A directional drilling track optimization method based on a GRU-MPC fusion model

Abstract

The present application relates to the oil and gas drilling engineering technical field, specifically relates to a kind of directional drilling track optimization method based on GRU-MPC fusion model, constructs the time series prediction model of drill bit track state in the process of while drilling, introduce the rolling prediction of track state in the well to be drilled well length by gate cycle unit network, and the prediction result is as state prediction model in model predictive control.On this basis, in combination with engineering constructability constraint, construct the multi-objective evaluation function of fusing track deviation, trajectory smoothness, drilling efficiency and control stability, and realize the adaptive decision of track optimization under different working conditions by dynamic weight adjustment mechanism.System carries out prediction evaluation to candidate control sequence in each control cycle, only executes the optimal control decision of current cycle, and repeats the above-mentioned prediction and optimization process after obtaining new while drilling data, to realize the rolling multi-objective optimization of directional drilling track control.

Description

Technical Field

[0001] This invention relates to the field of oil and gas drilling engineering technology, and in particular to a method for optimizing directional drilling trajectories based on the GRU-MPC fusion model. Background Technology

[0002] Directional and horizontal well technologies have become important tools for developing oil and gas reservoirs in deep, ultra-deep, and complex structural areas. The rationality of well trajectory control directly affects target area hit rate, drilling safety, and overall operational efficiency. As well types develop towards larger displacement, longer horizontal sections, and higher trajectory complexity, the well trajectory control process faces challenges such as frequent changes in operating conditions, complex constraints, and diversified control objectives, significantly increasing the difficulty of trajectory control decisions.

[0003] Currently, on-site track control decisions mainly rely on engineers' experience, which suffers from high subjectivity and poor consistency, making it difficult to guarantee control effectiveness under complex operating conditions. Furthermore, due to the large number of feasible correction paths, traditional optimization methods are limited by downhole computing resources, making real-time global search and decision-making difficult.

[0004] While existing research attempts to introduce physical models or data-driven methods for prediction and optimization, these methods often have limitations. For example, physical models struggle to adapt to dynamic changes downhole, while purely data-driven intelligent methods lack a systematic closed-loop control framework. Furthermore, in current technologies, trajectory state prediction, optimization calculation, and control execution are often treated as independent processes, lacking a unified rolling decision-making mechanism. This leads to a disconnect between optimization results and actual execution, making it difficult to establish a continuously updated closed-loop control process during drilling. Under multi-objective constraints and limited computational resources, achieving forward-looking evaluation and real-time optimization of trajectory control decisions remains a critical technical problem that urgently needs to be solved in the field of directional drilling trajectory control.

[0005] To address the aforementioned issues, there is an urgent need for a trajectory control decision-making method that integrates data-driven prediction capabilities with model predictive control concepts during drilling. This method should enable rolling optimization decisions under multi-objective constraints without relying on precise downhole dynamics analytical models, thereby improving the stability, adaptability, and engineering applicability of the trajectory control process. Summary of the Invention

[0006] The purpose of this invention is to provide a directional drilling trajectory optimization method based on the GRU-MPC fusion model, which aims to solve the problem of how to select the optimal path with smooth trajectory, high mechanical drilling speed and compliance with engineering constraints in real time from among countless possible correction paths after the actual drilling trajectory deviates from the design trajectory.

[0007] To achieve the above objectives, this invention provides a method for directional drilling trajectory optimization based on the GRU-MPC fusion model, comprising the following steps: At the current well depth, obtain the drill bit trajectory state parameters and construct a state vector representing the wellbore trajectory attitude and drilling state. Construct an orbital state prediction model based on historical orbital state sequences; Track candidate generation and rolling optimization decision based on constructability constraints; Execution result feedback and adaptive update of model parameters.

[0008] The step of "obtaining drill bit trajectory state parameters at the current well depth and constructing a state vector representing the wellbore trajectory attitude and drilling state" includes the following steps: Collect drilling directional and logging-while-drilling data, surface logging data, geological and engineering design parameters, and historical well trajectory data of the same formation; The collected raw data is time-aligned, outlier removed, and filtered and smoothed. The trajectory parameters are resampled based on a unified well depth step size to construct a set of state parameters that characterize the current wellbore trajectory control.

[0009] The "Constructing an orbital state prediction model based on historical orbital state sequences" includes the following steps: The GRU trajectory state prediction model is trained offline based on historical well measurement-while-drilling data to learn the evolution of trajectory state under different drilling conditions. During the drilling application phase, the model parameters remain fixed or are periodically updated based on feedback information; Whenever new drilling data points are acquired, the historical state sequence is updated and the GRU prediction model is called again to generate a new prediction state sequence, thus realizing the rolling update of the prediction process.

[0010] The "Trajectory Candidate Generation and Rolling Optimization Decision Based on Constructability Constraints" includes the following steps: Rolling predictions of the predicted orbital state are performed for multiple sets of candidate control sequences within the prediction time domain. The weight coefficients of each sub-objective in the multi-objective comprehensive evaluation function are dynamically adjusted according to the current drilling conditions or historical performance to balance the relationship between trajectory smoothness, correction efficiency and drilling efficiency, and to achieve adaptive decision-making under different working conditions.

[0011] The "Execution Result Feedback and Model Parameter Adaptive Update" section includes the following steps: After the trajectory control command is executed, actual drilling data is continuously collected during the drilling process; The actual track status is compared and analyzed with the predicted track status, and the deviation between the actual execution result and the prediction result is calculated. The weight coefficients in the multi-objective comprehensive evaluation function are adaptively adjusted based on the accumulated performance evaluation results, and the parameters of the GRU orbital state prediction model are updated when the control cycle reaches the set threshold.

[0012] This invention discloses a directional drilling trajectory optimization method based on a GRU-MPC fusion model. This method constructs a time-series prediction model of the drill bit trajectory state during drilling, introduces a gated cyclic unit (GRU) network to perform rolling predictions of the trajectory state within several future well depth steps, and uses the prediction results as the state prediction model in model predictive control (MPC). Based on this, and combined with engineering constructability constraints, a multi-objective evaluation function integrating trajectory deviation, trajectory smoothness, drilling efficiency, and control stability is constructed, and adaptive decision-making under different operating conditions is achieved through a dynamic weight adjustment mechanism. In each control cycle, the system predicts and evaluates candidate control sequences, executes only the optimal control decision for the current cycle, and repeats the above prediction and optimization process after acquiring new drilling data, thereby achieving rolling multi-objective optimization of directional drilling trajectory control. This invention achieves a close integration of trajectory state prediction and optimization decision-making by embedding a data-driven time-series prediction model into the model predictive control framework, improving the foresight, stability, and engineering applicability of trajectory correction decisions without relying on an exact analytical dynamic model. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is the overall flowchart of the present invention.

[0015] Figure 2 This is a flowchart of a directional drilling trajectory optimization method based on a GRU (Gated Cyclic Unit) - MPC (Model Predictive Control) fusion model.

[0016] Figure 3 This is the GRU rolling prediction model construction process.

[0017] Figure 4 This is the MPC rolling optimization decision-making flowchart.

[0018] Figure 5This is an example of how the GRU-MPC model of this invention corrects deviations in a certain well. Detailed Implementation

[0019] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0020] Please see Figures 1 to 5 This invention provides a method for optimizing directional drilling trajectories based on a GRU-MPC fusion model, comprising the following steps: S1 obtains the drill bit trajectory state parameters at the current well depth position and constructs a state vector representing the wellbore trajectory attitude and drilling state; S11 collects drilling directional and logging-while-drilling data, surface logging data, geological and engineering design parameters, and historical well trajectory data in the same formation; S12 performs time alignment, outlier removal, and filtering and smoothing on the collected raw data; S13 resamples the trajectory parameters based on a unified well depth step size to construct a set of state parameters that characterize the current wellbore trajectory control.

[0021] Specifically, the system first acquires multi-source raw data during on-site drilling operations. Data sources include MWD and LWD data, surface logging data, geological and engineering design parameters, and historical well trajectory data from the same formation. MWD data is provided by the MWD and LWD systems and includes well depth (MD), inclination angle (Inc), and azimuth angle (Azi). Auxiliary data includes drilling pressure (WOB), rotational speed (RPM), and torque. LWD data includes at least formation lithology coding parameters. The system preprocesses the acquired raw data. This preprocessing includes time alignment, outlier removal, missing value completion, and filtering and smoothing for depth-related time-series data to reduce measurement noise and the impact of asynchronous multi-source data on trajectory state modeling. Static data, such as geological information and drill string assembly types, are converted into categorical data that can be processed by machine learning using one-hot coding. In a preferred embodiment, the inclination angle (Inc) and azimuth angle (Azi) sequences are smoothed and filtered to suppress high-frequency random disturbances. After data preprocessing, the system resamples the trajectory parameters based on a unified well depth step size and constructs a set of state parameters characterizing the current wellbore trajectory control. In implementation, this set of state parameters can be represented as a feature vector of wellbore trajectory state and drilling conditions: in Used to comprehensively characterize the wellbore trajectory status and drilling conditions at the current well depth location; For depth measurement; The deviation value of the orbital parameters, i.e., the first... The Euclidean distance between this point and the corresponding design point at a depth of meters; The well inclination angle; It is the azimuth angle; For drilling pressure; Rotational speed; Torque; This refers to the mechanical drilling speed.

[0022] Where k is the discrete sampling index, representing the k-th measurement point (or the k-th update time) along the well depth sequence, therefore Indicates in The wellbore trajectory status and drilling condition feature vector at the location.

[0023] In addition, the system reads the pre-set design trajectory information and the corresponding target window parameters, wherein the target window radius is denoted as... These are used as reference constraints in subsequent track control decisions. It is not only used to determine whether the trajectory deviation exceeds the limit in real time, but also serves as one of the reference parameters in the multi-objective evaluation function of model predictive control, which is used to constrain the control effect of the predicted trajectory within the design target area.

[0024] Finally, the system combines the wellbore trajectory status with the drilling condition feature vector. Together with the reference constraint parameters, they serve as inputs for subsequent track state prediction models and model predictive control decision calculations, providing basic state information for rolling multi-objective track control decisions.

[0025] S2 constructs an orbital state prediction model based on historical orbital state sequences; The S21 GRU trajectory state prediction model is trained offline based on historical well measurement-while-drilling data to learn the evolution of trajectory state under different drilling conditions. S22 keeps the model parameters fixed or updates them periodically based on feedback information during the drilling application phase. Whenever new drilling data points are acquired, S23 updates the historical state sequence and calls the GRU prediction model again to generate a new prediction state sequence, thus realizing the rolling update of the prediction process.

[0026] Specifically, after completing the trajectory control state construction in step one, the system constructs a trajectory state prediction model based on historical trajectory state sequences to predict the evolution trend of wellbore trajectory state within several future well depth steps. The trajectory state prediction model is implemented using a gated cyclic unit network (GRU) to characterize the temporal correlation and nonlinear evolution characteristics of wellbore trajectory state during drilling. The specific operation is as follows: At the current well depth At this point, a recent historical data window of length H meters is extracted as the model input. The H-meter historical wellbore trajectory status and drilling condition characteristic sequence is represented as follows: in Using the current point k as the "end of the window", we look back H (or within a range of approximately H meters) historical states and drilling condition feature vectors to form a "state sequence", which is the input sequence used by the time series model GRU. The feature vectors of wellbore trajectory status and drilling conditions constructed in step one are used to comprehensively characterize the current wellbore trajectory status and drilling conditions.

[0027] The GRU orbital state prediction model uses historical state sequences. As input, the output is a future prediction window of length N meters.

[0028] The historical window length H and the prediction window length N are set according to the well section length and data sampling interval. Preferably, both H and N are between 20 meters and 50 meters. This represents the GRU state prediction function after training is complete; This represents the predicted state sequence value for the next N prediction windows (from k+1 to k+N) starting from the current index k.

[0029] The GRU prediction model can be trained offline based on historical well measurement-while-drilling (MWD) data to learn the evolution of trajectory states under different drilling conditions. During the MWD application phase, the model parameters remain fixed or are periodically updated based on feedback information. Whenever new MWD data points are acquired, the system updates the historical state sequence and re-invokes the GRU prediction model to generate a new predicted state sequence, thus achieving rolling updates in the prediction process.

[0030] It should be noted that the predicted state sequence is not directly used as the final output of trajectory control or correction operations. Instead, it serves as a state prediction model in model predictive control, used to evaluate the impact of different control decisions on the evolution of wellbore trajectory state in the prediction time domain. By introducing the state prediction model constructed using GRU into the model predictive control framework, the system can achieve a forward-looking characterization of future trajectory state changes without relying on an accurate downhole dynamics analytical model, providing a predictive basis for subsequent multi-objective optimization decisions.

[0031] S3 Track candidate generation and rolling optimization decision based on constructability constraints; S31 generates multiple candidate control sequences through discrete sampling in the prediction time domain, and uses the track state prediction model to perform rolling prediction of the track state corresponding to each candidate control sequence; The weight coefficients of each sub-objective in the S32 multi-objective comprehensive evaluation function are dynamically adjusted according to the current drilling conditions or historical performance to balance the relationship between trajectory smoothness, correction efficiency and drilling efficiency, and to achieve adaptive decision-making under different working conditions.

[0032] Specifically, after completing the trajectory state prediction model construction in step two, the system at the current well depth position... Based on the current trajectory control state and predicted state information, a model predictive control problem for directional drilling trajectory control is constructed to achieve rolling optimization decision-making under multi-objective constraints.

[0033] (1) Model predictive control problem construction: During the current control cycle, the system uses the trajectory state and drilling condition feature vector constructed in step S1. As the initial state for model predictive control, and using the predicted state sequence obtained in step S2 based on GRU. As a reference model for predicting state evolution in the time domain.

[0034] The system constructs a set of candidate control sequences within a prediction time domain length N. The candidate control sequences consist of several control variables arranged in steps according to well depth. Composition. In implementation, the single-step control vector can be expressed as: Where k is the control step index, which is related to the well depth step size. Corresponding to, that is, each advancement Update the control vector once. Indicates the tool face angle control amount. This indicates the dogleg control amount within the control step. This represents the sliding drilling ratio. SR ranges from [0,1], representing the proportion of the sliding drilling length to the total drilling length in this control step. The range of values ​​for the control variable is preset based on the drill string's structural characteristics and engineering experience.

[0035] In the prediction time domain, the candidate control sequence can be represented as: in, It represents a sequence of future control actions starting from the current control step k within a prediction time domain length N. A control step represents the smallest decision unit that updates the control quantity once and executes a small well depth increment. The candidate control sequence is used to characterize different trajectory control strategies that may be adopted within the prediction time domain.

[0036] In the prediction time domain, the m-th candidate control sequence is represented as: in, The number of candidate control sequences is and the set of candidate control sequences is . Candidate control sequences are used to characterize different orbit control strategies that may be adopted in the prediction time domain.

[0037] (2) Construction of multi-objective comprehensive evaluation function for candidate orbit schemes: In order to quantify the comprehensive performance of different candidate modified orbits, a multi-objective evaluation function is constructed to conduct preliminary evaluation and screening of candidate orbits. The evaluation function is composed of multiple sub-objectives weighted together.

[0038] Define evaluation function Used for quantizing orbits The advantages and disadvantages. Sub-objectives are defined as: Total dogleg angle of the track is the sum of the dogleg angles of the entire well section.

[0039] : Target distance, the distance between the end point of the track and the designed target.

[0040] : Frequency of tool face angle change, based on The sequence statistics show that the tool face angle variation within a unit well depth exceeds The number of times.

[0041] ,in This represents the average mechanical drilling speed.

[0042] .

[0043] To enhance the adaptive capability of optimization decision-making under different operating conditions, the weighting coefficients... to The settings can be configured based on engineering experience or initialized based on the statistical characteristics of historical successful well sections across various evaluation indicators. During drilling, the system can statistically analyze the contribution frequency of different optimization objectives to successful corrections based on the current trajectory correction effect, and periodically update the weight parameters in conjunction with the current drilling status to balance the relationship between trajectory smoothness, correction efficiency, and drilling efficiency. The weight vector serves as a parameter in the MPC cost function and is updated within each control cycle.

[0044] (3) Modeling and screening of track engineering constraints for constructability: To ensure the constructability of the track correction scheme, during the MPC optimization process, constraints on any control step in the prediction time domain are selected. The following engineering constraints are imposed on the control variables: in, Pre-set based on the structural characteristics of the drilling tool and engineering experience.

[0045] To limit the rate of change of wellbore attitude per unit well depth, the following constraints are applied to the changes in inclination and azimuth angles in the prediction time domain:

[0046] in This represents the maximum allowable rate of change in the drilling tool's directional and turning capabilities.

[0047] To prevent the track condition from entering an unworkable zone, a safety domain constraint is imposed on the predicted track condition:

[0048] in, This represents the set of feasible trajectory states determined by the combined factors of formation conditions, wellbore structure, and construction safety requirements.

[0049] In the prediction time domain, if a certain candidate control sequence If any of the above engineering constraints are violated during the prediction process, the control sequence is determined to be an infeasible solution and is eliminated during the candidate trajectory selection process; candidate control sequences that meet the constraints can be used as initial control sequences for subsequent continuous optimization solutions.

[0050] (4) Rolling solution and online execution mechanism for trajectory optimization decision: Within the prediction time domain length N, the system is in the current state Using these as initial conditions, and based on a multi-objective comprehensive evaluation function and engineering constraints, a rolling prediction of the trajectory state is performed: in, This represents the orbital state prediction model used for model predictive control.

[0051] For each candidate control sequence The system constructs a multi-objective integrated cost function in the prediction time domain: in: Predicting the target deviation in the well section; : The degree of sycophancy; Energy loss due to friction around the wellbore; Average mechanical drilling speed; Tool face angle change; : Dynamic weighting coefficient, whose value is adaptively adjusted based on the evaluation of the effects of similar historical cases.

[0052] The evaluation function performs a preliminary evaluation of the candidate orbitals and selects the candidate control sequence with better evaluation indicators as the initial control sequence for continuous optimization.

[0053] To further improve the optimization accuracy of candidate trajectories to meet the real-time requirements of drilling and to efficiently handle continuous control variables, this invention uses the selected candidate control sequences as initial values ​​and employs gradient descent, sequential quadratic programming, or other continuous optimization methods as online optimization solvers to continuously optimize the control sequences. Taking gradient descent as an example, this method approximates the optimal solution iteratively. The specific steps are as follows: 1. Initialization: Set the initial iteration count q=0. Guess the value using a feasible initial control sequence. Start the iteration; the initial value can be set to continue the optimal control sequence from the previous cycle.

[0054] 2. Iterative optimization a. Gradient calculation: In the q-th iteration, calculate the objective function. Relative to the current control sequence gradient The gradient indicates the direction in which the objective function decreases most rapidly.

[0055] b. Gradient Update: Update the control sequence along the inverse direction of the gradient (i.e., the descent direction of the objective function): in, The learning rate for the q-th iteration is used to control the update magnitude. Its value can be adaptively adjusted according to the iteration process or decayed according to a predetermined rule to balance convergence speed and stability.

[0056] 3. Constraint Handling (Projected Gradient Method): After each gradient update, the resulting control sequence... This may violate the engineering constraints defined in step (3). To ensure the constructability of the solution, the projected gradient method is used. Projecting back into the feasible region: in, For projection operators, This represents the set of feasible solutions comprised of all linear and nonlinear engineering constraints. The projection operation ensures that the updated sequence satisfies all constraints.

[0057] 4. Convergence Criterion: Calculate the change in the objective function value before and after the update. .like Less than the preset convergence threshold Or the number of iterations q reaches the preset maximum value. If the iteration terminates, the current sequence will be changed. As the optimal control sequence Output. Otherwise, let Return to step 2 and continue iterating.

[0058] 5. Instruction output and rolling execution: Obtaining the optimal sequence Then, only the first control vector is selected as the control command for the current well depth position and executed:

[0059] After executing the control command, the system acquires new drilling data, updates the track control status, and enters the next control cycle, repeating the above prediction and optimization process, thereby realizing a track control decision mechanism based on model predictive control, which involves rolling prediction, rolling optimization, and rolling execution.

[0060] S4 execution result feedback and adaptive update of model parameters.

[0061] S41 continuously collects actual drilling data during the drilling process after the trajectory control command is executed; S42 compares and analyzes the actual track status with the predicted track status, and calculates the deviation between the actual execution result and the prediction result. S43 adaptively adjusts the weight coefficients in the multi-objective comprehensive evaluation function based on the accumulated performance evaluation results, and updates the parameters of the GRU orbital state prediction model when the control cycle reaches the set threshold.

[0062] Specifically, after the trajectory control command obtained in step S3 is executed, the system continuously collects actual drilling data during drilling to evaluate the effectiveness of the control decision under real-world conditions. The collected data includes at least parameters such as well inclination angle, azimuth angle, tool face angle, drilling pressure, torque, and drilling efficiency. The system compares and analyzes the actual acquired trajectory state with the predicted trajectory state obtained in step S2, calculating the deviation between the actual execution result and the predicted result to evaluate the effectiveness and stability of the trajectory control decision within the current control cycle. The evaluation result serves as feedback information for model predictive control calculations in subsequent control cycles. In a preferred embodiment, the system can adaptively adjust relevant parameters in model predictive control based on the accumulated execution effect evaluation results. These parameters include at least the weighting coefficients in the multi-objective comprehensive evaluation function to adapt to changes in the importance of control objectives under different formation conditions and drilling states. This achieves dynamic trade-offs in multi-objective control decisions during drilling. In another preferred embodiment, the system can update the parameters of the trajectory state prediction model used for model predictive control when predetermined conditions are met. These predetermined conditions include the control cycle reaching a set threshold or a significant change in drilling conditions. The model parameters can be updated offline based on historical drilling data, or periodically if on-site computing resources permit, to improve the predictive model's adaptability to different working conditions. Through the above-mentioned execution result feedback and parameter update mechanism, the system continuously corrects the trajectory state prediction and control decisions during drilling, enabling the trajectory optimization decisions based on model predictive control to gradually improve stability, adaptability, and engineering applicability.

[0063] The directional drilling trajectory optimization method based on the GRU-MPC fusion model proposed in this invention aims to solve the core engineering problem of how to automatically and in real time select the optimal path with smooth trajectory, high mechanical drilling rate, and compliance with engineering constraints from countless possible correction paths after the actual drilling trajectory deviates from the design trajectory. Specifically, this model has direct and efficient application value in the following typical drilling scenarios: (1) Real-time Adaptive Deviation Correction Decision in Complex Formations: In actual drilling, when encountering complex formations with strong heterogeneity and alternating soft and hard surfaces, the drill bit stress and tool build-up performance will fluctuate unpredictably, leading to frequent small deviations from the trajectory. Using this model, the system can dynamically predict the trajectory development trend of the next 20-50 meters of well section based on real-time collected time-series data such as well inclination, azimuth, drilling pressure, rotational speed, and torque via a GRU network. When the prediction indicates that the deviation will exceed the allowable threshold, the MPC controller will immediately solve for an optimal combination sequence of tool face angle, well inclination angle, azimuth angle, and build-up rate from numerous candidate control schemes within milliseconds. This scheme not only ensures a smooth regression of the designed trajectory with the smallest possible puglegant margin but also maintains a relatively high mechanical drilling rate during the deviation correction process by optimizing parameters such as the sliding drilling ratio, achieving a dynamic balance between "deviation correction" and "efficient drilling," replacing the traditional manual deviation correction method that relies on the driller's experience and has a delayed response.

[0064] (2) Multi-objective trajectory smoothing control in long horizontal drilling sections: In long horizontal drilling sections of horizontal wells or extended reach wells, the core objective of trajectory control is to maintain the stability of the wellbore's absolute attitude and the absolute smoothness of the trajectory, so as to reduce frictional torque and prevent complex downhole conditions. In this scenario, the model can adaptively configure the weights of the multi-objective evaluation function, significantly improving the weights of "trajectory smoothness" and "control stability". Through MPC rolling optimization, the system can proactively plan a "trajectory" with a gradual change in tool face angle and an extremely low absolute build-up rate. For example, when the measurement while drilling shows a slow leftward drift in the azimuth angle, the model will not wait for the deviation to accumulate, but will calculate and execute a small, continuous right-handed tool face correction command several meters in advance, achieving a smooth, gradual heading maintenance similar to "automatic navigation", effectively ensuring high-quality and safe drilling in ultra-long horizontal sections.

[0065] (3) Dynamic target tracking and optimization during geological steering: In geological steering drilling, the designed trajectory is often updated in real time according to the formation model, and the target window may be dynamically adjusted. This model can work in deep collaboration with the geological steering system. When the geologist redetermines a better underground target point based on logging-while-drilling data, the coordinates of the new target point and the designed trajectory will be input into the system immediately. The GRU-MPC model will take the current bottom hole state as the starting point, regard the new trajectory as a special "correction path", and quickly re-optimize the global system. Under the premise of satisfying formation boundary constraints, it can generate a steering path that takes into account "target accuracy", "wellbore quality" and "drilling efficiency" within seconds, and decompose it into executable control commands to guide the drill bit to drill accurately and economically to the new geological target, which greatly improves the intelligence level and decision-making efficiency of steering operations.

[0066] (4) Predictive Risk Warning and Preventive Control: The GRU prediction module of this model has the ability to predict future trajectory states and engineering parameters (such as expected dogleg degree and tool face load). Therefore, its application can go beyond "post-mortem correction" and be upgraded to "pre-mortem warning". During drilling, the system continuously assesses the construction risks of a future well depth. For example, when the prediction results show that a well section with a dogleg degree close to the engineering limit or requiring frequent and significant tool face adjustments is about to be entered, the system can issue an early warning to the engineer. At the same time, the MPC optimizer can initiate preventive control strategies before the risky well section actually arrives, such as adjusting drilling parameters in advance or optimizing the stress on the bottom drill string assembly, to pass through the risky area in a smoother and safer manner, minimizing the risks of keyway and stuck drill bit, and achieving a leap from passive response to active optimization.

[0067] By applying it in the aforementioned core scenarios, this GRU-MPC fusion model closely combines the advantages of data-driven prediction and model predictive control, providing a real-time, forward-looking, adaptive, and multi-objective intelligent trajectory control solution for directional drilling, significantly improving the drilling accuracy, efficiency, and safety of wells with complex structures.

[0068] The above-disclosed embodiments are merely preferred embodiments of the directional drilling trajectory optimization method based on the GRU-MPC fusion model of the present invention. Of course, they should not be construed as limiting the scope of the present invention. Those skilled in the art can understand that all or part of the processes of the above embodiments can be implemented, and equivalent changes made in accordance with the claims of the present invention are still within the scope of the invention.

Claims

1. A method for optimizing directional drilling trajectories based on a GRU-MPC fusion model, characterized in that, Includes the following steps: At the current well depth, obtain the drill bit trajectory state parameters and construct a state vector representing the wellbore trajectory attitude and drilling state. Construct an orbital state prediction model based on historical orbital state sequences; Track candidate generation and rolling optimization decision based on constructability constraints; Execution result feedback and adaptive update of model parameters.

2. The directional drilling trajectory optimization method based on the GRU-MPC fusion model as described in claim 1, characterized in that, The step of "obtaining drill bit trajectory state parameters at the current well depth and constructing a state vector representing the wellbore trajectory attitude and drilling state" includes the following steps: Collect drilling directional and logging-while-drilling data, surface logging data, geological and engineering design parameters, and historical well trajectory data of the same formation; The collected raw data is time-aligned, outlier removed, and filtered and smoothed. The trajectory parameters are resampled based on a unified well depth step size to construct a set of state parameters that characterize the current wellbore trajectory control.

3. The directional drilling trajectory optimization method based on the GRU-MPC fusion model as described in claim 1, characterized in that, The "Construction of an Orbit State Prediction Model Based on Historical Orbit State Sequences" includes the following steps: The GRU trajectory state prediction model is trained offline based on historical well measurement-while-drilling data to learn the evolution of trajectory state under different drilling conditions. During the drilling application phase, the model parameters remain fixed or are periodically updated based on feedback information; Whenever new drilling data points are acquired, the historical state sequence is updated and the GRU prediction model is called again to generate a new prediction state sequence, thus realizing the rolling update of the prediction process.

4. The directional drilling trajectory optimization method based on the GRU-MPC fusion model as described in claim 1, characterized in that, The "Trajectory Candidate Generation and Rolling Optimization Decision Based on Constructability Constraints" includes the following steps: Rolling predictions of the predicted orbital state are performed for multiple sets of candidate control sequences within the prediction time domain. The weight coefficients of each sub-objective in the multi-objective comprehensive evaluation function are dynamically adjusted according to the current drilling conditions or historical performance to balance the relationship between trajectory smoothness, correction efficiency and drilling efficiency, and to achieve adaptive decision-making under different working conditions.

5. The directional drilling trajectory optimization method based on the GRU-MPC fusion model as described in claim 1, characterized in that, The "Execution Result Feedback and Model Parameter Adaptive Update" section includes the following steps: After the trajectory control command is executed, actual drilling data is continuously collected during the drilling process; The actual track status is compared and analyzed with the predicted track status, and the deviation between the actual execution result and the prediction result is calculated. The weight coefficients in the multi-objective comprehensive evaluation function are adaptively adjusted based on the accumulated performance evaluation results, and the parameters of the GRU orbital state prediction model are updated when the control cycle reaches the set threshold.

Images