load at each depth point changes with time and tends to fluctuate stably within a certain range. This law is consistent with the law of change in the dynamics of torsion pendulum drilling. The axial load change trend calculated by the theoretical model is consistent with the prediction of the PINN model.

[0156] (2) Similarly, observing the trend of circumferential load changes, the circumferential load that changes with depth at each moment indicates the depth of torque transmission. It can be found that the torque change at the wellhead presents a regular reciprocating fluctuation. This pattern is consistent with the law of the torsion angle control mode. When the torque reaches the amplitude, the depth of the torque transmission is the deepest, but there will be a time lag of several seconds.

[0157] (3) The relative error cloud diagrams of the axial load and circumferential load predicted by the theoretical model and the PINN model show that the results are consistent with those obtained by the physical model in terms of regularity and the numerical differences are also small. The relative errors are all controlled within 10%, indicating that the PINN model can effectively solve and predict the axial load and circumferential load.

[0158] In order to determine the optimal weight coefficient, the influence of different weight values ​​on the prediction results was analyzed. In this embodiment, the calculated λ value is between 0 and 1 (when λ is 0, it means there are only data constraints; when λ is 1, it means there are only physical constraints. When there are only physical constraints, the model lacks the real target value (HL, TOR) constraints and cannot correctly learn the relevant information, so this condition is excluded). The interval is 0.1. The hook load and torque are tested under 10 conditions. The mean square error (MSE) between the predicted value and the true value is as follows: Figure 17 shown.

[0159] from Figure 17 It can be seen that (1) when λ is 0.3, the model's prediction relative error is the smallest, indicating that this weight configuration can provide the best prediction effect, making the predicted values ​​of the hook load and torque closest to the true value; (2) The relative error distribution of the hook load is relatively uniform, maintaining within the range of 5% to 10%, among which the relative error is lower when λ is 0.1 and 0.3, and the relative error is highest when λ is 0.8; (3) The relative error of the torque is maintained within the range of 5% to 11%, and the relative error is lower when λ is 0.3 and 0.4, and is concentrated in the range of 6% to 9%; (4) When the weight coefficient λ is selected as 0.3, that is, the data constraint term coefficient λ is 0.7, the PINN model reaches a balanced state when predicting the hook load and torque, minimizing the overall prediction error.

[0160] In addition, under the above-determined optimal weight coefficient conditions, the prediction performance of the PINN model of the present invention was verified, and the hook load and torque prediction results obtained by this model were compared with those of the artificial neural network (ANN), CNN-LSTM-Attention model and drill string dynamics model. In order to verify the effectiveness and comparability of each model, the prediction results of the test set were compared. The prediction results of each model are shown in Figure 2. Figure 18 and Figure 19 shown.

[0161] from Figure 18It can be seen that: (1) From the perspective of the change in the hook load of the actual test set, the change range of the hook load is between 906kN and 911kN. The measured values ​​of the hook load of this test set were interpolated within one minute to match the number of samples obtained by the drill string dynamics model. From the figure, the change trend of the drill string dynamics model calculation results and the actual values ​​is not well matched. This is the result of not training with real field hook load data, but the numerical difference is not large. The other models all showed good fitting effects, and the error was basically kept within the range of 1.5kN. (2) For the ANN model, although it can predict the change trend of the hook load well, there is still a large deviation in the numerical value of the previous sample, which is significantly different from the actual value. (3) The CNN-LSTM-Attention model, with the advantage of fitting multi-dimensional time series, performs well in both numerical and trend fitting. (4) The PINN model shows excellent prediction performance. It can be found that the prediction accuracy of this model is the highest, but the curve is not the smoothest and has a certain amount of noise. This is because the model combines the drill string dynamics physical model with field data to learn the hook load change characteristics. The prediction error of this model fluctuates within 0.8kN.

[0162] from Figure 19 It can be seen that: (1) From the torque variation of the actual test set, the torque variation range is between 0kN·m and 8kN·m. Since the torque recorded in the field is always non-negative, even when the drill bit rotates in the opposite direction, the torque is positive. When the theoretical model is calculated, the torque obtained by the reverse rotation of the drill bit is negative. Therefore, in order to match the field data, the results calculated by the physical model are taken as absolute values. From the figure, it can be seen that the results calculated by the drill string dynamics model are consistent with the variation trend of the actual torque, but there are some differences in the torque amplitude. This is because the boundary condition control of the model is based on the trapezoidal wave angle control, and the torque is inversely calculated based on the angle, so there are certain differences. The torque values ​​predicted by each model are not much different from the measured values, and all show good fitting effects, with the error basically kept within the range of 1.2kN·m; (2) For the ANN model, it can better predict the change trend of torque, but the prediction of torque amplitude has a large deviation, indicating that the change of torque amplitude is not well learned; (3) The CNN-LSTM-Attention model still performs well in both numerical and trend fitting for torque prediction, but lacks learning of physical features and has limited generalization; (4) The PINN model, with its excellent prediction performance and model generalization performance, not only learns the change law of field data but also the prediction results are corrected by the physical model, and the prediction error is kept within 1.0kN·m.

[0163] The error distribution histograms and fitting curves of the hook load and torque predicted by the above models are shown in the figure below: Figure 20 As shown in the figure. This figure is a visualization method for analyzing the error characteristics of the model to better understand the prediction performance of each model. The vertical coordinate represents the counted samples, the horizontal coordinate represents the absolute error, the bar chart represents the error distribution, and the fitted curve represents the distribution law of the error. The peak value is the mean value, and the closer it is to 0, the better the prediction effect of the model.

[0164] From Figure 20 It can be seen that: (1) When predicting the hook load, the mean prediction errors of the PINN model, the CNN-LSTM-Attention model, and the ANN model are relatively close, all within 0.4 kN, while the mean prediction error of the drill string dynamics model is about 1.2 kN. When predicting the torque, the mean prediction errors of the PINN model and the CNN-LSTM-Attention model are relatively close, both below 0.2 kN·m, and the prediction errors of the ANN model and the drill string dynamics model are within 0.6 kN·m; (2) By comparing and analyzing the hook load and torque errors predicted by each model, although there are certain differences in the prediction results of different models, the order of the prediction performance is the same, and the error performance is PINN < CNN-LSTM-Attention < ANN < drill string dynamics model.

[0165] In summary, the present invention can accurately predict the hook load and torque. Compared with the prior art, the present invention has made significant progress.

[0166] The above is only a preferred embodiment of the present invention and does not impose any form of limitation on the present invention. Although the present invention has been disclosed as above with a preferred embodiment, it is not intended to limit the present invention. Any person skilled in the art can make some changes or modifications to the above-disclosed technical content to obtain equivalent embodiments with equivalent changes, but as long as the content does not depart from the technical solution of the present invention, any simple modification, equivalent change, and modification made to the above embodiments based on the technical essence of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A physics-data collaborative driven hook load torque prediction method, characterized in that: The following steps are involved: S1: Prepare physical constraint data and data constraint data; S2: Constructing a PINN model, wherein the PINN model includes a physical constraint item, a data constraint item, and a loss function; the physical constraint item is a drill string dynamics model, and the data constraint item is a CNN-LSTM-Attention multidimensional temporal network model; S3: training the PINN model using the physical constraint data and the data constraint data to obtain a trained PINN model; S4: Using the trained PINN model to predict the hook load and torque.

2. The method for predicting hook load torque by physics-data collaborative driving according to claim 1, characterized in that: In step S1, the physical constraint data include boundary conditions and initial conditions of the drill string dynamics model, and the data constraint data include actual drilling characteristic data and measured hook load and torque.

3. The method for predicting hook load torque driven by physics-data collaboration according to claim 2, characterized in that: The boundary conditions include the axial displacement and yaw angle of the wellhead, and the initial conditions include the initial axial force and initial torque at the bottom of the well; the actual drilling characteristic data includes time domain data, non-time series text data, and depth domain data; The time domain data includes weight on bit, standpipe pressure, rotary table speed, total pump stroke, inlet temperature, outlet temperature, inlet conductance, outlet conductance, inlet density, outlet density, inlet flow rate, outlet flow rate, total pool volume, mechanical penetration rate, tool face angle, hook load and torque; The non-time-series text data includes drilling fluid system and drilling tool assembly; The depth domain data includes depth measurement, well inclination angle and azimuth angle.

4. The physics-data collaborative driving hook load torque prediction method according to claim 1, characterized in that: In step S2, the drill string dynamics model is: Where: R1 represents the equation established in the axial direction of the drill string based on Newton's second law; T t is the internal force component acting in the axial direction of the cross section, N; s is the distance from the drill string to the drill bit, m; μ a is the axial friction coefficient component; N is the normal pressure, MPa; c1 is the axial drilling fluid resistance coefficient; u is the axial displacement, m; t is the time, s; q s is the drill pipe line weight, N / m; is the average well inclination, rad; ρ is the drill string density, kg / m 3 ; A is the cross-sectional area of ​​the drill string, m 2 ; R2 represents the equation established in the circumferential direction of the drill string according to Newton's second law; M t is the torque, N·m; μ t is the tangential friction coefficient component; D j is the diameter of the drill string joint, m; c2 is the drilling fluid resistance coefficient in the torsional direction; θ is the angular displacement, rad; I p is the polar moment of inertia, m 4 .

5. The method for predicting hook load torque driven by physics-data collaboration according to claim 4, characterized in that: The axial friction coefficient component and the tangential friction coefficient component are respectively calculated by the following formulas: Where: μ is the friction coefficient between the drill string and the well wall; v a and v t are the velocity components in the axial and torsional directions, m / s, respectively.

6. The method for predicting hook load torque driven by physics-data collaboration according to claim 5, characterized in that: The friction coefficient between the drill string and the well wall is calculated by the following formula: Where: μ* is the steady-state reference value of the friction coefficient μ; a, b, and c are dimensionless model parameters; ε is a dimensionless parameter to eliminate model singularities; V * is the characteristic velocity, m / s.

7. The method for predicting hook load torque driven by physics-data collaboration according to claim 4, characterized in that: The positive pressure is calculated by the following formula: M b =NO a k b (9) Where: N n and N b are the main direction and secondary normal pressure components, N; k b is the wellbore curvature, rad / m; k α is the rate of change of well inclination, rad / m; α is the well inclination, rad; k n is the natural curvature of the drill string centerline, rad / m; M b is the bending moment, N·m; is the azimuth angle change rate, rad / m; E is the elastic modulus of the drill string, GPa; I a is the drill string section moment of inertia, m 4 ; G is the shear modulus, GPa.

8. The method for predicting hook load torque driven by physics-data collaboration according to claim 4, characterized in that: The axial drilling fluid resistance coefficient and the torsional drilling fluid resistance coefficient are respectively calculated by the following formulas: Where: D i 、D o and D b are the inner diameter, outer diameter and borehole diameter of the drill string, m; τ0 is the yield limit point, Pa; dx is the incremental length of the drill string, m; μ pv is the plastic viscosity, mPa·s; v r is the relative velocity between the drill string and the drilling fluid.

9. The method for predicting hook load torque driven by physics-data collaboration according to claim 1, characterized in that: In step S2, the CNN-LSTM-Attention multidimensional temporal network model includes an input layer, a CNN layer, an activation function layer, an LSTM layer, an Attention layer, a fully connected layer, and a discard layer connected in sequence.

10. The physics-data collaborative driving hook load torque prediction method according to claim 1, characterized in that: In step S2, the loss function is: L PINN (b)=λ·L phy (b)+(1-λ)·L data (b) (13) Where: L PINN (β) is the total loss function; λ is the weight coefficient; L phy (β) is the physical loss function; L data (β) is the data loss function; N' is the number of sample points; R1 represents the equation established in the axial direction of the drill string based on Newton's second law; R2 represents the equation established in the circumferential direction of the drill string based on Newton's second law; is the hook load predicted by the CNN-LSTM-Attention multi-dimensional temporal network model, N; HL is the measured hook load, N; is the torque predicted by the CNN-LSTM-Attention multi-dimensional temporal network model, N·m; TOR is the measured torque, N·m.

Images