Dynamic scheduling optimization method for low-altitude logistics distribution network

Abstract

The invention discloses a dynamic scheduling optimization method for a low-altitude logistics distribution network, and the method comprises the steps: a server side builds a multi-source sensing network through satellite remote sensing, an unmanned plane airborne sensor and ground traffic monitoring, fuses meteorological data, airspace control data and order distribution data which are collected in real time, and generates a four-dimensional space-time grid map; based on the four-dimensional space-time grid map, the server side adopts a TD3-GA hybrid intelligent algorithm to carry out path planning of the logistics distribution network; the edge calculation end generates an optimal scheduling scheme of the unmanned aerial vehicle group through a multi-objective optimization function based on the global path; and the server side performs security risk assessment on the optimal scheduling scheme by using a Bayesian network model, and dynamically adjusts a space-time routing strategy of the unmanned aerial vehicle cluster according to an assessment result. According to the method, in a large-scale unmanned aerial vehicle concurrent scheduling scene, the scheduling efficiency can be effectively improved, the response time delay is reduced, the risk prediction accuracy is improved, and the timeliness and safety of a low-altitude distribution network are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of low-altitude logistics delivery technology, and more specifically to a dynamic scheduling optimization method for low-altitude logistics delivery networks. It integrates proactive safety control and emergency response mechanisms to achieve dynamic scheduling optimization of low-altitude logistics, and is applicable to improving the timeliness and safety of large-scale urban drone delivery scenarios. Background Technology

[0002] With the continuous maturation of unmanned aerial vehicle (UAV) technology and the widespread deployment of 5G communication networks, logistics delivery based on low-altitude airspace is becoming an important direction for smart city construction and industrial upgrading. On the one hand, advanced flight platforms have advantages such as vertical take-off and landing, high maneuverability, and controllable costs, which greatly improves the efficiency of "last mile" delivery; on the other hand, the high bandwidth, low latency, and wide coverage provided by 5G networks lay a solid communication foundation for UAV swarm control, real-time monitoring, and multi-aircraft collaborative scheduling.

[0003] However, most current low-altitude logistics scheduling systems still rely primarily on static planning, mainly optimizing information such as orders, routes, and resources in advance to form fixed flight routes and schedules. While this approach can meet the needs when order volume is low and delivery areas are limited, it encounters problems such as slow response, route conflicts, and insufficient safety redundancy when facing large-scale, multi-source, heterogeneous orders and highly dynamic urban airspace environments. Furthermore, traditional static scheduling lacks the ability to rapidly replan in response to weather changes, airspace restrictions, sudden equipment failures, or emergency tasks, making it difficult to guarantee the continuity and safety of delivery missions.

[0004] Furthermore, with the expansion of delivery networks and the diversification of business models, the coupling effect between subsystems such as drone swarm formation coordination, air traffic management (UTM), battery energy consumption and range management, dynamic obstacle avoidance, and safety protection is becoming increasingly prominent. Existing technologies are mostly limited to single algorithms or single-machine simulations, failing to fully utilize emerging technologies such as edge computing, cloud-edge collaboration, artificial intelligence, and big data analytics to achieve real-time, global, and adaptive comprehensive management of scheduling optimization in dynamic environments. As a result, the overall system's resource utilization, service level, and security capabilities have not reached ideal levels.

[0005] Therefore, there is an urgent need for a dynamic scheduling optimization method based on real-time data perception and intelligent decision-making that can respond quickly to sudden orders in large-scale orders, complex airspace and variable environments, while taking into account flight safety and system robustness, so as to achieve efficient, reliable and sustainable operation of the low-altitude logistics and distribution network. Summary of the Invention

[0006] In view of this, the present invention provides a dynamic scheduling optimization method for low-altitude logistics distribution networks, which can respond to dynamic orders in real time, support large-scale multi-UAV collaboration, and has online risk assessment capabilities.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a dynamic scheduling optimization method for low-altitude logistics distribution networks, comprising the following steps:

[0009] S1. The server constructs a multi-source perception network through satellite remote sensing, UAV airborne sensors and ground traffic monitoring, and integrates real-time collected meteorological data, airspace control data and order distribution data to generate a four-dimensional spatiotemporal grid map.

[0010] S2. The server-side uses the TD3-GA hybrid intelligent algorithm for path planning of the logistics distribution network based on the four-dimensional spatiotemporal grid map; wherein, the TD3 algorithm handles dynamic disturbance events, and the genetic algorithm optimizes the global path;

[0011] S3. Based on the global path, the edge computing terminal generates the optimal scheduling scheme for the drone swarm through a multi-objective optimization function; the function includes a weighted combination of three dimensions: delivery time, energy consumption cost, and security risk.

[0012] S4. The server uses a Bayesian network model to conduct a security risk assessment of the optimal scheduling scheme and dynamically adjusts the spatiotemporal routing strategy of the drone cluster based on the assessment results.

[0013] S5. The UAV receives scheduling instructions from the server, fine-tunes its flight attitude in real time using a lightweight LSTM model, and feeds back local sensor data to the server.

[0014] Further, in step S1, the four-dimensional spatiotemporal grid map includes:

[0015] The three-dimensional geographic space is divided into 1000m×1000m×100m cubic units, and sliced ​​along the time axis at 5-minute intervals. The formula for dividing the spatiotemporal unit is as follows:

[0016] (1)

[0017] Among them, L x L y L z , n represent the longitude grid number, latitude grid number, height layer number, and time slice number, respectively; t0 is the initial timestamp; Δx=1000m, Δy=1000m, Δz=100m, Δt=5min;

[0018] The unit feature is encoded as follows:

[0019] (2)

[0020] Among them, W t P is the meteorological disturbance coefficient. a For airspace control level, O d For order density, E c This refers to the capacity of the charging station.

[0021] Furthermore, in step S2, the TD3-GA hybrid intelligent algorithm includes:

[0022] a) The state space of the TD3 algorithm is defined as follows:

[0023] (3)

[0024] in, Let B be the eigenvector of the unit at time t. remain Q represents the remaining battery power of the drone. pending This is a queue of orders awaiting processing.

[0025] b) The chromosome encoding of the genetic algorithm uses a three-dimensional waypoint sequence:

[0026] (4)

[0027] Where x, y, and z are geographic coordinates, and t is the arrival time;

[0028] Fitness function:

[0029] (5)

[0030] T total E represents the total estimated time for the route, in minutes. cost Energy cost is calculated as electricity consumption × electricity price, unit: yuan; R risk The risk coefficient is a value between 0 and 1, calculated based on historical accident data; α is the time weight, β is the energy cost weight, and γ is the risk penalty weight.

[0031] Furthermore, the reward function of the TD3 algorithm is:

[0032] (6)

[0033] Where, ΔQ processed E represents the change in the number of orders processed per unit of time. consumed Energy consumption value, unit is kilowatt-hour; I violate D is a counter for the number of airspace rule violations. deviation This represents the path deviation.

[0034] Furthermore, the genetic algorithm includes innovation operators:

[0035] a) Dynamic crossover operator:

[0036] (7)

[0037] p cross For dynamic crossover probability, This represents the ratio of the current iteration count to the maximum iteration count.

[0038] b) Adaptive mutation operator:

[0039] (8)

[0040] p mut For adaptive variability rate, Fitness avg The average fitness value of the population. min The minimum fitness value of the population, Fitness max This represents the maximum fitness value of the population.

[0041] c) Elite preservation strategy: In each generation of evolution, a number of individuals with the highest fitness are retained and directly enter the next generation to ensure the inheritance of excellent genes.

[0042] Furthermore, in step S3, the multi-objective optimization function is:

[0043] (9)

[0044] In the formula, T i For the delivery time of the i-th drone, T max The maximum allowable time is represented by w1, w2, and w3, which are weighting coefficients; E i E represents the energy consumption cost of the i-th drone. budget The maximum energy consumption allowed for a single drone to perform a single delivery task;

[0045] The constraints include:

[0046] (10)

[0047] (11)

[0048] (12)

[0049] In the formula, The total service time for processing the j-th order for the i-th drone; Let i be the latest allowed delivery completion time for the i-th order; B represents the real-time remaining battery percentage of the i-th drone. safe The minimum safe return-to-home battery threshold; v i Let v be the three-dimensional real-time velocity vector of UAV i. nominal The nominal cruise speed vector, Δv max This represents the maximum permissible speed deviation.

[0050] Further, in step S4, the Bayesian network model includes:

[0051] Risk points:

[0052] (13)

[0053] W h W is the hidden layer weight matrix; f The feature fusion matrix; is the input 4-dimensional feature vector; tanh is the hyperbolic tangent activation function; softmax is the normalized exponential function;

[0054] The dynamic adjustment strategy is as follows:

[0055] When P(Risk>0.7)>0.5, path replanning is triggered:

[0056] (14)

[0057] Where ⊕ represents the heading angle correction operation, Δθ is the base heading angle correction amount; k is the range attenuation coefficient, and D risk Euclidean distance for risk points.

[0058] Furthermore, when a sudden change in weather or a change in airspace control level is detected that exceeds a threshold, the four-dimensional spatiotemporal grid map automatically reduces the time slice interval to 2 minutes and re-divides the spatial grid density to 200m×200m×20m until the environment stabilizes.

[0059] Furthermore, in step S5, the lightweight LSTM model is implemented through the following steps:

[0060] a) Use channel pruning techniques to remove redundant neurons;

[0061] b) Perform 8-bit fixed-point quantization on the weight matrix;

[0062] c) Fine-tune model parameters online using historical UAV flight data.

[0063] As can be seen from the above technical solution, compared with the prior art, the present invention has the following technical advantages:

[0064] This invention can effectively improve scheduling efficiency, reduce response latency, and improve the accuracy of risk prediction in large-scale drone concurrent scheduling scenarios, thereby significantly improving the timeliness and safety of low-altitude delivery networks.

[0065] 1. Real-time environmental perception: By integrating multi-source data to construct a four-dimensional spatiotemporal grid map, it breaks through the perception limitations of traditional static planning, accurately captures dynamic interference factors such as sudden weather changes and airspace restrictions, and significantly improves response speed;

[0066] 2. Intelligent collaborative decision-making: The TD3-GA hybrid algorithm combines dynamic perturbation handling (TD3) and global path optimization (GA) to solve the path conflict problem of large-scale UAV swarms, improve path planning efficiency, and reduce energy consumption;

[0067] 3. Proactive security defense: Based on Bayesian network risk prediction and dynamic routing adjustment, it further reduces the incidence of sudden incidents and improves security redundancy;

[0068] 4. Efficient resource utilization: Multi-objective optimization at the edge computing end and lightweight LSTM closed-loop control at the drone end enable load balancing and real-time attitude fine-tuning of drone swarms, thereby improving the overall resource utilization of the delivery network.

[0069] This invention effectively overcomes the shortcomings of existing technologies, such as poor dynamic adaptability, weak multi-machine collaboration, and insufficient safety assurance, through a closed-loop control system of environmental perception, global planning, edge scheduling, risk assessment, and execution feedback. It provides a highly reliable and flexible intelligent scheduling solution for urban low-altitude logistics. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0071] Figure 1 This is a schematic diagram of the architecture of the low-altitude logistics distribution network provided by the present invention.

[0072] Figure 2 The flowchart illustrates the dynamic scheduling optimization method for low-altitude logistics distribution networks provided by this invention.

[0073] Figure 3 The flowchart for the dynamic update of the four-dimensional spatiotemporal grid provided by this invention.

[0074] Figure 4 This is a diagram of the TD3-GA hybrid algorithm architecture provided by the present invention. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] This invention discloses a dynamic scheduling optimization method for low-altitude logistics distribution networks, referring to... Figure 1 As shown, this invention involves the interaction between cloud servers, edge computing nodes, and drone swarms; it also involves the data acquisition layer collecting data from weather radar and order management systems. Specifically, the cloud server runs the TD3-GA algorithm and Bayesian networks to manage global path planning; the edge nodes deploy multi-objective optimization modules to handle local scheduling; and the drone terminals perform lightweight control and provide real-time data feedback. This invention employs a layered computing architecture, and the collaborative processing of the cloud, edge, and terminal reduces system response time to one-fifth of that of traditional centralized architectures.

[0077] This scheduling optimization method refers to... Figure 2 As shown, the process includes the following steps S1 to S5:

[0078] Spatiotemporal coupling mechanism: The four-dimensional spatiotemporal grid map established in step S1 ensures that the decisions in S2-S4 are consistent over time.

[0079] Computational load distribution: S2 global planning (cloud), S3 local scheduling (edge), and S5 real-time control (endpoint) constitute a three-tier computing architecture.

[0080] Safety control loop: The risk assessment results of S4 are used to modify the constraints of S3 through formula (14) to form a feedforward-feedback composite control;

[0081] Data-driven iteration: Flight data fed back from S5 continuously optimizes the grid encoder of S1 and the TD3 policy network of S2.

[0082] Specifically as follows:

[0083] S1. The server constructs a multi-source perception network through satellite remote sensing, UAV airborne sensors and ground traffic monitoring, and integrates real-time collected meteorological data, airspace control data and order distribution data to generate a four-dimensional spatiotemporal grid map.

[0084] In step S1, a four-dimensional spatiotemporal grid map is constructed:

[0085] 1) Data Acquisition: The server acquires large-scale geographic information (macro-topography, weather cloud images, etc.) through satellite remote sensing, while UAV airborne sensors (LiDAR, optical cameras, infrared detectors, used to acquire obstacle and micro-meteorological information in real time) acquire local environmental data, and the ground traffic monitoring system acquires ground traffic information.

[0086] 2) Data Fusion: The collected meteorological data (wind speed, wind direction, temperature, humidity, etc.), airspace control data (restricted flight areas, flight altitude restrictions, etc.) and order distribution data (order quantity, location, time, etc.) are fused after multi-source time synchronization and spatial registration by applying Kalman filtering.

[0087] 3) Spatiotemporal unit division: The three-dimensional geographic space is divided into 1000m × 1000m × 100m cubic units, and sliced ​​along the time axis at 5-minute intervals to form a four-dimensional spatiotemporal grid map. The feature code of each spatiotemporal unit includes information such as meteorological disturbance coefficient, airspace control level, order density, and charging station capacity.

[0088] The four-dimensional spatiotemporal grid map includes:

[0089] (1)

[0090] Among them, L x L y L z , n represent the longitude grid number, latitude grid number, height layer number, and time slice number, respectively; t0 is the initial timestamp; Δx=1000m, Δy=1000m, Δz=100m, Δt=5min;

[0091] The unit feature is encoded as follows:

[0092] (2)

[0093] Among them, W t The meteorological disturbance coefficient is calculated using meteorological radar data, and the formula is as follows:

[0094] ;

[0095] P a Airspace control levels (Levels 1-5, Level 1 is open, Level 5 is no-fly zone);

[0096] O d Order density (unit: order / cube unit), updated in real time through the order management system;

[0097] E cThe charging station capacity (unit: drone charging times / hour) is fed back from the status of the ground charging station.

[0098] 4) Dynamic Updates: Based on real-time data, the feature codes of spatiotemporal units are dynamically updated. For example, when there are sudden weather changes or changes in airspace control levels, the time slice interval is automatically reduced to 2 minutes, and the spatial grid density is re-divided to 200m × 200m × 20m. Bayesian network evaluation results are then injected until the environment stabilizes. (See the overall description for reference.) Figure 3 As shown.

[0099] In this embodiment, a wavelet transform-CNN hybrid encoder can be used to reduce the dimensionality of the spatiotemporal unit feature vector, compressing the original 4-dimensional feature vector into a 32-dimensional latent representation, improving storage efficiency by 76% while maintaining feature fidelity ≥92%.

[0100] S2. The server-side uses the TD3-GA hybrid intelligent algorithm to plan the path of the logistics distribution network based on the four-dimensional spatiotemporal grid map; wherein, the TD3 algorithm handles dynamic disturbance events and the genetic algorithm optimizes the global path;

[0101] The hybrid algorithm process includes:

[0102] 1) In the initial stage, a genetic algorithm is used to generate an initial path planning scheme.

[0103] 2) The TD3 algorithm dynamically adjusts the path based on real-time environment and order changes to handle unexpected situations.

[0104] 3) The genetic algorithm optimizes the global path based on the adjustment results of the TD3 algorithm, thereby improving efficiency and reducing energy consumption.

[0105] The TD3-GA hybrid architecture achieves dual guarantees of "global optimization and dynamic response," such as... Figure 4 As shown:

[0106] The state space of the TD3 algorithm is defined as follows:

[0107] (3)

[0108] in, Let B be the eigenvector of the unit at time t. remain Q represents the remaining battery level of the drone (0-100%). pending This is a queue of orders awaiting processing.

[0109] The chromosome encoding in the genetic algorithm uses a three-dimensional waypoint sequence:

[0110] (4)

[0111] Where x, y, and z are geographic coordinates, and t is the arrival time;

[0112] The fitness function considers factors such as delivery time, energy costs, and risk coefficients, and the weighting coefficients can be adjusted according to actual conditions.

[0113] (5)

[0114] T total E represents the total estimated time for the route, in minutes. cost Energy cost is calculated as electricity consumption × electricity price, unit: yuan; R risk The risk coefficient is calculated based on historical accident data, ranging from 0 to 1; α is the time weight, for example, 0.7; β is the energy cost weight, for example, 0.2; and γ is the risk penalty weight, for example, 0.1.

[0115] That is, Figure 4 As shown, the real-time features of the four-dimensional spatiotemporal grid map are injected into the TD3-GA hybrid algorithm for collaborative optimization and strategy fusion, outputting three-dimensional coordinates and timestamps to obtain a spatiotemporal joint path scheme.

[0116] The reward function of the TD3 algorithm described above is designed based on factors such as the number of orders processed, energy consumption, number of airspace rule violations, and path deviation, encouraging drones to complete tasks efficiently and safely. (See below.)

[0117] (6)

[0118] Where, ΔQ processed The change in the number of orders processed per unit of time (e.g., +10 points for processing 1 order every 5 minutes); E consumed This is the energy consumption value, measured in kilowatt-hours (kWh). For example, 0.5 points are deducted for every 1 kWh consumed. violate This is a counter for violations of airspace rules, such as incurring a 20-point deduction for entering a no-fly zone; D deviation This represents the path deviation.

[0119] Innovative operators in genetic algorithms:

[0120] a) Dynamic crossover operator: The crossover probability is dynamically adjusted based on the number of iterations to avoid getting trapped in local optima.

[0121] (7)

[0122] p cross For dynamic crossover probability, This represents the ratio of the current iteration count to the maximum iteration count; as the number of iterations increases, the crossover probability decreases, thus avoiding overexploration.

[0123] b) An adaptive mutation operator dynamically adjusts the mutation rate based on the population fitness distribution to maintain population diversity:

[0124] (8)

[0125] p mut For adaptive variability rate, Fitness avg The average fitness value of the population. min The minimum fitness value of the population, Fitness max This represents the maximum fitness value of the population. When the fitness difference between populations is large, the mutation rate is reduced to retain superior individuals; when the difference is small, the mutation rate is increased to increase diversity.

[0126] c) Elite Preservation Strategy: In each generation of evolution, a few individuals with the highest fitness are retained and directly enter the next generation to ensure the inheritance of superior genes. Retaining the individuals with the highest fitness accelerates the convergence speed. For example, the top 10% of high-fitness individuals in each generation are retained and directly enter the next generation.

[0127] S3. Based on the global path, the edge computing terminal generates the optimal scheduling scheme for the drone swarm through a multi-objective optimization function; the function includes a weighted combination of three dimensions: delivery time, energy consumption cost, and security risk.

[0128] The multi-objective optimization function considers three dimensions: delivery time, energy cost, and safety risk, and sets weight coefficients according to the actual situation:

[0129] (9)

[0130] In the formula, T i For the delivery time of the i-th drone, T max The maximum allowable time is defined by w1, w2, and w3, which are weighting coefficients, for example, 0.5 / 0.3 / 0.2 respectively. w1 is adjusted based on the urgency of real-time orders; w2 is adjusted based on peak and off-peak electricity prices; and w3 is adjusted based on historical incident frequency. E i E represents the energy consumption cost of the i-th drone. budget The maximum energy consumption allowed for a single drone to perform a single delivery task;

[0131] The constraints include:

[0132] (10)

[0133] (11)

[0134] (12)

[0135] In the formula, The total service time for processing the j-th order for the i-th drone; Let i be the latest allowed delivery completion time for the i-th order; B represents the real-time remaining battery percentage of the i-th drone. safe The minimum safe return battery level (e.g., 20% for a safe return); v i Let v be the three-dimensional real-time velocity vector of UAV i. nominal The nominal cruise speed vector, Δv max This represents the maximum permissible speed deviation.

[0136] At the edge computing end, a multi-objective optimization function is used to find the optimal scheduling scheme under the premise of satisfying the constraints, so as to achieve load balancing and efficient resource utilization of the drone swarm.

[0137] S4. The server uses a Bayesian network model to conduct a security risk assessment of the optimal scheduling scheme and dynamically adjusts the spatiotemporal routing strategy of the drone cluster based on the assessment results.

[0138] Bayesian network models can improve the spatiotemporal predictive capabilities of risk assessment. Input features: a feature vector F of a four-dimensional spatiotemporal grid (including meteorological, airspace, order density, and charging station capacity, etc.); risk nodes:

[0139] (13)

[0140] W h W is the hidden layer weight matrix (trained using historical accident data); f This is a feature fusion matrix (compressing multidimensional features into risk-related dimensions). is the input 4-dimensional feature vector; tanh is the hyperbolic tangent activation function; softmax is the normalized exponential function;

[0141] The dynamic adjustment strategy is as follows:

[0142] When P(Risk>0.7)>0.5, path replanning is triggered:

[0143] (14)

[0144] Where ⊕ represents the heading angle correction operation, Δθ is the basic heading angle correction amount (e.g., ±5° correction when the deviation from the risk point is 100m); k is the distance attenuation coefficient, k=5°×risk level; D risk The Euclidean distance to the risk point. (The above...) .

[0145] The Bayesian network model establishes a network structure that includes risk nodes and feature nodes, and trains parameters based on historical accident data and expert experience.

[0146] In this step, the drone scheduling plan is used as input, and a Bayesian network model is used to calculate the probability of safety risks. When the risk probability exceeds a preset threshold, path replanning is triggered, and the heading angle is corrected based on the location and distance of the risk point to ensure flight safety.

[0147] For example, when any of the following conditions are met: sudden wind speed change (ΔV ≥ 3 m / s within 10 seconds); sudden obstacle detection (radar detects a new obstacle within 20 m); or battery warning (SOC ≤ 25% and distance to the target point > 500 m), the TD3 strategy network is activated to generate an emergency obstacle avoidance path. In other words, when the risk probability exceeds a threshold, re-optimization or a return-to-base approach is triggered.

[0148] The server uses a Bayesian network model to assess the security risks of the optimal scheduling scheme. When a sudden change in weather or abnormal airspace control is detected, the following emergency response mechanism is automatically triggered:

[0149] 1. Proactive prevention and control: Based on historical accident data and real-time characteristics (Wt, Pa, Od), the risk probability P(Risk) is dynamically calculated. If P(Risk>0.7)>0.5, a three-level early warning signal (yellow / orange / red) is generated and pushed to the management platform.

[0150] 2. Emergency dispatch: Using the heading angle correction formula Path new =Path old ⊕(Δθ*e −kDrisk Dynamically avoid high-risk areas, while simultaneously adjusting the risk coefficient R. risk Feedback is sent to the edge computing end, and the multi-objective optimization weight w3 is adjusted accordingly (Formula 9).

[0151] 3. Disaster Response: When the airspace control level is raised to Level 5 (no-fly zone) or the weather disturbance factor W... t When the value exceeds 1.2, the nearest return-to-base strategy will be forcibly activated, and the backup charging station capacity E will be activated. c .

[0152] The risk assessment result of step S4 is used to correct the path constraint of S3 using formula (14), and Rrisk is injected as a feedback quantity into the weights of the multi-objective optimization function (formula 9):

[0153] w3′ (w3′ = w3⋅(1+k⋅P(Risk))) forms a dynamic closed loop of “risk assessment → parameter adjustment → path regeneration”.

[0154] S5. The UAV receives scheduling instructions from the server, fine-tunes its flight attitude in real time using a lightweight LSTM model, and feeds back local sensor data to the server.

[0155] Among them, the lightweight LSTM model:

[0156] Channel pruning is used to remove redundant neurons and reduce model size. For example, removing neurons with activation values ​​below a threshold (such as 0.1) in an LSTM reduces computation by 40%.

[0157] The weight matrix is ​​quantized to 8-bit fixed-point to reduce the computational cost of the model; the weight matrix is ​​compressed from 32-bit floating-point to 8-bit fixed-point, reducing memory usage by 75%.

[0158] The model parameters are fine-tuned online using historical flight data of the drone, and local sensor data (such as gyroscope and barometer) is uploaded to the server every 10 seconds to improve the model's adaptability and robustness.

[0159] Real-time control: The UAV receives scheduling commands from the server and fine-tunes its flight attitude in real time using a lightweight LSTM model to ensure it flies along the planned path. For example, inputs include current attitude data (heading angle, altitude, and speed); outputs include fine-tuning commands (e.g., adjusting pitch angle ±2°, altitude layer ±1 layer); and execution is performed every 0.5 seconds to ensure flight stability in dynamic environments.

[0160] Data feedback: The drone feeds back local sensor data (position, speed, attitude, etc.) to the server in real time, providing data support for subsequent path planning and risk assessment.

[0161] Through the above steps, this invention achieves dynamic scheduling optimization of low-altitude logistics delivery networks, effectively improving delivery efficiency, reducing energy consumption costs, enhancing safety, and enabling collaborative operations of drone swarms. It significantly improves scheduling response efficiency and path optimization, supports rapid replanning and task insertion in response to emergencies, and provides a highly reliable and flexible intelligent scheduling solution for urban low-altitude logistics. It can adapt to large-scale task scheduling at the city or regional level.

[0162] Taking urban logistics and delivery as an example, let's say we deploy 200 logistics drones in a core urban area, serving a radius of 20 kilometers. The server updates the four-dimensional spatiotemporal grid every 5 minutes, dynamically adjusting grid parameters (e.g., densifying the grid to 500m×500m in areas with high order density). Edge computing nodes optimize the scheduling scheme based on real-time traffic data, with weighting coefficients w1, w2, and w3 set to 0.6:0.2:0.2 during peak hours and adjusted to 0.4:0.3:0.3 at night.

[0163] When the Bayesian network detects a risk probability exceeding 0.7 in a certain area, it triggers a heading angle correction (Δθ=10°) and replans the path using the TD3 algorithm. The UAV-side LSTM model adjusts the pitch and roll angles at a frequency of 20Hz, with the error controlled within ±0.5°.

[0164] Algorithm parameter configuration:

[0165] TD3 algorithm: learning rate α = 0.001, discount factor γ = 0.99;

[0166] Genetic algorithm: population size 200, maximum number of iterations 100; crossover rate 0.7, mutation rate 0.05.

[0167] In this embodiment, a perception network is constructed using satellites, drones, and ground monitoring to integrate meteorological, airspace, and order data in real time, improving the accuracy of environmental perception. The TD3-GA algorithm is used to collaboratively process dynamic disturbances and optimize the overall system, balancing real-time response and long-term efficiency. Additionally, edge collaborative scheduling generates the optimal drone swarm solution based on multi-objective optimization, balancing delivery timeliness, energy consumption, and safety. Finally, a Bayesian network is used to predict risks and adjust routes in real time, enhancing system security and robustness. Closed-loop control optimization is achieved: drones adjust their flight in real time using a lightweight LSTM, and feedback data drives strategy iteration, forming an adaptive closed loop.

[0168] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0169] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A dynamic scheduling optimization method for a low-altitude logistics distribution network, characterized in that, Includes the following steps: S1. The server constructs a multi-source perception network through satellite remote sensing, UAV airborne sensors and ground traffic monitoring, and integrates real-time collected meteorological data, airspace control data and order distribution data to generate a four-dimensional spatiotemporal grid map. S2. The server-side uses the TD3-GA hybrid intelligent algorithm for path planning of the logistics distribution network based on the four-dimensional spatiotemporal grid map; wherein, the TD3 algorithm handles dynamic disturbance events, and the genetic algorithm optimizes the global path; S3. Based on the global path, the edge computing terminal generates the optimal scheduling scheme for the drone swarm through a multi-objective optimization function; the function includes a weighted combination of three dimensions: delivery time, energy consumption cost, and security risk. S4. The server uses a Bayesian network model to conduct a security risk assessment of the optimal scheduling scheme and dynamically adjusts the spatiotemporal routing strategy of the drone cluster based on the assessment results. S5. The UAV receives scheduling instructions from the server, fine-tunes its flight attitude in real time using a lightweight LSTM model, and feeds back local sensor data to the server.

2. The dynamic scheduling optimization method for a low-altitude logistics distribution network according to claim 1, characterized in that, In step S1, the four-dimensional spatiotemporal grid map includes: The three-dimensional geographic space is divided into 1000m×1000m×100m cubic units, and sliced ​​along the time axis at 5-minute intervals. The formula for dividing the spatiotemporal unit is as follows: (1)； Among them, L x L y L z , n represent the longitude grid number, latitude grid number, height layer number, and time slice number, respectively; t0 is the initial timestamp; Δx=1000m, Δy=1000m, Δz=100m, Δt=5min; The unit feature is encoded as follows: (2)； Among them, W t P is the meteorological disturbance coefficient. a For airspace control level, O d For order density, E c This refers to the capacity of the charging station.

3. The dynamic scheduling optimization method for a low-altitude logistics distribution network according to claim 2, characterized in that, In step S2, the TD3-GA hybrid intelligent algorithm includes: a) The state space of the TD3 algorithm is defined as follows: (3)； in, Let B be the eigenvector of the unit at time t. remain Q represents the remaining battery power of the drone. pending This is a queue of orders awaiting processing. b) The chromosome encoding of the genetic algorithm uses a three-dimensional waypoint sequence: (4)； Where x, y, and z are geographic coordinates, and t is the arrival time; Fitness function: (5)； T total E represents the total estimated time for the route, in minutes. cost Energy cost is calculated as electricity consumption × electricity price, unit: yuan; R risk The risk coefficient is a value between 0 and 1, calculated based on historical accident data; α is the time weight, β is the energy cost weight, and γ is the risk penalty weight.

4. The dynamic scheduling optimization method for a low-altitude logistics distribution network according to claim 3, characterized in that, The reward function of the TD3 algorithm is: (6)； Where, ΔQ processed E represents the change in the number of orders processed per unit of time. consumed Energy consumption value, unit is kilowatt-hour; I violate D is a counter for the number of airspace rule violations. deviation This represents the path deviation.

5. The dynamic scheduling optimization method for a low-altitude logistics distribution network according to claim 4, characterized in that, The genetic algorithm includes innovation operators: a) Dynamic crossover operator: (7)； p cross For dynamic crossover probability, This represents the ratio of the current iteration count to the maximum iteration count. b) Adaptive mutation operator: (8)； p mut For adaptive variability rate, Fitness avg The average fitness value of the population. min The minimum fitness value of the population, Fitness max This represents the maximum fitness value of the population. c) Elite retention strategy: In each generation of evolution, the individuals with the highest fitness are retained and directly enter the next generation.

6. The dynamic scheduling optimization method for a low-altitude logistics distribution network according to claim 5, characterized in that, In step S3, the multi-objective optimization function is: (9)； In the formula, T i For the delivery time of the i-th drone, T max The maximum allowable time is represented by w1, w2, and w3, which are weighting coefficients; E i E represents the energy consumption cost of the i-th drone. budget The maximum energy consumption allowed for a single drone to perform a single delivery task; The constraints include: (10)； (11)； (12)； In the formula, The total service time for processing the j-th order for the i-th drone; Let i be the latest allowed delivery completion time for the i-th order; B represents the real-time remaining battery percentage of the i-th drone. safe The minimum safe return-to-home battery threshold; v i Let v be the three-dimensional real-time velocity vector of UAV i. nominal The nominal cruise speed vector, Δv max This represents the maximum permissible speed deviation.

7. The dynamic scheduling optimization method for a low-altitude logistics distribution network according to claim 6, characterized in that, In step S4, the Bayesian network model includes: Risk points: (13)； W h W is the hidden layer weight matrix; f The feature fusion matrix; is the input 4-dimensional feature vector; tanh is the hyperbolic tangent activation function; softmax is the normalized exponential function; The dynamic adjustment strategy is as follows: When P(Risk>0.7)>0.5, path replanning is triggered: (14)； Where ⊕ represents the heading angle correction operation, Δθ is the base heading angle correction amount; k is the range attenuation coefficient, and D risk Euclidean distance for risk points.

8. The dynamic scheduling optimization method for a low-altitude logistics distribution network according to claim 7, characterized in that, When a sudden change in weather or a change in airspace control level is detected that exceeds a threshold, the four-dimensional spatiotemporal grid map automatically reduces the time slice interval to 2 minutes and re-divides the spatial grid density to 200m×200m×20m until the environment stabilizes.

9. The dynamic scheduling optimization method for a low-altitude logistics distribution network according to claim 1, characterized in that, In step S5, the lightweight LSTM model is implemented through the following steps: a) Use channel pruning techniques to remove redundant neurons; b) Perform 8-bit fixed-point quantization on the weight matrix; c) Fine-tune model parameters online using historical UAV flight data.

Images