A humanoid robot collaborative control method based on embodied intelligence

Abstract

The application discloses a kind of based on somatic intelligence human-shaped robot cooperative control method, belong to somatic intelligence and human-shaped robot control technical field, for human-shaped robot to the grabbing and handover operation of goods on dynamic conveying equipment, this method obtains the preliminary six-dimensional pose estimation value of goods according to visual data, obtains its instantaneous speed and acceleration estimation value according to conveying equipment encoder pulse signal, and then generates the cooperative grabbing and handover trajectory with time stamp, obtains end desired acceleration instruction according to trajectory path point and end contact force data, obtains zero-moment point constraint condition according to plantar pressure data, the whole body joint torque instruction is obtained by solving secondary planning to the joint execution of both, and is based on simulation environment pre-training strategy and real environment contact feedback deviation, strategy parameters are corrected by online residual error learning network, the application can improve the grabbing success rate and handover flexibility of human-shaped robot in dynamic conveying scene.

Description

Technical Field

[0001] This invention relates to the field of embodied intelligence and humanoid robot control technology, specifically to a humanoid robot collaborative control method based on embodied intelligence. This method is applicable to the control process of humanoid robots in smart warehousing, advanced logistics equipment, and high-end intelligent manufacturing equipment scenarios, for real-time collaborative grasping, handling, and handover of goods moving on dynamic conveying equipment. In particular, it relates to a collaborative control method that integrates visual, tactile, and proprioceptive information to achieve dynamic pose prediction, joint solution of whole-body dynamics, and cross-scenario adaptive generalization. Background Technology

[0002] With the rapid development of smart logistics and advanced manufacturing industries, material sorting and loading / unloading systems centered on conveyor lines and automated warehouses are undertaking the task of continuous flow of massive amounts of materials. These systems are typically composed of dynamic conveying devices such as conveyor belts, roller conveyors, and elevators. Goods move continuously on the conveying devices at a constant or variable speed. To achieve automated grasping, handling, and handover of goods at the end of the production line, existing technologies generally adopt two types of solutions: one is a traditional industrial robotic arm with a fixed base, which completes the grasping when the goods pass through a fixed workstation by offline calibration of the conveyor belt's motion parameters and combining a geometric kinematic model; the other is a wheeled automated guided vehicle equipped with special tooling fixtures, which completes the transfer and handover of goods based on a preset path and standardized pallets. Humanoid robots, with their workspace highly similar to humans and bipedal mobility, can flexibly switch workstations and adapt to non-standardized operating scenarios without large-scale modifications to the production line, and are considered an important development direction for flexible warehousing and manufacturing.

[0003] The closest existing technical solution to this invention is a conveyor belt collaborative gripping method based on a fixed-base robotic arm or a wheeled automated guided vehicle. Its general process is as follows: Images of goods on the conveyor belt are acquired using a vision sensor, and the time it takes for the goods to reach the gripping station is calculated based on a pre-calibrated uniform motion model of the conveyor belt. The end effector of the robotic arm moves to the predicted position according to a pre-planned fixed trajectory to perform the gripping action. After gripping, the robotic arm completes the handling and placement with the support of the reaction force provided by the fixed base. The entire process does not require consideration of the dynamic balance of the base itself. This solution comprises a vision acquisition unit, a kinematic calibration and trajectory planning unit, a robotic arm body fixed to the ground or production line support, and an end effector in the form of a gripper or suction cup. The vision acquisition unit transmits image data to the trajectory planning unit, which generates a fixed gripping trajectory based on an offline calibrated conveyor belt speed model and drives the robotic arm body and end effector to complete the gripping action.

[0004] The existing technology has the following drawbacks: Firstly, its grasping trajectory is pre-planned based on an offline calibrated conveyor belt uniform motion model. When the conveyor belt operates at variable speed or the goods are in a non-standard shape, a deviation will occur between the fixed trajectory and the actual goods pose, leading to an increased grasping failure rate. This makes it difficult to adapt to long-tail non-standardized scenarios and the needs of flexible production lines requiring rapid reconfiguration. Secondly, it is generally believed in the field that end-effector force control and bipedal dynamic balance should adopt a hierarchical independent control strategy to reduce system coupling complexity (technical bias). However, this application breaks through this inherent understanding by incorporating end-effector compliance control and bipedal zero-moment point constraints into a unified whole-body dynamics quadratic programming framework for joint solution, thereby timely offsetting end-effector nonlinear disturbances at the moment of grasping and handover. Momentum transfer caused by dynamic forces: When a humanoid robot performs dynamic load grasping under bipedal non-fixed base conditions, the nonlinear disturbance force of the conveying equipment and goods on the end effector will directly cause changes in the robot's overall momentum. If the existing isolated end control strategy of the robotic arm or the simple center of mass follow strategy is adopted, it will be impossible to counteract the impact of this momentum transfer on the bipedal support stability domain in time, which may easily lead to the robot tipping over or grasping failure, and it will also be impossible to achieve smooth handover with dynamic equipment. In addition, existing solutions usually rely on a large amount of real machine data from a single production line for strategy adjustment. Without a large amount of real machine data, it is difficult to make the control strategy quickly generalize to adapt to cross-scenario application needs of different materials, different equipment models and unknown cargo shapes. Summary of the Invention

[0005] The technical problem this invention aims to solve is: in scenarios where humanoid robots replace traditional fixed robotic arms or wheeled automated guided vehicles in dynamic conveyor line grasping, handling, and handover operations, how to resolve the strong coupling conflict between the dynamic heavy-load operation of the upper limbs and the dynamic balance of the lower limbs on a non-fixed base. Specifically, this includes: how to predict the pose changes of dynamically operating conveying equipment and non-standard goods in real time to generate a spatiotemporally synchronized collaborative grasping trajectory; how to resist the continuous nonlinear disturbance force generated by the conveying equipment and goods load on the robot's end effector at the moment of grasping and handover, avoiding instability or even tipping of the two feet due to momentum transfer; and how to enable the control strategy to quickly generalize and adapt to cross-scenario application needs of different materials, different equipment models, and unknown goods forms in the absence of massive amounts of real machine data.

[0006] To address the aforementioned technical problems, this invention provides a humanoid robot collaborative control method based on embodied intelligence. This method first synchronously acquires image data from the humanoid robot's head vision sensor, end-effector contact force data from the end-effector six-dimensional torque sensor, body state data from the body inertial measurement unit and joint encoder, and plantar pressure data from the plantar pressure sensor. It also receives encoder pulse signals from a dynamic conveying device. Subsequently, it obtains a preliminary six-dimensional pose estimate of the cargo based on the image data, and instantaneous velocity and acceleration estimates of the dynamic conveying device based on the encoder pulse signals. Based on these estimates, it generates a timestamped collaborative grasping and handover trajectory. Then, it uses the path points in the collaborative grasping and handover trajectory... The desired end-effector acceleration command is obtained from the end-effector contact force data, and the zero-moment constraint condition is obtained from the plantar pressure data. The desired end-effector acceleration command and the zero-moment constraint condition are combined to perform a quadratic programming solution to obtain the torque command of the humanoid robot's joints. This torque command is then sent to the bi-arm joint actuators and bi-foot joint actuators. Finally, basic strategy parameters are obtained through pre-training in a physical simulation environment. Based on the deviation between the actual contact feedback collected by the humanoid robot in the real environment and the expected contact feedback in the simulation environment, a strategy correction amount is obtained through an online residual learning network. The strategy correction amount is added to the basic strategy parameters to obtain the corrected strategy parameters, which are then used for the quadratic programming solution to achieve continuous online calibration of the control strategy.

[0007] The technical effects that this invention can achieve include: 1. By integrating visual, tactile and proprioceptive information, a dynamic transport flow spatiotemporal synchronous prediction mechanism is constructed to generate a timestamped collaborative grasping and handover trajectory. This enables the grasping trajectory to be dynamically updated according to the actual movement state of the transport equipment and goods, reducing the deviation between the fixed trajectory and the actual position of the goods, and improving the grasping success rate in dynamic scenarios. 2. By jointly solving the end-effector admittance control of the two arms and the zero-moment point stability domain constraint of the two feet through quadratic programming, the end-effector compliance operation and the dynamic balance of the two feet can be solved in a unified manner at the whole-body dynamics level. This can promptly offset the momentum transfer caused by the nonlinear disturbance of the end-effector at the moment of grasping and handing over, reduce the risk of robot instability and tipping due to dynamic load, and improve the compliance of the cargo handing over process. 3. By constructing a physical simulation environment that includes various materials and friction coefficients and combining it with an online residual learning network based on real tactile feedback, the control strategy can complete parameter calibration within a millisecond timescale, reducing the dependence on the collection of massive real machine data from a single production line and enabling rapid cross-scenario generalization deployment of the control strategy across different equipment models and different cargo forms. Attached Figure Description

[0008] Figure 1This is an overall flowchart of the humanoid robot collaborative control method of the present invention; Figure 2 This is a flowchart of the spatiotemporal synchronization prediction and trajectory generation process of the present invention; Figure 3 This is a flowchart of the whole-body dynamics joint control and strategy correction of the present invention. Detailed Implementation

[0009] The following is in conjunction with the appendix Figure 1-3 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other. The method described in the present invention is applied to the scenario of humanoid robots grasping and handing over goods on dynamic conveying equipment. It is executed by the control system configured on the humanoid robot. The method as a whole consists of four mutually cooperating processing processes: synchronous data acquisition, spatiotemporal synchronous prediction, whole-body dynamic joint control, and cross-scenario adaptive generalization.

[0010] During the data synchronous acquisition process, the control system simultaneously acquires image data from the humanoid robot's head vision sensor, end-effector contact force data from the end-effector six-dimensional torque sensor, angular velocity and acceleration data from the body inertial measurement unit, joint angle data from each joint encoder, and plantar pressure data from the plantar pressure sensor at fixed intervals. It also receives encoder pulse signals from the programmable logic controller of the dynamic conveying equipment in real time via a distributed communication bus. The above data are aligned with a unified timestamp to form a synchronous multimodal data frame, which serves as the input for subsequent processing. The encoder pulse signals reflect the current operating status of the dynamic conveying equipment, and their combination with the conveyor belt pulley circumference reflects the motion characteristics of the conveyor belt. The end-effector contact force data reflects the contact force state exerted by the goods and conveying equipment on the end effector. The body status data reflects the posture and joint motion state of the humanoid robot body, and the plantar pressure data reflects the contact force distribution state between the humanoid robot's feet and the ground.

[0011] In the spatiotemporal synchronous prediction process, the control system first obtains a preliminary six-dimensional pose estimate of the cargo based on image data from the synchronous multimodal data frames. Specifically, the image data is input into a pre-trained vision-language-action model. This model identifies the category and boundaries of the target cargo based on image semantic information and, combined with depth information from the image data, calculates the three-dimensional position coordinates and three-dimensional attitude angles of the target cargo in the camera coordinate system, outputting a preliminary six-dimensional pose estimate of the cargo, denoted as... ,in , , This represents the three-dimensional position components of the cargo in the camera coordinate system. , , These represent the roll angle, pitch angle, and yaw angle of the cargo, respectively.

[0012] Based on the preliminary six-dimensional pose estimation, the control system further obtains the instantaneous velocity and acceleration estimates of the dynamic conveyor based on the encoder pulse signal. In one embodiment, optical flow tracing is performed on the feature points on the conveyor belt surface in continuous image frames to obtain the instantaneous displacement vector of the feature points on the image plane. This displacement vector, together with the measured velocity value fed back by the encoder pulse signal, is used as an observation and input into a Kalman filter for state estimation. The state prediction equation of the filter is: ; in, Indicates the first Time based on the first The prior estimate of the current state from the current state at time step, the state vector containing the instantaneous velocity and acceleration of the conveyor belt. This is the state transition matrix, describing the kinematic transmission relationship between velocity and acceleration. For the first State estimate at time 10:00 To control the input matrix, For the first The control input at time t is such that the filter's state update equation is: ; in, For the first The posterior state estimate at time t. For the first The Kalman gain at time step is used to balance the reliability of the predicted value versus the observed value. For the first The actual observations at any given time include the optical flow displacement conversion velocity and the encoder pulse feedback velocity. The observation matrix is ​​used to map the state vector to the observation space. Through the above recursive calculation, the instantaneous velocity estimate of the conveying equipment within the future time window is obtained. Compared with acceleration estimates In another implementation, the control system directly establishes a high-frequency data interface with the programmable logic control system of the dynamic conveying equipment, reads the encoder pulse count, and calculates the instantaneous speed and acceleration estimates by combining the known pulley circumference. In this case, the vision module is only used for the preliminary six-dimensional pose estimation of the goods and does not need to perform optical flow tracking. This implementation is suitable for production line scenarios where the programmable logic controller communication interface is open and the real-time requirements are high.

[0013] After obtaining preliminary six-dimensional pose estimates, instantaneous velocity estimates, and acceleration estimates, the control system, combined with preset future time window parameters, constructs a predictive model of the cargo pose evolution over time and calculates future moments. Predicted pose of cargo: ; in, Indicates the predicted time The six-dimensional pose of the cargo. For a moment Instantaneous velocity component in the direction of the conveying equipment. The acceleration estimate is given by the integral term, which represents the cumulative offset of the cargo position caused by the movement of the conveying equipment. The term representing half of this offset multiplied by the product of acceleration and the square of time represents the additional impact of acceleration on the cargo position. Based on this, and combined with the current joint state of the humanoid robot, a series of time-stamped discrete path points are planned to form the collaborative grasping and handover trajectory. ,in Indicates the first The predicted timestamps corresponding to each path point are used as the input reference for the subsequent whole-body dynamic joint control process.

[0014] In the whole-body dynamics joint control process, the control system uses the path points in the cooperative grasping and handover trajectory as the reference for the desired end-effector pose, and defines the deviation between the actual end-effector pose and the desired pose as... ,in This represents the actual pose of the end effector. Given the desired pose of the trajectory, establish the end-effector admittance control equations: ; in, The pre-defined virtual inertia matrix at the end is used to adjust the response speed of the end to external impact forces. A pre-defined virtual damping matrix is ​​used to suppress end oscillations. A pre-defined virtual stiffness matrix is ​​used to constrain the degree to which the end deviates from the desired trajectory. and These represent the acceleration and velocity of the pose deviation, respectively. The end contact force data obtained by the end six-dimensional torque sensor includes the nonlinear disturbance force exerted by the conveying equipment and cargo on the end effector. By solving this equation, the desired end acceleration command that meets the compliance requirements is obtained.

[0015] Meanwhile, the control system establishes the system's centroid dynamic equation using the humanoid robot's center of mass as a reference point. Combined with the plantar pressure data determining the bipedal support region, it calculates the allowable zero-moment point boundary range of the current bipedal support polygon, thus obtaining the zero-moment point constraint conditions. ; in, The zero-moment point position is calculated in real time based on the robot's current center of mass acceleration and rate of change of angular momentum. and These are the two boundary values ​​of the current bipedal support polygon in the preset direction. This constraint is used to ensure that the robot remains within the dynamic stable domain throughout the entire grasping and handover process.

[0016] The control system combines the desired end-effector acceleration command with the zero-moment point constraint to perform a quadratic programming solution, constructing a unified optimization problem: ; The constraints are: ; in, Let be the angular acceleration vectors of all joints in the body that need to be solved. The joint space desired acceleration reference is derived from the end-admittance control equation. The output torque vectors for each joint to be solved are: and This is the corresponding weighting matrix, used to balance tracking accuracy and torque energy consumption. For joint angle The relevant whole-body inertia matrix, Coriolis force and gravity terms, The transpose of the end-effector Jacobian matrix is ​​used to map the end-effector contact force to the joint space. Solving the above optimization problem yields the whole-body joint torque command that satisfies the dynamic equilibrium constraint. The control system sends the solved joint torque command to the dual-arm and dual-foot joint actuators at a control frequency of not less than 1000 Hz to drive the humanoid robot to complete the grasping and handover actions. At the same time, the end-effector contact force data and plantar pressure data after execution are re-acquired and used as inputs for calculating the end-effector desired acceleration command and zero-moment point constraint conditions in the next control cycle, forming a closed-loop control.

[0017] In the cross-scenario adaptive generalization process, the control system pre-establishes physical property models of goods containing rigid, flexible, and fragile materials in the physical simulation engine. Multiple sets of friction coefficients, belt speed ranges, and load distribution parameters are set for different conveyor equipment models, forming a set of simulation scenarios covering various working conditions. Within this set of simulation scenarios, a pre-set machine learning training method is used to pre-train the control strategy parameters corresponding to the aforementioned whole-body dynamics joint control process, obtaining basic strategy parameters applicable to various simulation working conditions, denoted as... The pre-trained basic strategy parameters are deployed to a real humanoid robot. During contact with real conveying equipment and real goods, the feedback signal from the end effector's six-dimensional torque sensor is collected in real time, and the deviation between the measured feedback signal and the expected feedback signal under the same working conditions in the simulation environment is calculated. The deviation and the current state observation are input into a lightweight online residual learning network, which outputs the correction to the basic policy parameters at a millisecond-level update frequency. The corrected strategy parameters are: ; in, The basic policy parameters are obtained through pre-training in the simulation environment. This refers to the correction value output in real time by the residual learning network based on real-time contact feedback. The corrected strategy parameters are used in real robot control. The control system uses the corrected strategy parameters in real time to solve the aforementioned quadratic programming problem, replacing the original basic strategy parameters. In subsequent control cycles, it continuously collects real contact feedback to iteratively update the residual learning network, so as to achieve continuous generalization and adaptation to different lighting conditions, different equipment models and unknown cargo forms.

[0018] In another implementation of the whole-body dynamic joint control process, instead of using a unified quadratic programming joint solution method, a hierarchical control architecture can be adopted. That is, the end-effector admittance controller independently calculates the desired end-effector torque, and then the torque is used as an external disturbance input to the bipedal balance controller. The bipedal balance controller solves the bipedal joint torque that satisfies the zero-moment point constraint separately. By using a hierarchical rather than joint solution method, the near decoupling control of end-effector compliance and bipedal dynamic balance is achieved. This implementation method has certain applicability in scenarios where hardware computing resources are limited.

[0019] Through the coordinated efforts of the four processing steps—synchronous data acquisition, spatiotemporal synchronous prediction, whole-body dynamic joint control, and cross-scenario adaptive generalization—the method of this invention enables humanoid robots to accurately predict, stably grasp, and compliantly transfer goods in dynamic transport scenarios. This improves the success rate of grasping while reducing the risk of instability and tipping, and also has strong cross-scenario generalization capabilities.

[0020] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, any modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall fall within the protection scope defined by the claims of the present invention.

Claims

1. A humanoid robot collaborative control method based on embodied intelligence, used for humanoid robots to perform grasping and handover operations on dynamic conveying equipment, characterized in that, include: The system collects image data from the humanoid robot's head vision sensor, end contact force data from the end-effector six-dimensional torque sensor, body state data from the body inertial measurement unit, joint encoder, and foot pressure sensor, and receives encoder pulse signals from the dynamic conveying device. Based on the image data, a preliminary six-dimensional pose estimate of the cargo is obtained; Based on the encoder pulse signal, the instantaneous speed estimate and acceleration estimate of the dynamic conveying device are obtained; Based on the preliminary six-dimensional pose estimation value, the instantaneous velocity estimation value, and the acceleration estimation value, a timestamped collaborative grasping and handover trajectory is generated, which includes multiple timestamped path points. Based on the path points in the collaborative grasping and handover trajectory and the end contact force data, the desired end acceleration command is obtained; Based on the plantar pressure data in the body state data, the zero-moment point constraint condition is obtained; The desired end-effector acceleration command and the zero-moment point constraint are combined to perform a quadratic programming solution, wherein the quadratic programming solution is performed using the angular acceleration vector of each joint of the humanoid robot and the output torque vector of each joint as unified optimization variables to obtain the torque command of the joints of the humanoid robot. The torque command is sent to the dual-arm joint actuators and dual-leg joint actuators of the humanoid robot to drive the humanoid robot to perform grasping and handover actions; Obtain the basic policy parameters pre-trained in the physical simulation environment; Based on the deviation between the actual contact feedback collected by the humanoid robot in the real environment and the expected contact feedback in the physical simulation environment, the policy correction amount is obtained through an online residual learning network. The policy correction amount is added to the basic policy parameters to obtain the corrected policy parameters, and the corrected policy parameters are used for solving the quadratic programming problem.

2. The method according to claim 1, characterized in that, The step of obtaining a preliminary six-dimensional pose estimate of the cargo based on the image data includes: inputting the image data into a pre-trained vision-language-action model, identifying the cargo category and boundaries through the vision-language-action model, and combining the depth information in the image data to calculate the three-dimensional position coordinates and three-dimensional attitude angles of the cargo in the camera coordinate system to obtain the preliminary six-dimensional pose estimate.

3. The method according to claim 1, characterized in that, The step of obtaining the instantaneous velocity estimate and acceleration estimate of the dynamic conveying device based on the encoder pulse signal includes: performing optical flow tracing on the surface feature points of the dynamic conveying device in continuous image frames to obtain the displacement vector of the feature points on the image plane, using the displacement vector and the measured velocity value corresponding to the encoder pulse signal as observations, inputting them into a Kalman filter for state estimation, and obtaining the instantaneous velocity estimate and the acceleration estimate.

4. The method according to claim 1, characterized in that, The generation of the timestamped collaborative grasping and handover trajectory includes: calculating the predicted six-dimensional pose of the cargo at multiple moments within the future time window based on the preliminary six-dimensional pose estimate, the instantaneous velocity estimate, the acceleration estimate, and preset future time window parameters. Each predicted six-dimensional pose value is the sum of the integral term of the preliminary six-dimensional pose estimate and the velocity of the conveying equipment at the corresponding moment, plus half of it multiplied by the product of the acceleration estimate and the square of the corresponding moment. Combining the predicted six-dimensional pose with the corresponding timestamp, and combining it with the current joint state of the humanoid robot body, the predicted six-dimensional pose is obtained by combining each predicted six-dimensional pose with the corresponding timestamp.

5. The method according to claim 1, characterized in that, The step of obtaining the desired end-effector acceleration command based on the path points in the collaborative grasping and handover trajectory and the end-effector contact force data includes: determining the deviation between the actual pose of the humanoid robot end effector and the desired pose corresponding to the path points; establishing an end-effector admittance control equation based on a pre-set virtual inertia matrix, virtual damping matrix, and virtual stiffness matrix, as well as the deviation, the first derivative of the deviation, the second derivative of the deviation, and the end-effector contact force data; solving the end-effector admittance control equation to obtain the desired end-effector acceleration command.

6. The method according to claim 1, characterized in that, The step of obtaining the zero-moment point constraint condition based on the plantar pressure data in the body state data includes: calculating the position of the zero-moment point based on the current center of mass acceleration and rate of change of angular momentum of the humanoid robot; determining the boundary range of the bipedal support polygon in a preset direction based on the plantar pressure data; and limiting the position of the zero-moment point within the boundary range to obtain the zero-moment point constraint condition.

7. The method according to claim 1, characterized in that, The step of jointly performing a quadratic programming solution on the desired end-effector acceleration command and the zero-moment point constraint includes: using the angular acceleration vectors of each joint of the humanoid robot and the output torque vectors of each joint as optimization variables, using the weighted norm of the deviation between the joint angular acceleration vector and the joint space desired acceleration reference calculated based on the desired end-effector acceleration command, and the weighted norm of the joint output torque vector as optimization objectives, and using the rigid body dynamics equations of the humanoid robot and the zero-moment point constraint as constraints to solve the problem and obtain the torque command.

8. The method according to claim 1, characterized in that, After the torque command is sent to the dual-arm joint actuators and dual-leg joint actuators of the humanoid robot, the method further includes: continuously sending the torque command at a control frequency of not less than 1000 Hz, and after each control cycle, collecting the end contact force data and plantar pressure data after execution, as inputs to obtain the end desired acceleration command and zero-moment point constraint conditions in the next control cycle, thus forming a closed-loop control.

9. The method according to claim 1, characterized in that, Before obtaining the basic strategy parameters pre-trained in the physical simulation environment, the method further includes: establishing a physical property model in the physical simulation environment that includes goods made of rigid materials, flexible materials, and fragile materials, and setting at least two different sets of parameter combinations consisting of friction coefficients, belt speed ranges, and load distribution parameters for different conveying equipment models to form a set of simulation scenarios. In the set of simulation scenarios, a preset machine learning training method is used to pre-train the control strategy parameters corresponding to the quadratic programming solution to obtain the basic strategy parameters.

Images