Intelligent selective firing device for oil and gas well and control method

Abstract

The present application relates to the technical field of oil and gas well engineering, and particularly relates to an intelligent selection and launching device for oil and gas wells and a control method thereof, which comprises the following steps: collecting real-time posture data of a selection and launching switch through an orientation measurement module, and obtaining a multi-dimensional performance optimization target; generating a candidate execution time sequence according to the real-time posture data; for each candidate execution time, predicting an expected performance vector generated by the execution of the candidate execution time on the multi-dimensional performance optimization target through a preset random forest model; receiving the expected performance vector through a multi-objective performance evaluator, and calculating a comprehensive benefit score of each candidate execution time based on a comprehensive benefit function; selecting a candidate execution time with the highest comprehensive benefit score as an optimal execution window, and driving the selection and launching switch main body to execute a preset operation task when an actual running state enters the window. The present application can significantly improve the success rate and economic benefits of directional perforation by real-time optimization of the best execution time.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of oil and gas well engineering, in particular to an intelligent shooting device for oil and gas wells and a control method. BACKGROUND

[0002] In modern oil and gas field development, complex structure wells such as horizontal wells and extended reach wells have become the mainstream way of efficient development of reserves. In order to maximize oil and gas production, it is often necessary to accurately attack the reservoir in a specific direction of the wellbore, that is, directional perforation. The directional perforation technology delivers a downhole tool such as a perforator to a predetermined depth, and detonates the perforating charge on the perforator to a specific formation direction, that is, a tool face angle, thereby forming an oil and gas flow channel with a specific direction in the reservoir. The successful implementation of the directional perforation technology is of decisive significance for improving the seepage conditions of the reservoir, avoiding water flooding or gas channeling, and improving the single well production and ultimate recovery. The posture of the downhole tool is affected by the torsion of the drill string, the geometry of the well trajectory, the disturbance of the downhole fluid, and the stick-slip effect of the tool itself, etc. during the lowering and positioning process, resulting in the tool face angle being in a continuous dynamic change, showing irregular swinging or slow rotation. Therefore, how to accurately capture the shooting window when the tool face angle is exactly aligned with the target direction, and timely and reliably trigger the operation instruction at this moment, is the core technical problem to ensure the success of the downhole directional operation such as directional perforation. The existing directional perforation operation mainly has the following deficiencies in decision execution time. On the one hand, the current operation mode highly depends on the experience judgment of the ground operator. The operator needs to monitor the tool face angle data transmitted from the downhole to the ground in real time, and manually detonates the command according to the fluctuation of the data curve. Not only is it subject to the subjective experience and fatigue state of the operator, but also it is easy to misjudge. Moreover, there is a significant communication delay between the ground command and the actual response of the downhole tool. When the ground thinks it is the best time, the posture of the downhole tool may have deviated from the target, resulting in the perforation direction deviating from the preset target. On the other hand, although some existing semi-automatic systems try to set a fixed angle tolerance window to trigger automatically, the decision logic is too simple. Such semi-automatic systems only judge whether the current angle is qualified, but completely ignore the dynamic trend of the tool. For example, they cannot distinguish whether the tool is stably staying in the window or is sweeping through the window at high speed. Moreover, in the case of long-term instability of the downhole tool posture, this passive waiting mode often leads to an infinite extension of the operation time, greatly increasing the time cost of oil field operation. SUMMARY

[0003] The present application aims to provide an intelligent shooting device for oil and gas wells and a control method, which uses a downhole intelligent decision and execution unit to use a prediction model and an optimization algorithm to optimize the best execution time in real time, thereby maximizing the perforation accuracy, operation efficiency and dynamic stability at the same time.

[0004] The application is achieved by the following technical solutions:

[0005] An intelligent firing device for oil and gas wells, comprising a firing switch main body and an azimuth measurement module for measuring real-time attitude data of the firing switch, further comprising:

[0006] A downhole intelligent decision-making and execution unit electrically connected with the azimuth measurement module and the firing switch main body, the downhole intelligent decision-making and execution unit being configured to:

[0007] Collecting real-time attitude data output by the azimuth measurement module and obtaining multi-dimensional performance optimization targets including execution accuracy, operation efficiency and rotation stability;

[0008] Generating a candidate execution time sequence according to the real-time attitude data;

[0009] For each candidate execution time, predicting an expected performance vector generated by its execution on the multi-dimensional performance optimization targets through a prediction model;

[0010] Receiving the expected performance vector through a multi-objective performance evaluator and calculating a comprehensive benefit score of each candidate execution time based on a comprehensive benefit function corresponding to the multi-dimensional performance optimization targets;

[0011] Selecting a candidate execution time with the highest comprehensive benefit score as an optimal execution window, and driving the firing switch main body to execute a preset operation task when an actual running state enters the optimal execution window.

[0012] Optionally, the real-time attitude data includes a current tool face angle, angular velocity and angular acceleration.

[0013] Optionally, the prediction model is a random forest model, and its construction process is as follows:

[0014] Defining an input feature set of the random forest model, the input feature set including a current tool face angle, angular velocity, angular acceleration and a historical time sequence of the tool face angle;

[0015] Defining an output of the random forest model as the expected performance vector;

[0016] Setting the number of decision trees as N, and generating independent training subsets for each decision tree through random sampling with replacement;

[0017] When each node of the decision tree is split, randomly selecting a feature subset from the input feature set for optimal feature selection;

[0018] Repeat the previous step until the preset tree growth stopping condition is met, complete the construction of N decision trees, and form the random forest model.

[0019] Optionally, the training process of the random forest model is as follows:

[0020] Obtain a training data set including multiple groups of historical operation data, each group of historical operation data including an input feature set and a corresponding verification data set;

[0021] Perform data cleaning, normalization, and feature engineering preprocessing on the training data set;

[0022] Train N decision trees in the random forest model based on the preprocessed training data set, respectively;

[0023] Perform performance evaluation on the trained random forest model through the verification data set, and perform hyperparameter optimization on the tree number, tree depth, and node split standard of the random forest model until the prediction accuracy of the random forest model meets the preset standard, and complete the training.

[0024] Optionally, the downhole intelligent decision and execution unit specifically adopts a particle swarm optimization algorithm to search for the candidate execution time and determine the time with the highest comprehensive benefit score, which is specifically as follows:

[0025] Step a: initialize the particle swarm, wherein each particle represents a candidate execution time of a unit quantity and is assigned a random initial position and speed;

[0026] Step b: take the position of each particle as the candidate execution time, calculate the corresponding comprehensive benefit score as the fitness value of the particle;

[0027] Step c: compare the current fitness value of each particle with its historical optimal fitness value, and update its individual optimal position;

[0028] Step d: compare the individual optimal fitness value of all particles with the global optimal fitness value, and update the global optimal position;

[0029] Step e: update the speed and position of each particle according to the individual optimal position and the global optimal position;

[0030] Repeat steps b to e until the maximum number of iterations is met, and determine the final global optimal position as the optimal execution time.

[0031] Optionally, the formula for updating the speed and position of each particle is as follows:

[0032]

[0033]

[0034] wherein, is the index of the current iteration number, is the maximum iteration number, is the index of the particle in the particle swarm, , are the velocity and position of the i-th particle at the j-th iteration, respectively, is the index of the current iteration number, represents the candidate execution time, , are the maximum and minimum values of the inertia weight, respectively, , are the learning factors, , are random numbers between [0, 1], is the individual optimal position experienced by the i-th particle at the end of the j-th iteration, is the global optimal position experienced by the i-th particle at the end of the j-th iteration. Optionally, the comprehensive benefit function is used to calculate the comprehensive benefit score at the candidate execution time

[0035]

[0036]

[0037] wherein, is the comprehensive benefit score at time is the absolute error of the predicted tool face angle at time and the target tool face angle, is the predicted waiting time from the start of the task to time is the predicted angular velocity at time is the predicted tool face angle acceleration at time , , are the weight coefficients of accuracy, efficiency and stability, respectively, , , are positive real constants for adjusting the sensitivity of each performance component, is the satisfaction interval threshold of the angular error, , ​​​​​​​​​​ω and αω are the weight coefficients of angular velocity and angular acceleration in the evaluation of rotational instability, respectively.

[0038] Optionally, the constraint conditions of the comprehensive benefit function include: an angle absolute tolerance constraint, a rotational speed absolute constraint, and a maximum execution delay constraint.

[0039] The calculation formula of the angle absolute tolerance constraint is:

[0040]

[0041] The calculation formula of the rotational speed absolute constraint is:

[0042]

[0043] The calculation formula of the maximum execution delay constraint is:

[0044]

[0045] wherein, is a preset maximum allowed angle error, is a predicted angular velocity, is a preset maximum allowed angular velocity, is a preset maximum allowed job waiting time.

[0046] A control method of an intelligent firing device for oil and gas wells, which is applied to the intelligent firing device for oil and gas wells, and the steps of the method include:

[0047] Collecting real-time posture data of a firing switch through an orientation measurement module, and obtaining multi-dimensional performance optimization targets including execution accuracy, job efficiency, and rotational stability;

[0048] Generating a candidate execution time sequence according to the real-time posture data;

[0049] For each candidate execution time, predicting an expected performance vector generated by the execution of the candidate execution time on the multi-dimensional performance optimization targets through a preset random forest model;

[0050] Receiving the expected performance vector through a multi-objective performance evaluator, and calculating a comprehensive benefit score of each candidate execution time based on a comprehensive benefit function;

[0051] Selecting a candidate execution time with the highest comprehensive benefit score as an optimal execution window, and driving a firing switch body to execute a preset job task when an actual running state enters the window.

[0052] Optionally, the preset job task is to detonate a perforating bullet to form a perforating hole in a specified direction of a wellbore.

[0053] The technical scheme of the present application has at least the following advantages and beneficial effects:

[0054] On the one hand, the present application upgrades the past decision mode relying on human brain judgment or simple logic gate to a forward-looking decision based on a random forest model, which can accurately predict the angle, angular velocity and other dynamic evolution trends of the tool in a future period of time based on the real-time posture data of the tool, and fundamentally overcomes the limitations of communication delay and human experience. On the other hand, the present application creatively constructs the decision process as a multi-objective optimization problem, and uses a particle swarm optimization algorithm to globally search for an optimal solution in the predicted time series, instead of passively waiting for a qualified opportunity, but actively seeking a golden moment that maximizes the comprehensive benefits. The comprehensive benefit function set by the present application integrates the three core demands of perforating accuracy, operation efficiency and dynamic stability, ensuring that each decision is the best choice that can be achieved under the current conditions. BRIEF DESCRIPTION OF DRAWINGS

[0055] Figure 1 A principle schematic diagram of an intelligent selection and launching device for oil and gas wells is provided for the present application;

[0056] Figure 2 A flowchart schematic diagram of a control method of the intelligent selection and launching device for oil and gas wells is provided for the present application. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical scheme and advantages of the embodiments of the present application clearer, the technical scheme in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. The components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0058] The present application proposes an intelligent selection and launching device for oil and gas wells and a control method thereof, aiming to revolutionarily solve the fundamental problem of execution timing decision in downhole directional operation, especially in directional perforation. The traditional operation mode highly depends on the experience judgment of ground operators and is subject to significant communication delay, resulting in difficulty in ensuring perforating accuracy; and the existing simple automatic logic often leads to low operation efficiency or mis-triggering in unstable state due to the lack of prediction ability of tool dynamic trend. The present application deploys an advanced prediction model and a multi-objective optimization algorithm in the downhole tool in advance, and constructs a closed-loop intelligent system capable of autonomous perception, forward-looking decision and continuous self-optimization, so as to actively seek and capture a golden execution window considering accuracy, efficiency and stability without human intervention, and realize the intelligentization and optimization of downhole operation.

[0059] Embodiment one

[0060] As Figure 1 shown, the embodiment provides a complete system architecture and working principle of an intelligent firing device for oil and gas wells. The intelligent firing device is designed to be seamlessly integrated into conventional perforating, bridge plug co-operation or other downhole tool strings that require precise azimuthal control. Its physical structure mainly includes a firing switch main body, an azimuth measurement module and a downhole intelligent decision and execution unit as the core of the invention.

[0061] In this embodiment, the firing switch main body is the final execution mechanism of the command in the device. Its core function is to be able to conduct a large current or output a specific form of high voltage pulse instantaneously after receiving the trigger command from the downhole intelligent decision and execution unit. It can be understood that the output electrical signal can be directly used to drive various types of operating devices connected at the end. In directional perforating applications, the electrical signal can be directly used to detonate electric blasting caps, magnetic blasting caps or digital electronic blasting caps, thereby completing the precise initiation of the perforating charge. The design of the firing switch main body fully considers the extremely harsh working environment downhole. All electronic components inside, including power switches, capacitors, controllers, etc., are selected from industrial-grade or higher-grade military-grade high-temperature-resistant, high-pressure-resistant and strong-impact-resistant models. The overall circuit is packaged in a solid metal shell and protected by mature sealing technology and efficient damping structure to ensure its high reliability and long-term stability during the entire operation process.

[0062] The azimuth measurement module is responsible for real-time sensing of the attitude of the firing switch itself. It is internally sealed with a set of high-precision inertial measurement units, which at least include a three-axis accelerometer and a three-axis magnetometer. The three-axis accelerometer is used to measure the orientation of the tool relative to the earth's gravitational field under static or quasi-static conditions. By calculating the inverse tangent of the three axial gravity components, the tool's inclination angle and gravity tool face angle can be accurately obtained. The three-axis magnetometer is used to measure the orientation of the tool relative to the earth's magnetic field. By combining the inclination angle and gravity tool face angle data, the tool's magnetic azimuth angle and magnetic tool face angle can be calculated through coordinate transformation.

[0063] Specifically, during the actual movement of the downhole tool, it can be understood that due to factors such as well trajectory changes and cable twisting, the tool will rotate or swing irregularly. At this time, the azimuth measurement module continuously collects three-axis acceleration and three-axis magnetic field data at a preset high sampling frequency, which can be understood as 100 Hz or higher. After preliminary processing such as filtering and coordinate conversion of these raw data streams inside the module, a series of key parameter streams describing the dynamic attitude of the tool are generated and output in real time. These real-time attitude data form the basis for downhole intelligent decision-making, and their specific content includes: current tool face angle, tool face angle velocity, and tool face angle acceleration.

[0064] In the present embodiment, the downhole intelligent decision-making and execution unit is the core control hub of the present application, is the downhole brain for realizing intelligent decision-making, and is responsible for completing all logical operations from data processing, future prediction to optimal decision-making. The unit is physically manifested as a solid and durable downhole controller, the core of which is one or more high-performance, low-power, high-temperature-resistant digital signal processors (DSPs) or field programmable gate arrays (FPGAs). The controller is equipped with a non-volatile memory, which can be understood as a flash memory, for solidifying and storing core algorithm programs and pre-trained prediction models; and is also equipped with a high-speed random access memory (RAM) for temporarily storing real-time collected data and complex intermediate calculation results during operation. The unit is closely electrically connected with the orientation measurement module and the firing switch main body through an internal high-speed data bus, forming a high-efficiency downhole information processing and instruction execution hub. The internal firmware can be logically divided into:

[0065] The data acquisition and preprocessing module is the starting point of the intelligent decision-making process, responsible for receiving the original attitude data stream from the orientation measurement module and converting it into standardized features suitable for the prediction model. Its processing flow is as follows:

[0066] The received high-frequency data stream is cleaned in real time. By applying median filtering, Kalman filtering, and other algorithms, abnormal noise spikes caused by downhole severe vibration, electromagnetic interference, or sensor itself are effectively filtered out, ensuring the purity of the data source.

[0067] The cleaned data is normalized. Since the collected physical quantities such as angle, angular velocity, and angular acceleration have different dimensions and value ranges, direct use will affect the training efficiency and prediction accuracy of the prediction model. This step uniformly scales these parameters of different dimensions to a standard numerical interval, which can be understood as [0, 1] or [-1, 1], thereby eliminating the dimension difference.

[0068] Feature engineering is performed to extract deeper dynamic information. This module extracts statistical features such as mean, variance, standard deviation, trend slope, kurtosis, and skewness that better reflect the tool's dynamic behavior and trends by applying sliding window calculation to the tool face angle data over a period of time.

[0069] The random forest model module is the key to realizing "forward-looking" decision-making, with a pre-trained random forest model embedded inside. As an advanced ensemble learning algorithm, random forest is composed of a large number of different decision trees. The final prediction result is determined by the output of all decision trees through voting or averaging, so that the random forest model has high prediction accuracy and excellent robustness, and can effectively resist overfitting.

[0070] The input feature set of the random forest model includes the tool face angle, tool face angle velocity, tool face angle acceleration at the current time, and the tool face angle history time series and its statistical features generated by the pre-processing module, containing the past several seconds or tens of seconds.

[0071] For the future series of candidate execution time generated by the downhole intelligent decision and execution unit , the output of the model is a multi-dimensional expected performance vector. This vector accurately quantifies the key performance indicators that are expected to be generated if the job is executed at time . Specifically, the expected performance vector at least includes:

[0072] Expected angle error : the absolute difference between the actual tool face angle and the target tool face angle at the time of execution .

[0073] Expected waiting time : the time span from the current time to the candidate execution time .

[0074] Expected angular velocity : the instantaneous tool face angle velocity at time .

[0075] Expected angular acceleration : the instantaneous tool face angle acceleration at time .

[0076] The construction and training process of the random forest model is completed on the ground using powerful computing resources, and then the trained model is solidified into the downhole unit. The process follows the standard machine learning process, and the specific steps are as follows:

[0077] Obtain a training data set including multiple sets of historical job data. Historical job data is derived from a large number of field measurements, or generated by high-fidelity physical simulation software. Each set of historical job data contains a complete sequence of input feature sets, and the corresponding true performance results verified or simulated accurately after the fact, i.e. the validation data set.

[0078] Data cleaning, normalization and feature engineering preprocessing are performed on the collected training data set, and the same algorithm and parameters as the pre-processing module of the downhole unit are used to ensure data consistency when the model is trained and applied.

[0079] The N decision trees in the random forest model are trained based on the pre-processed training dataset. During the training process, for each decision tree, an independent training subset is generated from the total training set by random sampling with replacement. When splitting at an internal node of each decision tree, the algorithm does not search for the optimal split point among all input features, but first randomly selects a feature subset from the total input feature set, and then selects the optimal feature from this subset for splitting. The double randomness of sample and feature is the key to the robustness and high performance of the random forest algorithm. This process is repeated until the preset tree growth stopping condition is met, which can be understood as reaching the maximum depth or the number of node samples being less than a threshold, and finally the construction of N decision trees is completed to form a complete random forest model.

[0080] Finally, the trained random forest model is comprehensively evaluated by the validation dataset. Techniques such as cross-validation are used to systematically search or Bayesian optimize the key hyperparameters of the random forest model, such as the number of decision trees N, the maximum depth of each tree, and the node splitting criteria (such as Gini impurity or information gain), until the prediction accuracy indicators of the random forest model, such as root mean square error or determination coefficient, meet the pre-set engineering application standards. At this time, the model training is completed.

[0081] In this embodiment, the multi-objective performance evaluator works with the optimal execution time searching module to form the core engine of the decision. Its task is to find the optimal execution time that maximizes the overall benefit among all future possibilities given by the prediction model through mathematical evaluation and optimization search.

[0082] The multi-objective performance evaluator has a built-in configurable comprehensive benefit function , which is used to quantitatively score the comprehensive score of each candidate execution time . The design of this function aims to unify the three core performance optimization targets of execution accuracy, job efficiency and rotation stability, which are mutually restrictive in actual work, into a single scalar score, and its calculation formula is:

[0083]

[0084] Among them, is the comprehensive benefit score at time , the higher the score, the better the time, is the absolute error of the predicted tool face angle at time and the target tool face angle, is the predicted waiting time from the start of the task to time , is the predicted angular velocity at time , the predicted tool face angle acceleration at time , , , are the weight coefficients for accuracy, efficiency and stability, respectively, , , are positive real constants for adjusting the sensitivity of each performance component, is the satisfaction interval threshold of the angle error, , are the weight coefficients for angular velocity and angular acceleration in the evaluation of rotational instability, respectively.

[0085] In this embodiment, the upper half of the comprehensive benefit function represents the accuracy score, which is an S-shaped function. When the predicted angle error is less than or equal to the preset satisfaction interval threshold , the score is close to the weight ; when the error increases and exceeds , the score will rapidly decrease with the growth of the exponential term, and the parameter is used to adjust the sensitivity of the function to the error exceeding the limit, the larger , the steeper the score decreases. The left structure of the lower half of the comprehensive benefit function represents the efficiency score. As the waiting time increases, the denominator slowly increases, resulting in a decrease in the efficiency score, thereby imposing a nonlinear penalty on excessively long waiting time, and the parameter adjusts the sensitivity to time. The right structure of the lower half of the comprehensive benefit function represents the stability score, and sech is the hyperbolic secant function, which is shaped like a bell curve, taking the maximum value 1 when the independent variable is 0. The independent variable here is the square root of the weighted square sum of the angular velocity and the angular acceleration , which comprehensively reflects the rotational instability of the tool. When the tool tends to be stationary, i.e., the angular velocity and the angular acceleration approach 0, the score of this term is the highest, equal to the weight ; when the tool rotates or shakes violently, the score rapidly decays to 0, and the parameter controls the speed of decay, , are used to balance the relative importance of angular velocity and angular acceleration in the evaluation of instability. Specifically, , , are the three core weight coefficients, and their sum is usually normalized to 1. The operator can configure according to the task requirements (e.g., accuracy first for thin reservoirs, and efficiency first for time-critical operations) before operation, thereby flexibly adjusting the focus of the decision.

[0086] Before calculating the comprehensive benefit score, the system first performs constraint checking on each candidate execution time. The constraints set in this embodiment are the safety and quality bottom line that must be met, and any time that violates the constraints will be directly eliminated and will not participate in the subsequent benefit evaluation. The constraint conditions specifically include:

[0087] Absolute angle tolerance constraint: the predicted absolute error of the angle must be less than or equal to a preset maximum allowed angle error . The calculation formula is: .

[0088] Absolute rotation speed constraint: the predicted absolute value of the angular velocity must be less than or equal to the preset maximum allowed angular velocity to prevent false triggering when the tool rotates at high speed across the target window. The calculation formula is: .

[0089] Maximum execution delay constraint: the predicted waiting time must be less than or equal to the preset maximum allowed job waiting time to avoid endless waiting due to continuous instability of the well conditions. The calculation formula is: .

[0090] Specifically, in order to efficiently search for the time with the highest comprehensive benefit score in the complex, nonlinear, and possibly multi-local optimal solution multidimensional search space defined by the comprehensive benefit function, this embodiment specifically uses the particle swarm optimization algorithm. The implementation logic in the downhole intelligent decision and execution unit is specifically:

[0091] Step a: initialize the particle swarm. When the algorithm starts, a particle swarm is generated in memory, where each particle represents a unit quantity of candidate execution time and is assigned a random initial position (i.e., a specific candidate time value) within the preset search time range and a random initial speed.

[0092] Step b: fitness evaluation. In each iteration, the position of each particle is taken as the candidate execution time, which is substituted into the prediction model and the comprehensive benefit function to calculate the corresponding comprehensive benefit score, and this score is taken as the fitness value of the particle.

[0093] Step c: update the individual optimal position. For each particle, compare its current fitness value with its historical optimal fitness value. If the current value is better, update its individual optimal position.

[0094] Step d: Update the global optimal position. Compare the individual optimal fitness values ​​of all particles with the currently recorded global optimal fitness value. If any particle's individual optimal fitness value exceeds the current global optimal fitness value, then update the global optimal position.

[0095] Step e: Update velocity and position. Based on the individual optimal position and the global optimal position, update the velocity and position of each particle according to a set formula, guiding the entire particle swarm to move towards a better solution region. The formula for updating the velocity and position of each particle is as follows:

[0096]

[0097]

[0098] in, This is the index of the current iteration number. The maximum number of iterations, This is the index of the particle in the particle swarm. , The first The particle in the first Velocity and position at the next iteration Represents the candidate execution time. , These are the maximum and minimum values ​​of the inertia weight, respectively. , These are learning factors, , These are random numbers between [0,1]. For the first The particle in the first The optimal position of an individual at the end of the next iteration. For the first The particle in the first The globally optimal position reached at the end of the iteration.

[0099] Repeat steps b to e until the termination conditions such as the maximum number of iterations or no significant improvement in the global optimal solution over multiple consecutive iterations are met. Finally, the obtained global optimal position is determined as the optimal execution time.

[0100] Further, the embodiment is designed with self-adaptive online optimization mechanism to make the system have the ability of self-learning and self-evolution. After each preset job task is completed, the module will automatically record the expected performance vector of this decision and the real performance vector obtained through high-frequency data analysis afterwards. This pair of data composed of "expected-real" is stored in the non-volatile memory of the downhole intelligent decision and execution unit as a new and valuable incremental training sample. It can be understood that when the stored new samples reach a predetermined number or receive a ground instruction, the module will call these incremental sample pairs to update or retrain the random forest model in the downhole. Through this mechanism, the decision model can continuously learn from the actual operation and continuously adapt to the dynamic characteristics of the current well section to achieve self-adaptive performance improvement.

[0101] Embodiment Two

[0102] As shown in Figure 2 , the embodiment describes a complete control method for a typical directional perforation operation using the above-mentioned intelligent selection and firing device.

[0103] Step One: System Downhole and Target Setting

[0104] The perforating tool string integrated with the intelligent selection and firing device of the present application is lowered into the well through the cable to the predetermined operation interval. The surface system supplies power to the downhole tool through the cable and establishes communication. At this time, the downhole intelligent decision and execution unit is activated, and the real-time attitude data of the selection and firing switch is collected by the high-frequency azimuth measurement module inside. The ground operator can set and issue multi-dimensional performance optimization targets including execution accuracy, operation efficiency and rotation stability according to the specific requirements of this operation, such as reservoir thickness, operation timeliness requirements, etc. This is specifically embodied in the configuration of the weight coefficients 、 、 in the comprehensive benefit function and the hard constraint parameters.

[0105] Step Two: Generate Candidate Execution Time Sequence

[0106] After receiving the start instruction, the downhole intelligent decision and execution unit generates a discrete candidate execution time sequence within a preset time window, which can be understood as 30 seconds, in the future, with a fixed time step, which can be understood as 0.1 seconds, according to the current time.

[0107] Step Three: Predict Expected Performance Vector

[0108] For each candidate execution moment in the sequence, the downhole intelligent decision and execution unit calls its internally preset random forest model. The model takes the current and historical pose data of the tool as input, and performs forward-looking calculation for each candidate moment to predict the expected performance vector generated by its execution on the multi-dimensional performance optimization target.

[0109] Step four: optimization to determine the optimal execution window.

[0110] The downhole intelligent decision and execution unit starts its internal multi-objective performance evaluator. The evaluator receives all predicted expected performance vectors, and based on the comprehensive benefit function issued from the ground, calculates the comprehensive benefit score of each candidate execution moment that meets the hard constraints. Next, the optimal execution moment searching module is started, which uses the particle swarm optimization algorithm to perform efficient global search among all candidate moments, and finally selects the candidate execution moment with the highest comprehensive benefit score as the only, optimal execution window for this operation.

[0111] Step five: autonomous execution of the operation task.

[0112] The downhole intelligent decision and execution unit enters a high-precision monitoring state, continuously monitoring the real-time clock inside the well. When the actual running state, i.e. the system time, accurately enters the optimal execution window determined in the previous step, the unit immediately and autonomously sends a high-level trigger signal to the selected firing switch body. The selected firing switch body responds to this signal at the nanosecond or microsecond level, and completes the ignition operation instantaneously. In directional perforation applications, this operation will detonate the perforating charge to form a precise perforation hole in the specified direction of the wellbore, thereby completing a high-precision, high-efficiency, and high-stability preset operation task.

[0113] In summary, the present application deploys complex prediction models and optimization algorithms in downhole tools in advance, and builds an intelligent operation system that can autonomously perceive, make forward-looking decisions, and continuously evolve. It fundamentally overcomes the drawbacks of traditional operation methods, and provides a powerful technical means for efficient, precise, and intelligent development of oil and gas fields. The ideas and solutions of the present application are not limited to directional perforation, but can also be applied to any downhole operation scenario that requires precise timing control, and have broad application prospects.

[0114] The above is only the preferred embodiment of the present application, and is not intended to limit the present application. For those skilled in the art, the present application can have various modifications and changes. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. An intelligent firing selection device for oil and gas wells, comprising a firing selection switch body and an azimuth measurement module, wherein the azimuth measurement module is used to measure the real-time attitude data of the firing selection switch, characterized in that, Also includes: The downhole intelligent decision-making and execution unit is electrically connected to the azimuth measurement module and the selective switch body, respectively. The downhole intelligent decision-making and execution unit is configured as follows: The system collects real-time attitude data output by the orientation measurement module and obtains multi-dimensional performance optimization targets, including execution accuracy, operation efficiency, and rotational stability. Based on the real-time attitude data, a candidate execution time sequence is generated; For each candidate execution time, the expected performance vector that it would generate for the multidimensional performance optimization objective is predicted by the prediction model. The multi-objective performance evaluator receives the expected performance vector and calculates the comprehensive benefit score for each candidate execution time based on the comprehensive benefit function corresponding to the multi-dimensional performance optimization objective. The candidate execution time that achieves the highest comprehensive benefit score is selected as the optimal execution window, and when the actual operating state enters the optimal execution window, the main body of the selection switch is driven to execute the preset operation task. The comprehensive benefit function is used to calculate the efficiency at candidate execution times. The comprehensive benefit score is calculated using the following formula: in, For at any time The overall benefit score, For the prediction at time The absolute error between the tool face angle and the target tool face angle. For the prediction from the start of the task to time... The waiting time For the prediction at time angular velocity, For the prediction at time Tool face angular acceleration, , , These are the weighting coefficients for accuracy, efficiency, and stability, respectively. , , These are positive real constants used to adjust the sensitivity of each performance component. This is the threshold value for the satisfactory range of angle error. , These are the weighting coefficients for angular velocity and angular acceleration in the assessment of rotational instability, respectively.

2. The intelligent selective firing device for oil and gas wells according to claim 1, characterized in that, The real-time attitude data includes: current tool face angle, angular velocity, and angular acceleration.

3. The intelligent selective firing device for oil and gas wells according to claim 2, characterized in that, The prediction model is specifically a random forest model, and its construction process is as follows: Define the input feature set of the random forest model, which includes the current tool face angle, angular velocity, angular acceleration, and the historical time series of the tool face angle; Define the output of the random forest model as the expected performance vector; Set the number of decision trees to N, and generate an independent training subset for each decision tree through random sampling with replacement; When splitting at a node of each decision tree, a subset of features is randomly selected from the input feature set for optimal feature selection; Repeat the previous step until the preset tree growth stopping condition is met, and complete the construction of N decision trees to form the random forest model.

4. The intelligent selective firing device for oil and gas wells according to claim 3, characterized in that, The training process of the random forest model is as follows: Obtain a training dataset that includes multiple sets of historical job data. Each set of historical job data includes an input feature set and a corresponding validation dataset. The training dataset is preprocessed by data cleaning, normalization, and feature engineering. The N decision trees in the random forest model are trained based on the preprocessed training dataset. The performance of the trained random forest model was evaluated using a validation dataset. Hyperparameters such as the number of trees, tree depth, and node splitting criteria were tuned until the prediction accuracy of the random forest model met the preset standards, thus completing the training.

5. The intelligent selective firing device for oil and gas wells according to claim 4, characterized in that, The downhole intelligent decision-making and execution unit specifically employs a particle swarm optimization algorithm to search for candidate execution times and determine the time that yields the highest overall benefit score. Specifically: Step a: Initialize the particle swarm, where each particle represents a unit number of candidate execution times and is assigned a random initial position and velocity; Step b: Take the position of each particle as a candidate execution time and calculate its corresponding comprehensive benefit score as the particle's fitness value; Step c: Compare the current fitness value of each particle with its historical best fitness value, and update its individual best position; Step d: Compare the individual optimal fitness values ​​of all particles with the global optimal fitness values, and update the global optimal position; Step e: Update the velocity and position of each particle based on the individual optimal position and the global optimal position; Repeat steps b through e until the maximum number of iterations is met, and then determine the final globally optimal position as the optimal execution time.

6. The intelligent selective firing device for oil and gas wells according to claim 5, characterized in that, The formula for updating the velocity and position of each particle is as follows: in, This is the index of the current iteration number. The maximum number of iterations, This is the index of the particle in the particle swarm. , The first The particle in the first Velocity and position at the next iteration Represents the candidate execution time. , These are the maximum and minimum values ​​of the inertia weight, respectively. , These are learning factors, , These are random numbers between [0,1]. For the first The particle in the first The optimal position of an individual at the end of the next iteration. For the first The particle in the first The globally optimal position reached at the end of the iteration.

7. The intelligent selective firing device for oil and gas wells according to claim 6, characterized in that, The constraints of the comprehensive benefit function include: absolute angle tolerance constraint, absolute rotation speed constraint, and maximum execution delay constraint. The formula for calculating the absolute tolerance constraint of the angle is as follows: The absolute constraint on rotational speed is calculated using the following formula: The maximum execution latency constraint is calculated using the following formula: in, The preset maximum allowable angle error, For the predicted angular velocity, The preset maximum permissible angular velocity, This is the preset maximum allowed job waiting time.

8. A control method for an intelligent selective firing device for oil and gas wells, characterized in that, It is applied to the intelligent selective firing device for oil and gas wells as described in any one of claims 1-7, and the method includes the following steps: The orientation measurement module collects real-time attitude data of the selective firing switch and obtains multi-dimensional performance optimization targets including execution accuracy, operation efficiency and rotational stability. Based on the real-time attitude data, a candidate execution time sequence is generated; For each candidate execution time, a pre-set random forest model is used to predict the expected performance vector that would be generated for the multi-dimensional performance optimization objective if it were executed. The expected performance vector is received by the multi-objective performance evaluator, and the comprehensive benefit score for each candidate execution time is calculated based on the comprehensive benefit function. The candidate execution time that achieves the highest overall benefit score is selected as the optimal execution window, and when the actual operating state enters this window, the selector switch body is driven to execute the preset task.

9. The control method for the intelligent selective firing device for oil and gas wells according to claim 8, characterized in that, The preset task is to detonate the perforating shell to form a perforation hole at a designated location in the wellbore.

Images