Path planning and energy management cooperative regulation and control method for multi-agent unmanned aerial vehicle

Abstract

The invention provides a path planning and energy management cooperative regulation and control method for a multi-agent unmanned aerial vehicle, and the method comprises the following steps: building a kinetic model based on a six-degree-of-freedom motion equation of a four-rotor unmanned aerial vehicle, and building an energy consumption model through combining the power of a motor and the discharge characteristics of a battery; sensing the state of the unmanned aerial vehicle in real time, defining a state space and an action space of a deep reinforcement learning algorithm, and generating navigation points by using the deep reinforcement learning algorithm; receiving a delivery request, and dynamically allocating tasks according to the state of the unmanned aerial vehicle by adopting a deep reinforcement learning algorithm; performing path planning on the unmanned aerial vehicle for task allocation by adopting an A * algorithm according to the navigation points in combination with the multi-dimensional cost, and obtaining an optimal path; dispersing the optimal path into a speed instruction, and dynamically adjusting the course angle and the flight speed of the unmanned aerial vehicle through a PID controller; through combination of deep reinforcement learning and an A * algorithm, the problems of path planning and energy management of the multi-agent quad-rotor unmanned aerial vehicle logistics distribution system in complex terrains are effectively solved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of unmanned aerial vehicle control planning, and particularly relates to a path planning and energy management collaborative regulation method for multi-agent unmanned aerial vehicles. BACKGROUND

[0002] With the rapid growth of e-commerce and instant delivery demand, unmanned aerial vehicle logistics delivery systems show significant advantages in mountainous areas, islands and complex urban terrains, especially in mountainous areas, traditional ground transportation is limited by rugged terrain and long-distance transportation efficiency, while unmanned aerial vehicle clusters can directly deliver to delivery points through airspace, greatly shortening the delivery time. However, complex environments such as variable wind fields and steep terrains pose serious challenges to the path planning and energy management of unmanned aerial vehicle clusters.

[0003] Current multi-agent quadrotor unmanned aerial vehicle logistics systems mostly use pure electric drive and rely on limited battery capacity to perform tasks. In complex scenarios such as mountainous areas, unmanned aerial vehicles need to frequently adjust flight altitudes to optimize energy, and need to resist strong crosswinds and headwinds, resulting in increased motor power fluctuations and significantly increased energy consumption. In addition, when multiple machines are used for collaborative delivery, traditional methods often separate path planning and energy management, and only use static task allocation strategies such as fixed priority or polling mechanisms, ignoring the real-time power status of unmanned aerial vehicles and environmental dynamics. For example, if a low-power unmanned aerial vehicle is assigned a long-distance task, it may run out of power midway through the task, resulting in task failure; and if a high-power unmanned aerial vehicle is idle for a long time, it will reduce the overall system efficiency.

[0004] In the prior art, the A* algorithm can generate the shortest path, but its cost function is not deeply coupled with the energy consumption model, making it difficult to avoid high wind resistance areas or optimize flight altitudes; deep reinforcement learning such as DQN has dynamic decision-making capabilities, but it is mostly focused on single-machine path optimization and lacks multi-machine collaborative task allocation and global energy balancing mechanisms. In addition, task allocation does not take into account delivery point priority and unmanned aerial vehicle power decay characteristics, which can lead to task accumulation or resource waste, especially when the cluster size is expanded, the above problems are further magnified, and the system is difficult to balance energy consumption optimization and task completion rate. SUMMARY

[0005] In view of this, the present application proposes a path planning and energy management collaborative regulation method for multi-agent unmanned aerial vehicles, which can ensure that the unmanned aerial vehicle cluster efficiently and reliably completes the logistics delivery task in complex terrain.

[0006] The technical solution of the present application is as follows: A path planning and energy management collaborative regulation method for multi-agent unmanned aerial vehicles, comprising the following steps: Step S1, a dynamic model is established based on the six-degree-of-freedom motion equation of a quadrotor unmanned aerial vehicle, and an energy consumption model is constructed by combining motor power and battery discharge characteristics; Step S2, real-time perception of the state of the UAV and definition of the state space and action space of the deep reinforcement learning algorithm, generation of navigation points by using the deep reinforcement learning algorithm; Step S3, receiving a delivery request, and dynamically assigning tasks according to the state of the UAV by using the deep reinforcement learning algorithm; Step S4, path planning for the UAV assigned with the task by using the A* algorithm according to the navigation points combined with multi-dimensional cost, and obtaining an optimal path; Step S5, discretizing the optimal path into speed instructions, and dynamically adjusting the heading angle and flight speed of the UAV by using a PID controller.

[0007] Preferably, the expression of the dynamic model in step S1 is:

[0008]

[0009]

[0010]

[0011] wherein, is the relative position with the earth as the reference system, respectively represent the attitude angles of the UAV, respectively represent the air damping coefficients of the UAV on the axis, respectively represent the moments of inertia of the UAV body around the body coordinate system, respectively represent the total lift, roll, pitch, and yaw moment inputs, and the expression is:

[0012] wherein is the distance from the motor to the center of mass, is the motor torque coefficient, is the lift coefficient, is the rotation speed of the four rotors of the UAV.

[0013] Preferably, the expression of the energy consumption model in step S1 is:

[0014] wherein is the rotor aerodynamic loss coefficient, is the motor torque coefficient, is the power consumption of the auxiliary equipment, is the rotation speed of the i-th rotor.

[0015] Preferably, the specific steps of step S2 are: Step S21, obtaining the current position, remaining power, real-time wind speed and task queue state of the UAV; Step S22, defining the state space as wherein is the remaining power, is the real-time wind speed vector, is the current position coordinate of the UAV, is the task point coordinate of the task queue state, is the task weight; Step S23, the action space output includes task point selection and flight speed decision, the deep reinforcement learning algorithm selects a target point from the task pool , generates a recommended flight speed according to the remaining power and the task weight:

[0016]

[0017] wherein is the index of the selected task point, is the recommended flight speed, is the maximum flight speed; Step S24, the deep reinforcement learning algorithm outputs the navigation point through the built-in policy network.

[0018] Preferably, the policy network in step S24 is optimized and adjusted through a reward and punishment mechanism, and the reward and punishment mechanism is: energy-saving reward, expressed as wherein is the actual energy consumption at the current time, is the baseline energy consumption, is the energy-saving reward coefficient; power depletion penalty, expressed as wherein is the power penalty coefficient, is the remaining power of the UAV, is the total power of the UAV; path deviation penalty, expressed as wherein is the Euclidean distance between the current position and the target path point in the optimal path generated by the A* algorithm, is the penalty coefficient, is the threshold value.

[0019] Preferably, the specific steps of step S3 are: Step S31, after receiving the delivery request, collecting the global information of the UAV, the global information including the remaining power of the UAV, the speed of the UAV, the task queue state and the real-time wind speed; Step S32, the deep reinforcement learning algorithm generates dynamic weights by the policy network, and assigns tasks, and the calculation formula of the dynamic weights is:

[0020] wherein represents the unmanned aerial vehicle assigns priorities to tasks , the remaining power of the unmanned aerial vehicle k, the total power of the unmanned aerial vehicle k, the delivery distance of the task j, the maximum delivery distance, the real-time wind speed, the maximum wind speed.

[0021] Preferably, the specific steps of the step S4 are: Step S41, initializing the starting point, target point and environmental constraints, calculating the multi-dimensional cost function value of the node and the terminal point, taking the node with the minimum multi-dimensional cost function value as the starting node and putting it into the openlist; Step S42, judging whether the openlist is empty, if it is empty, the starting node needs to be adjusted, and if it is not empty, the starting node is regarded as the current node; Step S43, if the current node is the terminal point, the search is ended to generate the optimal path, and if the current node is not the terminal point, the current node is stored in the closelist and the current node in the openlist is deleted; Step S44, querying the node with the minimum multi-dimensional cost function value and putting it into the openlist, and repeating the above steps until the target node is reached and the optimal path is obtained.

[0022] Preferably, the multi-dimensional cost function in the step S4 includes: path length cost , which is expressed as the Euclidean distance or actual flight distance from node i to node j, and its expression is: , , , ) and ( , , ) represent the three-dimensional coordinates of nodes i and j, respectively; flight height cost , and its expression is , wherein is the height energy consumption coefficient, is the absolute flight height of the unmanned aerial vehicle, is the basic height offset.​ terrain elevation cost , the expression of which is , wherein is a terrain penalty coefficient, is a horizontal movement distance, is a vertical height difference; no-fly zone penalty cost , the expression of which is:

[0023] , wherein is a no-fly zone penalty coefficient, is the shortest distance from the current position of the UAV to the boundary of the no-fly zone, is a safety buffer distance threshold;

[0024] The expression of the comprehensive cost function is: , wherein represents the actual cost from the starting point to the current node n, and the calculation formula is: , is a heuristic function, which is the Euclidean distance from the current node to the end point, is the weight of the cost in different dimensions.

[0025] Preferably, the expression of the PID controller in the step S5 for dynamic adjustment is:

[0026] , wherein is a correction amount of the motor speed of the UAV, are proportional adjustment coefficients, differential adjustment coefficients and integral adjustment coefficients, respectively, is the Euclidean distance deviation of the current position of the UAV from the target path point.

[0027] Preferably, the PID controller feeds back the motor speed of the UAV, the actual flight speed and the tracking error to the deep reinforcement learning algorithm, and the deep reinforcement learning algorithm updates the navigation points.

[0028] Compared with the prior art, the present application has the following advantages: The path planning and energy management collaborative regulation method of the multi-agent UAV of the present application effectively solves the path planning and energy management problems of the multi-agent four-rotor UAV logistics distribution system in complex terrain, combines the A* algorithm for planning the path by fusing the constraints of terrain, wind field, etc. and the navigation points dynamically generated by the deep reinforcement learning model, realizes low-energy consumption optimization of the optimal path, and avoids energy-consuming areas such as high wind resistance and high elevation; The task allocation weight mechanism based on the deep reinforcement learning model combines the real-time power of the unmanned aerial vehicle, the task priority and the environmental wind speed, realizes dynamic task allocation of multi-machine cooperation, avoids the failure caused by the low-power unmanned aerial vehicle undertaking a long-distance task, simultaneously improves the utilization rate of the high-power unmanned aerial vehicle, and balances the global energy consumption. Through dynamic adjustment of the path tracking error by the PID controller, the unmanned aerial vehicle can stably track the planned path, and invalid energy consumption is reduced, so that the task completion rate and efficiency of logistics distribution are improved, and the method can be applied to many fields such as agricultural plant protection and military reconnaissance, and has wide engineering application value. BRIEF DESCRIPTION OF DRAWINGS

[0029] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced as follows. Obviously, the drawings in the following description are only preferred embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0030] Fig. 1 A flow chart of a path planning and energy management collaborative regulation method of a multi-agent unmanned aerial vehicle of the present application; Fig. 2 A schematic diagram of a quad-rotor unmanned aerial vehicle dynamics model of a path planning and energy management collaborative regulation method of a multi-agent unmanned aerial vehicle of the present application; Fig. 3 A flow chart of generating an optimal path by A* algorithm of a path planning and energy management collaborative regulation method of a multi-agent unmanned aerial vehicle of the present application; DETAILED DESCRIPTION In order to better understand the technical content of the present application, a specific embodiment is provided below, and the present application is further described in combination with the drawings.

[0031] Referring to Figs. 1 to 3 The path planning and energy management collaborative regulation method of a multi-agent unmanned aerial vehicle provided by the present application comprises the following steps: Step S1, a dynamics model is established based on the six-degree-of-freedom motion equation of the quad-rotor unmanned aerial vehicle, and an energy consumption model is constructed in combination with the motor power and the battery discharge characteristics; Step S2, the state of the unmanned aerial vehicle is sensed in real time, and the state space and the action space of the deep reinforcement learning algorithm are defined, and the navigation points are generated by using the deep reinforcement learning algorithm; Step S3, a distribution request is received, and the deep reinforcement learning algorithm is used to dynamically allocate tasks according to the state of the unmanned aerial vehicle; Step S4, the A* algorithm is used to plan a path for the unmanned aerial vehicle allocated with the task according to the navigation points in combination with multi-dimensional costs, and an optimal path is obtained; Step S5, discretize the optimal path into speed instructions, dynamically adjust the heading angle and flight speed of the UAV through the PID controller.

[0032] The present application aims at the problems of path planning and energy management being split, poor adaptability to dynamic environment, and insufficient energy consumption optimization in the existing multi-agent quadrotor UAV logistics distribution system, and proposes a path planning and energy management collaborative control method for multi-agent UAVs, generates navigation points based on the real-time state of the multi-agent UAVs through the DQN algorithm, uses the navigation points as the heuristic input of the A* algorithm, optimizes the multi-dimensional cost function weight distribution of the global path generation layer, ensures that the path planning prioritizes low energy consumption routes, and in addition, collaboratively optimizes energy consumption through global energy management to avoid premature depletion of single machine power, and ensures efficient task execution.

[0033] When path planning and energy management are performed, a dynamics model and an energy consumption model of the UAV need to be constructed first, and the present application uses a quadrotor UAV as the main body, after the dynamics model and the energy consumption model are constructed, a deep reinforcement learning algorithm can be introduced, the deep learning reinforcement algorithm can perceive the state of the UAV, and generate navigation points through the Q-learning strategy, the navigation points are used as the heuristic input of the A* algorithm, the priority of the target node of the A* algorithm is adjusted, the optimal path planned by the A* algorithm is corrected, and multi-dimensional control parameters such as flight speed and motor power distribution are optimized to minimize instantaneous energy consumption and cumulative power consumption, in addition, the deep learning reinforcement algorithm can dynamically allocate tasks according to the state of the UAV after receiving a distribution request, for example, high-power UAVs are preferentially allocated long-distance tasks, through task allocation, UAVs with higher priority can be obtained, then the A* algorithm can plan a path for the UAVs with higher priority, when planning the path, terrain elevation, no-fly zones, distribution point priorities and other constraints can be fused, and multi-dimensional cost can be combined for path planning to generate an optimal path, avoid high energy consumption areas, optimize path length and flight height, reduce basic power consumption, and existing lightweight communication protocols can also be used to share the key state positions of the UAV, after the optimal path is obtained, the optimal path can be discretized into speed instructions through the PID controller, and then the heading angle and flight speed of the UAV can be dynamically adjusted according to the speed instructions, and the PID controller can also feed back the actual flight data to the deep reinforcement learning algorithm to adjust the strategy network of the deep reinforcement learning algorithm, update the navigation points, and form a closed-loop optimization.

[0034] Preferably, the expression of the dynamics model in step S1 is:

[0035]

[0036]

[0037]

[0038] wherein, is the relative position with respect to the earth, respectively represent the attitude angles of the UAV, respectively represent the air damping coefficients of the UAV on the axis, respectively represent the rotational inertia of the UAV body around the body coordinate system, respectively represent the total lift, roll, pitch, yaw moment inputs, and the expressions are:

[0039] wherein is the distance from the motor to the center of mass, is the motor torque coefficient, is the lift coefficient, is the rotational speed of the four rotors of the UAV.

[0040] Preferably, the expression for establishing the instantaneous energy consumption model based on the motor power and the battery discharge characteristics in the step S1 is:

[0041] wherein is the rotor aerodynamic loss coefficient, is the motor torque coefficient, is the auxiliary equipment power consumption, is the rotational speed of the i-th rotor.

[0042] The dynamic model quantifies the relationship between the UAV position, attitude and rotor speed based on the motion equation of the quad-rotor UAV, and provides a basis for analyzing the flight characteristics in complex terrain. The energy consumption model combines the motor power and the battery characteristics to calculate the instantaneous energy consumption. Through the dynamic model and the energy consumption model, quantitative basis can be provided for energy consumption optimization in path planning, DQN reward function design and PID control parameter adjustment, supporting efficient energy management and reliable task execution of the UAV cluster in complex environments.

[0043] Preferably, the specific steps of the step S2 are: Step S21, acquiring the current position, remaining power, real-time wind speed and task queue state of the UAV; Step S22, defining the state space as wherein is the remaining power, is the real-time wind speed vector, is the current position coordinate of the UAV, Task point coordinates for task queue state, Task weight; Step S23, the action space output includes task point selection and flight speed decision, the deep reinforcement learning algorithm selects a target point from the task pool , generates a recommended flight speed according to the remaining power and the task weight:

[0044]

[0045] wherein is the index of the selected task point, is the recommended flight speed, high-weight tasks are accelerated, and low-weight tasks maintain an economic speed, is the maximum flight speed; Step S24, the deep reinforcement learning algorithm outputs a navigation point through the built-in policy network, and filters the candidate points based on the energy consumption prediction model, ensuring that the new navigation point is located in a low wind resistance and low elevation area, while the policy network is optimized and adjusted through a reward and punishment mechanism. The core of the reward and punishment function design is to minimize instantaneous energy consumption, punish the deviation of the path and the risk of power depletion, and the reward and punishment mechanism is: Energy-saving reward, optimize flight path and motor power distribution, reduce instantaneous and cumulative energy consumption, expression is wherein is the actual energy consumption at the current time (based on the motor power model calculation), is the baseline energy consumption (theoretical energy consumption under static path planning), Energy-saving reward coefficient, for example, can be selected ; Power depletion penalty, to prevent the UAV from interrupting the task due to insufficient power, expression is wherein is the power penalty coefficient, for example , is the remaining power of the UAV, total power of the UAV, when the remaining power of the UAV is less than 20%, the penalty value increases linearly with the decrease of power, if the power is depleted, the penalty value reaches the maximum, that is, it is positively related to the priority of the task; Path deviation penalty, to ensure that the UAV strictly follows the global optimal path and reduces invalid detours, expression is wherein is the Euclidean distance between the current position and the target path point in the optimal path generated by the A* algorithm, is the penalty coefficient (such as ), which determines the penalty gradient, is the threshold value, the numerical value is 2m, the penalty increases in proportion to the excess.

[0046] The state space contains the real-time state of the UAV, environmental information, and task queue information, which can be used to evaluate the current decision-making scenario, and the action space outputs the task point selection and flight speed decision, thereby generating the navigation point, and the reward function is used to guide the UAV to generate a low-energy and high-reliability navigation point, and the punishment mechanism is used to avoid invalid paths or task failures, and the generated navigation point can be used as the input of the A* algorithm.

[0047] Preferably, the specific steps of step S3 are: Step S31, after receiving the distribution request, collecting global information of the UAV, the global information including the remaining power of the UAV, the speed of the UAV, the task queue state and the real-time wind speed; Step S32, the deep reinforcement learning algorithm generates dynamic weights through a policy network to allocate tasks, and the calculation formula of the dynamic weights is:

[0048] Among them, represents the UAV the priority of the task allocation, is the remaining power of the UAV k, is the total power of the UAV k, is the distribution distance of task j, is the maximum distribution distance, is the real-time wind speed, is the maximum wind speed.

[0049] The deep reinforcement learning algorithm of the application can perform multi-machine cooperation and dynamic task allocation for a UAV cluster. In order to realize the cooperative distribution of multi-agent UAVs, a "centralized training-distributed execution" framework is adopted. In the centralized training stage, the global information of the UAV is collected, and the task allocation strategy and path planning weight are optimized through the deep reinforcement learning algorithm. In the distributed execution stage, each UAV makes independent decisions based on local observations, dynamically adjusts the task execution and path tracking strategy, and obtains the allocation priority value of different UAVs for different tasks through dynamic weight calculation. The higher the output value, the higher the priority. Ultimately, the UAV with the highest dynamic weight can be selected to execute the task. If the weights of multiple machines are similar, the task can be allocated to the UAV with high power.

[0050] Preferably, the specific steps of step S4 are: Step S41, initializing the starting point, target point and environmental constraints, calculating the multi-dimensional cost function value of the node and the end point, taking the node with the minimum multi-dimensional cost function value as the starting node, and putting it into the open list; Step S42, judging whether the openlist is empty, if empty, the starting node needs to be adjusted, if not empty, the starting node is regarded as the current node; Step S43, if the current node is the end point, the search is ended to generate the optimal path, if the current node is not the end point, the current node is stored in the closelist, and the current node in the openlist is deleted; Step S44, the node with the minimum multi-dimensional cost function value is inquired and put into the openlist, the above steps are repeated until the target node is reached and the optimal path is obtained.

[0051] The A* algorithm can perform path planning on the unmanned aerial vehicle assigned with a task, and in the path planning, a multi-dimensional cost function is combined for evaluation, all nodes can be obtained by continuously inquiring the nodes and storing and deleting operations in the openlist and the closelist, and the optimal path can be obtained after connecting the nodes.

[0052] Preferably, the multi-dimensional cost function in the step S4 comprises: path length cost , which is expressed as the Euclidean distance or the actual flight distance from the node i to the node j, and the expression is: , , , ) and ( , , ) represent the three-dimensional coordinates of the nodes i and j respectively; flight height cost , the flight height and the energy consumption are in a nonlinear relationship, the low-altitude type can reduce air resistance, the high-altitude flight needs to overcome the gravitational potential energy, and is more affected by the wind field, and the expression is , wherein is the height energy consumption coefficient, is the absolute flight height of the unmanned aerial vehicle, is the basic height offset; terrain elevation cost , the greater the terrain slope, the more energy the unmanned aerial vehicle needs to consume to maintain the attitude, and the expression of the terrain elevation cost is , wherein is the terrain penalty coefficient, is the horizontal movement distance, is the vertical height difference; Traditional no-fly zone processing usually adopts "one size fits all" unlimited punishment, but in actual scenarios, the path may only approach the no-fly zone boundary rather than directly crossing it, for this purpose, the invention proposes a hierarchical no-fly zone punishment mechanism, combining distance sensitivity and violation risk, the following cost function is designed. Definition: let the no-fly zone be a polygon area, the nearest distance from the current position of the unmanned aerial vehicle to the boundary of the no-fly zone be , and the safety buffer distance threshold be (such as 50 meters), then the no-fly zone punishment cost , its expression is:

[0053] Among them, is the no-fly zone punishment coefficient; The above multi-dimensional cost is fused into a comprehensive cost function by weighting to guide the node expansion direction of the A* algorithm, its expression is: , wherein represents the actual cost from the starting point to the current node n, the calculation formula is: , is the heuristic function, taking the Euclidean distance from the current node to the end point, or the estimated energy consumption modified in combination with factors such as wind field and terrain, is the weight of different dimension cost, in different terrains, the weight distribution is different, for example, in flat urban areas, , in mountainous strong wind areas, .

[0054] Preferably, the expression of the PID controller in step S5 for dynamic adjustment is:

[0055] Among them, is the correction amount of the motor speed of the unmanned aerial vehicle, are the proportional adjustment coefficient, the differential adjustment coefficient and the integral adjustment coefficient respectively, is the Euclidean distance deviation of the current position of the unmanned aerial vehicle from the target path point, that is, the tracking error.

[0056] The PID controller adjusts the heading angle and flight speed of the unmanned aerial vehicle for the discrete speed command, and its adjustment coefficient can be adjusted in real time according to the remaining power and wind speed, for example, the proportional gain is reduced to save energy when the power is low, and the differential gain is increased to suppress oscillation when the wind is strong, so as to minimize the path tracking error.

[0057] Preferably, the PID controller feeds back the motor speed, actual flight speed and tracking error of the unmanned aerial vehicle to the deep reinforcement learning algorithm, and updates the navigation point by the deep reinforcement learning algorithm.

[0058] The PID controller uploads the motor speed, the actual flight speed and the tracking error to the deep reinforcement learning algorithm in real time, the deep reinforcement learning algorithm updates the navigation point generation strategy by using the data, and the A* algorithm corrects the prediction model of the global path according to the actual flight trajectory, so that closed-loop optimization is realized.

[0059] The above merely describes preferred embodiments of the present application but should not be used to limit the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A multi-agent UAV path planning and energy management coordinated control method, characterized in that: The following steps are involved: Step S1: establishing a dynamic model based on the six-degree-of-freedom motion equation of the quadrotor drone, and building an energy consumption model based on the motor power and battery discharge characteristics; Step S2: Real-time perception of the drone state and definition of the state space and action space of the deep reinforcement learning algorithm, and use of the deep reinforcement learning algorithm to generate navigation points; Step S3: Receive delivery requests and dynamically assign tasks based on the drone status using a deep reinforcement learning algorithm; Step S4: Using the A* algorithm to plan the path of the assigned UAV based on the navigation points and multi-dimensional costs, and obtain the optimal path; Step S5: Discretize the optimal path into speed instructions, and dynamically adjust the heading angle and flight speed of the UAV through the PID controller.

2. The method for coordinated control of path planning and energy management of a multi-agent UAV according to claim 1, characterized in that: The expression of the kinetic model in step S1 is: in, is the relative position with respect to the Earth, Represent the attitude angle of the UAV, Respectively indicate that the drone is Air damping coefficient on the axis, They represent the moment of inertia of the UAV body around the body coordinate system, They represent the total lift, roll, pitch, and yaw moment inputs respectively, and are expressed as: in is the distance from the motor to the center of mass, is the motor torque coefficient, is the lift coefficient, is the rotation speed of the four rotors of the UAV.

3. The method for coordinated control of path planning and energy management of a multi-agent UAV according to claim 1, characterized in that: The expression of the energy consumption model in step S1 is: in is the rotor aerodynamic loss coefficient, is the motor torque coefficient, is the power consumption of auxiliary equipment, is the rotation speed of the i-th rotor.

4. The method for coordinated control of path planning and energy management of a multi-agent UAV according to claim 1, characterized in that: The specific steps of step S2 are: Step S21: Obtain the current position, remaining battery power, real-time wind speed, and task queue status of the drone; Step S22: define the state space as ,in is the remaining power, is the real-time wind speed vector, is the current position coordinate of the drone, The coordinates of the task point in the task queue state. is the task weight; Step S23: The action space output includes task point selection and flight speed decision. The deep reinforcement learning algorithm selects a target point from the task pool. , generates the recommended flight speed based on the remaining power and mission weight: in is the index of the selected task point, For the recommended flight speed, is the maximum flight speed; Step S24: The deep reinforcement learning algorithm outputs navigation points through the built-in policy network.

5. The method for path planning and energy management coordinated control of a multi-agent UAV according to claim 4, characterized in that: The policy network in step S24 is optimized and adjusted through a reward and penalty mechanism, and the reward and penalty mechanism is: Energy saving reward, expressed as ,in is the actual energy consumption at the current moment, is the baseline energy consumption, Energy saving bonus coefficient; The power depletion penalty is expressed as ,in is the power penalty coefficient, The remaining battery power of the drone. The total battery power of the drone; Path deviation penalty, expressed as ,in It is the Euclidean distance between the current position and the target path point in the optimal path generated by the A* algorithm. is the penalty coefficient, is the threshold.

6. The method for coordinated control of path planning and energy management of a multi-agent UAV according to claim 1, characterized in that: The specific steps of step S3 are: Step S31: After receiving the delivery request, collect the drone's global information, including the drone's remaining battery power, drone speed, task queue status, and real-time wind speed; Step S32: The deep reinforcement learning algorithm generates dynamic weights through the policy network to assign tasks. The calculation formula for generating dynamic weights is: in Indicates drone Task The allocation priority of is the remaining power of drone k, is the total power of UAV k, is the delivery distance of task j, is the maximum delivery distance, is the real-time wind speed, is the maximum wind speed.

7. The method for path planning and energy management coordinated control of a multi-agent UAV according to claim 1, characterized in that: The specific steps of step S4 are: Step S41: Initialize the starting point, target point, and environmental constraints, calculate the multi-dimensional cost function values ​​of the nodes and the end point, and select the node with the smallest multi-dimensional cost function value as the starting node and put it into the openlist; Step S42: determine whether the openlist is empty. If it is empty, the starting node needs to be adjusted. If it is not empty, the starting node is regarded as the current node. Step S43: If the current node is the end point, the search is terminated to generate the optimal path. If the current node is not the end point, the current node is stored in the closelist and the current node in the openlist is deleted. Step S44: Query the node with the smallest multi-dimensional cost function value and put it into the openlist, repeat the above steps until the target node is reached and the optimal path is obtained.

8. The method for coordinated control of path planning and energy management of a multi-agent UAV according to claim 1, characterized in that: The multi-dimensional cost function in step S4 includes: Path length cost , which is the Euclidean distance or actual flight distance from node i to node j, and its expression is: , ( 、 、 ) and( 、 、 ) represent the three-dimensional coordinates of nodes i and j respectively; Flight altitude cost , whose expression is ,in is the height energy consumption coefficient, is the absolute flight altitude of the UAV, is the base height offset; Terrain elevation cost , whose expression is ,in is the terrain penalty coefficient, is the horizontal movement distance, is the vertical height difference; No-fly zone penalty costs , whose expression is: in is the no-fly zone penalty coefficient, is the shortest distance from the current position of the drone to the boundary of the no-fly zone, is the safety buffer distance threshold; The expression of the comprehensive cost function is: ,in Represents the actual cost from the starting point to the current node n, and the calculation formula is: , As the heuristic function, take the Euclidean distance from the current node to the end point, is the weight of the cost of different dimensions.

9. The method for coordinated control of path planning and energy management of a multi-agent UAV according to claim 1, characterized in that: The expression for dynamic adjustment of the PID controller in step S5 is: in is the correction value of the UAV motor speed, They are proportional adjustment coefficient, differential adjustment coefficient, and integral adjustment coefficient respectively. is the Euclidean distance deviation between the current position of the UAV and the target path point.

10. The method for coordinated control of path planning and energy management of a multi-agent UAV according to claim 1, characterized in that: The PID controller feeds the drone's motor speed, actual flight speed, and tracking error back to the deep reinforcement learning algorithm, which then updates the navigation points.

Images