A self-adaptive cooperative scheduling method for distributed tasks in a multi-field environment

Abstract

The application discloses a kind of multi-airport environment under distributed task's adaptive cooperative scheduling method, it is related to task scheduling technical field, including: according to the priority of to-be-executed task and task geographical position, generate preliminary flight path and extract time sequence meteorological feature sequence according to space-time range;According to the time sequence change rate of meteorological element in time sequence meteorological feature sequence, calculate path meteorological complexity coefficient;Predict the predicted power consumption by the pre-constructed power consumption prediction model array;The real-time available power of corresponding nest is compared with predicted power consumption, when real-time available power is greater than or equal to the product of predicted power consumption and preset safety factor, preliminary flight path is used as confirmed task, otherwise, to-be-executed task is marked as overflow task, trigger multi-airport collaborative negotiation process.The technical problem that task scheduling response speed is slow, efficiency is poor, and it is difficult to cross regional scheduling resources in prior art due to task redistribution relies on manual intervention is solved.

Description

Technical Field

[0001] This invention relates to the field of task scheduling technology, specifically to an adaptive collaborative scheduling method for distributed tasks in a multi-airport environment. Background Technology

[0002] In power distribution network inspection scenarios, by deploying multiple drone nests / smart airports within the inspection area, drones can take off and land nearby, charge autonomously, and continue their tasks, effectively expanding the coverage of a single operation and improving inspection efficiency.

[0003] However, most existing multi-airport drone scheduling technologies adopt a model of allocating tasks first and then assessing power consumption, without fully considering the impact of meteorological factors on flight energy consumption. At the same time, when an airport is unable to accept new tasks due to insufficient power, drone malfunction, or sudden task backlog, task redistribution relies on manual intervention, resulting in slow response speed, low efficiency, and difficulty in achieving optimal resource allocation and load balancing across regions.

[0004] Therefore, there is a need for an intelligent scheduling method that can comprehensively consider multiple factors such as task priority, real-time power consumption, and weather conditions in a multi-airport distributed environment, and achieve adaptive task matching, dynamic collaborative scheduling, and global resource optimization. Summary of the Invention

[0005] This invention provides an adaptive collaborative scheduling method for distributed tasks in a multi-airport environment, which solves the technical problems in the prior art where task redistribution relies on manual intervention, resulting in slow task scheduling response speed, poor efficiency, and difficulty in scheduling resources across regions.

[0006] This invention provides an adaptive cooperative scheduling method for distributed tasks in a multi-airport environment, the method comprising: Based on the priority and geographical location of the task to be executed, a preliminary flight path is generated for the task to be executed. Based on the spatiotemporal range traversed by the preliminary flight path, a time-series meteorological feature sequence distributed along the path is extracted from the local meteorological grid data of the multi-airport area. The time-series meteorological feature sequence includes the wind speed vector and precipitation intensity on each path segment. The path meteorological complexity coefficient is calculated based on the temporal change rate of meteorological elements in the temporal meteorological feature sequence. Based on the path meteorological complexity coefficient and the time-series meteorological feature sequence, the predicted power consumption is obtained by using a pre-constructed power consumption prediction model array. The predicted power consumption is compared with the real-time available power of the corresponding airport. When the real-time available power is greater than or equal to the product of the predicted power consumption and the preset safety factor, the preliminary flight path is assigned as a confirmation task. Otherwise, the task to be executed is marked as an overflow task, triggering a multi-airport collaborative negotiation process to reallocate the overflow task to other available airports.

[0007] Preferably, a preliminary flight path is generated for the task to be executed based on its priority and geographical location, including: Starting from the location of the airport nest corresponding to the task to be executed and ending at the geographical location of the task to be executed, a search tree is initialized in the three-dimensional airspace, with the starting point as the root node of the search tree; Candidate nodes are randomly sampled and generated in the three-dimensional spatial domain, and the candidate nodes are guided and expanded in the direction of the endpoint to generate new nodes; The new node is added to the search tree, and the distance between the new node and the destination is checked. When the distance is less than a preset termination distance threshold, the new node is connected to the destination to generate a preliminary route connecting the starting point and the destination. Collision detection is performed on the preliminary route to determine whether the preliminary route intersects with the restricted area defined by the static geofence constraint. If it intersects, random sampling, node expansion and collision detection are performed again until a route that does not intersect with the restricted area is generated, thus obtaining a collision-free route. The final collision-free route will be used as the initial flight path.

[0008] Preferably, based on the spatiotemporal range traversed by the preliminary flight path, a time-series meteorological feature sequence distributed along the path is extracted from the local meteorological grid data of the multi-airport area, including: The preliminary flight path is divided into multiple continuous path segments according to a preset spatial interval, and the geographical coordinate range and expected flight time period corresponding to each path segment are determined. Based on the geographic coordinate range and expected flight time of each path segment, the matching spatiotemporal grid cell is retrieved from the local meteorological grid data of the multi-airport area, and the meteorological element values ​​stored in the spatiotemporal grid cell are extracted, wherein the meteorological element values ​​include the direction angle and magnitude of the wind speed vector and the precipitation intensity. The meteorological element values ​​corresponding to each path segment are arranged sequentially according to the flight direction from the starting point to the end point to form a time-series meteorological characteristic sequence. When a single path segment crosses multiple meteorological grid units, the meteorological element values ​​of each meteorological grid unit are weighted and averaged according to the proportion of the overlap area between the path segment and each meteorological grid unit, and this average is used as the meteorological element value of the single path segment.

[0009] Preferably, when a single path segment crosses multiple meteorological grid units, the meteorological element values ​​of each meteorological grid unit are weighted and averaged according to the proportion of the overlap area between the path segment and each meteorological grid unit, and this average is used as the meteorological element value of the path segment, including: The intersection calculation is performed between the geographic projection range of the single path segment in the local meteorological grid data and the spatial range of each meteorological grid unit to determine each meteorological grid unit covered by the single path segment. Calculate the overlap area between each individual path segment and each meteorological grid unit; The proportion of the overlapping area of ​​each meteorological grid unit to the total area of ​​the single path segment is calculated and used as the weighting coefficient of each meteorological grid unit. The meteorological element values ​​for a single path segment are obtained by multiplying the wind speed vector direction angle, wind speed vector magnitude, and precipitation intensity stored in each meteorological grid unit by their respective weighting coefficients and then summing them up.

[0010] Preferably, the path meteorological complexity coefficient is calculated based on the temporal change rate of meteorological elements in the temporal meteorological feature sequence, including: Calculate the change in the directional angle of the wind speed vector between adjacent path segments in the time-series meteorological feature sequence, and arrange them in the temporal order of the path segments to obtain the sequence of directional change. Calculate the magnitude change of wind speed vector between adjacent path segments in the time-series meteorological feature sequence, and arrange them according to the temporal order of the path segments to obtain the magnitude change sequence; Calculate the absolute value of the difference in precipitation intensity between adjacent path segments in the time-series meteorological feature sequence, arrange them in the temporal order of the path segments, and obtain the precipitation intensity change sequence. Calculate the coefficient of variation of the direction change sequence, the coefficient of variation of the magnitude change sequence, and the coefficient of variation of the precipitation intensity change sequence. Normalize each coefficient of variation to obtain the first dynamic weight of the direction change, the second dynamic weight of the magnitude change, and the third dynamic weight of the precipitation intensity change. Assign the first dynamic weight, the second dynamic weight, and the third dynamic weight to the sequence of directional changes, the sequence of magnitude changes, and the sequence of precipitation intensity changes, respectively, perform weighted summation to obtain a weighted sum value, and normalize the weighted sum value to the interval between 0 and 1 to obtain the path meteorological complexity coefficient.

[0011] Preferably, based on the path meteorological complexity coefficient and the time-series meteorological feature sequence, the predicted power consumption is obtained by predicting the power consumption through a pre-constructed power consumption prediction model array, including: Obtain a pre-built power consumption prediction model array, which contains N power consumption prediction models, where N is a positive integer greater than or equal to 1; Calculate the product of the path meteorological complexity coefficient and N, and then round the product up to obtain the number of models K to be called in this prediction, where 1≤K≤N; K power consumption prediction models are randomly selected from the power consumption prediction model array as the K calling models for this prediction; Extract the path features of the preliminary flight path, which include the total path length, spatial distance of each path segment, flight altitude, preset flight speed, and mission payload weight; The path features and the time-series meteorological feature sequence are concatenated and uniformly encoded according to a preset format to generate a standardized input feature vector. The input feature vector is then input into the K calling models respectively, and K initial predicted power consumptions are output. Calculate the arithmetic mean of the K initial predicted power consumptions, and use the arithmetic mean as the predicted power consumption.

[0012] Preferably, obtaining a pre-built array of power consumption prediction models includes: Multiple sets of historical flight data samples are acquired. Each set of historical flight data samples includes historical path characteristics, historical time-series meteorological characteristic sequences, and corresponding historical actual power consumption. Based on the long short-term memory network architecture, construct N power consumption prediction models; The multiple sets of historical flight data samples are divided into N folds. Each fold is used as the validation set and the remaining N-1 folds are used as the training set. Supervised training is performed on the N power consumption prediction models until the validation convergence is achieved, resulting in N trained power consumption prediction models. The N trained power consumption prediction models are combined into a power consumption prediction model array.

[0013] Preferably, the task to be executed is marked as an overflow task, triggering a multi-airport collaborative negotiation process to reallocate the overflow task to other available airports, including: Based on the overflow task's geographical location, the predicted power consumption, the task priority, and the task time window constraints, a standardized task proposal is generated. The task tender is broadcast to all available airports in the region except the original airport through a multi-airport collaborative communication network. Within a preset bidding time window, the available airports submit bidding information in response to the task tender. The bidding information includes the comprehensive cost price calculated by each available airport for executing the overflow task, wherein the comprehensive cost price is generated by each available airport through a locally pre-trained bidding decision model. From all the received bids, the available airport with the lowest overall cost price is selected as the winning airport; The task data and preliminary flight path of the overflow task are sent to the winning airport to complete the redistribution of the overflow task.

[0014] Preferably, the pre-training process of the bidding decision model includes: Obtain historical bidding data for this airport. Each set of historical bidding data includes historical task bids, the status information of this airport at the time the historical task bids were received, and the corresponding optimal comprehensive cost price tag. The status information of this airport includes the number of available drones at this airport, the real-time power of the drone nest, the estimated power consumption and estimated completion time of each task in the scheduled task queue. A training sample set is constructed using the historical task tender documents and the airport status information as input features, and the optimal comprehensive cost pricing label as a supervision label. An initial bidding decision model is constructed based on reinforcement learning; The initial bidding decision model is trained using the training sample set. During the training process, the network parameters of the initial bidding decision model are updated with the goal of maximizing the cumulative expected value of the reward function until convergence is verified, thus obtaining the pre-trained bidding decision model.

[0015] Preferably, the network parameters of the initial bidding decision model are updated with the training objective of maximizing the cumulative expected value of the reward function, including: When the airport successfully executes the overflow task it has undertaken, the reward function provides a positive reward. The positive reward is positively correlated with the task priority of the overflow task and negatively correlated with the direct energy consumption cost of executing the overflow task. When accepting the overflow task causes delays in other scheduled tasks at the airport, the reward function provides a negative reward, which is positively correlated with the delay duration. When the overflow task fails due to insufficient real-time power in the airport's hangar after being accepted, the reward function provides the maximum negative reward. The network parameters of the initial bidding decision model are updated with the training objective of maximizing the cumulative expected value of the positive reward, the negative reward, and the maximum negative reward, until convergence is verified, thus obtaining a pre-trained bidding decision model.

[0016] One or more technical solutions provided in this invention have at least the following technical effects or advantages: This invention provides an adaptive collaborative scheduling method for distributed tasks in a multi-airport environment. First, a preliminary flight path is generated based on task priority and geographical location, ensuring that high-priority tasks receive priority path planning resources. Second, the preliminary flight path is spatiotemporally matched with local meteorological grid data to extract time-series meteorological feature sequences distributed along the path, providing a meteorological data input basis for subsequent power consumption prediction. Third, based on the time-series change rate of meteorological elements in the meteorological feature sequence, the path meteorological complexity coefficient is calculated to comprehensively reflect the severity of dynamic meteorological changes faced by the entire flight path, avoiding insufficient accuracy or wasted computational resources caused by using the same prediction model under all meteorological conditions. Fourth, based on the path meteorological complexity coefficient and the time-series meteorological feature sequence, a pre-constructed power consumption prediction model array is built to predict the predicted power consumption. The path meteorological complexity coefficient determines the number of models called in this prediction, solving the problem of insufficient generalization ability of a single prediction model under different meteorological conditions. Finally, the predicted power consumption is compared with the real-time available power of the corresponding airport, realizing a task allocation decision mechanism, effectively solving the airport resource scheduling problem, and improving the overall task execution success rate and resource utilization efficiency.

[0017] Furthermore, before accepting any overflow tasks, the winning airport performs an independent verification step. This involves adding the predicted power consumption of the overflow task to the remaining estimated power consumption of each task in the airport's already scheduled task queue. Acceptance is only confirmed if the total estimated power consumption after acceptance does not exceed the product of the real-time available power of the terminal and a preset proportional threshold; otherwise, the airport actively rejects the task and returns a reason for rejection. This verification step serves as a safety layer after the bidding decision model output. It effectively avoids the risk of power depletion due to model prediction errors or uncontrolled model output under extreme conditions, prevents the cascading effect of accepting new tasks causing delays in already scheduled tasks, and improves the reliability and task fulfillment rate of multi-airport collaborative scheduling in actual operation. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating an adaptive collaborative scheduling method for distributed tasks in a multi-airport environment provided by the present invention. Figure 2 This is a logical schematic diagram of an adaptive collaborative scheduling method for distributed tasks in a multi-airport environment provided by the present invention. Detailed Implementation

[0019] This invention provides an adaptive collaborative scheduling method for distributed tasks in a multi-airport environment, which solves the problems in the prior art.

[0020] The present invention will now be described in detail with reference to the accompanying drawings.

[0021] In one embodiment, such as Figure 1 , Figure 2 As shown, this invention provides an adaptive cooperative scheduling method for distributed tasks in a multi-airport environment, the method comprising: S100: Based on the priority and geographical location of the task to be executed, a preliminary flight path is generated for the task to be executed, and based on the spatiotemporal range traversed by the preliminary flight path, a time-series meteorological feature sequence distributed along the path is extracted from the local meteorological grid data of the multi-airport area. The time-series meteorological feature sequence includes the wind speed vector and precipitation intensity on each path segment. In this embodiment of the invention, the task to be executed is the inspection operation task that needs to be performed by the UAV in the power distribution network UAV autonomous inspection and scheduling system; priority is a quantitative indicator that sorts multiple tasks to be executed according to their importance and urgency, and the higher the priority, the earlier the task should be allocated resources and executed. The task's geographical location refers to the position coordinates of the inspection target point of the task to be executed in three-dimensional space; the preliminary flight path is the initial route formed by connecting the waypoints in three-dimensional space in sequence from the UAV's nest position to the task's geographical location.

[0022] Specifically, the location of the airport's hive for the task to be executed is obtained as the flight start point, and the geographical location of the task is obtained as the flight destination. A fast randomized expanding tree algorithm is used for path search in the three-dimensional airspace. Candidate nodes are randomly sampled and expanded towards the destination, gradually generating new nodes and adding them to the search tree to generate an initial flight path. Collision detection is then performed to determine whether the path intersects with restricted areas defined by static geofence constraints.

[0023] Step S100 in the method of this embodiment of the invention includes: Starting from the location of the airport nest corresponding to the task to be executed and ending at the geographical location of the task to be executed, a search tree is initialized in the three-dimensional airspace, with the starting point as the root node of the search tree; Candidate nodes are randomly sampled and generated in the three-dimensional spatial domain, and the candidate nodes are guided and expanded in the direction of the endpoint to generate new nodes; The new node is added to the search tree, and the distance between the new node and the destination is checked. When the distance is less than a preset termination distance threshold, the new node is connected to the destination to generate a preliminary route connecting the starting point and the destination. Collision detection is performed on the preliminary route to determine whether the preliminary route intersects with the restricted area defined by the static geofence constraint. If it intersects, random sampling, node expansion and collision detection are performed again until a route that does not intersect with the restricted area is generated, thus obtaining a collision-free route. The final collision-free route will be used as the initial flight path.

[0024] In this embodiment of the invention, firstly, the search tree is rooted at the starting point, with the airport nest location corresponding to the task to be executed as the starting point and the task's geographical location as the ending point. Here, the nest is the storage, charging, maintenance, and take-off / landing base for the UAV, while the airport nest location is the geographical coordinate of the UAV nest deployment. The fast expanding random tree algorithm is a path planning algorithm based on random sampling, which explores feasible paths by randomly sampling in the state space and gradually expanding the tree structure.

[0025] Secondly, based on the initial search tree constructed in the three-dimensional spatial domain, candidate nodes are randomly sampled and generated in the three-dimensional spatial domain. These candidate nodes are then guided to expand towards the endpoint, with a fixed step size to generate new nodes. Specifically, the initial search tree includes: setting the starting point as the root node, setting its parent node to null, and setting the accumulated cost to 0; adding the root node to the initial search tree; and setting the current iteration counter to 0.

[0026] The steps for generating new nodes include: random sampling, finding the nearest node, and directional guidance for expansion. Random sampling generates a random number between 0 and 1. A target bias probability is set, and during random sampling, the endpoint is directly used as the sampling point with a certain probability, guiding the tree to grow towards the target. The target bias probability is then compared with the generated random number. If the random number is less than or equal to the target bias probability, this sampling point is directly set as the endpoint; otherwise, uniform random sampling continues within the three-dimensional search space.

[0027] Then, the nearest node is found. In the initial search tree, the tree node with the closest Euclidean distance to the sampling point is found. After the nearest node is found, directional guidance expansion is performed: calculate the unit direction vector from the nearest node to the endpoint, and then move forward one step along this direction from the nearest node to generate a new node, that is, new node = nearest node + step × (end point - nearest node) / |end point - nearest node|.

[0028] Next, after generating a valid new node, it is added to the search tree, and the nearest node is set as its parent node. The cumulative cost is updated to the cumulative cost of the nearest node plus the Euclidean distance from the nearest node to the new node. Then, the Euclidean distance between the new node and the destination is checked. If this Euclidean distance is less than a preset termination distance threshold, such as 20 meters, the termination condition is met, and the new node is directly connected to the destination, forming a preliminary route from the starting point to the destination. If the distance is not less than the preset termination distance threshold, the iteration continues, searching for other nodes of the new node and attempting to reconnect them to reduce the cumulative cost. If a better reconnection is found, the parent node is updated.

[0029] The preset termination distance threshold can be set based on the task's requirements for endpoint accuracy, the inherent error of the UAV's navigation and positioning system, and the safety margin of the surrounding environment of the target point. If the task requires the UAV to accurately reach the designated coordinate point for high-precision detection, the preset termination distance threshold should be set relatively small, but must be greater than the positioning error to avoid the algorithm failing to converge due to noise interference. If the task only needs to fly over the target area for inspection, the preset termination distance threshold can be appropriately relaxed to improve planning efficiency. For example, if the task requires the UAV to hover above the base of a tower for temperature measurement, and the GPS positioning error is about 5 meters, the preset termination distance threshold can be set to 10 meters. If it only needs to fly over a designated area for overview inspection, the preset termination distance threshold can be relaxed to 50 meters to accelerate the path search process.

[0030] Furthermore, after generating the initial flight path, collision detection is required to ensure that the entire flight path does not intersect with restricted areas defined by static geofencing constraints. Restricted areas include buildings, power transmission towers, high-voltage line corridors, residential areas, airport airspace protection zones, etc.

[0031] Collision detection is performed on the initial route to determine whether it intersects with restricted areas defined by static geofencing constraints. Specifically, the initial route is discretized into a series of segments, each a line segment connecting two adjacent path points. For each segment, it is checked whether it intersects with the polyhedral model of the restricted area in 3D space. For example, a hierarchical bounding box can be used for coarse detection, followed by precise collision detection using a triangular mesh. If any segment intersects with a restricted area, a collision is determined, and the current initial route is unusable. At this point, the path point where the collision occurred and its subsequent branches are marked as invalid, and random sampling, node expansion, and collision detection are performed again, continuing to grow from the current tree and attempting to generate new paths around obstacles. This process is repeated until a collision-free route that does not intersect with any restricted areas is generated.

[0032] Finally, the generated collision-free route is used as the preliminary flight path. The preliminary flight path is represented by a series of three-dimensional waypoints, with estimated flight distances and times for each segment.

[0033] Step S100 in the method of this embodiment of the invention further includes: The preliminary flight path is divided into multiple continuous path segments according to a preset spatial interval, and the geographical coordinate range and expected flight time period corresponding to each path segment are determined. Based on the geographic coordinate range and expected flight time of each path segment, the matching spatiotemporal grid cell is retrieved from the local meteorological grid data of the multi-airport area, and the meteorological element values ​​stored in the spatiotemporal grid cell are extracted, wherein the meteorological element values ​​include the direction angle and magnitude of the wind speed vector and the precipitation intensity. The meteorological element values ​​corresponding to each path segment are arranged sequentially according to the flight direction from the starting point to the end point to form a time-series meteorological characteristic sequence. When a single path segment crosses multiple meteorological grid units, the meteorological element values ​​of each meteorological grid unit are weighted and averaged according to the proportion of the overlap area between the path segment and each meteorological grid unit, and this average is used as the meteorological element value of the single path segment.

[0034] In this embodiment of the invention, the initial flight path is first divided into multiple continuous path segments according to a preset spatial interval. Specifically, a complete waypoint sequence of the initial flight path is obtained. Starting from the starting point, the flight distance is accumulated along the path direction. Each time the accumulated distance reaches a preset spatial interval, the current accumulated endpoint is taken as a segment breakpoint. The last segment may be smaller than the spatial interval, and its natural endpoint is taken as the path endpoint. After division, the original path is divided into L continuous path segments, each segment containing its starting point coordinates, ending point coordinates, and spatial distance.

[0035] Subsequently, for each path segment, the line connecting its starting and ending points is used as the centerline. Starting from the starting point, the flight time of each segment is accumulated in flight sequence, and each flight time is the expected flight time period for each path segment. Simultaneously, the path is expanded outwards with a preset width to form a rectangular spatial coverage area, which represents the geographical coordinate range of that segment on the horizontal plane. In the vertical direction, the average Z-coordinate of the two endpoints of the segment is taken as the flight altitude, and combined with the vertical meteorological grid layer height, the vertical coordinate range is determined.

[0036] Secondly, based on the geographic coordinate range and expected flight time period of each path segment, matching spatiotemporal grid cells are retrieved from the local meteorological grid data of the multi-airport area. The coordinates of each path segment are converted to the coordinates used by the local meteorological grid. Then, the geographic coordinate range is intersected with the horizontal spatial grid of the local meteorological grid to determine which meteorological grid cells spatially overlap with that segment. Simultaneously, based on the expected flight time period of the segment, matching grid cells are searched in the time series of the meteorological grid data. Finally, the meteorological element values ​​stored in the spatiotemporal grid cells are extracted, including the direction angle and magnitude of the wind speed vector and the precipitation intensity.

[0037] Next, the meteorological element values ​​corresponding to each path segment are arranged sequentially according to the flight direction from the starting point to the end point to form a time-series meteorological characteristic sequence.

[0038] Ultimately, when a single path segment spans multiple meteorological grid units, the meteorological value of a single grid unit cannot be simply selected as the representative value for that segment, as this would introduce significant errors. Therefore, a weighted average of the meteorological element values ​​from each meteorological grid unit is calculated based on the proportion of overlap between the path segment and each meteorological grid unit, and this average is used as the meteorological element value for a single path segment.

[0039] In step S100 of the method of this embodiment of the invention, when a single path segment crosses multiple meteorological grid units, a weighted average of the meteorological element values ​​of each meteorological grid unit is calculated according to the proportion of the overlap area between the path segment and each meteorological grid unit, and this average is used as the meteorological element value of the path segment. This includes: The intersection calculation is performed between the geographic projection range of the single path segment in the local meteorological grid data and the spatial range of each meteorological grid unit to determine each meteorological grid unit covered by the single path segment. Calculate the overlap area between each individual path segment and each meteorological grid unit; The proportion of the overlapping area of ​​each meteorological grid unit to the total area of ​​the single path segment is calculated and used as the weighting coefficient of each meteorological grid unit. The meteorological element values ​​for a single path segment are obtained by multiplying the wind speed vector direction angle, wind speed vector magnitude, and precipitation intensity stored in each meteorological grid unit by their respective weighting coefficients and then summing them up.

[0040] In this embodiment of the invention, firstly, the geographic projection range of a single path segment in the local meteorological grid data is intersected with the spatial range of each meteorological grid unit. For each meteorological grid unit, the system calculates the intersection between its spatial range and the projection range of the path segment. If the intersection area is greater than zero, the grid unit is recorded as a meteorological grid unit covered by the path segment.

[0041] Secondly, the overlap area between each individual path segment and each meteorological grid cell is calculated. The overlap area can be calculated using a polygon clipping algorithm.

[0042] Specifically, the coordinates of the four corner points of the path segment coverage zone are arranged counterclockwise to form a convex polygon. The rectangular area of ​​each candidate meteorological grid cell is also represented as a polygon. The area of ​​the intersection of the polygons is calculated to obtain the overlapping area. Standard geometric algorithms can be used for this calculation, such as decomposing the intersection polygon into several triangles and then summing the areas of each triangle. When there is partial overlap between the path segment coverage zone and the grid cell boundary, the clipping algorithm can accurately calculate the area of ​​the overlapping region.

[0043] Next, the proportion of the overlapping area of ​​each meteorological grid unit to the total area of ​​a single path segment is calculated as the weighting coefficient of each meteorological grid unit. The weighting coefficient = overlapping area of ​​each meteorological grid unit / total area of ​​a single path segment. Since the sum of the overlapping areas equals the total area of ​​the coverage zone, the sum of all weighting coefficients equals 1. A larger weighting coefficient indicates a greater impact of the meteorological conditions of that grid unit on the entire flight path; a smaller weighting coefficient indicates that the meteorological conditions of that grid unit only affect the edge of the path segment, with limited impact.

[0044] Finally, the wind speed vector's direction angle, magnitude, and precipitation intensity stored in each meteorological grid cell are multiplied by their respective weighting coefficients. A linear weighted average is then applied directly to the precipitation intensity. However, due to the periodicity of the wind speed direction angle, linear weighted averaging of the angle values ​​can lead to calculation errors. Therefore, a vector synthesis method is used: first, the wind speed direction angle and magnitude of each grid cell are synthesized into a horizontal wind speed vector; then, each component is weighted and averaged to obtain the weighted averaged wind speed vector components; finally, the weighted averaged wind speed magnitude and direction angle are obtained.

[0045] In this embodiment of the invention, the RRT algorithm is used to quickly generate collision-free flight paths in a complex three-dimensional airspace, improving path planning efficiency. Furthermore, target-guided expansion and collision detection ensure path safety and search convergence. Then, the initial flight path is divided, and meteorological grid spatiotemporal matching is performed to extract meteorological element values ​​corresponding to each path segment. These values ​​are arranged according to the flight direction to generate a time-series meteorological feature sequence. Subsequently, the overlap area between each individual path segment and each meteorological grid unit is calculated as the weight coefficient of each meteorological grid unit. The meteorological element values ​​of each path segment are obtained through weighted summation, improving the accuracy of meteorological feature extraction and providing accurate input for energy consumption prediction.

[0046] S200: Calculate the path meteorological complexity coefficient based on the temporal change rate of meteorological elements in the temporal meteorological feature sequence; In this embodiment of the invention, the spatiotemporal range refers to the spatial and temporal range experienced by the UAV along the initial flight path from the starting point to the destination; the local meteorological grid data is three-dimensional meteorological gridded data covering the area where multiple airports are located, and each grid cell stores the meteorological element values ​​of the corresponding geographical area at a specific time or time period; the time-series meteorological feature sequence is the time-series data formed by arranging the meteorological element values ​​corresponding to each path segment in sequence according to the flight direction from the starting point to the destination after the initial flight path is spatially segmented.

[0047] Specifically, the initial flight path is divided into multiple continuous path segments according to a preset spatial interval, and the geographic coordinate range and expected flight time period corresponding to each path segment are determined. Subsequently, based on the geographic coordinate range and expected flight time period of each path segment, matching spatiotemporal grid units are retrieved from the local meteorological grid data of the multi-airport area, and the meteorological element values ​​stored in the grid unit are extracted. Finally, the meteorological element values ​​corresponding to each path segment are arranged sequentially according to the flight direction from the starting point to the end point to form a time-series meteorological feature sequence.

[0048] Step S200 in the method of this embodiment of the invention includes: Calculate the change in the directional angle of the wind speed vector between adjacent path segments in the time-series meteorological feature sequence, and arrange them in the temporal order of the path segments to obtain the sequence of directional change. Calculate the magnitude change of wind speed vector between adjacent path segments in the time-series meteorological feature sequence, and arrange them according to the temporal order of the path segments to obtain the magnitude change sequence; Calculate the absolute value of the difference in precipitation intensity between adjacent path segments in the time-series meteorological feature sequence, arrange them in the temporal order of the path segments, and obtain the precipitation intensity change sequence. Calculate the coefficient of variation of the direction change sequence, the coefficient of variation of the magnitude change sequence, and the coefficient of variation of the precipitation intensity change sequence. Normalize each coefficient of variation to obtain the first dynamic weight of the direction change, the second dynamic weight of the magnitude change, and the third dynamic weight of the precipitation intensity change. Assign the first dynamic weight, the second dynamic weight, and the third dynamic weight to the sequence of directional changes, the sequence of magnitude changes, and the sequence of precipitation intensity changes, respectively, perform weighted summation to obtain a weighted sum value, and normalize the weighted sum value to the interval between 0 and 1 to obtain the path meteorological complexity coefficient.

[0049] In this embodiment of the invention, firstly, the change in the directional angle of the wind speed vector between adjacent path segments in the time-series meteorological feature sequence is calculated, and the segments are arranged in temporal order to obtain a sequence of directional change. The change in directional angle... The value ranges from 0° to 180°, where θ i+ 1 and θ i It represents the change in two adjacent direction angles.

[0050] Secondly, the magnitude change of wind speed vector between adjacent path segments in the time-series meteorological characteristic sequence is calculated, and the segments are arranged in temporal order to obtain a sequence of magnitude changes. The magnitude change of wind speed vector is calculated as |wind speed vector of the previous adjacent segment - current wind speed vector|, with units of m / s.

[0051] Next, the absolute value of the difference in precipitation intensity between adjacent path segments in the time-series meteorological characteristic sequence is calculated, with the unit being mm / h. The segments are then arranged in temporal order to obtain the precipitation intensity change sequence.

[0052] Furthermore, the coefficients of variation for the directional change series, the magnitude change series, and the precipitation intensity change series are calculated. Specifically, for each series, its mean and standard deviation are calculated, and the ratio of the standard deviation to the mean is taken as the coefficient of variation for each series. Each coefficient of variation is then normalized by dividing the coefficient of variation of each series by its mean. This yields the first dynamic weight for the directional change, the second dynamic weight for the magnitude change, and the third dynamic weight for the precipitation intensity change. The sum of these three dynamic weights is 1.

[0053] Finally, the sequence of directional change, the sequence of magnitude change, and the sequence of precipitation intensity change are normalized to eliminate the dimensional differences among them. Then, the normalized sequence of directional change, the sequence of magnitude change, and the sequence of precipitation intensity change are assigned a first dynamic weight, a second dynamic weight, and a third dynamic weight, respectively, and then weighted and summed to obtain a weighted sum value. The weighted sum value is then normalized to the interval between 0 and 1 to obtain the path meteorological complexity coefficient.

[0054] Specifically, the normalization of the sequence of changes in direction, magnitude, and precipitation intensity can be achieved using extreme value normalization to determine the maximum and minimum values ​​of each change sequence. Theoretical upper limits or the maximum values ​​from historical statistical data can be used. For example, based on actual flight data, the maximum change in azimuth angle should not exceed 180° and the minimum should not be less than [a certain value], the maximum change in wind speed magnitude should not exceed 15 m / s, and the maximum change in precipitation intensity should not exceed 10 mm / h.

[0055] Therefore, the normalization upper limits are set as follows: maximum change in azimuth angle = 180°, maximum change in wind speed = 10 m / s, and maximum change in precipitation intensity = 5 mm / h. The minimum change value can be uniformly set to 0, because each change is an absolute value, and its physical lower limit is zero, corresponding to the case where there is no change in meteorological elements between adjacent segments; therefore, during normalization, only the maximum change value of each sequence needs to be determined as the denominator, and the change value can be directly divided by this maximum value to map to the [0,1] interval. The mean of each change value sequence is calculated, and then the mean of each change value sequence is multiplied by its respective dynamic weight and summed to obtain the path meteorological complexity coefficient, where the normalized change value = mean of the change value / change of the maximum change value.

[0056] In this embodiment of the invention, the coefficients of variation of the direction change sequence, the magnitude change sequence, and the precipitation intensity change sequence are calculated, and their dynamic weights are obtained by normalizing the coefficients of variation. Subsequently, each sequence is normalized, and then the normalized change sequences are weighted to make the meteorological element values ​​of each segment more accurately reflect the actual spatial distribution, improve the continuity and accuracy of the time-series meteorological characteristic sequences, and provide a reliable data foundation for power consumption prediction.

[0057] S300: Based on the path meteorological complexity coefficient and the time-series meteorological feature sequence, the predicted power consumption is obtained by predicting the power consumption through a pre-constructed power consumption prediction model array; In this embodiment of the invention, the temporal change rate is the degree of change of meteorological element values ​​between adjacent path segments in a temporal meteorological feature sequence; the path meteorological complexity coefficient refers to a comprehensive index used to quantify the complexity of meteorological conditions on the entire flight path; the coefficient of variation is the ratio of the standard deviation to the mean, used to measure the degree of dispersion of a set of data relative to its mean.

[0058] Specifically, firstly, the change in the directional angle of the wind speed vector between adjacent path segments in the time-series meteorological feature sequence is calculated, and these segments are arranged in temporal order to obtain the sequence of directional changes. Simultaneously, the change in the magnitude of the wind speed vector between adjacent path segments is calculated to obtain the sequence of magnitude changes. The absolute value of the difference in precipitation intensity between adjacent path segments is then calculated to obtain the sequence of precipitation intensity changes. Subsequently, the coefficients of variation for the directional change sequence, the magnitude change sequence, and the precipitation intensity change sequence are calculated respectively. Each coefficient of variation is normalized, assigned a corresponding dynamic weight, and then weighted and summed to obtain the path meteorological complexity coefficient.

[0059] Step S300 in the method of this embodiment of the invention includes: Obtain a pre-built power consumption prediction model array, which contains N power consumption prediction models, where N is a positive integer greater than or equal to 1; Calculate the product of the path meteorological complexity coefficient and N, and then round the product up to obtain the number of models K to be called in this prediction, where 1≤K≤N; K power consumption prediction models are randomly selected from the power consumption prediction model array as the K calling models for this prediction; Extract the path features of the preliminary flight path, which include the total path length, spatial distance of each path segment, flight altitude, preset flight speed, and mission payload weight; The path features and the time-series meteorological feature sequence are concatenated and uniformly encoded according to a preset format to generate a standardized input feature vector. The input feature vector is then input into the K calling models respectively, and K initial predicted power consumptions are output. Calculate the arithmetic mean of the K initial predicted power consumptions, and use the arithmetic mean as the predicted power consumption.

[0060] In this embodiment of the invention, firstly, a pre-constructed power consumption prediction model array is obtained. The power consumption prediction model array contains N power consumption prediction models, where N is a positive integer greater than or equal to 1. Each model in this array is trained separately based on different combinations of meteorological conditions, and each is adept at accurately predicting flight power consumption under specific meteorological scenarios.

[0061] Secondly, the path meteorological complexity coefficient represents the degree of drastic change in meteorological conditions along the entire flight path. A lower coefficient indicates a more gradual change in meteorological conditions along the path, resulting in more consistent predictions from various power consumption prediction models for the same input, requiring only a small number of models to obtain reliable predictions. Conversely, a higher coefficient indicates more drastic changes in meteorological conditions, leading to greater differences in predictions from different power consumption prediction models under different meteorological sub-scenarios, necessitating the integration of more power consumption prediction models to reduce prediction bias.

[0062] Therefore, the path meteorological complexity coefficient is multiplied by N, and the product is rounded up to obtain the number of power consumption prediction models K to be called in this prediction, i.e., K = ceil(path meteorological complexity coefficient × N). Here, 1 ≤ K ≤ N, which ensures that at least one power consumption prediction model participates in the prediction under the simplest meteorological conditions, avoiding the invalid state of zero model calls.

[0063] Next, K power consumption prediction models are randomly selected from the power consumption prediction model array as the K models to be used in this prediction. Random sampling without replacement can be used to extract K power consumption prediction models from the power consumption prediction model array to ensure that the same model is not selected repeatedly.

[0064] Furthermore, path features of the preliminary flight path are extracted. These features include the total path length, spatial distances of each path segment, flight altitude, preset flight speed, and mission payload weight. The total path length is the total length of the complete route from the starting point to the end point. The spatial distances of each path segment are the spatial distances of each segment after dividing the preliminary flight path according to preset spatial intervals, reflecting the length distribution of the preliminary flight path. The flight altitude is the average flight altitude of each segment of the preliminary flight path; since flight altitude affects air density and motor efficiency, it can affect power consumption. The preset flight speed is the flight speed preset by the UAV during the cruise phase, where flight speed directly affects motor output power and flight duration. The mission payload weight is the weight of the equipment carried by the mission, which directly affects lift requirements and motor power consumption.

[0065] Furthermore, the path features and time-series meteorological feature sequences are concatenated and uniformly encoded according to a preset format to generate standardized input feature vectors. These input feature vectors are then input into K calling models, and K initial predicted power consumption values ​​are output.

[0066] Finally, after obtaining K initial predicted power consumptions, the arithmetic mean of the K initial predicted power consumptions is calculated, and the arithmetic mean is used as the predicted power consumption.

[0067] Step S300 in the method of this embodiment of the invention, obtaining a pre-constructed power consumption prediction model array, includes: Multiple sets of historical flight data samples are acquired. Each set of historical flight data samples includes historical path characteristics, historical time-series meteorological characteristic sequences, and corresponding historical actual power consumption. Based on the long short-term memory network architecture, construct N power consumption prediction models; The multiple sets of historical flight data samples are divided into N folds. Each fold is used as the validation set and the remaining N-1 folds are used as the training set. Supervised training is performed on the N power consumption prediction models until the validation convergence is achieved, resulting in N trained power consumption prediction models. The N trained power consumption prediction models are combined into a power consumption prediction model array.

[0068] In this embodiment of the invention, firstly, multiple sets of historical flight data samples are obtained from a historical flight log database. Each set of historical flight data samples includes historical path features, a historical time-series meteorological feature sequence, and the corresponding historical actual power consumption. The historical path features record the actual flight path parameters of the UAV during a historical flight mission; the historical time-series meteorological feature sequence records the actual meteorological conditions encountered along the flight path during the historical flight mission; and the historical actual power consumption is the actual battery consumption recorded by the battery management system after the flight mission ends. For example, suppose 1200 sets of historical flight data samples from a power distribution network inspection area over the past year are obtained from the historical flight log database. Each sample contains complete path information, a meteorological condition sequence along the path, and an actual power consumption record. From these, 1000 sets of samples of acceptable quality are selected as the base dataset for model training.

[0069] Secondly, based on the long short-term memory network architecture, N power consumption prediction models are constructed.

[0070] Each power consumption prediction model has the same network architecture, but different initialization parameters. Specifically, the input layer dimension of each LSTM model is equal to the dimension of the input feature vector, i.e., the total dimension after concatenating the path features and the time-series meteorological features. The output layer is a scalar value, i.e., the predicted power consumption. The structure of the power consumption prediction model includes an input layer, an LSTM layer, a fully connected layer, and an output layer. The input layer receives the standardized input feature vector and includes 5 neurons; the LSTM layer captures the long-short-term dependencies in the time-series meteorological feature sequence, and can be set to two LSTM layers, each with 64 hidden units, using the tanh activation function, and a dropout rate of 0.2 to prevent overfitting; the fully connected layer maps the high-dimensional features output by the LSTM to the final power consumption prediction value, and includes 32 neurons using the ReLU activation function; the output layer contains 1 neuron, has no activation function, and directly outputs the regression value of the predicted power consumption.

[0071] Furthermore, to enable the N power consumption prediction models to learn different data feature distributions and form diverse prediction capabilities during subsequent training, each power consumption prediction model is initialized with different random seeds. Each power consumption prediction model can use slightly different combinations of hyperparameters, such as learning rate and dropout rate, during training to further enhance the diversity among the power consumption prediction models.

[0072] In a specific embodiment, N=5 is set, that is, five LSTM power consumption prediction models are constructed, denoted as M1, M2, M3, M4, and M5 respectively. Each model adopts a two-layer LSTM network structure, but uses different random initialization seeds, for example, randomly initialized seeds seed=42,123,456,789,1011, to ensure that the initial values ​​of the parameters are different, laying a differentiated foundation for subsequent cross-validation training.

[0073] Next, the historical flight data samples are divided into N folds. Each fold is used as the validation set and the remaining N-1 folds are used as the training set. Supervised training is performed on the N power consumption prediction models until the validation convergence is achieved, resulting in N trained power consumption prediction models.

[0074] Specifically, all historical flight data samples are divided into N equal subsets (N folds) according to chronological order or random method. Each fold contains 1000 sample groups divided by N. If the total number of samples is not divisible by N, random sampling is used to balance the number of samples in each fold, ensuring that the number of sample groups for each power consumption prediction model differs by no more than one group. After the division, N data folds are obtained.

[0075] For the j-th power consumption prediction model M j F, the corresponding j-th fold sample j The validation set is used as the training set, and all other samples are merged into it. During training, historical path features and historical time-series meteorological feature sequences from the training set are used as input feature vectors, and the corresponding historical actual electricity consumption is used as supervision labels. Mean squared error is used as the loss function, and the Adam optimizer is used to update the parameters. After each training epoch, the validation loss is calculated on the validation set of the current model. When the validation loss no longer decreases for several consecutive epochs, the model is considered to have converged, training is stopped, and the current model parameters are saved.

[0076] Because LSTM networks are sensitive to the order of input sequences, the original order of the time-series meteorological feature sequences needs to be maintained during training. In batch training, padding or variable-length sequence techniques are used to process samples of different lengths to ensure that each sample can be correctly input into the network.

[0077] In one feasible embodiment, 1000 sets of historical flight data samples are randomly divided into 5 folds, with 200 samples in each fold. When training the power consumption prediction model M1, F1 is used as the validation set, and the sum of F2, F3, F4, and F5 is used as the training set. During training, the batch size is set to 32, and the initial learning rate is 0.001. If the validation loss does not decrease after every 50 batches, the learning rate is reduced to 0.5 times the original value. The validation loss decreases to 0.032 and stabilizes after approximately 120 batches, at which point training stops, and the parameters of the power consumption prediction model M1 are saved. When training the power consumption prediction model M2, F2 is used as the validation set, and the remaining 4 folds are used as the training set, repeating the above process. This process is repeated for power consumption prediction models M3, M4, and M5. Because the validation sets for each power consumption prediction model are different, the parameters obtained from training are different, forming a differentiated model array.

[0078] Finally, the N trained power consumption prediction models are combined into a power consumption prediction model array. The resulting power consumption prediction model array is then deployed to the model inference engine of the scheduling center. All models are preloaded into memory at startup to ensure rapid invocation during actual task scheduling.

[0079] In this embodiment of the invention, the temporal variations of wind speed direction, wind speed magnitude, and precipitation intensity are differentially weighted and fused to construct a path meteorological complexity coefficient, providing a quantitative decision basis for the dynamic invocation of subsequent power consumption prediction models. Based on this complexity coefficient, the number of models to be invoked is adaptively determined, and a corresponding number of models are randomly selected from the pre-constructed model array for multimodal fusion prediction. This avoids the prediction bias of a single model under specific meteorological scenarios, improves the robustness and generalization ability of power consumption prediction, provides accurate energy consumption data support for task allocation decisions, and ensures the safety and economy of the scheduling scheme.

[0080] S400: Compare the predicted power consumption with the real-time available power of the corresponding airport. When the real-time available power is greater than or equal to the product of the predicted power consumption and the preset safety factor, the preliminary flight path is assigned as a confirmation task. Otherwise, the task to be executed is marked as an overflow task, triggering a multi-airport collaborative negotiation process to reallocate the overflow task to other available airports.

[0081] In this embodiment of the invention, the power consumption prediction model array is a set of multiple power consumption prediction models. Each model is built on a long short-term memory network architecture and is trained using different training data partitioning schemes.

[0082] Specifically, a pre-built array of power consumption prediction models is obtained. Based on the calculated path meteorological complexity coefficient, the number of models to be called for this prediction is determined. K models are randomly selected from the power consumption prediction model array as the power consumption prediction models to be called. The path features of the preliminary flight path are extracted, and a standardized input feature vector is generated. These vectors are then input into the K power consumption prediction models, and K initial predicted power consumption values ​​are output to obtain the final predicted power consumption.

[0083] Step S400 in the method of this embodiment of the invention includes: Based on the overflow task's geographical location, the predicted power consumption, the task priority, and the task time window constraints, a standardized task proposal is generated. The task tender is broadcast to all available airports in the region except the original airport through a multi-airport collaborative communication network. Within a preset bidding time window, the available airports submit bidding information in response to the task tender. The bidding information includes the comprehensive cost price calculated by each available airport for executing the overflow task, wherein the comprehensive cost price is generated by each available airport through a locally pre-trained bidding decision model. From all the received bids, the available airport with the lowest overall cost price is selected as the winning airport; The task data and preliminary flight path of the overflow task are sent to the winning airport to complete the redistribution of the overflow task.

[0084] In this embodiment of the invention, firstly, a standardized task proposal is generated based on the overflow task's geographical location, predicted power consumption, task priority, and task time window constraints. The task proposal serves as a standardized information carrier for all potential receiving airports, and its content must completely and accurately describe all the requirements of the task to be executed. The task proposal includes, for example, the task's geographical location, predicted power consumption, task priority, task time window constraints, preliminary flight path, task payload requirements, and the validity period of the task proposal. After generation, the proposal needs to be encoded according to a preset standardized format so that decoders at various airports can parse it uniformly. In this embodiment, JSON format is used for serialization, and key fields use a unified naming convention to ensure semantic consistency across airports. After encoding, a digital signature from the initiating airport is attached to the proposal to prevent tampering during transmission.

[0085] Secondly, the task proposal is broadcast to all available airports in the region, excluding the original airport, through a multi-airport collaborative communication network. Each airport, upon receiving the proposal, must send a confirmation signal back to the originating airport.

[0086] Specifically, the broadcast uses a one-to-many multicast or broadcast communication mode to ensure that all online and available airports can receive the tender information simultaneously. The broadcast range can be dynamically adjusted based on the geographical location of the overflow task. A maximum broadcast radius can be set, such as 200 kilometers. An airport will only receive the task tender if the straight-line distance between it and the task target point is less than this radius.

[0087] Next, within the preset bidding time window, the system receives bidding information submitted by each available airport in response to the task tender. The bidding information includes the comprehensive cost price calculated by each available airport for executing the overflow task. The comprehensive cost price is generated by each available airport through a locally pre-trained bidding decision model.

[0088] Specifically, the bidding decision-making models deployed at each airport are pre-trained regression or reinforcement learning models. Their inputs include two parts: external information contained in the task proposal and internal state information of the airport. External information includes factors such as geographical location, predicted power consumption, task priority, and time window constraints. Internal state information includes factors such as the current real-time battery level of the drone nest, existing scheduled task queues, available number of drones, and local weather forecasts. The output of the bidding decision-making model is the comprehensive cost price offered by the airport for undertaking the task.

[0089] The lower the overall cost price, the better the economy and timeliness of the airport in undertaking the task. The overall cost price may include: direct energy consumption cost, ferry flight cost, task preemption cost, and task priority adaptation cost. Direct energy consumption cost is the predicted power consumption required for the UAV to fly from the airport to the target point and perform the inspection; its calculation method is the same as the predicted power consumption calculation method. Ferry flight cost is the power consumption corresponding to the UAV's flight distance from the airport to the task target point. Task preemption cost is the cost incurred because the airport already has a scheduled task queue, and accepting a new task may delay existing tasks. Task priority adaptation cost is the increased task cost if a high-priority task is undertaken by a distant airport, as the response time may not meet the urgent requirements.

[0090] After calculating the comprehensive cost price, each available airport packages the bidding information, including the airport identifier, the bid price, the number of available drones at the airport, and the estimated response time, and then transmits the encrypted bidding information back to the initiating airport through the communication network.

[0091] Furthermore, the prices of all received bids from the initiating airport can be sorted, and the airport with the lowest price can be selected as the winning airport. If multiple airports have the same lowest price, the expected response time obtained from the bid information of each airport can be compared, and the airport with the shortest response time can be selected as the winning airport; if they are still the same, one of them can be randomly selected.

[0092] Finally, the overflow mission data and preliminary flight paths are sent to the winning airport. Upon receiving the complete mission data, the winning airport performs local verification. This includes checking the integrity of the mission data, verifying the compliance of the preliminary flight paths within its local geofence, and assessing whether the current nesting status meets the mission requirements. After successful verification, the winning airport sends a mission acceptance confirmation message back to the initiating airport, completing the redistribution of the overflow mission.

[0093] Step S400 of the method in this embodiment of the invention, the pre-training process of the bidding decision model, includes: Obtain historical bidding data for this airport. Each set of historical bidding data includes historical task bids, the status information of this airport at the time the historical task bids were received, and the corresponding optimal comprehensive cost price tag. The status information of this airport includes the number of available drones at this airport, the real-time power of the drone nest, the estimated power consumption and estimated completion time of each task in the scheduled task queue. A training sample set is constructed using the historical task tender documents and the airport status information as input features, and the optimal comprehensive cost pricing label as a supervision label. An initial bidding decision model is constructed based on reinforcement learning; The initial bidding decision model is trained using the training sample set. During the training process, the network parameters of the initial bidding decision model are updated with the goal of maximizing the cumulative expected value of the reward function until convergence is verified, thus obtaining the pre-trained bidding decision model.

[0094] In this embodiment of the invention, firstly, the historical bidding data of the airport is obtained. Each set of historical bidding data includes historical task bids, the status information of the airport at the time the historical task bids were received, and the corresponding optimal comprehensive cost price tag. The status information of the airport includes the number of available drones at the airport, the real-time power of the drone nest, the estimated power consumption and estimated completion time of each task in the scheduled task queue.

[0095] Specifically, the number of available drones refers to the number of drones currently on standby and available to perform new missions; the real-time power level of the drone nests is the current remaining power of the airport's drone nests; the estimated power consumption of each mission in the scheduled mission queue is the estimated power consumption of each pending mission in the airport's existing mission queue, used to calculate mission preemption costs; the estimated completion time of each mission in the scheduled mission queue is the estimated completion time of each pending mission in the airport's existing mission queue, used to assess whether accepting new missions will cause delays to existing missions; and the local weather forecast information is the current and short-term weather forecast for the area where the airport is located, used to assess the impact of weather conditions on flight missions.

[0096] The optimal comprehensive cost price label is the actual bid price that was ultimately verified as the best for the airport in this set of historical data. The optimal comprehensive cost price label is determined as follows: during the actual execution of the historical task, if an airport wins the bid and successfully completes the task, its actual bid price is recorded as a valid bid price; if the airport does not win the bid, a counterfactual evaluation is conducted to compare the airport's bid price with the bid prices of other bidding airports at that time and the actual execution results, thus estimating the airport's theoretically optimal bid price in that scenario. Each set of historical data is assigned an optimal cost price label as the training objective for supervised learning.

[0097] In a feasible implementation, assume that Airport A participated in 200 bidding events over the past 6 months, and each bid is recorded as a set of historical data samples. Taking one set of samples as an example: at a certain historical moment t, Airport A received a task proposal, in which the task target point is at coordinates (500, 300), the predicted power consumption is 1500mAh, the task priority is high, and the time window is 30 minutes. When receiving the task proposal, Airport A's status is: 2 drones available, real-time drone battery power is 3200mAh, there are 3 tasks in the scheduled task queue, and the local weather forecast is light wind and no rain.

[0098] Secondly, using historical mission proposals and local airport status information as input features, and the optimal comprehensive cost label as the supervision label, a training sample set is constructed. This training sample set is then divided into a training set and a validation set according to a preset ratio, such as 8:2. The training set is used for model parameter updates, the validation set for early stopping and hyperparameter tuning, and the test set for final model performance evaluation.

[0099] Next, an initial bidding decision model is constructed based on reinforcement learning. This initial bidding decision model can be built using a deep Q-network. A deep Q-network is an algorithm that combines deep learning and reinforcement learning, capable of learning the optimal decision strategy from high-dimensional input features.

[0100] Specifically, the structure of a deep Q-network includes an input layer, hidden layers, and an output layer. The input layer receives input features, with an input dimension equal to the dimension of the input features, and contains two neurons. The hidden layers are fully connected layers with 64 neurons, using the ReLU activation function. The output layer contains one neuron, using a linear activation function, and directly outputs the predicted value of the comprehensive cost price. Before training, all trainable parameters of the deep Q-network are randomly initialized. A Xavier uniform initialization method is used to ensure consistent variance across layers, which helps accelerate convergence. For example, the Adam optimizer is used, with an initial learning rate of 0.001. The network parameters are initialized using Xavier uniform initialization, with a random seed of 123 to ensure reproducibility. The price range output by the model is constrained to the [50, 500] interval by a limiting operation following the output layer, avoiding the output of extremely unreasonable prices.

[0101] Finally, the initial bidding decision model is trained using the training sample set. During the training process, the training objective is to maximize the cumulative expected value of the reward function. The network parameters are updated using the Q-learning algorithm based on temporal difference until the model converges on the validation set, thus obtaining the pre-trained bidding decision model.

[0102] Step S400 in the method of this embodiment of the invention, which updates the network parameters of the initial bidding decision model with the goal of maximizing the cumulative expected value of the reward function, includes: When the airport successfully executes the overflow task it has undertaken, the reward function provides a positive reward. The positive reward is positively correlated with the task priority of the overflow task and negatively correlated with the direct energy consumption cost of executing the overflow task. When accepting the overflow task causes delays in other scheduled tasks at the airport, the reward function provides a negative reward, which is positively correlated with the delay duration. When the overflow task fails due to insufficient real-time power in the airport's hangar after being accepted, the reward function provides the maximum negative reward. The network parameters of the initial bidding decision model are updated with the training objective of maximizing the cumulative expected value of the positive reward, the negative reward, and the maximum negative reward, until convergence is verified, thus obtaining a pre-trained bidding decision model.

[0103] In this embodiment of the invention, the reward function is the core driving force for reinforcement learning training, and its design should accurately reflect the actual business objectives of bidding decisions. In this embodiment, the reward function R can be set as: R = R 成功 +R 延误 +R 失败 Among them, R 成功Indicates successful reward execution, R 延误 Indicates a penalty for delay, R 失败 The system indicates penalties for failure. Successful execution rewards a positive reward when the airport successfully completes an assigned task; delay penalties result in a negative reward if accepting the task causes delays to other scheduled tasks at the airport; and failure penalties result in the maximum negative reward if the task fails due to insufficient battery power after being accepted.

[0104] Specifically, when the airport successfully executes an overflow task, the reward function provides a positive reward. This positive reward is positively correlated with the task priority of the overflow task and negatively correlated with the direct energy cost of executing the overflow task, i.e., R0. 成功 =α × Task Priority Value - β × Direct Energy Cost, where the task priority value can be set as Normal Priority = 1, High Priority = 2, Emergency Priority = 3, and the direct energy cost is positively correlated with the predicted power consumption. It is the predicted power consumption required to fly from this airport to the target point and perform inspections. Its calculation method is the same as the predicted power consumption calculation method. α and β are weighting coefficients. α is the positive weighting coefficient of the task priority item, which is set based on the value corresponding to the priority difference. Based on scheduling business experience, the three priorities, Normal, High, and Emergency, are mapped to the values ​​1, 2, and 3, respectively. α is set to 50, indicating that the reward difference between adjacent priorities is 50 reward units. β is the negative weighting coefficient of the direct energy cost, which is set to scale the power consumption to a reward scale comparable to the priority reward item. Statistical analysis of historical flight data shows that setting β=0.1 can make the typical task energy cost item fall between 50 and 300. Where α=50 and β=0.1, higher rewards are obtained when undertaking high-priority tasks or low-energy-consumption tasks.

[0105] When accepting overflow tasks causes delays to other scheduled tasks at the airport, the reward function provides a negative reward, which is positively correlated with the delay duration. Specifically, Where γ is the penalty coefficient, set to γ=20, its setting is based on the quantification and conversion of the value of task delay time. Delay time is in minutes, and γ=20 means that every minute of delay generates 20 units of negative reward, that is, a 10-minute delay generates 200 points of negative penalty. T c T represents the delayed estimated completion time. d The original deadline is the deadline. The longer the delay, the greater the negative reward.

[0106] When an overflow task fails due to insufficient real-time power in the airport's hangar after being accepted, the reward function provides the maximum negative reward; R 失败=-M, where M is a large constant, much larger than the absolute values ​​of positive and negative rewards under normal circumstances, effectively preventing the model from winning bids at low prices when the battery is insufficient. represents the maximum negative reward when a task fails. Its setting is based on the premise that this value must be much larger than the cumulative upper limit of the absolute values ​​of all positive and negative rewards under normal decision-making scenarios. For example, setting M to 500 means that the penalty for a single task failure due to insufficient battery is equivalent to the positive reward of failing three or more consecutive emergency tasks.

[0107] Finally, with the training objective of maximizing the cumulative expected value of positive rewards, negative rewards, and the maximum negative reward, the network parameters of the initial bidding decision model are updated until convergence is verified, thus obtaining the pre-trained bidding decision model.

[0108] For example, a batch of historical samples is randomly sampled from the training sample set. For each sample, the input state is fed into the current Q-network to obtain the predicted price. The training batch size is set to 16, and the target network is updated every 100 steps. The coefficients of the reward function are set to α=50, β=0.1, γ=20, and M=500. The target value is calculated based on the actual subsequent results in the samples, and the loss function is calculated using the mean squared error. The network parameters are updated using stochastic gradient descent. Every fixed number of training epochs, the current network parameters are copied to the target network. After each training epoch, the average reward and average price error of the model on the validation set are calculated. For example, during training, the model performance is evaluated on the validation set every 50 batches. When the average reward on the validation set no longer improves for 20 consecutive batches, or the improvement is less than a preset threshold, the model is considered to have converged, training is stopped, and the current model parameters are saved. For example, the preset threshold is set to 0.1%.

[0109] Furthermore, before undertaking the overflow task, the pre-trained bidding decision model performs the following verification steps: Step 1: Add the predicted power consumption of the overflow task to the remaining estimated power consumption of each task in the airport's scheduled task queue to obtain the total estimated power consumption after acceptance. Step 2: Determine whether the total estimated power consumption after the takeover exceeds the preset percentage threshold of the real-time available power of the airport's terminal. Step 3: If the preset ratio threshold is exceeded, the overflow task will be rejected and a message with the reason for rejection will be returned to the main airport; otherwise, the overflow task will be accepted.

[0110] In this embodiment of the invention, although the pre-trained bidding decision model indirectly reflects the constraints of existing tasks on the acceptance capacity during the bidding calculation stage through input features such as the expected power consumption and expected completion time of each task in the scheduled task queue, an independent hard safety check is still required before actual acceptance to ensure the feasibility of the acceptance decision under power safety constraints. This check step serves as a safety protection layer after the output of the bidding decision model, preventing the risk of insufficient power due to model prediction deviations, input feature quantization errors, or model output runaway under extreme conditions.

[0111] The specific implementation of step one is as follows: extract the predicted power consumption of the overflow task. The predicted power consumption is obtained by the power consumption prediction model array and represents the predicted amount of power consumption required to execute the overflow task; traverse each task to be executed in the scheduled task queue of this airport, obtain the remaining estimated power consumption of each task to be executed, and add the predicted power consumption of the overflow task to the remaining estimated power consumption of all scheduled tasks to obtain the total estimated power consumption of this airport after accepting the overflow task.

[0112] The specific implementation of step two is as follows: Obtain the real-time available power of the airport's hub, which represents the total remaining available power of the hub; set a preset ratio threshold, which is a positive number not greater than 1, to limit the maximum safe proportion of available power after accepting a new task, so as to reserve some power as a safety margin to deal with the execution deviation of scheduled tasks and unforeseen circumstances; determine whether the total expected power consumption after accepting the task exceeds the product of the real-time available power of the airport's hub and the preset ratio threshold; if it exceeds, it means that after accepting the overflow task, the total expected power consumption of the airport will exceed the safe occupancy limit, and accepting the task may lead to the depletion of hub power or the failure of existing tasks to be completed on schedule; if it does not exceed, it means that the total expected power consumption after accepting the overflow task is within the safe occupancy range and meets the acceptance conditions.

[0113] The preset percentage threshold refers to the maximum percentage of electricity that the airport is allowed to use when accepting overflow tasks. This is used to reserve a portion of the electricity as a safety margin to cope with deviations in the execution of scheduled tasks and unforeseen circumstances. The preset percentage threshold is a positive number not greater than 1, ranging from 0.7 to 0.9. The preset percentage threshold can be dynamically quantified based on the length of the airport's scheduled task queue. A longer queue indicates a heavier current load at the airport, and a greater risk of power depletion or delays to existing tasks due to accepting new tasks. In this case, the safety margin should be increased, and the preset percentage threshold should be decreased accordingly. Conversely, a shorter queue indicates a lighter current load at the airport, and a greater margin for accepting new tasks. In this case, the preset percentage threshold should be increased accordingly.

[0114] Preferably, the preset ratio threshold can be dynamically quantified as follows: Obtain the number of tasks currently awaiting execution in the airport's scheduled task queue, denoted as queue length M, and obtain the airport's maximum capacity, denoted as capacity limit Cmax (in flight counts). Calculate the queue occupancy rate R = M / Cmax. The preset ratio threshold is linearly mapped to the queue occupancy rate: when the queue occupancy rate is 0, the preset ratio threshold is the upper limit of 0.90; when the queue occupancy rate is greater than or equal to 1, the preset ratio threshold is the lower limit of 0.70; when the queue occupancy rate is between 0 and 1, the preset ratio threshold is determined through linear interpolation. The linear mapping formula is: Preset ratio threshold = 0.90 - 0.20 × R.

[0115] For example, if an airport's maximum mission capacity (Cmax) is 10 flights, and the current scheduled mission queue length is 3, then the queue occupancy rate (R) is 3 / 10 = 0.3, and the preset ratio threshold is 0.90 - 0.20 × 0.3 = 0.84. If the total estimated power consumption after accepting a new mission is 3200mAh, and the airport's real-time available power in the hangar is 4000mAh, then the safe occupancy limit is 4000 × 0.84 = 3360mAh. Since 3200mAh does not exceed 3360mAh, the mission is accepted. If the current scheduled mission queue length is 8, then R = 8 / 10 = 0.8, the preset ratio threshold is 0.90 - 0.20 × 0.8 = 0.74, and the safe occupancy limit is 4000 × 0.74 = 2960mAh. Since 3200mAh exceeds 2960mAh, the mission is rejected.

[0116] The specific implementation of step three is as follows: If the total estimated power consumption after accepting the task exceeds the product of the real-time available power of the airport's nest and a preset proportional threshold, it is determined that the airport does not currently have the conditions to safely accept the overflow task. A rejection message is returned to the main airport that initiated the task tender. The rejection message carries a rejection reason identifier, which is used to indicate that the task is rejected due to insufficient nest power safety margin, so that the main airport can continue to broadcast the task to other available airports. If the total estimated power consumption after accepting the task does not exceed the product of the real-time available power of the airport's nest and a preset proportional threshold, it is determined that the airport has the conditions to safely accept the overflow task. The airport confirms the acceptance of the overflow task and returns a confirmation message to the main airport.

[0117] Thus, the output of the bidding decision model undergoes an independent safety protection verification before actual execution. This verification is based on a hard judgment of the current real-time power of the nest and the actual power consumption of the scheduled task queue, which can effectively avoid the risk of overload due to model prediction deviation.

[0118] In this embodiment of the invention, a decentralized distributed collaborative scheduling architecture is constructed by achieving automatic transfer and dynamic reallocation of overflow tasks through standardized tender document generation, multi-airport broadcast communication, and the lowest comprehensive cost bidding mechanism. Based on historical bidding data and the real-time status of the airport, each airport cluster has the decision-making ability to autonomously generate comprehensive cost bids. By designing a reinforcement learning reward function with positive rewards for task priority, negative penalties for delays, and maximum penalties for failures, the self-healing ability and resource utilization efficiency of the multi-airport scheduling system under the impact of sudden tasks are improved, ensuring the task completion rate.

[0119] Through the above specific implementation methods, the embodiments of the present invention achieve the following technical effects: In this embodiment of the invention, the RRT algorithm is first used to quickly generate collision-free flight paths in a complex three-dimensional airspace, improving path planning efficiency. Target-guided expansion and collision detection further ensure path safety and search convergence. Then, the initial flight path is divided, and meteorological grid spatiotemporal matching is performed to extract meteorological element values ​​corresponding to each path segment. These values ​​are arranged according to the flight direction to generate a time-series meteorological feature sequence. Subsequently, the overlap area between each individual path segment and each meteorological grid unit is calculated as the weight coefficient of each meteorological grid unit. The meteorological element values ​​of each path segment are obtained through weighted summation, improving the accuracy of meteorological feature extraction and providing accurate input for energy consumption prediction.

[0120] Secondly, the coefficients of variation of the direction change series, magnitude change series, and precipitation intensity change series are calculated, and their dynamic weights are obtained by normalizing the coefficients of variation. Subsequently, each series is normalized, and then the normalized change series are weighted to make the meteorological element values ​​of each segment more accurately reflect the actual spatial distribution, improve the continuity and accuracy of the time-series meteorological characteristic series, and provide a reliable data foundation for power consumption prediction.

[0121] Furthermore, the temporal variations of wind speed direction, wind speed magnitude, and precipitation intensity are differentiated and weighted to construct a path meteorological complexity coefficient, providing a quantitative decision-making basis for the dynamic invocation of subsequent power consumption prediction models. Based on this complexity coefficient, the number of models to be invoked is adaptively determined, and a corresponding number of models are randomly selected from the pre-constructed model array for multimodal fusion prediction. This avoids the prediction bias of a single model under specific meteorological scenarios, improves the robustness and generalization ability of power consumption prediction, provides accurate energy consumption data support for task allocation decisions, and ensures the safety and economy of the scheduling scheme.

[0122] Ultimately, through standardized tender document generation, multi-airport broadcast communication, and a lowest comprehensive cost bidding mechanism, the automatic transfer and dynamic reallocation of overflow tasks are achieved, constructing a decentralized distributed collaborative scheduling architecture. Based on historical bidding data and the real-time status of the airport, each airport cluster is equipped with the decision-making ability to autonomously generate comprehensive cost bids. By designing a reinforcement learning reward function with positive rewards for task priority, negative penalties for delays, and maximum penalties for failures, the self-healing ability and resource utilization efficiency of the multi-airport scheduling system under the impact of sudden tasks are improved, ensuring the task completion rate.

Claims

1. An adaptive cooperative scheduling method for distributed tasks in a multi-airport environment, characterized in that, The method includes: Based on the priority and geographical location of the task to be executed, a preliminary flight path is generated for the task to be executed. Based on the spatiotemporal range traversed by the preliminary flight path, a time-series meteorological feature sequence distributed along the path is extracted from the local meteorological grid data of the multi-airport area. The time-series meteorological feature sequence includes the wind speed vector and precipitation intensity on each path segment. The path meteorological complexity coefficient is calculated based on the temporal change rate of meteorological elements in the temporal meteorological feature sequence. Based on the path meteorological complexity coefficient and the time-series meteorological feature sequence, the predicted power consumption is obtained by using a pre-constructed power consumption prediction model array. The predicted power consumption is compared with the real-time available power of the corresponding airport. When the real-time available power is greater than or equal to the product of the predicted power consumption and the preset safety factor, the preliminary flight path is assigned as a confirmation task. Otherwise, the task to be executed is marked as an overflow task, triggering a multi-airport collaborative negotiation process to reallocate the overflow task to other available airports.

2. The adaptive cooperative scheduling method for distributed tasks in a multi-airport environment according to claim 1, characterized in that, Based on the priority and geographical location of the task to be executed, a preliminary flight path is generated for the task to be executed, including: Starting from the location of the airport nest corresponding to the task to be executed and ending at the geographical location of the task to be executed, a search tree is initialized in the three-dimensional airspace, with the starting point as the root node of the search tree; Candidate nodes are randomly sampled and generated in the three-dimensional spatial domain, and the candidate nodes are guided and expanded in the direction of the endpoint to generate new nodes; The new node is added to the search tree, and the distance between the new node and the destination is checked. When the distance is less than a preset termination distance threshold, the new node is connected to the destination to generate a preliminary route connecting the starting point and the destination. Collision detection is performed on the preliminary route to determine whether the preliminary route intersects with the restricted area defined by the static geofence constraint. If it intersects, random sampling, node expansion and collision detection are performed again until a route that does not intersect with the restricted area is generated, thus obtaining a collision-free route. The final collision-free flight path will be used as the initial flight path.

3. The adaptive cooperative scheduling method for distributed tasks in a multi-airport environment according to claim 1, characterized in that, Based on the spatiotemporal range traversed by the preliminary flight path, time-series meteorological feature sequences distributed along the path are extracted from local meteorological grid data of the multi-airport area, including: The preliminary flight path is divided into multiple continuous path segments according to a preset spatial interval, and the geographical coordinate range and expected flight time period corresponding to each path segment are determined. Based on the geographic coordinate range and expected flight time of each path segment, the matching spatiotemporal grid cell is retrieved from the local meteorological grid data of the multi-airport area, and the meteorological element values ​​stored in the spatiotemporal grid cell are extracted, wherein the meteorological element values ​​include the direction angle and magnitude of the wind speed vector and the precipitation intensity. The meteorological element values ​​corresponding to each path segment are arranged sequentially according to the flight direction from the starting point to the end point to form a time-series meteorological characteristic sequence. When a single path segment crosses multiple meteorological grid units, the meteorological element values ​​of each meteorological grid unit are weighted and averaged according to the proportion of the overlap area between the path segment and each meteorological grid unit, and this average is used as the meteorological element value of the single path segment.

4. The adaptive cooperative scheduling method for distributed tasks in a multi-airport environment according to claim 3, characterized in that, When a single path segment spans multiple meteorological grid units, the meteorological element values ​​of each meteorological grid unit are weighted and averaged according to the proportion of overlap between the path segment and each meteorological grid unit, and this average is used as the meteorological element value of the path segment, including: The intersection calculation is performed between the geographic projection range of the single path segment in the local meteorological grid data and the spatial range of each meteorological grid unit to determine each meteorological grid unit covered by the single path segment. Calculate the overlap area between each individual path segment and each meteorological grid unit; The proportion of the overlapping area of ​​each meteorological grid unit to the total area of ​​the single path segment is calculated and used as the weighting coefficient of each meteorological grid unit. The meteorological element values ​​for a single path segment are obtained by multiplying the wind speed vector direction angle, wind speed vector magnitude, and precipitation intensity stored in each meteorological grid unit by their respective weighting coefficients and then summing them up.

5. The adaptive cooperative scheduling method for distributed tasks in a multi-airport environment according to claim 1, characterized in that, Based on the temporal change rate of meteorological elements in the aforementioned temporal meteorological feature sequence, the path meteorological complexity coefficient is calculated, including: Calculate the change in the directional angle of the wind speed vector between adjacent path segments in the time-series meteorological feature sequence, and arrange them in the temporal order of the path segments to obtain the sequence of directional change. Calculate the magnitude change of wind speed vector between adjacent path segments in the time-series meteorological feature sequence, and arrange them according to the temporal order of the path segments to obtain the magnitude change sequence; Calculate the absolute value of the difference in precipitation intensity between adjacent path segments in the time-series meteorological feature sequence, arrange them in the temporal order of the path segments, and obtain the precipitation intensity change sequence. Calculate the coefficient of variation of the direction change sequence, the coefficient of variation of the magnitude change sequence, and the coefficient of variation of the precipitation intensity change sequence. Normalize each coefficient of variation to obtain the first dynamic weight of the direction change, the second dynamic weight of the magnitude change, and the third dynamic weight of the precipitation intensity change. Assign the first dynamic weight, the second dynamic weight, and the third dynamic weight to the sequence of directional changes, the sequence of magnitude changes, and the sequence of precipitation intensity changes, respectively, perform weighted summation to obtain a weighted sum value, and normalize the weighted sum value to the interval between 0 and 1 to obtain the path meteorological complexity coefficient.

6. The adaptive cooperative scheduling method for distributed tasks in a multi-airport environment according to claim 1, characterized in that, Based on the path meteorological complexity coefficient and the time-series meteorological feature sequence, the predicted power consumption is obtained by predicting the power consumption through a pre-constructed power consumption prediction model array, including: Obtain a pre-built power consumption prediction model array, which contains N power consumption prediction models, where N is a positive integer greater than or equal to 1; Calculate the product of the path meteorological complexity coefficient and N, and then round the product up to obtain the number of models K to be called in this prediction, where 1≤K≤N; K power consumption prediction models are randomly selected from the power consumption prediction model array as the K calling models for this prediction; Extract the path features of the preliminary flight path, which include the total path length, spatial distance of each path segment, flight altitude, preset flight speed, and mission payload weight; The path features and the time-series meteorological feature sequence are concatenated and uniformly encoded according to a preset format to generate a standardized input feature vector. The input feature vector is then input into the K calling models respectively, and K initial predicted power consumptions are output. Calculate the arithmetic mean of the K initial predicted power consumptions, and use the arithmetic mean as the predicted power consumption.

7. The adaptive cooperative scheduling method for distributed tasks in a multi-airport environment according to claim 6, characterized in that, Obtain a pre-built array of power consumption prediction models, including: Multiple sets of historical flight data samples are acquired. Each set of historical flight data samples includes historical path characteristics, historical time-series meteorological characteristic sequences, and corresponding historical actual power consumption. Based on the long short-term memory network architecture, construct N power consumption prediction models; The multiple sets of historical flight data samples are divided into N folds. Each fold is used as the validation set and the remaining N-1 folds are used as the training set. Supervised training is performed on the N power consumption prediction models until the validation convergence is achieved, resulting in N trained power consumption prediction models. The N trained power consumption prediction models are combined into a power consumption prediction model array.

8. The adaptive cooperative scheduling method for distributed tasks in a multi-airport environment according to claim 1, characterized in that, The task to be executed is marked as an overflow task, triggering a multi-airport collaborative negotiation process to reallocate the overflow task to other available airports, including: Based on the overflow task's geographical location, the predicted power consumption, the task priority, and the task time window constraints, a standardized task proposal is generated. The task tender is broadcast to all available airports in the region except the original airport through a multi-airport collaborative communication network. Within a preset bidding time window, the available airports submit bidding information in response to the task tender. The bidding information includes the comprehensive cost price calculated by each available airport for executing the overflow task, wherein the comprehensive cost price is generated by each available airport through a locally pre-trained bidding decision model. From all the received bids, the available airport with the lowest overall cost price is selected as the winning airport; The task data and preliminary flight path of the overflow task are sent to the winning airport to complete the redistribution of the overflow task.

9. The adaptive cooperative scheduling method for distributed tasks in a multi-airport environment according to claim 8, characterized in that, The pre-training process of the bidding decision model includes: Obtain historical bidding data for this airport. Each set of historical bidding data includes historical task bids, the status information of this airport at the time the historical task bids were received, and the corresponding optimal comprehensive cost price tag. The status information of this airport includes the number of available drones at this airport, the real-time power of the drone nest, the estimated power consumption and estimated completion time of each task in the scheduled task queue. A training sample set is constructed using the historical task tender documents and the airport status information as input features, and the optimal comprehensive cost pricing label as a supervision label. An initial bidding decision model is constructed based on reinforcement learning; The initial bidding decision model is trained using the training sample set. During the training process, the network parameters of the initial bidding decision model are updated with the goal of maximizing the cumulative expected value of the reward function until convergence is verified, thus obtaining the pre-trained bidding decision model. Before accepting the overflow task, the pre-trained bidding decision model also performs the following verification steps: Step 1: Add the predicted power consumption of the overflow task to the remaining estimated power consumption of each task in the airport's scheduled task queue to obtain the total estimated power consumption after acceptance. Step 2: Determine whether the total estimated power consumption after the takeover exceeds the preset percentage threshold of the real-time available power of the airport's terminal. Step 3: If the preset ratio threshold is exceeded, the overflow task will be rejected and a message with the reason for rejection will be returned to the main airport; otherwise, the overflow task will be accepted.

10. The adaptive cooperative scheduling method for distributed tasks in a multi-airport environment according to claim 9, characterized in that, The network parameters of the initial bidding decision model are updated with the training objective of maximizing the cumulative expected value of the reward function, including: When the airport successfully executes the overflow task it has undertaken, the reward function provides a positive reward. The positive reward is positively correlated with the task priority of the overflow task and negatively correlated with the direct energy consumption cost of executing the overflow task. When accepting the overflow task causes delays in other scheduled tasks at the airport, the reward function provides a negative reward, which is positively correlated with the delay duration. When the overflow task fails due to insufficient real-time power in the airport's terminal block after being accepted, the reward function provides the maximum negative reward. The network parameters of the initial bidding decision model are updated with the training objective of maximizing the cumulative expected value of the positive reward, the negative reward, and the maximum negative reward, until convergence is verified, thus obtaining a pre-trained bidding decision model.

Images