Low-altitude logistics dynamic route planning method and system based on demand prediction driving

Abstract

The invention provides a low-altitude logistics dynamic route planning method and system based on demand prediction driving, and belongs to the technical field of low-altitude traffic management and intelligent logistics scheduling. The method comprises the following steps: predicting a logistics demand by using a space-time diagram neural network and generating a virtual task set; constructing a dynamic airspace state model fusing real-time weather, control and monitoring data; by taking the virtual task set as input and the dynamic model as constraint, executing global route planning to generate a preliminary plan; calculating a network congestion risk degree and a planned load balance degree index, triggering feedback adjustment according to an index threshold value, and performing iterative optimization to form a global reference route plan; and in the real-time operation stage, dynamic re-planning is carried out by taking the reference route plan as a baseline and combining a real-time order and the latest airspace state. Through cooperation of prediction, modeling, planning and feedback, conversion from passive response to active scheduling is realized, and the problems of planning lag, local congestion and insufficient global cooperation are solved.

Description

A Demand Forecast-Driven Dynamic Route Planning Method and System for Low-Altitude Logistics Technical Field

[0001] This invention relates to the field of low-altitude traffic management and intelligent logistics scheduling technology, and in particular to a method and system for dynamic route planning of low-altitude logistics based on demand forecasting. Background Technology

[0002] As an emerging delivery mode, low-altitude logistics relies heavily on precise route planning technology for its efficient operation. Current low-altitude logistics route planning schemes can be mainly divided into two categories: static global planning and real-time reactive planning.

[0003] Static global planning methods rely on fixed geographical information and pre-set flight rules, and are essentially pre-defined, rigid solutions. They cannot adapt to the dynamic changes in logistics needs in time and space during actual operations, nor can they respond to real-time airspace disturbances (such as sudden weather or temporary air traffic control), thus exhibiting poor robustness in dynamic environments.

[0004] Real-time reactive planning methods focus on in-process, passive responses. They typically perform online calculations and replanning based on real-time environmental information perceived by the UAV and currently generated orders. While this method has a certain degree of dynamic adaptability, its core logic is event-triggered, meaning that calculations only begin when an order actually arrives or a conflict is about to occur. This leads to three inherent and interrelated systemic defects: (1) Planning lag and congestion dilemma: Because planning always lags behind demand, the system lacks the ability to proactively respond to predictable future demand peaks (such as midday delivery peaks and holiday order surges). When demand surges, passive planning can easily lead to instantaneous overload of local airspace traffic, congestion and queuing at take-off and landing nodes, resulting in a sharp decline in overall network efficiency and a deterioration in service quality.

[0005] (2) Local Optimality and Global Efficiency Loss: The optimization objective of this type of method is usually limited to minimizing the flight distance or time of a single or a small number of tasks. This "short-sighted" decision-making mode lacks a global consideration of the overall set of tasks in the future. After the continuous accumulation of local optimal decisions, it often leads to conflicts in the utilization of system-level resources (such as airspace access capacity, take-off and landing pads, and charging facilities) and a loss of overall efficiency.

[0006] (3) Lack of coordination and predictability in resource allocation: Due to the inability to predict the scale and spatiotemporal distribution of future tasks, the system is unable to coordinate and optimize the allocation of limited, distributed physical resources (such as charging piles and loading / unloading points) across time periods and regions, resulting in uneven resource availability and an inability to support large-scale, high-density commercial operations.

[0007] In summary, the fundamental limitation of existing technologies lies in the disconnect between their planning process and future demand information, lacking a mechanism that deeply integrates forward-looking demand forecasting into route generation and resource scheduling decisions. Summary of the Invention

[0008] The purpose of this invention is to provide a dynamic route planning method and system for low-altitude logistics based on demand forecasting, in order to solve the problems that existing static planning cannot adapt to dynamic demand and airspace changes, as well as the problems of planning lag, congestion and lack of global coordination in real-time reactive planning. It achieves forward-looking route generation, dynamic conflict resolution and global resource optimization, thereby improving the overall operating efficiency, robustness and scalability of the low-altitude logistics system.

[0009] To achieve the above objectives, this invention provides a low-altitude logistics dynamic route planning method based on demand forecasting, comprising the following steps: Step S1, obtaining low-altitude logistics spatiotemporal demand forecast data for future target time periods; Step S2, generating a virtual task set for pre-planning based on the spatiotemporal demand forecast data; Step S3, constructing a dynamic airspace state model based on the spatiotemporal range covered by the spatiotemporal demand forecast data and combined with real-time airspace information sources; Step S4, performing initial global route planning using the virtual task set as the task set to be assigned and the dynamic airspace state model as the constraint condition to obtain a preliminary route plan; Step S5, calculating a first indicator reflecting future network congestion risk and a load balance indicator reflecting the plan execution based on the preliminary route plan. The second indicator; Step S6: Determine whether the first indicator exceeds the first threshold, or whether the second indicator is lower than the second threshold; Step S7: If the determination result of step S6 is yes, then determine the feedback adjustment strategy based on the comparison result of the first indicator and the first threshold, and / or the comparison result of the second indicator and the second threshold, and update the constraints or task set input to step S4 according to the feedback adjustment strategy, and then return to execute step S4; If the determination result is no, then determine the preliminary route plan output by the current step S4 as the global baseline route plan; Step S8: In the real-time operation phase, obtain real-time logistics orders, and combine them with the latest airspace status, using the global baseline route plan determined in step S7 as the adjustment baseline, and perform dynamic route replanning and scheduling.

[0010] Preferably, step S1 specifically includes: step S101, acquiring historical logistics order data, urban dynamic information data, and environmental data; historical logistics order data includes at least historical origin-destination pairs, cargo weight, order placement time, and delivery time; urban dynamic information includes at least regional population heatmaps, point-of-interest distribution, and event calendars; environmental data includes at least historical meteorological information for the same period; step S102, inputting the data acquired in step S101 into a pre-trained spatiotemporal graph neural network prediction model; the spatiotemporal graph neural network prediction model takes historical sequence data as input, and its output layer corresponds to the spatiotemporal units divided within the future target time period, with the output value being the predicted logistics demand; step S103, receiving the output of the spatiotemporal graph neural network prediction model, with dimensions of... Future OD demand tensor ,in Indicates the number of future time slices. Indicates the number of spatial units. Indicates the first Within a time slice, from spatial units To space unit The predicted freight volume.

[0011] Preferably, step S2 specifically includes: step S201, parsing the future OD demand tensor. For each of these non-zero elements This determines it as a predicted logistics flow direction. and its transport volume and associated time slices Step S202: For each logistics flow direction and its transport volume Based on the preset single-mission carrying capacity limit Break it down into One virtual task, among which is the floor function; is the th element after splitting. A virtual task Calculate its expected takeoff time With expected arrival time The calculation formula is as follows: ; ;in, and Time slices The start time and duration, For spatial units arrive The preset average service duration; Step S203: Summarize all generated virtual tasks into a virtual task set. Each of the tasks Includes attributes , Indicates task The initial space unit, Indicates task The target space unit, Indicates task Expected takeoff time Indicates task Expected arrival time Indicates the task index. This indicates the total number of virtual tasks.

[0012] Preferably, step S3 constructs a dynamic airspace state model, specifically including: step S301, based on the geographical area involved in the spatiotemporal demand prediction data in step S1. and future target period Initialize a four-dimensional spatiotemporal grid model, where, and These represent the minimum and maximum latitudes of the region, respectively. and These represent the minimum and maximum longitudes of the region, respectively. and These represent the start and end times of the future target time period, respectively; the four-dimensional spatiotemporal grid model is spatially divided into... The cells are divided into time segments. Each time period is denoted as (the four-dimensional unit is then defined as follows). ,in Indicates the spatial row index. Indicates a spatial column index. Indicates time index; Step S302: Access real-time meteorological data, airspace control notices, and automatic dependent surveillance broadcast data; Step S303: For each four-dimensional unit Dynamically calculate its traffic status weight The calculation formula is as follows: ;in, This indicates the weighting of weather impacts obtained by mapping real-time meteorological data. This represents the control constraint weights mapped from airspace control notices. This represents the real-time aircraft density calculated based on automatic dependent surveillance (ADS) data. , , These represent the normalization coefficients of each term; Step S304: Define the four-dimensional spatiotemporal grid model and the traffic state weights of each element together as the dynamic spatial domain state model. Its formal representation is: .

[0013] Preferably, step S4 specifically includes: step S401, setting up the virtual task set. Each task The origin and destination are mapped to the dynamic spatial state model. The nearest spatial grid node; Step S402: Construct the optimization objective function The calculation formula is as follows: ;in, , , All represent preset weighting coefficients; This represents the total flight distance cost calculated based on the planned path length; Indicates delivery time based on the task schedule. Compared to the expected arrival time The total delay penalty cost is calculated based on the deviation. This represents the estimated cost of congestion, calculated using the following formula: ;in, This represents the statistics based on the path allocation scheme of the current iteration. The expected flow, Representing spatial grid The static capacity threshold; step S403, in Under the constraints, a heuristic search algorithm is used to solve the optimization objective function. The minimum value is The task allocation specifies the flight path and takeoff time, and outputs a preliminary route plan. .

[0014] Preferably, step S5 specifically includes: step S501, analyzing the preliminary route plan. Statistics in the future Within a given time period, the plan is to utilize spatial grids. The number of drones is recorded as the official traffic flow. Step S502: Calculate the first indicator – network congestion risk. The calculation formula is as follows: Step S503: Calculate the second indicator – planned load balance. The calculation formula is as follows: ;in, Indicates task The length of the planned time window and its minimum required flight time The difference, Indicates task The planned flight duration.

[0015] Preferably, step S7 determines and updates the feedback adjustment strategy, specifically including: step S701, if step S6 determines that... ,in If the first threshold is met, then the first type of regulation strategy is generated: identification. All For each such cell, generate a demand splitting suggestion weight matrix. Its dimensions are , of which elements This indicates a suggestion regarding spatial units. For destination, from space unit The adjustment weights applied to the forecast demand The values ​​of satisfy: ; to matrix The spatiotemporal graph neural network prediction model fed back to step S1 is used to adjust the predicted values ​​of relevant OD pairs in the next round of prediction; step S702, if step S6 determines that... ,in If the second threshold is met, then a second type of regulation strategy is generated: In the middle, identify Calculation in progress Minimum task The main time period During this period Within, the traffic status weights of some spatial grids in non-core areas. Temporarily multiply by a relaxation factor less than 1 Step S703: After executing the adjustment strategy determined in step S701 or step S702, update the forecast data or update the... Feedback is sent to step S4, triggering a new round of planning.

[0016] Preferably, step S8 specifically includes: step S801, receiving a logistics order that arrives in real time. Transform it into an immediate planning task. Including the start and end points and latest delivery time Step S802: Obtain the set of statuses of all tasks for which instructions have been issued but not yet completed. ; Each active task The spatiotemporal grid cells corresponding to the air routes occupied during its planned execution period The area is marked as a temporary lockdown area, that is, in The passage status weights of these units Set to a maximum value Step S803: In the global baseline route plan, filter out the spatiotemporal range of all planned paths and... Origin and end points and time windows There are overlapping tasks, which constitute a set of tasks to be protected. ,in This represents the current system moment at which dynamic reprogramming is performed; aiming to minimize the impact on... Total increase in planned delivery time for all tasks To optimize the objective, among which The calculation formula is: ;in, Indicates the originally scheduled delivery time. Indicates the delivery time of the new plan; after updating the lockdown area. Under constraints, for Search for feasible insertion paths and times; Step S804: Output the updated output containing... The executable route instruction sequence.

[0017] Preferably, the method further includes step S9: collecting the actual task completion time generated during the actual operation in step S8. With planned time deviation ( And the actual traffic volume of airspace units obtained through monitoring data. Used for periodic calibration of the first threshold Second threshold The calibration principle is to make step S6 trigger the feedback adjustment frequency. Maintain within the preset reasonable range Inside, and These represent the preset minimum and maximum reasonable trigger frequencies, respectively.

[0018] This invention also provides a low-altitude logistics dynamic route planning system driven by demand forecasting, comprising: a demand forecasting unit configured to acquire historical logistics order data, urban dynamic information data, and environmental data, and input the acquired data into a pre-trained spatiotemporal graph neural network prediction model to output an OD demand tensor for the future target time period; and a set of virtual tasks generated based on the OD demand tensor, wherein each virtual task includes an origin spatial unit, a destination spatial unit, and expected takeoff and arrival time attributes; and an airspace state modeling unit configured to initialize... A four-dimensional spatiotemporal grid model is used; real-time meteorological data, airspace control notices, and automatic dependent surveillance broadcast (ADS-B) data are integrated to dynamically calculate the traffic status weight for each spatiotemporal grid unit, constructing a dynamic airspace status model; a global planning unit is configured to receive a set of virtual tasks and the dynamic airspace status model, map the origin and destination of virtual tasks to the nearest spatial grid node, construct an optimization objective function including total flight distance cost, total delay penalty cost, and estimated congestion cost, and solve for minimizing the optimization objective function using a heuristic search algorithm under the constraints of the dynamic airspace status model, outputting a preliminary route plan; an evaluation feedback unit is configured to analyze the preliminary route. The plan calculates a first indicator reflecting the future network congestion risk and a second indicator reflecting the planned load balance; it determines whether the first indicator exceeds a first threshold or the second indicator falls below a second threshold, and if the determination is yes, it generates a feedback adjustment strategy based on the comparison result. The feedback adjustment strategy includes outputting a demand diversion suggestion weight matrix to the demand forecasting unit to adjust the OD forecast value, or outputting a relaxation factor to the airspace state modeling unit to adjust the traffic status weights of non-core area grid units, triggering the global planning unit to execute a new round of planning, until the determination result is no, at which point a global baseline route plan is determined; a real-time replanning unit is configured to... The real-time operation phase receives real-time logistics orders and transforms them into instant planning tasks. It acquires the current set of incomplete task statuses and marks the spatiotemporal grid cells it plans to occupy as temporary blockade areas. Under the constraints of the updated dynamic airspace state model, with the optimization objective of minimizing the total increase in delay for planned tasks, it searches for feasible insertion paths and times for instant planning tasks and outputs an updated sequence of executable route instructions. The parameter calibration unit is configured to collect the deviation between the actual completion time and the planned time of the task and the actual traffic data of the airspace cells, and periodically calibrates the first threshold and the second threshold to keep the feedback adjustment trigger frequency within a preset reasonable range.

[0019] Therefore, the present invention adopts the above-mentioned low-altitude logistics dynamic route planning method and system based on demand forecasting, and the beneficial technical effects are as follows: (1) The present invention combines demand forecasting of spatiotemporal graph neural network with global path planning based on virtual task set, realizing the transformation from passive response to active scheduling. This scheme uses the virtual tasks generated by prediction to pre-plan before the actual arrival of the order, fundamentally solving the problem of local airspace congestion during peak periods caused by the response lag in the real-time reactive planning in the prior art; at the same time, the global planning module not only considers distance and delay when optimizing, but also incorporates the estimated congestion cost based on dynamic airspace status, so that the generated baseline route plan has the inherent ability to guide traffic, thereby improving the overall air traffic efficiency and stability of the network.

[0020] (2) This invention establishes an evaluation feedback mechanism based on network congestion risk and planned load balance, and uses the evaluation results to dynamically adjust the prediction model or airspace constraints, forming a closed-loop optimization system. This collaborative mechanism enables the system to proactively guide the diversion of relevant OD demand predictions or temporarily expand non-core airspace resources when the preliminary planning results show that a specific airspace unit is at risk of overload. This not only solves the problem of static planning being rigid and unable to adapt to dynamic needs, but also achieves global, adaptive, and collaborative optimization of airspace resources and task load through the linkage of various modules (prediction, planning, evaluation, and feedback) within the system, avoiding systemic efficiency losses caused by the superposition of local optimal decisions.

[0021] (3) This invention effectively improves the robustness of the system in dealing with uncertainties by constructing a dynamic airspace state model that integrates real-time information and performing real-time replanning under the constraints of this model. During the real-time operation phase, the system uses the global baseline plan as the adjustment baseline, marks occupied airways as temporary blockade areas to ensure the execution of tasks, and inserts new tasks with the goal of minimizing the total delay increment of planned tasks. This combination of technical features ensures that the system can respond quickly to real-time disturbances such as sudden orders and weather changes, while maximizing the maintenance of the integrity of existing plans and the overall performance of the system, thereby solving the core defects of poor dynamic environment adaptability and lack of cross-time coordination in resource allocation in the prior art. Attached Figure Description

[0022] Figure 1 is a flowchart of the low-altitude logistics dynamic route planning method based on demand forecasting of the present invention; Figure 2 is an architecture diagram of the low-altitude logistics dynamic route planning system based on demand forecasting of the present invention; Figure 3 is a schematic diagram of the global route planning and evaluation feedback closed loop. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0025] As shown in Figure 1, the first embodiment of the low-altitude logistics dynamic route planning method based on demand forecasting includes the following steps: Step S1: Obtain low-altitude logistics spatiotemporal demand forecast data for future target time periods.

[0026] This step aims to accurately predict the demand for low-altitude logistics transportation between different areas of the city within a future period (e.g., the next 2 hours).

[0027] Step S101: Obtain historical logistics order data, city dynamic information data, and environmental data.

[0028] Historical logistics order data: Order records for the past 30 days were obtained from a commercial low-altitude logistics operation platform. Each record includes: the sender's grid number (Origin), the recipient's grid number (Destination), the weight of the goods (kg), the user's order timestamp, and the system's promised delivery timestamp.

[0029] Urban dynamic information: Access the following information through the city's big data platform: Regional population heat map: Population density heat map with a 500-meter grid size at 15-minute intervals.

[0030] Points of Interest (POI) Distribution: Includes the location and category weight of POIs in categories such as business centers, residential areas, office buildings, and industrial parks.

[0031] Event Calendar: Information on events that may affect logistics demand, such as sporting events, large-scale performances, and traffic control measures, within the next 48 hours.

[0032] Environmental data: Meteorological information such as temperature, wind speed, precipitation, and visibility for the same period in history (the same month, day, and time period in the past three years) is obtained from the meteorological department.

[0033] Step S102: Data input and model prediction.

[0034] The multi-source heterogeneous data obtained in step S101 is cleaned, aligned, and feature-engineered before being input into a pre-trained spatiotemporal graph neural network (ST-GNN) model. This model takes historical demand sequences, urban dynamics, and environmental features from the past 24 hours as input. The model structure includes graph convolutional layers (for capturing spatial dependencies), temporal convolutional layers (for capturing temporal dependencies), and an attention mechanism (for focusing on key spatiotemporal nodes). The output layer corresponds to 12 time slices (10 minutes each) and 200 spatial units (1 km) within the next 2 hours. (1km grid).

[0035] Step S103: Receive the prediction results.

[0036] The model outputs a future OD demand tensor with dimensions [12, 200, 200]. .in, Indicates the prediction in the 1st Within a time slice (e.g., the third time slice, i.e., the next 20-30 minutes), from the spatial unit (e.g., grid ID101) to spatial unit The logistics demand for grid ID 205 is 5.2 standard transport units (e.g., corresponding to 52kg of goods, assuming a standard unit of 10kg). This tensor This will be used as input for subsequent steps.

[0037] Step S2: Based on the spatiotemporal demand prediction data, generate a set of virtual tasks for pre-planning.

[0038] Step S201: Traverse tensors Each non-zero element in the set. Parse out... Then it is determined as a predicted logistics flow direction: from grid 101 to grid 205, the predicted transport volume. Related time slices .

[0039] Step S202: For each logistics flow direction and its transport volume Based on the preset single-mission carrying capacity limit (Set to 1.0, i.e., 1 standard transport unit), split it into One virtual task, among which This represents the floor function; it is the first floor function after splitting. A virtual task Calculate its expected takeoff time With expected arrival time The calculation formula is as follows: ; ;in, and Representing time slices The start time and duration, Indicates from spatial unit arrive The preset average service duration.

[0040] Assume that time slice 3 starts at 10:20 and lasts for 10 minutes.

[0041] The preset average service time (including take-off, landing, loading and unloading) from grid 101 to 205 is 25 minutes.

[0042] This method smoothly distributes the concentrated demand of 5.2 across 10-minute time slices, generating 6 virtual missions with different expected take-off and landing times.

[0043] Step S203: Summarize the set of virtual tasks.

[0044] All generated virtual tasks are aggregated into a virtual task set. Each of the tasks Includes attributes , Indicates task The initial space unit, Indicates task The target space unit, Indicates task Expected takeoff time Indicates task Expected arrival time For task indexing, This represents the total number of virtual tasks. As input for step S4.

[0045] Step S3: Based on the spatiotemporal range covered by the spatiotemporal demand prediction data and combined with real-time airspace information sources, construct a dynamic airspace state model.

[0046] Constructing a dynamic spatial state model specifically includes: step S301, initializing a four-dimensional spatiotemporal grid.

[0047] Based on the geographical area covered by the spatiotemporal demand forecast data in step S1 and future target period Initialize a four-dimensional spatiotemporal grid model, where, and These represent the minimum and maximum latitudes of the region, respectively. and These represent the minimum and maximum longitudes of the region, respectively. and These represent the start and end times of the future target time period, respectively; the four-dimensional spatiotemporal grid model is spatially divided into... The cells are divided into time segments. (24) time periods, each four-dimensional unit is denoted as ,in Indicates the spatial row index. Indicates a spatial column index. Indicates a time index.

[0048] Step S302: Access the real-time data stream.

[0049] Real-time meteorological data: Access weather radar and ground station data to obtain real-time wind speed, rainfall intensity, and visibility for each grid.

[0050] Notification of Airspace Control (NOTAM): Analyzes and accesses information such as temporary airspace restrictions and no-fly zone updates issued by civil aviation authorities.

[0051] Automatic Dependent Surveillance-Broadcast (ADS-B) data: Receives real-time position, altitude, and speed information for all aircraft (including drones and other aircraft) in the airspace.

[0052] Step S303, for each four-dimensional unit Dynamically calculate its traffic status weight The calculation formula is as follows: ;in, This indicates the weather impact weights mapped from real-time meteorological data. For example, if wind speed < 5 m / s, no precipitation, and visibility > 5 km, then... =1.0 (optimal); if moderate rain occurs, then =2.5 (increased travel costs). This represents the control constraint weights obtained from the airspace control notice mapping, based on the NOTAM mapping. Normal airspace is 1.0, and temporarily restricted airspace is 10.0 (approximately prohibiting passage). This represents the real-time aircraft density calculated based on Automatic Dependent Surveillance-Broadcast (ADS-B) data, or calculated based on ADS-B data. The number of aircraft expected to appear in this grid during the time period is normalized. The higher the density, the more normal the number of aircraft. The higher the value, the more congested the traffic. , , The normalization coefficients for each term are 0.4, 0.3, and 0.3, respectively, in this embodiment.

[0053] Step S304: Define the dynamic spatial state model.

[0054] The four-dimensional spatiotemporal grid model and the passage state weights of each cell are collectively defined as the dynamic spatial state model. This model reflects the real-time ease of passage in each "spatial-temporal block" of airspace over the next two hours, serving as a constraint for route planning. Its formal representation is: ; This serves as a constraint for step S4.

[0055] Step S4: Using the virtual task set as the task set to be assigned and the dynamic airspace state model as the constraint, perform the initial global route planning to obtain the preliminary route plan.

[0056] Step S401: Mapping the start and end points of the task.

[0057] virtual task set Each task The origin and destination are mapped to the dynamic spatial state model. The nearest spatial grid node.

[0058] Step S402: Construct the optimization objective function.

[0059] Optimize objective function The calculation formula is as follows: ;in, , , These represent preset weighting coefficients, which are set to 0.3, 0.4, and 0.3 in this embodiment, respectively. This represents the total flight distance cost calculated based on the planned path length; Indicates delivery time based on the task schedule. Compared to the expected arrival time The total delay penalty cost is calculated based on the deviation. This represents the estimated cost of congestion, calculated using the following formula: ;in, This represents the statistics based on the path allocation scheme of the current iteration. The expected flow, Representing spatial grid The static capacity threshold.

[0060] Step S403: Solve and output the preliminary plan.

[0061] exist Under the constraints, a heuristic search algorithm is used to solve the optimization objective function. The minimum value is The task allocation specifies the flight path and takeoff time, and outputs a preliminary route plan. This includes the specific flight path (represented by a series of grid sequences) and planned takeoff time for each virtual mission. Solving the objective function follows existing techniques and will not be detailed here.

[0062] Step S5: Based on the preliminary route plan, calculate the first indicator reflecting the future network congestion risk and the second indicator reflecting the load balance of the plan execution.

[0063] Step S501: Statistical analysis of official traffic.

[0064] Analysis of preliminary route plans Statistics in the future Within a given time period, the plan is to utilize spatial grids. The number of drones is recorded as the official traffic flow. ,in This indicates that according to the preliminary route plan It is expected to be in the spatiotemporal grid unit The number of drones appearing in the scene.

[0065] Step S502: Calculate the first indicator – network congestion risk. The calculation formula is as follows: This indicator reflects the ratio of flow to capacity in the most congested grid in the plan; the higher the value, the higher the risk.

[0066] Step S503: Calculate the second indicator – planned load balance. The calculation formula is as follows: ;in, Indicates task The length of the planned time window and its minimum required flight time The difference, Indicates task The planned flight duration.

[0067] This indicator reflects the average tightness or looseness of the task plan. The closer the value is to 1, the more compact the plan, the heavier the load, and the less flexibility there is; the larger the value, the more buffer is reserved and the more balanced the load is.

[0068] Step S6: Determine whether the first indicator exceeds the first threshold or whether the second indicator is below the second threshold.

[0069] Step S7, as shown in Figure 3, if the judgment result of step S6 is yes, then the feedback adjustment strategy is determined based on the comparison result of the first indicator and the first threshold, and / or the comparison result of the second indicator and the second threshold, and the constraints or task set input to step S4 are updated according to the feedback adjustment strategy, and then the execution of step S4 is returned; if the judgment result is no, then the preliminary route plan output by the current step S4 is determined as the global baseline route plan.

[0070] Step S701, if step S6 determines that... ,in If the first threshold is met, then the first type of regulation strategy is generated: identification. All For each such cell, generate a demand splitting suggestion weight matrix. Its dimensions are , of which elements This indicates a suggestion regarding spatial units. (and (Corresponding) destination, from spatial unit Adjustment weights applied to the forecast demand ( ), The values ​​of satisfy: ; to matrix The spatiotemporal neural network prediction model fed back to step S1 is used to adjust the predicted values ​​of relevant OD pairs downward in the next round of prediction, thereby alleviating congestion at the source.

[0071] Step S702, if step S6 determines that... ,in If the second threshold is met, then a second type of regulation strategy is generated: In the middle, identify Calculation in progress Minimum task The main time period During this period Within, the traffic status weights of some spatial grids in non-core areas. Temporarily multiply by a relaxation factor less than 1 ( This means reducing the transit costs of these grids, encouraging planning algorithms to utilize these airspace resources, and alleviating the load pressure on the main air routes.

[0072] Step S703: After executing the adjustment strategy determined in step S701 or step S702, update the forecast data or update the... Feedback is sent to step S4, triggering a new round of planning, resulting in a new preliminary route plan, and then the evaluation is performed again in step S5. This loop iterates until the judgment result of step S6 is "no", thus outputting the optimized global baseline route plan.

[0073] Step S8: During the real-time operation phase, obtain real-time logistics orders and, in conjunction with the latest airspace status, perform dynamic route replanning and scheduling based on the global baseline route plan determined in Step S7.

[0074] Step S801: Receive real-time delivery orders. Transform it into an immediate planning task. Including the start and end points and latest delivery time .

[0075] Step S802: Obtain the set of statuses of all tasks for which instructions have been issued but not yet completed. ; Each active task The spatiotemporal grid cells corresponding to the air routes occupied during its planned execution period The area is marked as a temporary lockdown area, that is, in The passage status weights of these units Set to a maximum value .

[0076] Step S803: In the global baseline route plan, filter out the spatiotemporal range of all planned paths and Origin and end points and time windows There are overlapping tasks, which constitute a set of tasks to be protected. ,in This represents the current system moment at which dynamic reprogramming is performed; aiming to minimize the impact on... Total increase in planned delivery time for all tasks To optimize the objective, among which The calculation formula is: ;in, The original scheduled delivery time, Delivery time for the new plan; after updating the lockdown area. Under constraints, for Search for feasible insertion paths and times; Step S804: Output the updated output containing... The executable route command sequence is issued to the corresponding UAV for execution.

[0077] Step S9: Parameter self-calibration.

[0078] Collect the actual completion time of the task generated during the actual operation of step S8. With planned time deviation ( And the actual traffic volume of airspace units obtained through monitoring data. Used for periodic (every 24 hours) calibration of the first threshold. Second threshold The calibration principle is to make step S6 trigger the feedback adjustment frequency. Maintain within the preset reasonable range Within this timeframe, such as 5-10 times per day, adaptive optimization can be achieved for long-term system operation. and These represent the preset minimum and maximum reasonable trigger frequencies, respectively.

[0079] Example 2, as shown in Figure 2, describes a low-altitude logistics dynamic route planning system driven by demand forecasting. This system includes: a demand forecasting unit configured to acquire historical logistics order data, urban dynamic information data, and environmental data, and input the acquired data into a pre-trained spatiotemporal neural network prediction model to output an OD demand tensor for the future target time period; and a virtual task set generated based on the OD demand tensor, where each virtual task includes an originating spatial unit, a destination spatial unit, and expected takeoff and arrival time attributes; and an airspace state modeling unit configured to initially... The system initializes a four-dimensional spatiotemporal grid model; it integrates real-time meteorological data, airspace control notices, and automatic dependent surveillance broadcast (ADS-B) data, dynamically calculates traffic status weights for each spatiotemporal grid unit, and constructs a dynamic airspace state model; a global planning unit is configured to receive a set of virtual tasks and the dynamic airspace state model, map the origin and destination of virtual tasks to the nearest spatial grid node, construct an optimization objective function including total flight distance cost, total delay penalty cost, and estimated congestion cost, and solve for minimizing the optimization objective function using a heuristic search algorithm under the constraints of the dynamic airspace state model, outputting a preliminary route plan; an evaluation feedback unit is configured to analyze the preliminary route plan. The route planning process calculates a primary indicator reflecting the future network congestion risk and a secondary indicator reflecting the planned load balance. It determines whether the primary indicator exceeds a first threshold or the secondary indicator falls below a second threshold. If the determination is yes, a feedback adjustment strategy is generated based on the comparison results. This feedback adjustment strategy includes outputting a demand diversion suggestion weight matrix to the demand forecasting unit to adjust the OD forecast value, or outputting a relaxation factor to the airspace state modeling unit to adjust the traffic status weights of non-core area grid units, triggering the global planning unit to execute a new round of planning, until the determination result is no, at which point a global baseline route plan is determined. A real-time replanning unit is configured to... The real-time operation phase receives real-time logistics orders and transforms them into instant planning tasks. It acquires the current set of incomplete task statuses and marks the spatiotemporal grid cells it plans to occupy as temporary blockade areas. Under the constraints of the updated dynamic airspace state model, with the optimization objective of minimizing the total increase in delay for planned tasks, it searches for feasible insertion paths and times for instant planning tasks and outputs an updated sequence of executable route instructions. The parameter calibration unit is configured to collect the deviation between the actual completion time and the planned time of the task and the actual traffic data of the airspace cells, and periodically calibrates the first threshold and the second threshold to keep the feedback adjustment trigger frequency within a preset reasonable range.

[0080] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0081] Therefore, this invention adopts the aforementioned low-altitude logistics dynamic route planning method and system based on demand forecasting. By introducing logistics demand forecasting based on spatiotemporal graph neural networks, constructing a dynamic airspace state model that integrates real-time airspace information, and establishing a closed-loop optimization mechanism of prediction-planning-evaluation-feedback, it deeply integrates forward-looking demand information into route generation and scheduling decisions. This effectively solves the technical defects of existing static planning being rigid, real-time reactive planning being lagging and prone to congestion. It realizes forward-looking proactive planning, global collaborative optimization, and dynamic robust adjustment of low-altitude logistics routes, improving the overall efficiency of network operation, resource utilization, and system adaptability.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A dynamic route planning method for low-altitude logistics based on demand forecasting, characterized in that, Includes the following steps: Step S1: Obtain low-altitude logistics spatiotemporal demand forecast data for the future target time period; Step S2: Based on the spatiotemporal demand forecast data, generate a set of virtual tasks for pre-planning; Step S3: Based on the spatiotemporal range covered by the spatiotemporal demand forecast data and combined with real-time airspace information sources, construct a dynamic airspace state model; Step S4: Using the set of virtual tasks as the task set to be assigned and the dynamic airspace state model as the constraint, perform initial global route planning to obtain a preliminary route plan. Step S5: Based on the preliminary route plan, calculate the first indicator reflecting the future network congestion risk and the second indicator reflecting the load balance of the plan execution; Step S6: Determine whether the first indicator exceeds the first threshold or whether the second indicator is lower than the second threshold; Step S7: If the determination result of step S6 is yes, then determine the feedback adjustment strategy according to the comparison result of the first indicator and the first threshold, and / or the comparison result of the second indicator and the second threshold, and update the constraints or task set input to step S4 according to the feedback adjustment strategy, and then return to execute step S4. If the judgment result is negative, the preliminary route plan output in the current step S4 will be determined as the global baseline route plan. Step S8: During the real-time operation phase, obtain real-time logistics orders and, in conjunction with the latest airspace status, perform dynamic route replanning and scheduling based on the global baseline route plan determined in Step S7.

2. The low-altitude logistics dynamic route planning method based on demand forecasting as described in claim 1, characterized in that, Step S1 specifically includes: Step S101, acquiring historical logistics order data, urban dynamic information data, and environmental data; historical logistics order data includes at least historical origin-destination pairs, cargo weight, order placement time, and delivery time; urban dynamic information includes at least regional population heatmaps, point-of-interest distribution, and event calendars; environmental data includes at least historical meteorological information for the same period; Step S102, inputting the data acquired in Step S101 into a pre-trained spatiotemporal graph neural network prediction model; the spatiotemporal graph neural network prediction model takes historical sequence data as input, and its output layer corresponds to the spatiotemporal units divided within the future target time period, with the output value being the predicted logistics demand; Step S103, receiving the output of the spatiotemporal graph neural network prediction model, with dimensions of... Future OD demand tensor ,in Indicates the number of future time slices. Indicates the number of spatial units. Indicates the first Within a time slice, from spatial units To space unit The predicted freight volume.

3. The low-altitude logistics dynamic route planning method based on demand forecasting as described in claim 2, characterized in that, Step S2 specifically includes: Step S201, parsing the future OD demand tensor. For each of these non-zero elements This determines it as a predicted logistics flow direction. and its transport volume and associated time slices Step S202: For each logistics flow direction and its transport volume Based on the preset single-mission carrying capacity limit Break it down into One virtual task, among which is the floor function; is the th element after splitting. A virtual task Calculate its expected takeoff time With expected arrival time The calculation formula is as follows: ； ;in, and Time slices The start time and duration, For spatial units arrive The preset average service duration; Step S203: Summarize all generated virtual tasks into a virtual task set. Each of the tasks Includes attributes ， Indicates task The initial space unit, Indicates task The target space unit, Indicates task Expected takeoff time Indicates task Expected arrival time Indicates the task index. This indicates the total number of virtual tasks.

4. The low-altitude logistics dynamic route planning method based on demand forecasting as described in claim 3, characterized in that, Step S3 constructs a dynamic airspace state model, specifically including: Step S301, based on the geographical area involved in the spatiotemporal demand prediction data in Step S1. and future target period Initialize a four-dimensional spatiotemporal grid model, where, and These represent the minimum and maximum latitudes of the region, respectively. and These represent the minimum and maximum longitudes of the region, respectively. and These represent the start and end times of the future target time period, respectively; the four-dimensional spatiotemporal grid model is spatially divided into... The cells are divided into time segments. Each time period is denoted as (the four-dimensional unit is then defined as follows). ,in Indicates the spatial row index. Indicates a spatial column index. Indicates time index; Step S302: Access real-time meteorological data, airspace control notices, and automatic dependent surveillance broadcast data; Step S303: For each four-dimensional unit Dynamically calculate its traffic status weight The calculation formula is as follows: ;in, This indicates the weighting of weather impacts obtained by mapping real-time meteorological data. This represents the control constraint weights mapped from airspace control notices. This represents the real-time aircraft density calculated based on automatic dependent surveillance (ADS) data. 、 、 These represent the normalization coefficients of each term; Step S304: Define the four-dimensional spatiotemporal grid model and the traffic state weights of each element together as the dynamic spatial domain state model. Its formal representation is: 。 5. The low-altitude logistics dynamic route planning method based on demand forecasting as described in claim 4, characterized in that, Step S4 specifically includes: Step S401, setting up the virtual task set. Each task The origin and destination are mapped to the dynamic spatial state model. The nearest spatial grid node; Step S402: Construct the optimization objective function The calculation formula is as follows: ;in, 、 、 All represent preset weighting coefficients; This represents the total flight distance cost calculated based on the planned path length; Indicates delivery time based on the task schedule. Compared to the expected arrival time The total delay penalty cost is calculated based on the deviation. This represents the estimated cost of congestion, calculated using the following formula: ;in, This represents the statistics based on the path allocation scheme of the current iteration. The expected flow, Representing spatial grid The static capacity threshold; step S403, in Under the constraints, a heuristic search algorithm is used to solve the optimization objective function. The minimum value is The task allocation specifies the flight path and takeoff time, and outputs a preliminary route plan. 。 6. The low-altitude logistics dynamic route planning method based on demand forecasting as described in claim 5, characterized in that, Step S5 specifically includes: Step S501, analyzing the preliminary route plan. Statistics in the future Within a given time period, the plan is to utilize spatial grids. The number of drones is recorded as the official traffic flow. Step S502: Calculate the first indicator – network congestion risk. The calculation formula is as follows: Step S503: Calculate the second indicator – planned load balance. The calculation formula is as follows: ;in, Indicates task The length of the planned time window and its minimum required flight time The difference, Indicates task The planned flight duration.

7. The low-altitude logistics dynamic route planning method based on demand forecasting as described in claim 6, characterized in that, Step S7 determines and updates the feedback adjustment strategy, specifically including: Step S701, if step S6 determines that... ,in If the first threshold is met, then the first type of regulation strategy is generated: identification. All For each such cell, generate a demand splitting suggestion weight matrix. Its dimensions are , of which elements This indicates a suggestion regarding spatial units. For destination, from space unit The adjustment weights applied to the forecast demand The values ​​of satisfy: ; to matrix The spatiotemporal graph neural network prediction model fed back to step S1 is used to adjust the predicted values ​​of relevant OD pairs in the next round of prediction; step S702, if step S6 determines that... ,in If the second threshold is met, then a second type of regulation strategy is generated: In the middle, identify Calculation in progress Minimum task The main time period During this period Within, the traffic status weights of some spatial grids in non-core areas. Temporarily multiply by a relaxation factor less than 1 Step S703: After executing the adjustment strategy determined in step S701 or step S702, update the forecast data or update the... Feedback is sent to step S4, triggering a new round of planning.

8. The low-altitude logistics dynamic route planning method based on demand forecasting as described in claim 7, characterized in that, Step S8 specifically includes: Step S801, receiving real-time arriving logistics orders. Transform it into an immediate planning task. Including the start and end points and latest delivery time Step S802: Obtain the set of statuses of all tasks for which instructions have been issued but not yet completed. ; Each active task The spatiotemporal grid cells corresponding to the air routes occupied during its planned execution period The area is marked as a temporary lockdown area, that is, in The passage status weights of these units Set to a maximum value Step S803: In the global baseline route plan, filter out the spatiotemporal range of all planned paths and... Origin and end points and time windows There are overlapping tasks, which constitute a set of tasks to be protected. ,in This represents the current system moment at which dynamic reprogramming is performed; aiming to minimize the impact on... Total increase in planned delivery time for all tasks To optimize the objective, among which The calculation formula is: ;in, Indicates the originally scheduled delivery time. Indicates the delivery time of the new plan; after updating the lockdown area. Under constraints, for Search for feasible insertion paths and times; Step S804: Output the updated output containing... The executable route instruction sequence.

9. The low-altitude logistics dynamic route planning method based on demand forecasting as described in claim 1, characterized in that, The method further includes step S9: collecting the actual task completion time generated during the actual operation of step S8. With planned time deviation （ And the actual traffic volume of airspace units obtained through monitoring data. Used for periodic calibration of the first threshold Second threshold The calibration principle is to make step S6 trigger the feedback adjustment frequency. Maintain within the preset reasonable range Inside, and These represent the preset minimum and maximum reasonable trigger frequencies, respectively.

10. A low-altitude logistics dynamic route planning system based on demand forecasting, characterized in that, include: The demand forecasting unit is configured to acquire historical logistics order data, urban dynamic information data, and environmental data, and input the acquired data into a pre-trained spatiotemporal graph neural network prediction model to output the OD demand tensor for the future target period; and generate a set of virtual tasks based on the OD demand tensor, wherein each virtual task includes the starting spatial unit, the destination spatial unit, the expected take-off time, and the expected arrival time attributes. The airspace state modeling unit is configured to initialize a four-dimensional spatiotemporal grid model based on the geographical area covered by the demand forecasting unit and the future target time period. The system integrates real-time meteorological data, airspace control notices, and automatic dependent surveillance broadcast data. It dynamically calculates the traffic status weight for each spatiotemporal grid unit and constructs a dynamic airspace status model. The global planning unit is configured to receive a set of virtual tasks and the dynamic airspace status model, map the origin and destination of the virtual tasks to the nearest spatial grid node, construct an optimization objective function that includes total flight distance cost, total delay penalty cost, and estimated congestion cost, and solves the optimization objective function using a heuristic search algorithm under the constraints of the dynamic airspace status model to output a preliminary route plan. The evaluation feedback unit is configured to analyze the preliminary route plan, calculate a first indicator reflecting the future network congestion risk and a second indicator reflecting the planned load balance; determine whether the first indicator exceeds a first threshold or whether the second indicator is below a second threshold, and when the determination result is yes, generate a feedback adjustment strategy based on the comparison result. The feedback adjustment strategy includes outputting a demand diversion suggestion weight matrix to the demand forecasting unit to adjust the OD forecast value, or outputting a relaxation factor to the airspace status modeling unit to adjust the traffic status weight of non-core area grid units, triggering the global planning unit to execute a new round of planning, until the determination result is no, and then determining the global baseline route plan. The real-time replanning unit is configured to receive real-time logistics orders and convert them into immediate planning tasks during the real-time operation phase. It obtains the current set of incomplete task states and marks the spatiotemporal grid cells it plans to occupy as temporary blockade areas. Under the constraints of the updated dynamic airspace state model, with the optimization objective of minimizing the total delay increase of the planned tasks, it searches for feasible insertion paths and times for the immediate planning tasks and outputs the updated executable route instruction sequence. The parameter calibration unit is configured to collect the deviation between the actual completion time and the planned time of the task and the actual flow data of the airspace unit, and periodically calibrate the first threshold and the second threshold to keep the feedback adjustment trigger frequency within a preset reasonable range.

Images