Urban low-altitude unmanned aerial vehicle instant distribution scheduling and path planning method

Abstract

The invention provides an urban low-altitude unmanned aerial vehicle instant distribution scheduling and path planning method considering demand prediction. The method comprises the following steps: establishing a three-dimensional environment map of an unmanned aerial vehicle distribution system by adopting a grid-topology hybrid modeling strategy based on three-dimensional urban environment map data; according to the three-dimensional environment map of the unmanned aerial vehicle distribution system, the starting point and the target point of the unmanned aerial vehicle, the optimal flight path of the unmanned aerial vehicle, the time required by the optimal path and the transportation cost are obtained by using an unmanned aerial vehicle path planning algorithm of space-time improvement A *; and according to the order data, the unmanned aerial vehicle performance parameters, the optimal flight path of the unmanned aerial vehicle, the time required by the optimal path and the transportation cost, using the CNN-LSTM-Attention order prediction model and the task allocation model to output a task allocation scheme including a virtual order and an actual order. Accurate spatial data support is provided for path planning of the unmanned aerial vehicle, the reasonability of low-altitude logistics distribution path planning is improved, and finer data support is provided for urban airspace management.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) path planning technology, and in particular to a method for real-time delivery scheduling and path planning of urban low-altitude UAVs. Background Technology

[0002] Commercial operation of drone-based on-demand delivery is still in the exploratory stage, and a mature operating system has not yet been formed. Currently, most operators' route planning and scheduling strategies are still mainly based on pilot projects or experiments, which are insufficient to meet the multiple requirements of on-demand delivery in complex urban scenarios, such as task time windows, priorities, and power constraints. Existing route planning methods are mostly based on static or semi-dynamic environments, making it difficult to cope with uncertainties such as dynamically generated orders, temporary no-fly zones, and sudden weather events, thus reducing practical feasibility. At the same time, there is a general lack of forecasting and pre-scheduling mechanisms oriented towards delivery demand, making it difficult to respond promptly to peak periods or sudden tasks, and there is still significant room for improvement in collaborative efficiency and service levels.

[0003] Based on the widespread application and advantages of grid-based methods in urban low-altitude airspace, researchers generally employ comprehensive optimization methods for urban low-altitude UAV path planning, primarily considering factors such as energy consumption, safety, and noise. Studies typically utilize algorithms like A*, RTT, and ACO to implement specific path planning. However, most methods are still based on idealized path planning, lacking system modeling and response mechanisms for potential trajectory deviations and environmental disturbances during actual flight, making it difficult to fully adapt to the real-time delivery needs in complex urban environments.

[0004] In the optimization of drone logistics delivery, research focuses on transportation costs, delivery time, and customer satisfaction, primarily employing integer programming or heuristic algorithms to reduce drone operating costs. However, proactive demand forecasting and pre-scheduling of drones to improve delivery efficiency and reduce delivery time are rarely adopted. For order demand forecasting, existing research widely utilizes machine learning ensemble methods, combining historical data to improve order forecast accuracy. Summary of the Invention

[0005] Embodiments of the present invention provide a method for real-time delivery scheduling and route planning of urban low-altitude drones, so as to effectively improve the scheduling efficiency of drones and reduce operating costs.

[0006] To achieve the above objectives, the present invention adopts the following technical solution.

[0007] A method for real-time delivery scheduling and route planning of urban low-altitude drones considering demand forecasting includes:

[0008] Based on urban map data, 3D urban environment map data was obtained, and a 3D environment map of the drone delivery system was established using a "grid-topology" hybrid modeling strategy.

[0009] A spatiotemporally improved A* drone path planning algorithm is constructed. Based on the 3D environment map of the drone delivery system, the drone's starting point and target point, the spatiotemporally improved A* drone path planning algorithm is used to obtain the optimal flight path of the drone, the time required for the optimal path, and the transportation cost.

[0010] Based on historical order data, current order data, and drone performance parameters, as well as the optimal flight path of the drone, the time required for the optimal path, and transportation costs, the CNN-LSTM-Attention order prediction model and task allocation model are used to output a task allocation scheme that includes virtual orders and actual orders.

[0011] Preferably, the step of obtaining three-dimensional urban environment map data based on urban map data and establishing a three-dimensional environment map of the drone delivery system using a "grid-topology" hybrid modeling strategy includes:

[0012] Three-dimensional urban environment map data was obtained based on urban map data. This 3D urban environment map data includes urban zoning, obstacle data, and no-fly zone data. A "grid-topology" hybrid modeling strategy was used to establish the 3D environment map for the drone delivery system. The drone delivery flight network in urban airspace is represented by a directed graph G(V,E), where the node set V consists of the center points of the effective grids. It is assumed that the environment consists of N 3D grids, and the grid index set is G = {1,2,...,g}. The position of each grid i (i∈G) is represented by 3D coordinates (x,y). i ,y i ,z i Let E represent the connectivity between grid cells. If grid cells i and j are adjacent in 3D space, then there exists an edge (i, j) from i to j in the directed graph. In 3D space, the connectivity between grid cells is based on 26 possible adjacency directions, i.e., the relative coordinate difference |x| between grid cells. i -x j |≤1,|y i -y j |≤1, and |z i -z j |≤1 indicates their adjacency in the horizontal, vertical and diagonal directions.

[0013] Preferably, the spatiotemporally improved A* UAV path planning algorithm, based on the 3D environment map of the UAV delivery system, the UAV's starting point, and the target point, utilizes the spatiotemporally improved A* UAV path planning algorithm to obtain the optimal flight path of the UAV, the time required for the optimal path, and the transportation cost, including:

[0014] A spatiotemporally improved A* UAV path planning algorithm is constructed, and the actual cost function of the spatiotemporally improved A* UAV path planning algorithm is set to be the energy consumption cost C. energy Safety cost C safety and noise cost C noise The weighted average of the three sub-objectives is as follows:

[0015] The drone from the grid point (x i ,y i ,z i ) fly to (x j ,y j ,z j Energy consumption cost C energy Equation (1) represents:

[0016]

[0017] The drone at its current location (x) i ,y i ,z i Safety cost C safety The conditional probabilities of these stages are calculated by integration, and the specific formula is shown in (3):

[0018] C safety =P c E c

[0019] =P d P i (x,y)P f (x,y)E c (3)

[0020] We consider ballistic descent and uncontrolled glide as two descent modes. Assuming the probability of each descent mode occurring remains constant for different UAVs, the probabilities of the two descent modes are:

[0021]

[0022] The probability P of a drone u hitting a person on the ground i (x, y) represents the probability that the drone, after losing control and falling to the ground, will collide with at least one person. Specifically:

[0023] P i (x,y)=ρ(x,y)A exp (4)

[0024]

[0025] Probability of injury or death P f(x, y) represents the probability of injury or death due to a collision between a crashed drone and a person. The injury / death rate is related to the shielding coefficient S. f The kinetic energy E(x,y) generated during the collision is related to the kinetic energy E(x,y) generated during the collision. The calculation method is shown in equation (7):

[0026]

[0027] Noise cost of drones C noise It shall not exceed the maximum allowable standard for the corresponding area, and the calculation method is shown in equation (9):

[0028] C noise =E n [L-20lg(b / r0)+3-N(x,y)] (9)

[0029] In summary, the actual cost function g(n) is expressed as:

[0030]

[0031] in, and C respectively energy C safety and C noise Weighting coefficients;

[0032] By utilizing the spatiotemporal improvement of the A* UAV path planning algorithm, local optima are followed at each decision step, minimizing the subsequent single-step cost function g(n). A single path planning requires multiple iterative calculations. Through continuous optimization and adjustment of path selection, the algorithm gradually approaches a better path scheme that meets actual needs.

[0033] The symbols and definitions of the parameters and variables involved in the cost function are shown in Tables 2 and 3:

[0034] Table 2 Parameter Symbols and Definitions

[0035]

[0036] N(x,y) represents the upper limit of the urban standard noise threshold for the area where the geographical location (x,y) is located, in decibels.

[0037]

[0038] Preferably, the step of using a CNN-LSTM-Attention order prediction model and a task allocation model to output a task allocation scheme containing virtual orders and actual orders, based on historical order data, current stage order data, UAV performance parameters, the optimal flight path of the UAV, the time required for the optimal path, and transportation costs, includes:

[0039] Actual orders are generated after a customer places an order. The order information includes the order time, the location of the pickup locker, the location of the delivery locker, the weight of the goods, and the latest delivery time. Virtual orders are predicted based on historical data and prediction algorithms. The order information only includes the time the order occurred and the location of the pickup locker.

[0040] The CNN-LSTM-Attention order prediction model is a hybrid model that integrates a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network, and an Attention mechanism. First, the CNN layer uses 1D or 2D convolutional kernels to extract local spatiotemporal features from the order data, which are then reduced in dimensionality by pooling layers to obtain feature vectors. Next, the LSTM layer processes these feature vectors, using a gating mechanism to learn the long-term temporal dependencies of the order data. Then, an Attention layer is introduced to calculate weights for the hidden states at each time step of the LSTM output, highlighting the key time steps for prediction. Finally, a fully connected layer outputs the predicted number of actual orders within a set future timeframe. Virtual orders are generated by clustering low-probability but spatially adjacent grid points, and these virtual orders, along with the actual orders, form a task set.

[0041] The task allocation model treats virtual and actual orders as tasks. It includes constraints such as task priority, drone payload limits, maximum number of consecutively executed tasks, and node access order. With the goal of minimizing transportation costs and timeout penalties, the model uses a spatiotemporally improved A* drone path planning algorithm to calculate the optimal flight path, required time, and transportation cost for all tasks, serving as the foundational data for task allocation. Finally, using these pre-calculated path costs and times as input, and under the constraint of minimizing total transportation costs and timeout penalties, the task allocation model allocates the task set to drones, prioritizing actual orders and allocating remaining capacity to virtual orders, thus forming the final task allocation scheme.

[0042] As can be seen from the technical solutions provided by the embodiments of the present invention described above, the present invention employs a grid-based method to discretize the urban environment into regular three-dimensional grid units, thereby establishing a drone flight reachability analysis model. This method can adapt well to complex urban environments and allows for the addition of dynamic obstacles at any time. It characterizes the scenario of urban low-altitude drone instant delivery and refines key path planning factors such as energy consumption, safety, and noise. Simultaneously, by combining historical order data and based on demand forecasting, drones are pre-scheduled to optimize overall operating costs.

[0043] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and will become apparent from the description or may be learned by practice of the invention. Attached Figure Description

[0044] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 A flowchart illustrating the real-time delivery scheduling and route planning method for urban low-altitude unmanned aerial vehicles provided by this invention;

[0046] Figure 2 A schematic diagram of the flight direction of an unmanned aerial vehicle (UAV) provided by the present invention;

[0047] Figure 3 This invention provides a schematic diagram of an impact event in a ballistic descent mode.

[0048] Figure 4 This invention provides a schematic diagram of an impact event during a runaway gliding mode.

[0049] Figure 5 A schematic diagram illustrating the altitude at which a drone generates noise, provided by the present invention;

[0050] Figure 6 This invention provides a schematic diagram of the planning results of a typical path under different optimization objectives;

[0051] Figure 7 A schematic diagram of the masking coefficient distribution of UAV paths under different optimization objectives (mean of 50 random samples) provided by the present invention;

[0052] Figure 8 This invention provides a schematic diagram of the distribution of UAV paths and building distances under different optimization objectives (mean of 50 random samples);

[0053] Figure 9 A schematic diagram of the overall cost under different weights provided by the present invention;

[0054] Figure 10 A schematic diagram of order aggregation provided by the present invention;

[0055] Figure 11 This is a schematic diagram of the overall operation process of a scheduling platform system provided in an embodiment of the present invention;

[0056] Figure 12 This invention provides a schematic diagram for aggregating and analyzing orders for a specific month by date and time.

[0057] Figure 13A Pearson correlation heatmap provided for an embodiment of the present invention;

[0058] Figure 14 A schematic diagram illustrating the prediction results of an XGBoost model provided in an embodiment of the present invention;

[0059] Figure 15 A schematic diagram illustrating the prediction results of a CNN-LSTM-Attention model provided in an embodiment of the present invention;

[0060] Figure 16 This is a schematic diagram of the loss curves of a CNN-LSTM-Attention model on the training and testing sets, provided as an embodiment of the present invention.

[0061] Figure 17 This is a schematic diagram of the layout planning results of the take-off and landing center and the delivery cabinet cluster center provided in an embodiment of the present invention.

[0062] Figure 18 This is a schematic diagram of execution layer deviation conflict monitoring and path replanning provided in an embodiment of the present invention;

[0063] Figure 19 A flowchart illustrating a deviation correction and conflict avoidance mechanism provided in an embodiment of the present invention;

[0064] Figure 20 A flowchart of a "prediction-planning-execution" simulation system provided in this embodiment of the invention;

[0065] Figure 21 This is a schematic diagram comparing the cost and number of timed-out tasks under different strategies, provided as an embodiment of the present invention. Detailed Implementation

[0066] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0067] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or couplings. The term “and / or” as used herein includes any and all combinations of one or more of the associated listed items.

[0068] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as herein.

[0069] To facilitate understanding of the embodiments of the present invention, the following will provide further explanation and description with reference to the accompanying drawings and several specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0070] This invention utilizes rasterized 3D city map modeling to divide airspace into functional zones such as residential areas, commercial areas, and office areas, and labels different areas with corresponding parameters such as occlusion coefficients, noise thresholds, and population densities. A spatiotemporally improved UAV path planning algorithm based on A* is designed, considering energy consumption, safety, and noise costs in the cost function. Sensitivity analysis experiments of the weighting coefficients are conducted based on the raster map.

[0071] This invention establishes a multi-drone collaborative scheduling model with constraints such as drone payload and the upper limit of the number of consecutive tasks, aiming to minimize transportation costs and timeout penalty costs. Based on the concept of demand forecast aggregation, a pre-scheduling mechanism of virtual orders is introduced to improve the system's response efficiency and foresight.

[0072] A simulation experiment on real-time delivery demand forecasting and dynamic scheduling was conducted. First, based on historical order data, two models, XGBoost and CNN-LSTM-Attention, were used to predict the number of orders and evaluate performance indicators. Second, a deviation correction and conflict avoidance mechanism and a multi-UAV scheduling strategy based on virtual order aggregation were designed to construct a "prediction-planning-execution" simulation framework. Finally, the feasibility and operational effectiveness of the number of UAVs operating in the field, the virtual order experience threshold, and the flight deviation and conflict avoidance mechanism were analyzed.

[0073] The flowchart of a real-time delivery scheduling and route planning method for urban low-altitude drones that considers demand forecasting provided by an embodiment of the present invention is as follows: Figure 1 As shown, the processing steps include the following;

[0074] Step S10: Urban Environment Modeling. Input urban map data and output 3D urban environment map data, including urban zoning, obstacle data, and no-fly zone data.

[0075] This step provides a realistic flight environment for the UAV path planning in step S20.

[0076] Step S20: 3D Path Planning for Urban Low-Altitude UAVs. Input 3D urban environment map data (obstacles, no-fly zones), UAV starting point and target point, and use the spatiotemporally improved A* UAV path planning algorithm to output the optimal flight path of the UAV, the time required for the optimal path, and the transportation cost.

[0077] This step provides crucial input data, such as cost and time, for step S30 at the path level. The deviation correction and conflict avoidance mechanisms in this step are proposed in this invention and can dynamically adjust the UAV's flight path to ensure the safety and stability of low-altitude airspace.

[0078] Step S30: Multi-UAV task allocation and pre-scheduling based on demand forecasting. Based on historical order data, current order data, UAV performance parameters, optimal flight paths for UAVs, the time required for those paths, and transportation costs, a CNN (Convolutional Neural Network)-LSTM (Long Short-Term Memory)-Attention order prediction model and task allocation model are used to output a task allocation scheme that includes both virtual and actual orders.

[0079] The pre-scheduling mechanism in this step (i.e., generating virtual orders and arranging drones to arrive at virtual order points in advance to improve the system's response speed to future tasks and the foresight of scheduling decisions) is an innovative proposal of this invention.

[0080] The CNN-LSTM-Attention order prediction model, which integrates multi-dimensional features such as weather, traffic, and time, to achieve high-precision prediction of order demand in a short period of time, is also an innovative proposal of this invention.

[0081] Urban environments consist of various functional areas such as residential areas, commercial areas, and office areas. These areas vary significantly in building density and height. Static obstacles such as buildings and permanent no-fly zones, as well as dynamic obstacles such as temporary no-fly zones and obstructed areas created by emergencies or construction, all affect drone flight. Furthermore, dynamic obstacles are time-dependent and unpredictable. Simultaneously, drone flight energy consumption is affected by changes in path length and altitude; obstacle avoidance in densely built-up areas increases energy consumption. Different areas harbor varying accident risks; densely populated and building-rich residential and commercial areas pose a greater risk of accidents. Drone noise can also disturb residents, especially in residential areas.

[0082] In summary, urban low-altitude drone path planning needs to comprehensively consider the impact of static obstacles (such as buildings and permanent no-fly zones) and dynamic obstacles (such as temporary no-fly zones and shielded areas), while also taking into account factors such as energy consumption, safety, and noise interference, in order to achieve a balance between economic and social benefits.

[0083] The actual cost function of the spatiotemporally improved A* UAV path planning algorithm in this embodiment of the invention is composed of three sub-objectives: energy cost, safety cost, and noise cost. The estimated cost function uses the Euclidean distance from the current node to the target node as heuristic information to guide the search direction.

[0084] The invention content section has already described urban environment modeling, directed graph construction and cost function, see Table 2, Table 3 and Formulas (1)-(10) for details.

[0085] This invention introduces a "time" dimension into traditional three-dimensional path planning, constructing a spatiotemporal four-dimensional path. This means that each search node includes not only spatial coordinates but also a time step. Specifically, this is reflected in the algorithm as follows:

[0086] (1) Before expanding each adjacent node, it will first check whether the node is an obstacle at a given time step, including static obstacles such as fixed buildings and no-fly zones, as well as dynamic obstacles formed by other drone paths.

[0087] (2) Once a UAV completes path planning, each quadruple (x,y,z,t) in its planned path will be registered in the spacetime occupancy table (StTable) as a dynamic obstacle that other UAVs need to avoid in subsequent path planning, effectively preventing path intersections and collisions.

[0088] The following table shows the pseudocode of the spatiotemporally improved A* algorithm proposed in this invention:

[0089] Table 4. Pseudocode of the Spacetime Improved A* Algorithm

[0090]

[0091]

[0092] Example 1

[0093] This invention uses a real city map of Nantou Subdistrict, Nanshan District, Shenzhen for environmental modeling. The study area is a 2.5 × 2.25 square kilometer area bounded by Yue Liang Wan Avenue, Nanshan Avenue, and Taoyuan Road. This area is further subdivided into various functional zones, including residential areas, office areas, commercial areas, government offices, transportation hubs, schools, and hazardous areas, to more accurately simulate the operation of drones in a complex urban environment. The map is partitioned using CAD software and converted into a three-dimensional coordinate system with X, Y, and Z axes, as shown in Table 5.

[0094] Table 5. Schematic diagram of three-dimensional coordinates and regional attributes of urban low-altitude airspace.

[0095]

[0096] In Table 5, coordinates (1,1,5) and (1,1,6) are both residential area types. However, at altitude 5, coordinate (1,1,5) is an obstacle, and drones cannot pass through; while at altitude 6, coordinate (1,1,6) is no longer an obstacle, and drones can pass through. Furthermore, coordinate (2,2,5) is a school area. This point itself is not an obstacle and drones are allowed to pass through, but due to the dense population in this area, it is designated as a no-fly zone for safety reasons to avoid potential risks and safety hazards.

[0097] The purpose of this experiment is to analyze the path changes of a single UAV under different weights according to the three different objective functions in equation (10), and to explore the impact of each objective on path planning through sensitivity analysis. This analysis examines the impact of different optimization strategies on the UAV's energy consumption cost C. energy Safety cost C safety and noise cost C noise The path characteristics under different objective functions were analyzed to evaluate the applicability of different optimization strategies. The experiment first randomly generated 50 different start and end points on a 3D urban environment map. All start and end points were located in non-obstacle areas, and the straight-line distance between them was ensured to be greater than 2 kilometers. After determining the start and end points, the optimal path was planned separately considering only a single objective function. The parameters used in path planning are shown in Tables 6 and 7.

[0098] Table 6 Experimental Operation Parameters

[0099]

[0100] Table 7. Shading coefficient, population density, and noise threshold for different functional zones

[0101]

[0102] Multi-objective function comparison. A typical path (from point (64,47,5) to point (195,202,12)) is selected from the 50 generated optimized paths for visualization. Figure 6 This diagram illustrates the planning results for a typical path under different optimization objectives. Considering only the safety cost C... safety In such cases, Figure 6 (a) It can be seen that drones tend to choose park areas with low population density as their main flight path to reduce safety risks. However, considering only the noise cost C... noise In such cases, Figure 6 (b) It can be seen that the drone prioritizes the vertical climb strategy and quickly ascends to an altitude of 120 meters to reduce the noise impact on the ground.

[0103] After obtaining the optimization results of 50 paths, the path characteristics under different optimization objectives were statistically analyzed and compared horizontally. The statistical indicators included: ① flight time; ② energy cost; ③ safety cost; ④ noise cost; ⑤ area type traversed by the UAV path (obscurity coefficient); and ⑥ distance distribution between the path and buildings. The relevant results are shown in Table 8. Figure 7 and Figure 8 As shown.

[0104] Table 8. Comparative analysis of key cost indicators under different optimization objectives (mean of 50 random samples)

[0105]

[0106] Figure 7 This is a schematic diagram illustrating the masking coefficient distribution of a UAV path under different optimization objectives, provided by an embodiment of the present invention. Figure 8 This is a schematic diagram of the distribution of UAV paths and building distances under different optimization objectives provided in an embodiment of the present invention. As can be seen from the figure, there are significant differences in path characteristics under different optimization objectives.

[0107] Considering only energy consumption costs, safety and noise costs increase by 9.971% and 17.776% respectively, with 85% of the paths passing through low-obscurity areas and a low proportion flying above 70m. This means that simply optimizing energy consumption will cause drones to fly close to buildings or at low altitudes in densely populated areas, increasing safety risks and noise pollution.

[0108] Considering only safety costs, flight time, energy consumption, and noise costs increase by 6.342%, 8.396%, and 8.345%, respectively. 90.388% of the routes pass through low-obscurity areas, while high-obscurity areas account for only 3.871%, and 23.803% of flights are within 30 meters. This indicates that when safety is emphasized, drones will detour around densely built-up areas and choose open areas for flight, thus increasing flight time and energy consumption.

[0109] Considering only noise costs, safety costs increased by 10.466%, with high-obscurity areas accounting for 6.377%, medium-obscurity areas for 10.394%, and 55.43% of the flight distance exceeding 70 meters. This indicates that to reduce noise, drones often choose high-altitude flight, even flying over some densely built-up areas, leading to a decrease in safety.

[0110] Weighting Coefficient Configuration: Building upon the analysis of individual objective functions, this paper further explores the weight allocation for the three objective functions. The weights of each sub-objective are designed to start at 0.1 and gradually increase to 0.8 in increments of 0.1, and the comprehensive cost under each weight combination is calculated. During the calculation process, the three cost indicators are first normalized to eliminate the influence of dimensional differences on the weighted results. Subsequently, a weighted summation method is used to calculate the comprehensive cost to evaluate the path performance under different weight combinations.

[0111] Figure 9 A schematic diagram of the overall cost under different weights provided by the present invention, such as Figure 9 As shown, the size of a point is inversely proportional to its overall cost. By selecting the point with the lowest overall cost and drawing its perpendicular lines to the three coordinate axes, the corresponding weight combinations can be obtained as follows: energy cost weight ω1 = 0.6, safety cost weight ω2 = 0.2, and noise cost weight ω3 = 0.2.

[0112] The weighted integrated optimization path achieved a certain balance among various indicators. Compared with optimizing noise or safety costs individually, the energy consumption cost of the integrated optimization path increased by only 1.061%, while the noise cost increased by 6.731%. Its shielding coefficient and building distance distribution characteristics are quite close to those of the energy-optimal path, with flight proportions of 23% within 30m, 27% between 30 and 70m, and 49% above 70m, respectively. This indicates that the integrated optimization scheme can effectively balance different cost factors, providing a more balanced path selection strategy and avoiding extreme biases towards or away from buildings, thus achieving a better balance between economy, safety, and noise impact.

[0113] A multi-drone task allocation model based on demand forecasting: The urban air traffic system is managed by operators who are responsible for the immediate delivery of drones. The operating environment includes take-off and landing centers (providing docking, take-off and landing, charging, loading and unloading of goods and maintenance services) and delivery lockers (for users to store and retrieve themselves, with automatic loading and unloading functions, but not providing long-term docking or charging).

[0114] Figure 10 This is a schematic diagram of order aggregation provided by the present invention. System orders are divided into two categories: actual and virtual. Actual orders include the order placement time, pickup and delivery locker location, cargo weight, and latest delivery time. Actual orders can be further divided into locked orders that have been allocated to drones and unlocked orders that have not been allocated. If the number of drones that can be allocated during scheduling is less than the number of actual orders, the remaining unlocked orders will be included in the next scheduling and given higher priority.

[0115] Virtual orders are generated based on historical data and predictive algorithms. Order information only includes the time the order occurred and the location of the pickup / delivery locker. The algorithm can predict the probability of an order being generated at each locker, but a low probability will generate too many invalid virtual orders, affecting delivery efficiency and accuracy. Therefore, this invention uses an aggregation method to optimize order probabilities:

[0116] First, let P be the probability of an order being generated at locker d. d When P d ≥P min (P min When P is a given empirical value, a virtual order is generated directly; when P... d When the value is small, merge the n nearest drop-off lockers, such that... This forms a cluster of delivery lockers, C. d Considering regional attributes and road network scale, the number of merged delivery lockers also needs to be constrained (n≤N). max ). In determining cluster C d When determining the specific location, the settlement process uses the weighted center of each delivery locker as the location of the virtual order, thereby ultimately generating a series of high-probability virtual orders. Figure 10 The diagram illustrates the order aggregation. Figure (a) shows that some points on the map have a probability of generating orders based on predictions. When the empirical value is 70%, Figure (b) shows that a virtual order is generated at coordinates (3,4). At the same time, the orders at coordinates (4,2), (7,2), and (6,5) are aggregated to obtain the weighted center coordinate (5,3) with a probability of 72%, thus generating another virtual order.

[0117] Drone Description: Based on whether they are executing an order at the initial scheduling moment, drones can be divided into mission drones and standby drones. Mission drones are those currently executing an order and may be en route to a pickup or delivery point. Standby drones, on the other hand, are those not currently executing an order and may be at their take-off or landing point or have left the delivery point and are en route to their take-off or landing point.

[0118] Only one type of drone is used, and drones of the same model have identical battery capacity, range, and payload capacity. Each drone has one cargo bay, which can only serve a maximum of one order, and the total weight of the order cannot exceed the drone's maximum payload. Due to limited power, each drone can perform a maximum of N delivery missions consecutively. After exceeding this number, it must immediately return to the nearest take-off and landing center for battery swapping to ensure the normal execution of subsequent missions.

[0119] This invention addresses the drone scheduling optimization problem in urban aerial delivery systems, primarily solving three aspects: order reception and prediction, task allocation, and path planning. First, while receiving orders, the system needs to predict future order demand based on historical data and filter high-probability orders to improve scheduling efficiency. Second, under the premise of meeting drone payload constraints and the upper limit on the number of consecutive tasks, the system uses minimizing transportation costs and timeout penalties as the objective function to rationally allocate drones to execute actual order tasks or to wait at virtual order pickup points. Finally, the system needs to plan the optimal flight path for drones and monitor flight deviations and potential conflicts in real time.

[0120] The overall operation flow of a scheduling platform system provided in this embodiment of the invention is as follows: Figure 11 As shown, it covers the entire process from user order placement to drone mission execution.

[0121] The basic assumptions of the problem are as follows:

[0122] (I) Assumptions related to drones:

[0123] (1) When using the same type of drone, the parameters such as battery power, flight speed, and payload capacity are completely identical, and the values ​​of these indicators are known.

[0124] (2) The drone is equipped with a cargo compartment with a fixed capacity and can only serve one order;

[0125] (3) The UAV takes off and lands using a vertical take-off and landing method, and can fly at an angle after leaving the take-off and landing center;

[0126] (4) The speed and energy consumption of the UAV are different when it ascends, descends, flies horizontally, and flies diagonally, and the speed and the unit energy consumption required to maintain the speed are known.

[0127] (5) When a drone performs more than N consecutive tasks, it will be judged as having insufficient power and needs to go to the take-off and landing center for a battery swap. The battery swap time is fixed, and the drone will be fully charged each time it leaves the battery swap station. There is no limit to the number of drones that the battery swap station can maintain at the same time, and drones can receive a battery swap immediately;

[0128] (6) After the drone completes the delivery task, if there are no subsequent orders, it needs to return to the nearest delivery center to wait for the next order.

[0129] (7) The pickup and delivery times of the drone at the delivery locker are known;

[0130] (8) Interference from other equipment in the airspace is not considered;

[0131] (9) The setting and layout of the UAV take-off and landing center are not considered;

[0132] (10) Assume that the drone is not affected by extreme weather such as rain, lightning, and hail, and is always in an operational state;

[0133] (11) Assume that no accidents will occur during the flight of the drone to ensure that the delivery task can be completed on time.

[0134] (II) Order-related assumptions:

[0135] (1) The actual order’s pickup point, delivery point, weight of goods, and latest delivery time are known;

[0136] (2) The pickup point for the virtual order is known;

[0137] (3) If the delivery time of the drone is later than the latest delivery time of the order, a late delivery penalty fee will be incurred;

[0138] (4) Orders cannot be split and must be delivered by a single drone in one delivery.

[0139] (5) Except for locked orders, actual orders that were not completed in the previous scheduling need to be redistributed in this scheduling.

[0140] Model parameters

[0141] Tables 9 to 11 further list the definitions and explanations of the residual set, key parameters, and decision variables involved in the model.

[0142] Table 9. Set Symbols and Definitions

[0143]

[0144] Table 10 Parameter Symbols and Definitions

[0145]

[0146] Table 11 Decision Variables and Definitions

[0147]

[0148]

[0149] Objective function: Let represent the transportation cost when drone u performs task k. Equation (10) is used as the formula for calculating the transportation cost. Based on this objective function, and incorporating actual and virtual orders, we can obtain:

[0150]

[0151] Overtime penalty cost

[0152] This represents the penalty cost for drone u performing task k, incurred when the actual delivery time exceeds the latest delivery time. For services such as food delivery or emergency medical supply transport, time sensitivity is high; late delivery can directly impact cargo quality or the fulfillment of urgent needs, leading to decreased user trust and reduced market share.

[0153]

[0154] Therefore, the objective function can ultimately be expressed as:

[0155]

[0156] Constraints

[0157]

[0158]

[0159] Equation (14) is a task priority constraint, ensuring that tasks with higher priority are assigned first. Equation (15) is a drone state constraint, ensuring that drones are not assigned tasks repeatedly while performing tasks. Equations (16) and (17) are task assignment uniqueness constraints, ensuring that each task can be assigned to at most one drone. Equation (18) is a virtual order constraint, ensuring that virtual orders can only be assigned after all actual orders have been assigned. Equation (19) is a drone payload constraint, ensuring that the drone's payload capacity (maximum payload minus its own weight) must be greater than or equal to the weight of the task's cargo. Equations (20) to (22) are constraints on the number of tasks a drone can continuously execute. Equation (20) ensures that the number of tasks is accumulated after each task is completed by the drone, and Equation (21) ensures that the accumulated number of tasks is refreshed each time the drone leaves the take-off and landing center. Equation (23) is a path point flow balance constraint, ensuring that the drone's flight path between path points is continuous and that there is no path interruption or repeated passage through a certain path point. Equations (24) and (25) are take-off and landing center constraints, ensuring that the UAV can only take off from and return to one take-off and landing center at most when performing a mission. Equations (26) and (27) are drop-off locker constraints, ensuring that the UAV will only go to one drop-off locker to pick up goods and go to one drop-off locker to deliver goods when performing a mission. Equation (28) is a node traversal sequence constraint, ensuring that the UAV's path is continuous.

[0160] Simulation Experiment of On-Demand Forecasting and Dynamic Scheduling for Instant Delivery

[0161] This invention focuses on the real-time delivery scenario of urban low-altitude drones, conducting a simulation experiment covering the entire process from order prediction to dynamic scheduling optimization. A CNN-LSTM-Attention prediction model is used to predict short-term order demand, serving as decision support for drone pre-scheduling. Furthermore, an integrated "prediction-planning-execution" simulation framework is established, with simulation experiments focusing on three aspects: drone quantity configuration, pre-scheduling empirical threshold parameters, and conflict avoidance mechanisms. These experiments verify the effectiveness of the proposed method in improving scheduling efficiency, reducing system costs, and ensuring operational safety.

[0162] Order demand forecast

[0163] Based on two months of drone instant delivery operation data from a Shenzhen company, this dataset covers the entire process of order placement and delivery, and is highly representative and practical.

[0164] Correlation analysis

[0165] Aggregate and analyze orders for a specific month by date and time, such as... Figure 12As shown, a clear intraday periodic fluctuation in order volume can be observed. Order peaks are mainly concentrated in two time periods: 10:00 AM to 1:00 PM and 5:00 PM to 7:00 PM. To effectively represent the periodicity of time, a time-series feature engineering method can be used to map the daily hours (hour∈[0,23]) onto a unit circle, constructing its two-dimensional periodic expression using sine and cosine functions. The specific conversion formula is as follows:

[0166]

[0167] Where hour represents the number of hours in a day, and T is the length of the time period.

[0168] After completing data cleaning and feature construction, this invention performs Pearson correlation coefficient analysis on key numerical variables. This coefficient measures the linear correlation between two variables, with a value range between [-1, 1], where: ρ > 0 indicates a positive correlation, meaning that when one variable increases, the other variable also increases; ρ < 0 indicates a negative correlation, meaning that when one variable increases, the other variable decreases; ρ = 0 indicates that there is no linear correlation. Figure 13 The Pearson correlation heatmap provided in this embodiment of the invention shows that multiple features in the current dataset are correlated with future order volume to varying degrees, indicating that these factors have a certain influence in predicting future order volume.

[0169] Experimental environment and model training configuration

[0170] ① Experimental environment configuration

[0171] Order prediction primarily employs the CNN-LSTM-Attention model, which combines a convolutional neural network (CNN), a long short-term memory network (LSTM), and an attention mechanism to effectively handle complex patterns and long-term dependencies in time-series data. Simultaneously, for comparative analysis, this invention also uses the XGBoost model as a benchmark. By comparing the prediction performance of different models, the effectiveness and applicability of the models can be more comprehensively evaluated.

[0172] The programming environment for this experiment is Python and PyTorch 2.2.0, operating system: Windows 10, memory: 16GB, CPU: 11th Gen Intel(R) Core(TM) i5-1135G7@2.40GHz, GPU: Intel(R) Iris(R) XeGraphics.

[0173] ② Model training parameter settings

[0174] To further improve model performance, this invention systematically sets and optimizes the parameters of the XGBoost model and the CNN-LSTM-Attention model.

[0175] 1) For the XGBoost model, this invention employs GridSearchCV for hyperparameter tuning. The specific parameter grid is shown in Table 12. During the hyperparameter search process, 3-fold cross-validation is used to evaluate each parameter combination, and negative mean squared error is used as the evaluation index for model performance. Finally, the parameter combination with the smallest error is selected as the optimal solution.

[0176] Table 12 XGBoost Parameter Mesh

[0177]

[0178] 2) For the CNN-LSTM-Attention model, the parameters used during the training phase are as follows: Batch Size = 64, Epochs = 100, and Validation Split = 25%.

[0179] Furthermore, this invention introduces a learning rate scheduler to balance rapid convergence in the early stages of training with fine-grained optimization in the later stages. Specifically, the initial learning rate is kept constant for the first 10 rounds, and then decreased by 1% per round from the 11th round onwards. This improves convergence speed while enhancing the model's stability and generalization ability in later stages.

[0180] Order forecast results for the next 15 minutes

[0181] A total of 62 days of actual order data were selected, covering the period from 8:00 AM to 9:30 PM daily, with statistical analysis performed at 15-minute intervals, resulting in 3348 data samples. These samples were then divided into a 90% training set and a 10% prediction set. The prediction results based on the XGBoost model and the CNN-LSTM-Attention model are shown below. Figure 14 and Figure 15 As shown in the figure, the loss curves of the training set and test set during the training process of the CNN-LSTM-Attention model are as follows: Figure 16 As shown in Table 13, the predictive performance evaluation indicators are summarized in the table.

[0182] Table 13 Evaluation of Model Prediction Indicators

[0183]

[0184] Based on the prediction results, both the XGBoost model and the CNN-LSTM-Attention model can effectively capture the morning and evening peak trends of order demand. Figure 16 The fitting results show that the model performs consistently on both the training and test sets, without significant overfitting or underfitting, indicating that the CNN-LSTM-Attention model has good generalization ability. The CNN-LSTM-Attention model performs well in terms of mean squared error (MSE), root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²). 2 The CNN-LSTM-Attention model outperforms the XGBoost model in key metrics such as accuracy and fit, indicating that it performs better in overall prediction accuracy and fitting ability. Although the mean absolute percentage error (MAPE) of the CNN-LSTM-Attention model is relatively high, it is still better than the XGBoost model.

[0185] Design of a Simulation Framework for Scheduling a Multi-UAV Delivery System Based on Demand Forecasting

[0186] Design of a pre-scheduling method based on aggregation

[0187] ①Setup of take-off and landing center and delivery locker

[0188] In the aforementioned environmental map modeling, the map has been divided into several different functional areas, but the site selection for take-off and landing centers and drop-off lockers is still lacking. To simplify the site selection model and focus on the core issues of this invention, the site selection modeling mainly considers two factors: functional area attributes and population density within the area. Using the peak-period population density of each area as a reference standard, cluster analysis is used to aggregate the map into regions, thereby obtaining suitable candidate locations for setting up drop-off lockers and take-off and landing centers. The parameters and weights used in the clustering are shown in Table 14.

[0189] Table 14 Functional Zone Types and Population Density Ratios

[0190]

[0191] By combining the relative proportion of the number of grid points in each functional area with the population density of that area, the service demand weight of each area can be calculated. The deployment location of delivery lockers is determined using a weighted clustering method. Based on the clustering results, some lockers are placed on the ground floor, and some are placed on the rooftops of buildings to meet the needs of users in high-rise buildings. When the total population in a functional area exceeds a set threshold, a second delivery locker will be added to improve service coverage. Furthermore, the weighted cluster center of all delivery lockers is selected as the take-off and landing center for drones to achieve a more efficient delivery network layout. Figure 17 This is a schematic diagram of the layout planning results of the take-off and landing center and the delivery cabinet cluster center provided in an embodiment of the present invention.

[0192] ② Design of pre-scheduling method

[0193] like Figure 12 As shown, order generation exhibits temporal fluctuations. It is assumed that during the morning peak hours (10:00-13:00), approximately 80% of orders will be generated in the office area, and 20% in the living area; while during the evening peak hours (17:00-19:00), 80% of orders will be generated in the living area, and 20% in the office area. Orders during other times are assumed to be generated evenly. Furthermore, the office and living areas are divided by the delivery lockers generated in the previous section based on their functional attributes. Some delivery lockers have similar functional attributes and are therefore set as general-purpose delivery lockers. The specific delivery locker settings are as follows: 8 delivery lockers in the office area, 6 in the living area, and 8 in the general-purpose area. Based on this, virtual orders are generated using the aforementioned virtual order aggregation method, and pre-scheduled based on the remaining drone capacity. The specific method is as follows:

[0194] Step 1: Predicting Data

[0195] Assuming the current time is 9:45, based on historical data, the CNN-LSTM-Attention model predicts the number of orders in the next 15 minutes. The model predicts that 20 orders will be generated at 10:00. Assuming the time period, 80% of these orders will be generated in the office area, and 20% will be generated in the living area, resulting in 16 orders and 4 orders respectively.

[0196] Step Two: Order Allocation

[0197] Assume that the generation of virtual orders is random and not specific to any particular drop locker. Based on the number of drop lockers and their functional attributes, for the office area, the probability of each of the 16 drop lockers in all office and general areas generating an order is:

[0198]

[0199] For the 14 delivery lockers in all living and general areas within the venue, the probability of each locker generating an order is:

[0200]

[0201] The specific probabilities of generating an order for each delivery locker are as follows: the probability of generating an order in the office area is 64.39%, the probability of generating an order in the living area is 25.65%, and the probability of generating an order in the general area is 73.52%.

[0202] By setting an empirical value of 70%, when the probability of a delivery locker generating an order is greater than 70%, that locker will generate a virtual order and add it to the order collection. For delivery lockers with a lower probability, the orders generated therewill be further aggregated.

[0203] Step 3: Clustering and Aggregation

[0204] 1) Calculate the distance matrix: Based on the spatial coordinates of all low-probability orders, calculate the distance between pairs of them to obtain a distance matrix.

[0205] 2) Hierarchical Clustering: Based on the calculated distance matrix, hierarchical clustering is performed using the Ward method. By selecting an appropriate distance threshold (e.g., 300), low-probability orders are divided into several clusters. The size of each cluster is limited by a maximum number of nodes (e.g., M=4). When the number of orders in a cluster exceeds this limit, the cluster will be further subdivided into multiple subclusters.

[0206] 3) Cluster splitting: If the number of orders in a cluster exceeds the preset maximum number of points M, it is split by sub-level clustering to ensure that the number of orders in each sub-cluster does not exceed the maximum limit.

[0207] 4) Calculate the probability of each cluster: The overall probability of each cluster is obtained by calculating the probability of all orders within the cluster. The formula is as follows:

[0208]

[0209] Where, p i Let be the probability of the i-th order within the cluster. If the calculated cluster probability is greater than a set empirical threshold, then the cluster is determined to generate a virtual order.

[0210] The virtual order aggregation results differ depending on the value of M, and the number of virtual orders generated is shown in Table 15. This experiment uses the aggregation result when M=3 as the final generated result.

[0211] Table 15 shows the number of virtual orders generated with different M values.

[0212]

[0213] Step 4: Order Set Generation

[0214] At 9:45, the actual orders and virtual orders are merged, and the resulting set of scheduled orders is shown in Table 16.

[0215] Table 16 Set of scheduling orders generated at 9:45

[0216]

[0217]

[0218] This invention proposes a deviation correction and conflict avoidance mechanism to dynamically adjust the flight path of UAVs and ensure the safety and stability of low-altitude airspace. Figure 18 This is a schematic diagram of execution layer deviation conflict monitoring and path replanning provided in an embodiment of the present invention, as shown below. Figure 18 As shown, it is necessary to monitor the drone's position information in real time and make timely adjustments according to environmental changes to ensure that the drone flies along the optimal route. When the system detects that a drone's path may conflict, it must adjust the individual flight path in a timely manner to avoid collisions and ensure the stability and safety of the overall operation. Figure 19 A flowchart of a deviation correction and conflict avoidance mechanism provided in an embodiment of the present invention is shown below. Figure 19 As shown, this mechanism sorts all flying drones in the airspace at fixed time steps and detects and adjusts their flight paths according to priority. For the drone N with the highest priority in the current sorting... i First, it checks whether a path deviation has occurred. If the deviation exceeds a preset threshold, it replans the optimal path for the drone based on environmental information. If there is no deviation, it further predicts whether the drone will collide with the preceding drone N in the next T time steps. i-1 There is a potential trajectory conflict. If a conflict risk is detected, then for N... i The improved spatiotemporal A* algorithm described in Section 4.1.2 is used to replan paths to avoid collisions. This process continues iteratively until all UAVs have completed detection and adjustment, thereby optimizing overall airspace scheduling efficiency while ensuring individual safety.

[0219] The "prediction-planning-execution" simulation system provided in this embodiment of the invention consists of three main modules: a prediction layer, a planning layer, and an execution layer. The overall flowchart is shown below. Figure 20 As shown.

[0220] Simulation experiments and drone deployment analysis were conducted. This invention examines the drone deployment cost, transportation cost, and overdue penalty cost from three aspects: the number of drones gradually increasing from 4 to 24, analyzing their impact on total cost and the number of delayed tasks. Simultaneously, the marginal cost change rate was introduced as an auxiliary indicator to provide a quantitative basis for drone deployment decisions. The total operating cost of an on-demand delivery drone over its lifecycle is as low as 1400 yuan per month, so it is assumed that the daily deployment cost of a drone is 45 yuan / drone, and the overdue penalty coefficient δ is set to 0.4 yuan / minute. A Python implementation of a traversal method was used to obtain the exact solution of the task allocation model, and the costs under different drone numbers are shown in Table 17.

[0221] Table 17 Cost Comparison for Different Numbers of Drones (Unit: Yuan)

[0222]

[0223] Table 17 shows that the activation cost increases linearly with the number of drones, while the overtime penalty cost gradually decreases with the increase in the number of drones. Transportation costs are higher when the number of drones is small, gradually decreasing as the number increases, but begin to rise slowly again after the number exceeds a certain threshold. Considering all three types of costs, the total cost generally shows a trend of first decreasing and then increasing, reaching its lowest value when the number of drones is 10. Further analysis using the marginal cost change rate shows that when the number of drones exceeds 10, the marginal cost tends to increase, and the cost-effectiveness of the invention gradually weakens.

[0224] Meanwhile, the number of drones is also closely related to the number of timed-out tasks. When the number of drones is small, the task allocation pressure is high, and task backlog is likely to occur, causing some orders to fail to be completed within the specified time, resulting in more timeouts. As the number of drones increases, the system scheduling capability is enhanced, and the number of timed-out tasks decreases significantly. As shown in Table 18, when the number of drones is between 16 and 20, each increase brings a reduction in timeouts of over 50%, resulting in high marginal benefits.

[0225] Table 18 Comparison of Time-Out Tasks under Different Numbers of Drones

[0226]

[0227] In summary, considering both cost and mission timeout, a relatively balanced resource allocation is achieved when the number of drones is between 14 and 18. This invention recommends 16 drones as the optimal number for the number of drones on-site.

[0228] Predictive Threshold Analysis: To improve the foresight and response efficiency of system scheduling, this invention introduces a virtual order generation mechanism into the proposed aggregation-based pre-scheduling method, and determines whether to generate virtual orders by setting an empirical threshold. When the prediction probability exceeds 60%, virtual orders are mainly concentrated during peak order periods, verifying the effectiveness of this mechanism in capturing demand peaks.

[0229] We set experience thresholds from 60% to 95% and analyzed order data for a full day. (Table 19 and...) Figure 21 The effects of strategies that do not introduce a prediction mechanism and generate virtual orders under different probability thresholds were compared.

[0230] Table 19 Comparison of Costs and Number of Overdue Tasks under Different Forecasting Strategies

[0231]

[0232] Depend on Figure 21It can be seen that the strategy system without introducing a prediction mechanism performs consistent with the 95% experience threshold, because virtual orders are concentrated during peak hours, and the strain on system capacity makes pre-scheduling ineffective. When the experience threshold is too low, the number of timed-out tasks and penalty costs are significantly reduced, but transportation costs increase because too many virtual orders lead to ineffective drone round trips.

[0233] In summary, excessively low thresholds lead to resource waste, while excessively high thresholds make early intervention in scheduling difficult. Overall, the empirical threshold strategy of 70% and 75% is optimal. At this level, the number of virtual orders is moderate, efficiency is high, scheduling is forward-looking while avoiding resource waste, achieving dual optimization of cost and service quality, and verifying the effectiveness of the aggregated pre-scheduling method.

[0234] Deviation and Conflict Avoidance Analysis: When 16 UAVs were scheduled at this threshold, multiple spatiotemporal path overlaps occurred (points appearing in pairs of the same color at the same time were considered collision locations). Conflicts were mainly concentrated during peak scheduling periods, indicating a high probability of potential conflicts due to scarce spatiotemporal resources, frequent path intersections, and other factors.

[0235] Further analysis revealed that conflicts frequently occur in two critical phases before and after mission execution, particularly when the drone approaches the delivery locker and returns to the takeoff and landing center. Because these areas have fixed or similar paths, drone paths tend to converge, leading to spatial overlap and increasing the risk of conflict. Comparative experiments show a difference between scheduling without and without a path replanning mechanism: without an avoidance mechanism, although the drone can avoid actual collisions, adjusting back to its original path takes time, reducing mission efficiency. This invention imposes a 2-second penalty on each drone involved in a potential conflict to quantify the time loss. The differences in the number of path conflicts and overall scheduling cost between the two strategies are shown in Table 20.

[0236] Table 20: Number of path conflicts and cost of path adjustment under different strategies

[0237]

[0238] The results show that without the collision avoidance strategy, the system experienced 81 path conflicts, resulting in a total scheduling cost of 324 seconds. This indicates that in scenarios with dense UAV operations, path conflicts are frequent, and autonomous collision avoidance requires time for adjustments. In contrast, after introducing the collision avoidance strategy, the number of path conflicts decreased to 59, a reduction of approximately 27.2%. This reduction is not only due to the proactive avoidance of current conflicts but also because path adjustments affect subsequent flight trajectories, thereby reducing the likelihood of future conflicts. Under this strategy, although some path replanning led to an increase in scheduling time (a total of 368 seconds), it remains within an acceptable range overall.

[0239] In summary, conflict avoidance strategies are effective in reducing potential conflicts and improve operational stability without significantly increasing system costs. This method has certain application value in scheduling scenarios with scarce urban low-altitude airspace resources and intensive tasks.

[0240] In summary, this invention provides a more detailed rasterized model of urban low-altitude airspace and rationally divides urban airspace by considering the characteristics of functional areas such as residential areas, shopping malls, and office buildings. Compared to traditional two-dimensional path planning methods, this method better reflects the characteristics of urban three-dimensional space, providing more accurate spatial data support for UAV path planning. This method not only improves the rationality of low-altitude logistics delivery path planning but also provides more detailed data support for urban airspace management, to some extent filling the gaps in existing research on urban low-altitude spatial modeling. Furthermore, this modeling method has certain scalability and can be applied to fields such as air traffic management and low-altitude inspection.

[0241] This invention designs an improved spatiotemporal A* algorithm. Building upon existing research, it more comprehensively considers multi-objective optimization factors such as energy consumption, noise, and safety. Compared to traditional algorithms that only aim for the shortest distance or shortest flight time, this method plans paths that are closer to the actual operational needs of urban low-altitude logistics. Furthermore, deviation correction and conflict avoidance mechanisms can monitor the UAV's flight status in real time. When deviations from the preset path or potential collisions are detected, the system can replan the path according to the priority of the UAV's tasks, ensuring flight safety and efficient delivery execution.

[0242] This invention constructs a multi-UAV collaborative scheduling model that comprehensively considers factors such as task time windows, priorities, and the upper limit of the number of consecutive delivery tasks, with the optimization objective of minimizing transportation costs and timeout penalty costs. Compared to traditional static scheduling methods, this research introduces a machine learning prediction method based on historical data to predict potential future orders and pre-schedule UAVs accordingly, improving the system's response speed and resource utilization efficiency. This method can effectively enhance the collaborative operation capabilities of UAVs, making their scheduling more flexible and reliable in dynamic environments.

[0243] The "prediction-planning-execution" technical framework proposed in this invention covers the entire process from airspace modeling and path planning to multi-UAV scheduling, providing a complete research approach for optimizing urban low-altitude UAV instant delivery systems. This framework is not only applicable to urban logistics and distribution but can also be extended to scenarios such as medical emergency response and emergency material delivery. The research results can provide useful references for the design of future urban low-altitude airspace management and UAV scheduling systems, and promote the application of related technologies in smart city construction.

[0244] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of one embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing the present invention.

[0245] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that the present invention can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in various embodiments or some parts of the embodiments of the present invention.

[0246] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0247] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for real-time delivery scheduling and route planning of urban low-altitude unmanned aerial vehicles (UAVs) considering demand forecasting, characterized in that, include: Based on urban map data, 3D urban environment map data was obtained, and a 3D environment map of the drone delivery system was established using a "grid-topology" hybrid modeling strategy. A spatiotemporally improved A* drone path planning algorithm is constructed. Based on the 3D environment map of the drone delivery system, the drone's starting point and target point, the spatiotemporally improved A* drone path planning algorithm is used to obtain the optimal flight path of the drone, the time required for the optimal path, and the transportation cost. Based on historical order data, current order data, and drone performance parameters, as well as the optimal flight path of the drone, the time required for the optimal path, and transportation costs, the CNN-LSTM-Attention order prediction model and task allocation model are used to output a task allocation scheme that includes virtual orders and actual orders.

2. The method according to claim 1, characterized in that, The process of acquiring 3D urban environment map data from city map data and establishing a 3D environment map of the drone delivery system using a "grid-topology" hybrid modeling strategy includes: A 3D urban environment map is obtained based on city map data. This 3D urban environment map includes city zoning, obstacle data, and no-fly zone data. A "grid-topology" hybrid modeling strategy is used to build a 3D environment map for the drone delivery system. The drone delivery flight network in urban airspace is represented by a directed graph G(V,E), where the node set V consists of the center points of the effective grids. It is assumed that the environment consists of N 3D grids, and the grid index set is G = {1,2,...,g}. The position of each grid i (i∈G) is represented by 3D coordinates (x,y). i ,y i ,z i Let E represent the connectivity between grid cells. If grid cells i and j are adjacent in 3D space, then there exists an edge (i, j) from i to j in the directed graph. In 3D space, the connectivity between grid cells is based on 26 possible adjacency directions, i.e., the relative coordinate difference |x| between grid cells. i -x j |≤1,|y i -y j |≤1, and |z i -z j |≤1 indicates their adjacency in the horizontal, vertical and diagonal directions.

3. The method according to claim 2, characterized in that, The aforementioned spatiotemporally improved A* UAV path planning algorithm, based on the 3D environment map of the UAV delivery system, the UAV's starting point, and the target point, utilizes the spatiotemporally improved A* UAV path planning algorithm to obtain the optimal flight path of the UAV, the time required for the optimal path, and the transportation cost, including: A spatiotemporally improved A* UAV path planning algorithm is constructed, and the actual cost function of the spatiotemporally improved A* UAV path planning algorithm is set to be the energy consumption cost C. energy Safety cost C safety and noise cost C noise The weighted average of the three sub-objectives is as follows: The drone from the grid point (x i ,y i ,z i ) fly to (x j ,y j ,z j Energy consumption cost C energy Equation (1) represents: The drone at its current location (x) i ,y i ,z i Safety cost C safety The conditional probabilities of these stages are calculated by integration, and the specific formula is shown in (3): C safety =P c E c =P d P i (x,y)P f (x,y)E c (3) We consider ballistic descent and uncontrolled glide as two descent modes. Assuming the probability of each descent mode occurring remains constant for different UAVs, the probabilities of the two descent modes are: The probability P of a drone u hitting a person on the ground i (x, y) represents the probability that the drone, after losing control and falling to the ground, will collide with at least one person. Specifically: P i (x,y)=ρ(x,y)A exp (4) Probability of injury or death P f (x, y) represents the probability of injury or death due to a collision between a crashed drone and a person. The injury / death rate is related to the shielding coefficient S. f The kinetic energy E(x,y) generated during the collision is related to the kinetic energy E(x,y) generated during the collision. The calculation method is shown in equation (7): Noise cost of drones C noise It shall not exceed the maximum allowable standard for the corresponding area, and the calculation method is shown in equation (9): C noise =E n [L-20lg(b / r0)+3-N(x,y)] (9) In summary, the actual cost function g(n) is expressed as: in, and C respectively energy C safety and C noise Weighting coefficients; By utilizing the spatiotemporal improvement of the A* UAV path planning algorithm, local optima are followed at each decision step, minimizing the subsequent single-step cost function g(n). A single path planning requires multiple iterative calculations. Through continuous optimization and adjustment of path selection, the algorithm gradually approaches a better path scheme that meets actual needs. The symbols and definitions of the parameters and variables involved in the cost function are shown in Tables 2 and 3: Table 2 Parameter Symbols and Definitions Table 3 Decision Variables and Definitions 4. The method according to claim 3, characterized in that, The method described above, based on historical order data, current order data, UAV performance parameters, the optimal flight path of the UAV, the time required for the optimal path, and transportation costs, utilizes a CNN-LSTM-Attention order prediction model and a task allocation model to output a task allocation scheme that includes both virtual and actual orders. Actual orders are generated after a customer places an order. The order information includes the order time, the location of the pickup locker, the location of the delivery locker, the weight of the goods, and the latest delivery time. Virtual orders are predicted based on historical data and prediction algorithms. The order information only includes the time the order occurred and the location of the pickup locker. The CNN-LSTM-Attention order prediction model is a hybrid model that integrates a Convolutional Neural Network (CNN), a Long Short-Term Memory (LSTM) network, and an Attention mechanism. First, the CNN layer uses 1D or 2D convolutional kernels to extract local spatiotemporal features from the order data, which are then reduced in dimensionality by pooling layers to obtain feature vectors. Next, the LSTM layer processes these feature vectors, using a gating mechanism to learn the long-term temporal dependencies of the order data. Then, an Attention layer is introduced to calculate weights for the hidden states at each time step of the LSTM output, highlighting the key time steps for prediction. Finally, a fully connected layer outputs the predicted number of actual orders within a set future timeframe. Virtual orders are generated by clustering low-probability but spatially adjacent grid points, and these virtual orders, along with the actual orders, form a task set. The task allocation model treats virtual and actual orders as tasks. It includes constraints such as task priority, drone payload limits, maximum number of consecutively executed tasks, and node access order. With the goal of minimizing transportation costs and timeout penalties, the model uses a spatiotemporally improved A* drone path planning algorithm to calculate the optimal flight path, required time, and transportation cost for all tasks, serving as the foundational data for task allocation. Finally, using these pre-calculated path costs and times as input, and under the constraint of minimizing total transportation costs and timeout penalties, the task allocation model allocates the task set to drones, prioritizing actual orders and allocating remaining capacity to virtual orders, thus forming the final task allocation scheme.

Images