An online dynamic unmanned aerial vehicle hierarchical distribution method under weather interference

Abstract

The application discloses an online dynamic unmanned aerial vehicle hierarchical distribution method under weather interference, comprising the following steps: obtaining a to-be-executed task, a unmanned aerial vehicle requesting to receive a new task and weather environment information at a current scheduling time, and constructing a task snapshot and a unmanned aerial vehicle snapshot; adopting integer vector coding to express task attribution according to the snapshot; decoding the coding into attribution unmanned aerial vehicles of each task and performing pairwise service estimation to obtain an estimated service cost; performing multi-target evaluation according to the estimated cost and a task time limit, and searching to obtain a candidate solution set; selecting a final scheme from the candidate solution set according to a preference weight and decoding the final scheme into a promised task set of each unmanned aerial vehicle; each unmanned aerial vehicle outputs a discrete rule action according to the promised task set and a local observation state, and submits a request to a platform when outputting a request for a new task, and the platform collects the request, allocates and feeds back. The application can cope with continuous task arrival, sudden disturbance and large scale, and improves a task completion rate and a transport capacity utilization rate.

Description

Technical Field

[0001] This invention belongs to the field of drone scheduling technology, and particularly relates to an online dynamic drone tiered delivery method under weather interference. Background Technology

[0002] As low-altitude drone logistics gradually moves from single test flights and small-scale demonstrations to continuous operation, the scheduling objects and simulation environments are also shifting from relatively static to highly dynamic, gradually revealing the limitations of static scenarios. On the one hand, new logistics tasks continuously enter the task pool during operation, with varying start and end points, payload weights, and service time limits. On the other hand, the drone's location, battery level, payload, and available time change in real time during execution, and events such as weather changes, task timeouts, and return-to-base refueling constantly alter the correspondence between existing tasks and drone resources. Therefore, low-altitude logistics scheduling is no longer simply about finding feasible paths for a pre-defined set of tasks, but rather about continuously generating high-quality decisions amidst the ever-changing task flow, drone status, and service constraints.

[0003] For UAV task allocation and path planning, existing search-driven methods can perform combinatorial optimization under given task sets, flight environments, and constraints, and obtain high-quality scheduling schemes. For example, swarm intelligence optimization and mathematical programming methods can be used to handle UAV task allocation, path generation, and multi-UAV collaborative problems. However, research typically relies on relatively explicit task sets and constraints; when low-altitude logistics tasks continuously arrive, task states change, or external disturbances alter UAV availability, existing schemes often require frequent rearrangement, and the search space needs repeated reconstruction, thus increasing the computational pressure of rolling scheduling and affecting real-time scheduling performance.

[0004] To improve adaptability in dynamic scenarios, learning-driven methods have begun to be used for drone logistics scheduling. These methods can learn online decision-making strategies based on system state, responding to changes in tasks and resources. However, in large-scale low-altitude logistics scenarios, if reinforcement learning strategies directly handle the matching relationships between all tasks and all drones, the state and action space will grow synchronously with the number of tasks and drones, easily leading to high training costs, difficulties in policy convergence, and challenges in transfer learning. Especially when switching between different task scales, different numbers of drones, and different load intensities, existing learning strategies often struggle to maintain stable scheduling performance.

[0005] Therefore, while search-driven methods possess strong combinatorial optimization capabilities, they are insufficiently responsive to continuously changing execution environments; while learning-driven methods, although possessing some adaptive capabilities, face challenges such as action space expansion and migration difficulties in large-scale task-UAV matching. Existing search-learning fusion methods attempt to combine the advantages of both, but they mostly focus on path search, regional coverage, or strategy adjustment under fixed task sets, failing to adequately support scenarios in low-altitude logistics where three types of problems coexist: continuous task arrival, frequent sudden disturbances, and an ever-expanding scale of task-UAV candidate matching.

[0006] Therefore, low-altitude logistics scheduling needs a collaborative mechanism that is more suitable for complex operational scenarios and can cope with the scheduling challenges brought about by rolling updates of task flows, reconstruction of sudden disturbances, and large-scale combined decision-making. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention proposes an online dynamic drone-based layered delivery method under weather interference, thereby resolving the issues present in the prior art.

[0008] To achieve the above objectives, the present invention provides an online dynamic drone tiered delivery method under weather interference, comprising: Obtain the set of tasks to be executed at the current scheduling time, the set of drones requesting to receive new tasks, and weather and environmental information; construct task snapshots and drone snapshots. Based on the task snapshot and the UAV snapshot, an integer vector code is used to represent the attribution relationship between the task and the UAV, thus obtaining the integer vector code; The integer vector is encoded and decoded into the associated drone for each task. Task-drone pair service estimation is performed for each task and its associated drone to obtain the estimated service cost. Based on the estimated service cost and the time constraints of each task, the integer vector code is evaluated in multiple objectives, and a search algorithm is used to search the solution space of the integer vector code to obtain a candidate solution set. The final allocation scheme is selected from the candidate solution set according to the preset preference weights and decoded into a set of committed tasks for each UAV; During execution, when a UAV has a decision opportunity, it outputs a discrete rule action based on the committed task set and its own local observation state. When the output discrete rule action is a request for a new task, it submits a request to the scheduling platform to receive a new task. The scheduling platform collects the requests from each UAV in the request window, performs task allocation, and feeds back the allocation result to the committed task set of each UAV to complete the hierarchical delivery.

[0009] Optionally, the task snapshot includes the task start point, task end point, pending execution time, service deadline, task weight, and timeliness risk; the drone snapshot includes the drone's current location or current target location, available time, speed, battery level, payload, and existing committed tasks.

[0010] Optionally, the integer vector encoding uses the number of tasks to be assigned in the current scheduling window as the vector dimension. Each element in the integer vector encoding corresponds to a task, and the value of each element is any drone number in the set of drone numbers requesting to participate in the allocation. The integer vector encoding is only used to determine the drone to which each task belongs, and does not limit the order in which each drone picks up and delivers tasks in its committed task set.

[0011] Optionally, the task-drone pairwise service estimation includes: Based on the drone anchor location and available time, estimate the access time from the anchor location to the mission start point, the delivery time from the mission start point to the mission end point, the pickup service time, and the delivery service time. Calculate the single-task service cost based on the access time, the delivery time, the pickup service time, and the delivery service time. Calculate the estimated completion time of the task based on the available time of the drone and the single-task service cost; Calculate the estimated lateness based on the estimated completion time and deadline of the task; The timeliness risk after task allocation is calculated based on the task waiting time, the expected completion time of the task, and the task deadline.

[0012] Optionally, the objectives adopted in the multi-objective evaluation include service efficiency objectives, risk timeliness objectives, and load balancing objectives; The service efficiency target is the sum of the single-task service costs of all assigned tasks in the candidate scheme; The risk timeliness target is the weighted sum of the timeliness risk of each task in the candidate scheme and the expected amount of delay; The load balancing objective is the variance of the estimated workload of each UAV under the current scheduling window. The estimated workload is determined based on the sum of the remaining occupancy time of the existing tasks of the UAV and the service cost of the newly assigned tasks.

[0013] Optionally, the search algorithm is a multi-objective evolutionary algorithm.

[0014] Optionally, a final allocation scheme is selected from the candidate solution set according to preset preference weights, including: The multi-objective evaluation values ​​of each candidate solution in the candidate solution set are normalized respectively; The preference score for each candidate solution is calculated by weighting and summing the normalized target values ​​with the preset preference weights. The candidate scheme with the smallest preference score is selected as the final allocation scheme.

[0015] Optionally, the local observation state is a fixed-dimensional vector, including action availability features, committed task pressure features, requestable task pressure features, task type features corresponding to the rule prototype, and expected consequences features of different rule actions. The availability characteristics of the action include whether there is an executable task, whether there is a pickup candidate, whether there is a delivery candidate, whether a new task can be requested, and whether the requested action is available; The pressure characteristics of the committed tasks include the number of committed tasks, the weight of committed tasks, the onboard load, the pressure of the pickup time limit, and the pressure of the delivery time limit; The requested task pressure characteristics include the number of requested tasks, the optimal estimated completion time, the minimum estimated delay, the maximum timeliness risk, and the total weight.

[0016] Optionally, the discrete rule actions include selecting the feasible task with the shortest travel time, selecting the feasible task with the earliest service deadline, selecting the feasible task with the smallest expected service margin, prioritizing delivery, prioritizing pickup, requesting a new task, and waiting. When the output discrete rule actions are the first five types of actions, a set of feasible tasks is constructed based on the UAV's current location, battery level, payload, committed tasks, loaded tasks, and safe return constraints, and tasks that meet the corresponding rule conditions are selected from the set of feasible tasks for execution.

[0017] Optionally, the discrete rule actions are output by a reinforcement learning policy, and multiple drones share the same reinforcement learning policy network. Each drone asynchronously calls the reinforcement learning policy network during execution. The immediate reward of the reinforcement learning policy is constructed based on the number of tasks completed in one decision step, the number of tasks completed on time, the number of assigned tasks that failed, the number of unserved tasks at the end of the process, and the total late time.

[0018] Compared with the prior art, the present invention has the following advantages and technical effects: This invention decomposes low-altitude logistics scheduling into two collaborative layers: upper-level task allocation and lower-level local decision-making. This compresses the task-UAV matching problem from large-scale combinatorial optimization to a finite set of committed tasks for each UAV, reducing the search complexity of a single scheduling operation and minimizing the computational pressure of rolling rearrangement. The upper layer uses integer vector encoding to represent task affiliation and, combined with pairwise service estimation, only needs to determine task affiliation without fully traversing the pickup and delivery sequence, thus reducing the computational overhead of a single allocation. Simultaneously, it constructs a Pareto search with multiple objectives—service efficiency, risk-timeliness, and load balancing—to enable the system to obtain the optimal solution under different operational preferences, avoiding increased timeliness risk or load imbalance caused by single-objective optimization. The lower layer uses a fixed... The local observation state of the dimension compresses variable-scale task information into a fixed input, avoiding the expansion of the action space of the reinforcement learning policy as the number of tasks increases. Multiple drones share the same policy and call it asynchronously, making the learning policy reusable under different numbers of drones, reducing migration costs and retraining overhead. In addition, the upper and lower layers form a closed loop through order request-allocation feedback. During the execution process, the lower layer submits unmanageable tasks to the upper layer through request actions. The upper layer collects requests within the request window, reallocates them, and writes them back to the committed task set. This enables the system to maintain scheduling response continuously under the conditions of continuous task arrival, sudden disturbances, and expansion of candidate matching scale, improving task completion rate and timeliness stability, and enhancing the utilization efficiency of drone capacity. Attached Figure Description

[0019] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a technical roadmap of an embodiment of the present invention; Figure 2 This is a network structure diagram of PPO in the lower layer of the technology roadmap. Detailed Implementation

[0020] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0021] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0022] Example 1 like Figure 1As shown in the figure, this embodiment provides an online dynamic drone hierarchical delivery method under weather interference. The low-altitude logistics scheduling process is divided into two levels: upper-level swarm intelligence task allocation and lower-level reinforcement learning drone decision-making. The collaboration is achieved through an order request-allocation feedback mechanism.

[0023] The upper layer addresses the allocation problem among multiple drones for large-scale low-altitude logistics tasks. Its inputs include the tasks pending in the current scrolling window, the status of drones requesting new tasks, and low-altitude operating environment information such as weather. Its output is the task-drone allocation scheme. The upper layer's role is to compress the large-scale task pool into a finite set of committed tasks for each drone through the scrolling window, thereby preventing the lower layer from directly dealing with the entire task combination and resolving issues of continuity and scalability.

[0024] The lower layer addresses online decision-making issues for individual UAVs during execution. Its inputs are the UAV's local observation state, including action availability, committed task pressure, requestable task pressure, and consequences of candidate actions; its outputs are rule-based actions such as picking up cargo, delivering goods, requesting new tasks, or waiting. The lower layer's role is to select specific actions in real-time, responding to changes in battery power, payload, task time limits, and sudden disturbances.

[0025] The system is driven by an event queue. Events such as task arrival, task readiness, drone completion of pickup, drone completion of delivery, task timeout failure, weather changes, and return-to-home recharging completion will all trigger system status updates and affect subsequent upper-layer allocation and lower-layer execution.

[0026] I. Upper-level swarm intelligence task allocation; (1) Input and encoding; During scheduling Let the set of tasks to be executed be... The set of drones requesting to receive the new mission is Let the number of tasks currently participating in the allocation be... The requested number of drones is .

[0027] The upper layer first constructs mission snapshots and drone snapshots. The mission snapshot includes the mission start point, mission end point, pending execution time, service deadline, mission weight, and timeliness risk; the drone snapshot includes the current location or current target location, available time, speed, battery level, payload, and existing committed missions. Weather conditions are used to adjust flight time and energy consumption estimates.

[0028] For the task-drone affiliation relationships within the current window, integer vector encoding is used: in, This represents a candidate allocation scheme for the current window. Indicates task The drone's identification number. This indicates the number of tasks currently awaiting assignment. This indicates the number of drones currently requesting allocation. If This indicates the task. Assigned to drones This code only determines the task's attribution, not the subsequent pickup and delivery order, thus reserving room for adjustment for real-time execution at the lower level.

[0029] (2) Decoding and estimation; For a candidate individual ,according to The task Decoding to drones The newly added commitment task set. Its corresponding binary assignment variable is: in, Indicates task At any moment Whether assigned to a drone , where I(·) is the indicator function. When hour, =1, otherwise =0.

[0030] In low-altitude logistics environments, the location, available time, battery level, payload, and weather conditions of drones are constantly changing. If the upper layer were to perform a complete search of subsequent pickup and delivery sequences every time an assignment is made, the computational overhead would increase significantly. Therefore, this application uses task-drone pair service estimation to approximate the service time, delay pressure, and timeliness risks that may result from assigning a task to a specific drone in the current state, enabling swarm intelligence search to quickly compare different assignment schemes in dynamic environments.

[0031] make Indicates drone At the anchor point location assigned at a given time, if the UAV already has a target to execute, then the target's location is used; otherwise, the current location is used. Indicates drone The available time in the current window. For the current scheduling time, For drones The next available time.

[0032] For the task and drones Estimate the access time from the anchor point to the task start point. Delivery time from the start point of the mission to the end point of the mission. Pickup service hours and delivery service time The cost of a single-task service is: in, Indicates at time The task Assigned to drones The estimated service cost.

[0033] The estimated completion time for the task is: in, Indicates task By drone The estimated completion time during execution.

[0034] The estimated number of late arrivals is: in, Indicates the estimated late time. Indicates task The service deadline. If the estimated completion time does not exceed the service deadline, the lateness count is 0.

[0035] Simultaneously, based on task waiting time, estimated completion time, service deadline, and basic timeliness risk, the timeliness risk after task allocation is calculated. .in, Indicates task Assigned to drones Risks of timeout failure or service breach afterward.

[0036] (3) Multi-objective function; Low-altitude logistics scheduling inherently involves multiple conflicting objectives: to improve service efficiency, the system tends to choose allocation schemes with shorter distances and lower service costs; to reduce timeliness risks, the system needs to prioritize tasks with long waiting times or tight deadlines; and to maintain long-term stable operation, the system also needs to avoid over-concentration of tasks on a few drones. Therefore, directly using a single objective or a single-objective function with pre-fixed weights can easily bias the scheduling results towards a certain metric, leading to increased lateness risks or drone load imbalances. This application employs multi-objective modeling, enabling swarm intelligence search to retain Pareto candidate solutions among efficiency, risk-timeliness, and load balancing, and then selects the final scheme based on operational preferences.

[0037] The service efficiency target is: in, Indicate candidate solutions Total service cost Indicates task Assigned to its affiliated drone The estimated cost of subsequent services.

[0038] The risk timeliness target is: in, This indicates the risk, timeliness, and cost of the candidate solutions. Indicates task Assigned to drones The risk of delays afterward Indicates the expected amount of delay. This indicates the late normalization scale.

[0039] The load balancing target is first defined as the drone. Rolling workload: in, Indicates drone The estimated workload under the current window consists of two parts: the first part represents the remaining time occupied by the drone due to existing tasks, and the second part represents the sum of the service costs of newly assigned tasks to the drone.

[0040] The degree of load concentration is then represented by the variance of the workload of each drone: in, This represents the load balancing cost of the candidate solution, and Var(·) represents the variance function. The larger this value, the more concentrated the tasks are on a subset of drones.

[0041] Therefore, the upper-level multi-objective task allocation model is as follows: in, This is the multi-objective evaluation vector for the candidate solutions.

[0042] (4) Swarm intelligence search and solution selection; The upper layer allocates vectors using integers. The population is initialized for individuals, and candidate task assignment schemes are generated through crossover, mutation, and rounding repair. Each candidate individual is... , , A joint evaluation yields a multi-objective fitness vector.

[0043] In each iteration, the parent and child generations are merged and then sorted non-dominated to obtain different Pareto levels; then, crowding distance is used to maintain the diversity of solutions within the same level. After a preset number of iterations, the Pareto candidate solution set within the current rolling window is obtained. .

[0044] Since a rolling window can only write back one allocation result at a time, the final solution is selected from the Pareto solution set based on preference weights. Let the preference weights be: in, This represents the normalized preference weight vector. , , These correspond to efficiency costs, risk timeliness costs, and load balancing costs, respectively.

[0045] After normalizing the target value, the final solution is: in, This indicates the final allocation scheme selected in the current scrolling window. This represents the normalized target vector. This represents the preference rating function. Ultimately, it will... The newly committed tasks for each drone are decoded and written into the lower-level execution environment.

[0046] II. Lower-level reinforcement learning for drone decision-making; (1) MDP modeling; The lower layer models the local execution process of the drone as a Markov decision process: in, Representing the Markov decision process, For local observation space, For discrete rule action space, This is an event-driven state transition function. For instant reward function, This is the discount factor.

[0047] When drones When given a decision-making opportunity, the system constructs local observations. Shared reinforcement learning strategy output actions The environment decodes actions into specific tasks, order requests, or waiting events, and returns rewards and the next observation after the system state progresses. Multiple drones share the same strategy, but the strategy is invoked asynchronously at different event times.

[0048] (2) State space; Lower-level local observations are defined as 37-dimensional vectors: in, Indicates drone At any moment The local observation status, It represents a 37-dimensional real space.

[0049] This vector consists of five types of features: in, This indicates the availability of actions, including whether there are executable tasks, whether there are pickup candidates, whether there are delivery candidates, whether new tasks can be requested, and whether the requested action is available. This indicates the pressure on committed tasks, including the number of committed tasks, the weight of committed tasks, the onboard load, the pressure on pickup time limits, and the pressure on delivery time limits; This indicates the available task pressure, including the number of available tasks, the optimal estimated completion time, the minimum estimated delay, the maximum timeliness risk, and the total weight. Indicates the task type corresponding to the rule prototype; This indicates the expected consequences of different rule actions, including task completion time, final completion time, and expected delay.

[0050] This state design compresses a variable number of tasks into a fixed-dimensional input, ensuring that the reinforcement learning policy does not change the input structure as the task size changes.

[0051] (3) Action space; The lower-level action space is defined as seven types of discrete rule actions: in, This represents a set of actions, with "nearest" indicating the feasible task with the shortest travel time. This indicates the option to select the feasible task with the earliest service deadline. This indicates that the feasible task with the smallest expected service margin is selected; Indicates priority delivery; Indicates priority pickup; This indicates a request for a new task from the upper layer; wait indicates waiting for subsequent events.

[0052] The first five types of actions are rule-based actions and do not have directly assigned task numbers. The system constructs a set of feasible tasks based on the drone's current location, battery level, payload, committed tasks, loaded tasks, and safe return constraints. .in, Indicates drone At any moment A set of executable candidate tasks. Only tasks that meet the constraints of task lifecycle, load capacity, power consumption, and safe return are included in this set. When the policy selects a rule action, the task decoder... Select the corresponding task to execute.

[0053] (4) Reward function and training; The reward function is used to correlate local actions with the overall quality of low-altitude logistics services. Let the number of tasks completed in one decision step be... The number of tasks completed on time was The number of assigned tasks that have failed is The final number of unserved tasks is The total time of being late is The instant reward is: in, _ Indicates time Instant rewards to The weighting coefficients for each reward or penalty item. This serves as a standardized measure for lateness. The first two criteria encourage task completion and on-time completion, while the latter three penalize task failure, non-service, and lateness. The reward does not separately reward intermediate actions such as pickup or delivery; rather, it directly corresponds to task completion rate and service timeliness.

[0054] The lower layer adopts the PPO training sharing strategy, and its network structure is as follows: Figure 2 As shown, the policy receives observations at each local decision point, outputs the probability distribution of seven action categories, and updates parameters based on the rewards and state transitions returned by the environment. After training, each UAV asynchronously invokes the same shared policy during actual execution.

[0055] III. Upper and lower level collaboration mechanism; The upper and lower layers coordinate their operation through an order request-allocation feedback mechanism. When the lower layer outputs its strategy... At that time, the drone submits a request to the platform to receive new tasks; the platform collects drone requests within the request window and forms a set of requesting drones. When the window expires or the triggering condition is met, the upper layer determines the current set of tasks to be executed. and request drone assemblies Perform intelligent allocation of multiple target groups.

[0056] After the upper layer writes the allocation results back to the drone's committed task set, the lower layer does not consistently execute the complete path. Instead, it reconstructs the feasible task set at each local decision and selects actions such as pickup, delivery, requesting new tasks, or waiting based on the real-time status. In this way, the upper layer is responsible for global task allocation and candidate compression, while the lower layer is responsible for real-time execution and disturbance response.

[0057] When events such as task completion, task failure, weather changes, drone return for recharging, or waiting occur during the lower-level execution process, the system updates the global state and feeds the updated results back to the upper-level rolling allocation for the next round. Through this closed-loop mechanism, continuously arriving tasks are absorbed by the upper-level rolling window, the large-scale task pool is compressed by the upper layer into a set of limited commitment tasks, and sudden disturbances are responded to in real time by the lower layer and fed back to the subsequent scheduling process.

[0058] This invention decomposes large-scale low-altitude logistics task scheduling into two collaborative layers: an upper-level multi-objective group intelligent task allocation layer and a lower-level UAV local execution decision layer. At the upper layer, a rolling time window continuously integrates new tasks, incomplete tasks, and UAV status, and combines task timeliness, service risk, and load balancing objectives to generate task-UAV allocation results, thereby solving the problems of continuity and scalability in complex scenarios. At the lower layer, based on the real-time location, battery status, and other conditions of isomorphic UAVs and the set of executable tasks, a centrally trained model selects local execution actions such as picking up goods, delivering, receiving new tasks, or waiting, thereby addressing sudden environmental disturbances.

[0059] Through the above design, the present invention can reduce the combination complexity caused by direct matching of large-scale tasks and drones, avoid the learning strategy from directly facing the action space that changes with the task scale, and enable the local execution state of the drone to be fed back to the subsequent rolling allocation process, thereby improving the completion rate of low-altitude logistics tasks, enhancing the response capability to sudden disturbances, and improving the utilization efficiency of multi-drone capacity.

[0060] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for online dynamic drone-based layered delivery under weather interference, characterized in that, Includes the following steps: Obtain the set of tasks to be executed at the current scheduling time, the set of drones requesting to receive new tasks, and weather and environmental information; construct task snapshots and drone snapshots. Based on the task snapshot and the UAV snapshot, an integer vector code is used to represent the attribution relationship between the task and the UAV, thus obtaining the integer vector code; The integer vector is encoded and decoded into the associated drone for each task. Task-drone pair service estimation is performed for each task and its associated drone to obtain the estimated service cost. Based on the estimated service cost and the time constraints of each task, the integer vector code is evaluated in multiple objectives, and a search algorithm is used to search the solution space of the integer vector code to obtain a candidate solution set. The final allocation scheme is selected from the candidate solution set according to the preset preference weights and decoded into a set of committed tasks for each UAV; During execution, when a UAV has a decision opportunity, it outputs a discrete rule action based on the committed task set and its own local observation state. When the output discrete rule action is a request for a new task, it submits a request to the scheduling platform to receive a new task. The scheduling platform collects the requests from each UAV in the request window, performs task allocation, and feeds back the allocation result to the committed task set of each UAV to complete the hierarchical delivery.

2. The online dynamic drone layered delivery method under weather interference according to claim 1, characterized in that, The task snapshot includes the task start point, task end point, pending execution time, service deadline, task weight, and timeliness risk; the drone snapshot includes the drone's current location or current target location, available time, speed, battery level, payload, and existing committed tasks.

3. The online dynamic drone layered delivery method under weather interference according to claim 1, characterized in that, The integer vector encoding uses the number of tasks to be assigned in the current scheduling window as the vector dimension. Each element in the integer vector encoding corresponds to a task, and the value of each element is any drone number in the set of drone numbers requesting to participate in the allocation. The integer vector encoding is only used to determine the drone to which each task belongs, and does not limit the order in which each drone picks up and delivers tasks in its committed task set.

4. The online dynamic drone layered delivery method under weather interference according to claim 1, characterized in that, The mission-drone pair service estimation includes: Based on the drone anchor location and available time, estimate the access time from the anchor location to the mission start point, the delivery time from the mission start point to the mission end point, the pickup service time, and the delivery service time. Calculate the single-task service cost based on the access time, the delivery time, the pickup service time, and the delivery service time. Calculate the estimated completion time of the task based on the available time of the drone and the single-task service cost; Calculate the estimated lateness based on the estimated completion time and deadline of the task; The timeliness risk after task allocation is calculated based on the task waiting time, the expected completion time of the task, and the task deadline.

5. The online dynamic drone layered delivery method under weather interference according to claim 1, characterized in that, The objectives adopted in the multi-objective evaluation include service efficiency objectives, risk timeliness objectives, and load balancing objectives. The service efficiency target is the sum of the single-task service costs of all assigned tasks in the candidate scheme; The risk timeliness target is the weighted sum of the timeliness risk of each task in the candidate scheme and the expected amount of delay; The load balancing objective is the variance of the estimated workload of each UAV under the current scheduling window. The estimated workload is determined based on the sum of the remaining occupancy time of the existing tasks of the UAV and the service cost of the newly assigned tasks.

6. The online dynamic drone layered delivery method under weather interference according to claim 1, characterized in that, The search algorithm is a multi-objective evolutionary algorithm.

7. The online dynamic drone layered delivery method under weather interference according to claim 1, characterized in that, Selecting a final allocation scheme from the candidate solution set according to preset preference weights includes: The multi-objective evaluation values ​​of each candidate solution in the candidate solution set are normalized respectively; The preference score for each candidate solution is calculated by weighting and summing the normalized target values ​​with the preset preference weights. The candidate scheme with the smallest preference score is selected as the final allocation scheme.

8. The online dynamic drone layered delivery method under weather interference according to claim 1, characterized in that, The local observation state is a fixed-dimensional vector, including action availability features, committed task pressure features, requestable task pressure features, task type features corresponding to the rule prototype, and expected consequences features of different rule actions. The availability characteristics of the action include whether there is an executable task, whether there is a pickup candidate, whether there is a delivery candidate, whether a new task can be requested, and whether the requested action is available; The pressure characteristics of the committed tasks include the number of committed tasks, the weight of committed tasks, the onboard load, the pressure of the pickup time limit, and the pressure of the delivery time limit; The requested task pressure characteristics include the number of requested tasks, the optimal estimated completion time, the minimum estimated delay, the maximum timeliness risk, and the total weight.

9. The online dynamic drone layered delivery method under weather interference according to claim 1, characterized in that, The discrete rule actions include selecting the feasible task with the shortest travel time, selecting the feasible task with the earliest service deadline, selecting the feasible task with the smallest expected service margin, prioritizing delivery, prioritizing pickup, requesting a new task, and waiting. When the output discrete rule actions are the first five types of actions, a set of feasible tasks is constructed based on the UAV's current location, battery level, payload, committed tasks, loaded tasks, and safe return constraints. Then, a task that meets the corresponding rule conditions is selected from the set of feasible tasks for execution.

10. The online dynamic drone layered delivery method under weather interference according to claim 1, characterized in that, The discrete rule actions are output by the reinforcement learning policy. Multiple drones share the same reinforcement learning policy network, and each drone asynchronously calls the reinforcement learning policy network during execution. The immediate reward of the reinforcement learning strategy is constructed based on the number of tasks completed in a single decision step, the number of tasks completed on time, the number of assigned tasks that failed, the number of unserved tasks at the end of the process, and the total late time.

Images