A low-altitude infrastructure planning industrial simulation method and system based on digital twin dynamics feedback

Abstract

A low-altitude infrastructure planning industrial simulation method and system based on digital twin dynamics feedback, based on the passable three-dimensional space map, airspace restriction layer and risk grading results to establish a multi-objective site selection model of distribution efficiency, population safety risk and comprehensive cost; Fusion digital twin dynamics feedback improves NSGA-III iterative solution, calls simulation engine to obtain unmanned aerial vehicle energy consumption data and air flow disturbance data, calculates correction factors ΔE and ΔW for differentiated correction of target fitness: ΔE and ΔW correct distribution efficiency, ΔW correct population safety risk, ΔE correct comprehensive cost, Gaussian process proxy model selectively replaces the engine according to the prediction confidence threshold, and the proxy model or the engine is sorted and selected according to the corrected fitness, and the population is fed back to the simulation engine to iterate until the Pareto frontier converges; The optimal site selection scheme is output after the digital twin simulation engine simulates and verifies the Pareto solution set. The system contains the functional modules of the above method.

Description

Technical Field

[0001] This invention belongs to the emerging software development field in the new generation of information technology industry, specifically involving an industrial simulation method and system for low-altitude infrastructure planning based on digital twin dynamic feedback. Background Technology

[0002] With the rapid development of the low-altitude economy, the need for planning infrastructure such as urban low-altitude aircraft take-off and landing points and airway networks is becoming increasingly urgent. Traditional planning methods are mostly based on two-dimensional geographic information systems and empirical formulas, which are difficult to accurately characterize multi-dimensional physical constraints in three-dimensional space, such as building occlusion, micro-meteorological disturbances, and electromagnetic environment. This often leads to problems such as large navigation and positioning deviations and difficulties in multi-objective coordination in actual operation. Although existing technologies attempt to introduce optimization algorithms, they generally lack a deep coupling simulation verification link with the real physical environment. The theoretical optimal solution output by the algorithm often fails to meet the accuracy requirements of industrial-grade engineering. The underlying reason is that there is a common technical bias in existing technologies, which believes that the computational cost of dynamic simulation is too high and not suitable for embedding in the iterative inner loop of multi-objective optimization algorithms. Therefore, the simulation link is placed after the optimization solution as an outer loop verification method, resulting in a lack of physical environment feedback in the optimization search process. Therefore, there is an urgent need for a dedicated industrial software toolchain that can deeply integrate three-dimensional environment modeling, dynamic simulation, and multi-objective optimization to support high-precision and high-reliability planning of low-altitude infrastructure. Summary of the Invention

[0003] The purpose of this invention is to provide an industrial simulation design method and system for low-altitude infrastructure planning, which solves the technical problems of lack of physical constraints and disconnect between simulation and verification in traditional two-dimensional planning methods.

[0004] To achieve the above objectives, the technical solution of the present invention includes the following steps: Step S1: Multi-source data acquisition and preprocessing Collect urban 3D model data, airspace restriction data, population density data, logistics demand data, and geographic information data, and perform standardized preprocessing.

[0005] Specifically, it includes: Urban 3D model data: including urban building 3D models (including building height, outline, and use), terrain elevation models (DEM), and land cover data, which are sourced from urban CIM platforms or oblique photogrammetry data; Airspace restriction data: including no-fly zone boundaries, airport airspace protection zones, military control zones, high-voltage power line corridors, temporary restricted airspace, and other airspace control information, sourced from civil aviation authorities and geographic information platforms; Population density data includes the distribution of permanent residents, the thermal distribution of the floating population, and the location of densely populated places such as schools, hospitals, and commercial centers. The data is sourced from statistics departments and mobile phone signaling data. Logistics demand data includes historical order spatiotemporal distribution, delivery demand density, and user distribution characteristics, derived from logistics company operational data. Geographic information data includes basic geographic information such as road networks, land use, and distribution of transportation facilities.

[0006] Step S2: Construct a three-dimensional space accessibility analysis model Based on the city's 3D model, a 3D spatial accessibility analysis model for UAVs is constructed to assess the accessibility of each spatial location for UAV flight.

[0007] Specifically, it includes: S2.1: Utilize the city's 3D model to establish a 3D gridded spatial model. The grid resolution is dynamically adjusted according to the city's building density, with a 5m resolution in the city's core area and a 10m resolution in the suburbs. S2.2: For each grid cell, calculate the three-dimensional spatial accessibility index, taking into account factors such as building obstruction, terrain undulation, and electromagnetic interference. S2.3: Based on the three-dimensional space accessibility index, generate a passable three-dimensional space map, mark flight obstacle areas and recommend flight paths.

[0008] Step S3: Construct a spatial constraint model Based on airspace constraint data, an airspace constraint model is constructed to transform airspace constraint information into location constraints.

[0009] Specifically, it includes: S3.1: Perform spatial vectorization and rasterization on various airspace restriction areas to create an airspace restriction layer; S3.2: Set different restriction levels according to airspace type: No-fly zone (Level I, absolutely prohibited), restricted-fly zone (Level II, permitted under special conditions), warning zone (Level III, risk assessment required); S3.3: Establish a spatial buffer zone for airspace restrictions. A transition buffer zone of not less than 100 meters shall be set outside the no-fly zone. Gradient buffer zones shall be set in the airport airspace according to the airport class and runway direction.

[0010] Step S4: Construct a population security risk assessment model Based on population density data, a population safety risk assessment model is constructed to conduct safety risk assessments on candidate site selection areas.

[0011] Specifically, it includes: S4.1: Perform spatial overlay analysis of population density data and candidate site selection areas; S4.2: Calculate the population security risk index R for each candidate location. The calculation formula is as follows: R = w1 × D_pop + w2 × D_sensitive + w3 × D_activity Where: D_pop is the normalized population density, D_sensitive is the proximity of sensitive locations (schools, hospitals, etc.), D_activity is the intensity of human activity, and w1, w2, and w3 are weighting coefficients, which are 0.4, 0.35, and 0.25 respectively; S4.3: Based on the population security risk index, the candidate areas are divided into high-risk areas (R>0.7), medium-risk areas (0.3≤R≤0.7), and low-risk areas (R<0.3).

[0012] Step S5: Construct a multi-objective location optimization model Taking into account delivery efficiency, airspace constraints, population safety risks, and construction costs, a multi-objective site selection optimization model is constructed.

[0013] Specifically, it includes: S5.1: Define the optimization objective function: Objective 1: Maximize delivery efficiency max F1 = Σᵢ Σⱼ V_3D / Dᵢⱼ Where Dᵢⱼ is the three-dimensional spatial distance from take-off point i to demand point j, and V_3D is the three-dimensional accessibility coefficient; Objective 2: Minimize population security risks min F2 = Σᵢ Rᵢ × nᵢ Where Rᵢ is the population security risk index of the i-th candidate point, and nᵢ is the number of drones taking off and landing from that point; Objective 3: Minimize overall cost min F3 = C_construction + C_operation + C_risk Where C_construction is the construction cost, C_operation is the operating cost, and C_risk is the risk cost; S5.2: Set constraints: Airspace constraints: Candidate points must not be located within no-fly zones (Level I) or buffer zones; Population security constraints: The population security risk index R of the candidate sites is less than 0.7; Service coverage constraint: Each demand point must be covered by at least one take-off and landing point within the service radius; Capacity constraint: The daily processing capacity of each take-off and landing point shall not exceed its design capacity; Spacing constraint: The minimum distance between adjacent take-off and landing points shall not be less than 500 meters.

[0014] Step S6: Solve based on the improved multi-objective optimization algorithm An improved NSGA-III multi-objective optimization algorithm, which integrates digital twin dynamic feedback and surrogate model collaboration mechanism, is used to solve the model and obtain the Pareto optimal solution set.

[0015] Specifically, it includes: S6.1: The coding scheme is designed based on the layout of take-off and landing points, using real number coding. Each chromosome represents a set of take-off and landing point position coordinates and quantity. S6.2: Initialize the population: Combining spatial constraints and population safety constraints, an initial population is generated within the feasible region to improve the quality of the initial solution; S6.3: Differentiated Fitness Correction for Dynamic Feedback: For each individual, in addition to calculating the theoretical values ​​of the three objective functions, the digital twin dynamic simulation engine is invoked to obtain the energy consumption value and airflow disturbance amount from the simulation output. Based on the energy consumption value and airflow disturbance amount, the energy consumption correction factor ΔE and the airflow disturbance correction factor ΔW are calculated, and the fitness values ​​of the three objective functions are differentiatedly corrected. Specifically, the correction of the delivery efficiency objective function F1 introduces both the energy consumption correction factor ΔE and the airflow disturbance correction factor ΔW, while the correction of the safety objective function F2 only introduces the airflow disturbance correction factor ΔW. The cost objective function... The correction for F3 only introduces an energy consumption correction factor ΔE, ensuring the evaluation results reflect operational performance under real physical conditions. Simultaneously, a surrogate model based on a Gaussian process is constructed to selectively approximate the digital twin dynamics simulation engine. This surrogate model takes the coding parameters of the location scheme as input and outputs energy consumption and airflow disturbance, along with the predicted mean and variance. During iterative optimization, the prediction confidence of the surrogate model is calculated for each individual model; this confidence is determined based on the ratio of the prediction variance to the prediction mean. When the prediction confidence falls below a preset threshold, the digital twin engine is invoked. The mechanics simulation engine performs accurate evaluation and updates the surrogate model based on the evaluation results. When the prediction confidence is not lower than a preset threshold, the prediction results of the surrogate model are used as the basis for fitness correction, without needing to call the digital twin dynamics simulation engine. The initial stage of iterative optimization refers to the period from the first iteration until the number of training samples reaches a preset model initialization threshold. In this stage, the digital twin dynamics simulation engine is called for each individual to obtain accurate energy consumption values ​​and airflow disturbances. The input parameters and simulation output results are stored in the database as training samples, and then training is performed based on these training samples. The Gaussian process surrogate model; in subsequent iterations after the number of training samples reaches the model initialization threshold, the prediction confidence is determined based on the ratio of the prediction variance to the prediction mean output by the surrogate model. When the prediction confidence is lower than a preset threshold, the digital twin dynamics simulation engine is invoked for accurate evaluation, and the surrogate model is updated with the evaluation result. When the prediction confidence is not lower than the preset threshold, the prediction result of the surrogate model is used as the basis for fitness correction. Based on the corrected fitness value, the non-dominated sorting and selection process (see S6.4) is entered to guide the population to converge toward the Pareto optimal frontier. S6.4: Non-dominated sorting and reference point selection: The non-dominated sorting mechanism of NSGA-III and the reference point-based selection strategy are adopted to ensure the diversity and convergence of solutions; S6.5: Adaptive Crossover and Mutation: Adaptively adjusts the crossover and mutation probabilities according to the evolutionary stage of the population to avoid getting trapped in local optima; S6.6: Termination condition: Terminate when the maximum number of iterations is reached or the Pareto front converges and stabilizes.

[0016] Step S7: Verify the site selection scheme based on digital twin simulation Digital twin simulations are used to verify the location schemes in the Pareto optimal solution set, and the final location scheme is output.

[0017] Specifically, it includes: S7.1: Construct digital twin models of candidate site selection schemes to simulate the entire process of drone take-off, landing, flight, and delivery in a virtual environment; S7.2: Simulates operational performance under different weather conditions (wind, precipitation, visibility) and different workloads; S7.3: A comprehensive score is given for each option, using the following formula: Score = α × E_normalized + β × S_normalized + γ × C_normalized Where E_normalized is the normalized delivery efficiency, S_normalized is the normalized safety score, C_normalized is the normalized cost score, and α, β, and γ are weighting coefficients. S7.4: Output the site selection scheme with the highest comprehensive score as the final recommended scheme and generate a site selection report. Beneficial effects

[0018] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention reduces the navigation and positioning error of the planning scheme by 37% compared with the traditional two-dimensional method by three-dimensional environment coupling simulation under the conditions of the embodiment; the Pareto solution set hypervolume index of multi-objective evolutionary search is improved by 28%, which verifies the industrial-grade simulation accuracy.

[0019] (2) The present invention integrates various types of airspace restriction information such as no-fly zones, airport airspace clearance zones, and high-voltage line corridors, and establishes a hierarchical airspace constraint model to ensure the airspace compliance of the site selection scheme and achieve full coverage of airspace constraints.

[0020] (3) This invention innovatively incorporates population density, proximity to sensitive locations, and intensity of human activity into site selection assessment, and establishes a quantifiable population safety risk assessment model, which effectively reduces the safety risks of drone operation.

[0021] (4) This invention uses digital twin technology to simulate and verify the site selection scheme, simulates the real operation scenario in a virtual environment, ensures the reliability and robustness of the site selection scheme, and forms a digital twin verification closed loop.

[0022] (5) This invention, through the collaborative mechanism of digital twin dynamic simulation engine and proxy model, reduces the number of simulation engine calls to about 5%-10% of the total number of evaluations while ensuring the accuracy of fitness correction, significantly reducing the computational overhead and making the internal circulation dynamic feedback mechanism feasible in industrial applications.

[0023] In existing technologies, multi-objective optimization algorithms and dynamic simulations typically employ a serial outer loop architecture, where optimization search is completed first, followed by simulation verification. Simulation results are used only for post-hoc validation, not for real-time feedback. This invention embeds a digital twin dynamic simulation engine into the optimization iteration inner loop. It uses an energy consumption correction factor ΔE and an airflow disturbance correction factor ΔW to correct the fitness values ​​of the three objective functions in real time, ensuring that the optimization search direction is always guided by the constraints of the real physical environment. This fundamentally changes the technical paradigm of a serial, disconnected optimization-simulation process. However, the inner loop dynamic feedback mechanism faces a computational bottleneck: with a population size of 200 and 500 iterations, if the simulation engine is called for each fitness evaluation, 100,000 simulations are required, which is computationally unacceptable. This invention constructs a surrogate model based on Gaussian processes to selectively approximate the simulation engine. A prediction confidence threshold mechanism determines whether to call the simulation engine, reducing the number of simulation engine calls to approximately 5%-10% of the total evaluations. This ensures the accuracy of fitness correction while making the inner loop dynamic feedback mechanism computationally feasible. The difference between the above-mentioned surrogate model selective invocation mechanism and the existing EGO method is that: in the EGO method, the surrogate model is invoked in every generation and is only used to construct the alternative objective function, while in this invention, the surrogate model also serves the prediction of the fitness correction factor, and the simulation engine is only invoked to update the surrogate model when the prediction confidence is insufficient. The two are different in terms of technical purpose and implementation method. Attached Figure Description

[0024] Figure 1 Flowchart of an Industrial Simulation Method for Low-Altitude Infrastructure Planning Based on Digital Twin Dynamics Feedback Figure 2 Schematic diagram of multi-source data acquisition and preprocessing Figure 3 Flowchart for constructing a 3D spatial accessibility analysis model Figure 4 Schematic diagram of airspace restriction and constraint model (including hierarchical labeling of no-fly zones, restricted flight zones, and buffer zones) Figure 5 Schematic diagram of population security risk assessment model (including population density heat map and risk level classification) Figure 6 Multi-objective location optimization model framework diagram Figure 7 Improved NSGA-III Algorithm Flowchart Figure 8 Digital Twin Simulation Verification System Architecture Diagram Detailed Implementation

[0025] The system and method of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments, taking a core area of ​​a city (approximately 50 square kilometers) as an example.

[0026] This invention provides an industrial simulation system for low-altitude infrastructure planning based on digital twin dynamic feedback, comprising the following modules: (1) Data acquisition and preprocessing module The system is configured to acquire urban 3D model data, airspace restriction data, population density data, logistics demand data, and geographic information data via API interface or file import, and to perform coordinate system standardization, format standardization, and outlier removal on the data.

[0027] Specifically, a 3D urban model of the area was collected from the CIM platform, containing approximately 12,000 buildings with heights ranging from 3 to 280 meters. Airspace restriction data was obtained from the Civil Aviation Administration, including one airport airspace protection zone, three no-fly zones, and two high-voltage power line corridors. Population density data was obtained from the statistics department and operators, with a grid precision of 100 meters × 100 meters. Historical order data for the past six months, totaling approximately 500,000 records, was obtained from a logistics company. All data was standardized to the WGS-84 coordinate system, heterogeneous data was converted into a standardized format, and outliers were identified and removed using the 3σ principle.

[0028] (2) Three-dimensional accessibility analysis module The system is configured to dynamically set the grid resolution based on the city's building density, establish a three-dimensional gridded spatial model, and calculate the three-dimensional spatial accessibility index for each grid cell by taking into account factors such as building obstruction, terrain undulation, and electromagnetic interference. This generates a passable three-dimensional spatial map with marked areas of flight obstacles and recommended flight paths.

[0029] Specifically, a three-dimensional gridded spatial model was established for a 50-square-kilometer area. The core commercial area (approximately 15 square kilometers) used a 5m×5m×5m grid resolution, while the remaining areas used a 10m×10m×10m grid resolution. The three-dimensional spatial accessibility index for each grid cell was calculated. Electromagnetic interference factors were assessed by collecting location and intensity data of electromagnetic radiation sources such as urban high-voltage power transmission lines and communication base stations to evaluate their impact on UAV navigation signals. A passable three-dimensional spatial map was generated, marking areas with flight obstacles and recommended flight paths. Analysis results show that approximately 62% of the area is passable, with the main obstacles being tall buildings and undulating terrain.

[0030] The formula for calculating the three-dimensional spatial accessibility index is: P = α1 × P_building + α2 × P_terrain + α3 × P_EM, where P_building is the accessibility score for building obstruction, P_terrain is the accessibility score for terrain undulation, P_EM is the accessibility score for electromagnetic interference, and α1, α2, and α3 are weighting coefficients. P_building is obtained by calculating the visible sky ratio of the grid cell through view analysis, P_terrain is obtained by calculating the slope accessibility of the grid cell through a digital elevation model, and P_EM is obtained by calculating the signal interference level of the grid cell through an electromagnetic field attenuation model.

[0031] (3) Spatial Constraint Modeling Module The configuration is to perform spatial vectorization and rasterization processing on various airspace restricted areas to establish airspace restriction layers, and to divide the restriction levels into no-fly zones (level I, absolutely prohibited), restricted-fly zones (level II, permitted under special conditions) and warning zones (level III, risk assessment required). A transition buffer zone of not less than 100 meters is set for no-fly zones, and a gradient buffer zone is set for airport clear zones according to airport level and runway direction.

[0032] Specifically, the no-fly zone, airport airspace protection zone, and high-voltage power line corridor are vectorized and rasterized to create an airspace restriction layer. A 100-meter buffer zone is set outside the no-fly zone, and a gradient buffer zone is set for the airport airspace protection zone according to the airport's classification (4E). After airspace constraint screening, the usable land area in this region accounts for approximately 71%.

[0033] (4) Safety Risk Assessment Module The configuration involves spatially overlaying population density data with candidate site selection areas. The population safety risk index R for each candidate location is calculated using the formula R = w1 × D_pop + w2 × D_sensitive + w3 × D_activity. Based on the R value, the candidate areas are divided into high-risk, medium-risk, and low-risk zones. Here, D_pop is the normalized population density, D_sensitive is the proximity of sensitive locations, D_activity is the intensity of human activity, and w1, w2, and w3 are weighting coefficients, which are 0.4, 0.35, and 0.25, respectively.

[0034] Specifically, population density data was overlaid with candidate areas to calculate the population safety risk index R for each location. The results showed that high-risk areas (R>0.7) accounted for approximately 8%, mainly distributed around commercial centers and schools; medium-risk areas (0.3≤R≤0.7) accounted for approximately 35%; and low-risk areas (R<0.3) accounted for approximately 57%.

[0035] (5) Multi-objective optimization modeling module The configuration is designed to build an optimization model that aims to maximize delivery efficiency, minimize population safety risks, and minimize overall costs, and imposes airspace constraints, population safety constraints, service coverage constraints, capacity constraints, and spacing constraints.

[0036] Specifically, the three optimization objective functions are: Objective 1: Maximize delivery efficiency, max F1 = Σᵢ Σⱼ V_3D / Dᵢⱼ, where Dᵢⱼ is the three-dimensional spatial distance from the take-off point i to the demand point j, and V_3D is the three-dimensional accessibility coefficient; Objective 2: Minimize population security risk, min F2 = Σᵢ Rᵢ × nᵢ, where Rᵢ is the population security risk index of the i-th candidate point, and nᵢ is the number of drones taking off and landing from that point; Objective 3: Minimize overall cost, min F3 = C_construction + C_operation + C_risk, where C_construction is the construction cost, C_operation is the operating cost, and C_risk is the risk cost.

[0037] The constraints include: Airspace constraints: Candidate points must not be located within no-fly zones (Level I) and their buffer zones; Population security constraints: The population security risk index R of the candidate sites is less than 0.7; Service coverage constraint: Each demand point must be covered by at least one take-off and landing point within the service radius; Capacity constraint: The daily processing capacity of each take-off and landing point shall not exceed its design capacity; Spacing constraint: The minimum distance between adjacent take-off and landing points shall not be less than 500 meters.

[0038] The number of take-off and landing points is set to be between 5 and 15, with a service radius of 2 kilometers and a daily processing capacity of 200 flights per point.

[0039] (6) Multi-objective optimization solution module The configuration is based on the take-off and landing point layout design coding scheme. Real number coding is used to initialize the population. During the iterative process of the improved NSGA-III algorithm, the digital twin dynamics simulation engine is called to obtain the energy consumption value and airflow disturbance amount of the simulation output. Based on the energy consumption value and airflow disturbance amount, the energy consumption correction factor and airflow disturbance correction factor are calculated to correct the fitness values ​​of the three objective functions. A surrogate model based on Gaussian process is constructed to selectively approximate and replace the digital twin dynamics simulation engine. When the prediction confidence of the surrogate model is lower than the preset threshold, the digital twin dynamics simulation engine is called to update the surrogate model. When the prediction confidence of the surrogate model is not lower than the preset threshold, the prediction result of the surrogate model is used as the fitness correction basis. Based on the corrected fitness value, non-dominated sorting and selection are performed to drive the iterative evolution of the population. Combined with the reference point selection and adaptive crossover mutation strategy, the population is guided to converge toward the Pareto optimal frontier, and the Pareto optimal solution set is output. Specifically, an improved NSGA-III algorithm incorporating digital twin dynamics feedback is used to iteratively optimize candidate solutions. An encoding scheme is designed based on the takeoff and landing point layout, and the population is initialized using real-number encoding. In the initial stage of iterative optimization, from the first iteration until the number of training samples reaches a preset model initialization threshold, for each individual, in addition to calculating the theoretical values ​​of the three objective functions, the digital twin dynamics simulation engine is invoked to input the takeoff and landing point layout corresponding to that individual into the virtual environment to perform UAV takeoff, landing, and cruise dynamics simulations. Accurate energy consumption and airflow disturbance values ​​are obtained. Based on these energy consumption and airflow disturbance values, energy consumption correction factors and airflow disturbance correction factors are calculated to correct the fitness value. The input parameters and simulation output results are stored in the database as training samples. A surrogate model based on Gaussian processes is trained based on these training samples. The surrogate model uses a location selection method... The input parameters are the encoding parameters of the solution, and the output parameters are the energy consumption value and the airflow disturbance amount. The output parameters also include the predicted mean and predicted variance. In subsequent iterations after the number of training samples reaches the model initialization threshold, the prediction confidence of the surrogate model is calculated for each individual. The prediction confidence is determined based on the ratio of the predicted variance to the predicted mean. When the prediction confidence is lower than a preset threshold, the digital twin dynamics simulation engine is invoked for accurate evaluation, and the surrogate model is updated based on the evaluation results. When the prediction confidence is not lower than the preset threshold, the prediction results of the surrogate model are used as the basis for fitness correction, without invoking the digital twin dynamics simulation engine, guiding the population to converge toward the Pareto optimal front. The non-dominated sorting mechanism of NSGA-III and the reference point-based selection strategy are adopted to ensure the diversity and convergence of solutions. The crossover probability and mutation probability are adaptively adjusted according to the population evolution stage to avoid getting trapped in local optima. The population size is set to 200, the maximum number of iterations to 500, the initial crossover probability to 0.9 (adaptively adjusted within the range of 0.6-0.95), and the initial mutation probability to 0.1 (adaptively adjusted within the range of 0.05-0.3). The Pareto optimal solution set obtained contains 18 non-dominated solutions.

[0040] The energy consumption correction factor ΔE is calculated by normalizing the difference between the simulated energy consumption value and the theoretical energy consumption value, i.e., ΔE = (E_sim − E_theory) / E_theory, where E_sim is the simulated energy consumption value, and E_theory is the theoretical energy consumption value calculated based on the UAV's nominal parameters and planned flight path. The value range of ΔE is limited to [−1, +∞). When ΔE is negative, it indicates that the simulation shows that the actual energy consumption is lower than the theoretical expectation, and the corrected fitness value is lowered accordingly, so that the part of the theoretical fitness value that is artificially high due to high energy consumption estimation is corrected to a level closer to the real physical environment. Each correction factor (1 + λ) kThe lower bound of (×ΔE) is limited to not less than 0 to avoid negative fitness values; the airflow disturbance correction factor ΔW is obtained by normalizing the simulated airflow disturbance amount with a preset benchmark value, i.e., ΔW = W_sim / W_ref, where W_sim is the simulated airflow disturbance amount and W_ref is the benchmark airflow disturbance amount statistically calculated based on historical meteorological data; the value range of ΔW is limited to [0, ΔW_max], where ΔW_max is the preset maximum airflow disturbance correction threshold. When W_sim / W_ref exceeds ΔW_max, ΔW_max is taken to avoid excessive deviation of the fitness value by the correction factor under extreme airflow disturbance conditions; since ΔW ≥ 0, the correction factor (1 + λ) containing ΔW k × ΔW) is always not less than 1, so no additional lower bound is required.

[0041] (7) Digital Twin Dynamics Simulation Engine The system is configured to perform UAV take-off and landing dynamics simulation on the take-off and landing layout corresponding to individuals in the multi-objective optimization solution module and the site selection scheme in the digital twin simulation verification module, extract energy consumption values ​​and airflow disturbances, and output the energy consumption values ​​and airflow disturbances to the multi-objective optimization solution module or the digital twin simulation verification module.

[0042] Specifically, the digital twin dynamics simulation engine receives the take-off and landing point layout corresponding to the individual from the multi-objective optimization solution module, performs UAV take-off, landing and cruise dynamics simulation in the digital twin environment, extracts energy consumption values ​​and airflow disturbances, and outputs the energy consumption values ​​and airflow disturbances to the multi-objective optimization solution module. The multi-objective optimization solution module calculates the energy consumption correction factor ΔE and the airflow disturbance correction factor ΔW based on the energy consumption value and the airflow disturbance amount, and corrects the fitness values ​​of the three objective functions. The correction method is as follows: F1' = F1 × (1 + λ1 × ΔE) × (1 + λ2 × ΔW), F2' = F2 × (1 + λ3 × ΔW), F3' = F3 × (1 + λ4 × ΔE), where F1, F2, and F3 are the theoretical fitness values ​​of the three objective functions before correction, and λ1, λ2, λ3, and λ4 are adjustment coefficients with a value range of [0,1]. The correction of F2 only introduces the airflow disturbance correction factor ΔW, because increased airflow disturbance will lead to an increased probability of drone yaw, which directly affects population safety risks; the correction of F3 only introduces the energy consumption correction factor ΔE, because energy consumption deviation directly affects operating costs and risk costs. The calculation methods for the energy consumption correction factor ΔE and the airflow disturbance correction factor ΔW are the same as those described in the multi-objective optimization solution module above.

[0043] (8) Digital twin simulation verification module The system includes a full-condition simulation submodule and a scheme ranking submodule. The full-condition simulation submodule is configured to call the digital twin dynamics simulation engine to construct digital twin models for each site selection scheme in the Pareto optimal solution set, simulating the entire process of UAV take-off, landing, flight, and delivery under different weather conditions and workloads. The scheme ranking submodule is configured to calculate a comprehensive score according to the scoring formula Score = α × E_normalized + β × S_normalized + γ × C_normalized, and output the site selection scheme with the highest comprehensive score as the final recommended scheme, generating a site selection report. Here, E_normalized is the normalized delivery efficiency, S_normalized is the normalized safety score, C_normalized is the normalized cost score, and α, β, and γ are weighting coefficients, which are 0.4, 0.35, and 0.25, respectively.

[0044] Specifically, digital twin simulations were performed on each of the 18 proposed solutions, simulating 30 days of operation including three typical weather scenarios: sunny, rainy, and windy. The solution with the highest overall score was: setting up 8 take-off and landing points, with a delivery efficiency score of 0.89, a safety score of 0.92, a cost score of 0.85, and an overall score of 0.89. In the final site selection plan, all 8 take-off and landing points are located in low-risk or low-to-medium-risk areas, avoiding all no-fly zones and high-voltage power line corridors, and possessing good three-dimensional spatial accessibility. The system generates a site selection report based on the overall score results, which includes the location coordinates of each take-off and landing point, delivery efficiency score, safety score, cost score, and overall score.

[0045] To verify the technical effectiveness of the present invention, comparative experiments were conducted under the conditions of the above embodiments. Using the traditional two-dimensional GIS planning method as a control group, the navigation and positioning error of the method of the present invention was reduced from 0.82 meters to 0.52 meters, a reduction of approximately 37%. Using the standard NSGA-III algorithm (without dynamic feedback) as a control group, the Pareto solution set hypervolume index of the improved NSGA-III algorithm of the present invention was improved from 0.61 to 0.78, an improvement of approximately 28%.

[0046] The data flow relationships between the modules are as follows: the data acquisition and preprocessing module outputs the preprocessed data to the 3D accessibility analysis module, the spatial constraint modeling module, and the safety risk assessment module; the output results of the 3D accessibility analysis module, the spatial constraint modeling module, and the safety risk assessment module are input to the multi-objective optimization modeling module; the multi-objective optimization modeling module outputs the constructed optimization model to the multi-objective optimization solution module; during the iterative optimization process, the multi-objective optimization solution module outputs the take-off and landing point layout corresponding to each individual to the digital twin dynamics simulation engine; the digital twin dynamics simulation engine feeds back the simulation output energy consumption value and airflow disturbance amount to the multi-objective optimization solution module, which calculates the energy consumption based on the energy consumption value and airflow disturbance amount. The fitness values ​​of the three objective functions are corrected by correction factors and airflow disturbance correction factors. Based on the corrected fitness values, non-dominated sorting and selection are performed to drive the iterative evolution of the population. The multi-objective optimization solution module constructs a surrogate model based on a Gaussian process to selectively approximate the digital twin dynamics simulation engine. When the prediction confidence of the surrogate model is lower than a preset threshold, the digital twin dynamics simulation engine is called for accurate evaluation and the evaluation result is used to update the surrogate model. When the prediction confidence of the surrogate model is not lower than the preset threshold, the prediction result of the surrogate model is used as the basis for fitness correction. The multi-objective optimization solution module outputs the Pareto optimal solution set to the digital twin simulation verification module. The digital twin simulation verification module outputs the final recommended scheme.

Claims

1. An industrial simulation method for low-altitude infrastructure planning based on digital twin dynamic feedback, characterized in that, Includes the following steps: S1: Collect urban 3D model data, airspace restriction data, population density data, logistics demand data, and geographic information data, and perform coordinate system standardization, format standardization, and outlier removal on the data; S2: Using the city's 3D model, establish a 3D gridded spatial model, calculate the 3D spatial accessibility index, and generate a passable 3D spatial map; S3: Perform spatial vectorization and rasterization processing on various airspace restriction areas to establish airspace restriction layers and spatial buffers; S4: Spatial overlay analysis of population density data and candidate site selection areas is performed to calculate the population safety risk index and classify the risk level; S5: Based on the accessible 3D spatial map, airspace restriction layer and risk classification results, define the multi-objective optimization objective function and constraints, and establish a multi-objective site selection optimization model; S6: An improved reference-point-based non-dominated sorting genetic algorithm NSGA-III is adopted. Based on the take-off and landing point layout, an encoding scheme is designed and the population is initialized. The multi-objective location optimization model is iteratively solved. During the iteration process, the digital twin dynamics simulation engine is called to simulate the candidate location schemes and obtain the energy consumption value and airflow disturbance amount of the simulation output. Based on the energy consumption value and airflow disturbance amount, the energy consumption correction factor and airflow disturbance correction factor are calculated to correct the fitness values ​​of the three objective functions. A surrogate model based on Gaussian process is constructed to selectively approximate and replace the digital twin dynamics simulation engine. When the prediction confidence of the surrogate model is lower than the preset threshold, the digital twin dynamics simulation engine is called to update the surrogate model. When the prediction confidence of the surrogate model is not lower than the preset threshold, the prediction result of the surrogate model is used as the fitness correction basis. Based on the corrected fitness value, non-dominated sorting and selection are performed to drive the iterative evolution of the population and guide the population to converge toward the Pareto optimal frontier to obtain the Pareto optimal solution set. S7: Perform digital twin simulation verification on the candidate location schemes in the Pareto optimal solution set, and output the optimal location scheme.

2. The industrial simulation method for low-altitude infrastructure planning based on digital twin dynamic feedback as described in claim 1, characterized in that, The urban 3D model data includes urban building 3D models, terrain elevation models, and land cover data, sourced from urban CIM platforms or oblique photogrammetry data; the airspace restriction data includes no-fly zone boundaries, airport airspace protection zones, military management zones, high-voltage power line corridors, and temporary restricted airspace, sourced from civil aviation management departments and geographic information platforms; the population density data includes the distribution of permanent residents, the thermal distribution of transient populations, and the location of densely populated areas, sourced from statistical departments and mobile phone signaling data; the logistics demand data includes the spatiotemporal distribution of historical orders, delivery demand density, and user distribution characteristics, sourced from logistics company operational data; the geographic information data includes road networks, land use properties, and the distribution of transportation facilities; step S2 specifically includes: using the urban 3D model to establish a 3D gridded spatial model, with a grid resolution based on... According to the dynamic adjustment of urban building density, a 5m resolution is used in the urban core area and a 10m resolution is used in the suburbs; for each grid cell, a three-dimensional spatial accessibility index is calculated, taking into account factors such as building obstruction, terrain undulation and electromagnetic interference; based on the three-dimensional spatial accessibility index, a passable three-dimensional spatial map is generated, marking flight obstacle areas and recommended flight channels; step S3 specifically includes: spatial vectorization and rasterization of various airspace restricted areas to establish an airspace restriction layer; different restriction levels are set according to airspace type: no-fly zones are level I, i.e., absolutely prohibited; restricted-fly zones are level II, i.e., permitted under special conditions; warning zones are level III, i.e., risk assessment is required; a spatial buffer zone for airspace restrictions is established, with a transition buffer zone of not less than 100 meters set outside the no-fly zone, and a gradient buffer zone is set for the airport clearance zone according to the airport level and runway direction.

3. The industrial simulation method for low-altitude infrastructure planning based on digital twin dynamic feedback as described in claim 1, characterized in that, Step S4 specifically includes: performing spatial overlay analysis of population density data and candidate site selection areas; calculating the population safety risk index R for each candidate location, using the formula: R = w1 × D_pop + w2 × D_sensitive + w3 × D_activity; where D_pop is the normalized population density, D_sensitive is the proximity of sensitive locations, D_activity is the intensity of human activity, and w1, w2, and w3 are weighting coefficients, taking values ​​of 0.4, 0.35, and 0.25 respectively; and dividing the candidate areas into high-risk, medium-risk, and low-risk zones based on the population safety risk index. Step S5 specifically includes: defining optimization objective functions, with objective one being maximizing delivery efficiency: max F1 = Σᵢ Σⱼ V_3D / Dᵢⱼ, where Dᵢⱼ is the three-dimensional spatial distance from take-off / landing point i to demand point j, and V_3D is the three-dimensional accessibility coefficient; and objective two being minimizing population safety risk: min F2 = Σᵢ Rᵢ × nᵢ, where Rᵢ is the population safety risk index of the i-th candidate point, and nᵢ is the number of drones taking off and landing from that point; Objective 3 is to minimize the comprehensive cost: min F3 = C_construction + C_operation + C_risk, where C_construction is the construction cost, C_operation is the operation cost, and C_risk is the risk cost; constraints are set, including airspace constraints, population safety constraints, service coverage constraints, capacity constraints, and spacing constraints.

4. The industrial simulation method for low-altitude infrastructure planning based on digital twin dynamic feedback as described in claim 1, characterized in that, Step S6 specifically includes: designing an encoding scheme based on the take-off and landing point layout, using real number encoding, where each chromosome represents a set of take-off and landing point location coordinates and quantity; generating an initial population within the feasible domain by combining airspace constraints and population safety constraints; in the initial stage of iterative optimization, i.e., from the first iteration until the number of training samples reaches the preset model initialization threshold, calling the digital twin dynamics simulation engine for each individual, inputting the take-off and landing point layout corresponding to that individual into the virtual environment to perform UAV take-off, landing, and cruise dynamics simulation, obtaining the energy consumption value and airflow disturbance amount of the simulation output, calculating the energy consumption correction factor ΔE and the airflow disturbance correction factor ΔW based on the energy consumption value and airflow disturbance amount, and using the input parameters and simulation output results as training data. The samples are stored in the database, and the Gaussian process-based surrogate model is trained based on the training samples. The surrogate model takes the coding parameters of the site selection scheme as input and outputs energy consumption and airflow disturbance, while also outputting the prediction mean and prediction variance. In subsequent iterations after the number of training samples reaches the model initialization threshold, the prediction confidence is determined based on the ratio of the prediction variance to the prediction mean output by the surrogate model. When the prediction confidence is lower than a preset threshold, the digital twin dynamics simulation engine is invoked for accurate evaluation, and the surrogate model is updated with the evaluation results. When the prediction confidence is not lower than the preset threshold, the prediction results of the surrogate model are used as the basis for fitness correction. The corrected fitness value is calculated as: F1' = F1 × (1 + λ1 × ΔE) × (1 + λ2 × ΔW), F2' = F2 × (1 + λ3 × ΔW), F3' = F3 × (1 + λ4 × ΔE); where F1, F2, and F3 are the theoretical fitness values ​​of the three objective functions before correction, and λ1, λ2, λ3, and λ4 are adjustment coefficients with values ​​ranging from [0,1]. The correction of F2 only introduces the airflow disturbance correction factor ΔW, because increased airflow disturbance will lead to an increased probability of drone yaw, directly affecting population safety risks. The correction of F3 only introduces the energy consumption correction factor ΔE, because energy consumption deviation directly affects operating costs and risk costs. The crossover probability and mutation probability are adaptively adjusted according to the population evolution stage to avoid getting trapped in local optima. Based on the corrected fitness values, the non-dominated sorting mechanism of NSGA-III and the selection strategy based on reference points are used to drive the population iterative evolution until the maximum number of iterations is reached or the Pareto front converges and stabilizes.

5. The industrial simulation method for low-altitude infrastructure planning based on digital twin dynamic feedback as described in claim 4, characterized in that, The energy consumption correction factor ΔE is calculated by normalizing the difference between the simulated energy consumption value and the theoretical energy consumption value, i.e., ΔE = (E_sim − E_theory) / E_theory, where E_sim is the simulated energy consumption value, and E_theory is the theoretical energy consumption value calculated based on the UAV's nominal parameters and planned flight path; the value range of ΔE is limited to [−1, +∞), and each correction factor (1 + λ) k The lower bound of (×ΔE) is limited to not less than 0 to avoid negative fitness values; the airflow disturbance correction factor ΔW is obtained by normalizing the simulated airflow disturbance amount with a preset benchmark value, i.e., ΔW = W_sim / W_ref, where W_sim is the simulated airflow disturbance amount and W_ref is the benchmark airflow disturbance amount statistically calculated based on historical meteorological data; the value range of ΔW is limited to [0, ΔW_max], where ΔW_max is the preset maximum airflow disturbance correction threshold. When W_sim / W_ref exceeds ΔW_max, ΔW_max is taken to avoid excessive deviation of the fitness value by the correction factor under extreme airflow disturbance conditions; since ΔW ≥ 0, the correction factor (1 + λ) containing ΔW k (×ΔW) is always not less than 1, and no additional lower bound is required; The adaptive adjustment of crossover and mutation probabilities based on the population evolution stage specifically includes: using a higher crossover probability and a lower mutation probability in the early stages of iteration to promote population exploration. In the later stages of iteration, the crossover probability is reduced and the mutation probability is increased to enhance the local search capability; the termination condition is: reaching the preset maximum number of iterations, or the distribution change of the Pareto front for multiple consecutive generations is less than the preset convergence threshold.

6. The industrial simulation method for low-altitude infrastructure planning based on digital twin dynamic feedback as described in claim 1, characterized in that, Step S7 specifically includes: constructing digital twin models of candidate site selection schemes, simulating the entire process of drone take-off, landing, flight, and delivery in a virtual environment; simulating the operational performance under different weather conditions and different business loads, and obtaining the comprehensive performance indicators of each scheme; comprehensively scoring each scheme based on the comprehensive performance indicators, outputting the site selection scheme with the highest comprehensive score as the final recommended scheme, and generating a site selection report; the formula for calculating the comprehensive score is: Score = α × E_normalized + β × S_normalized + γ × C_normalized; where E_normalized is the normalized delivery efficiency, S_normalized is the normalized safety score, C_normalized is the normalized cost score, and α, β, and γ are weighting coefficients, which are 0.4, 0.35, and 0.25, respectively.

7. An industrial simulation system for low-altitude infrastructure planning based on digital twin dynamic feedback, characterized in that, The system, employing the method as described in any one of claims 1-6, comprises: The data acquisition and preprocessing module is configured to acquire urban 3D model data, airspace restriction data, population density data, logistics demand data, and geographic information data, and to perform coordinate system standardization, format standardization, and outlier removal on the data. The 3D accessibility analysis module is configured to use the city's 3D model to build a 3D gridded spatial model, calculate the 3D spatial accessibility index, and generate a 3D accessible spatial map. The airspace constraint modeling module is configured to perform spatial vectorization and rasterization processing on various airspace constraint areas, and to establish airspace constraint layers and spatial buffers; The safety risk assessment module is configured to perform spatial overlay analysis of population density data and candidate site selection areas, calculate the population safety risk index, and classify the risk level. The multi-objective optimization modeling module is configured to define multi-objective optimization objective functions and constraints based on the accessible 3D spatial map, airspace restriction layer and risk classification results, and establish a multi-objective site selection optimization model. The multi-objective optimization solution module is configured to use an improved reference-point-based non-dominated sorting genetic algorithm NSGA-III. It designs an encoding scheme based on the takeoff and landing point layout and initializes the population. It iteratively solves the multi-objective location optimization model. During the iteration process, it calls a digital twin dynamics simulation engine to simulate candidate location schemes. Based on the energy consumption and airflow disturbance output by the digital twin dynamics simulation engine, it calculates energy consumption correction factors and airflow disturbance correction factors to correct the fitness values ​​of the three objective functions. A surrogate model based on a Gaussian process is constructed to selectively approximate and replace the digital twin dynamics simulation engine. When the prediction confidence of the surrogate model is lower than a preset threshold, it calls the digital twin dynamics simulation engine to update the surrogate model. When the prediction confidence of the surrogate model is not lower than the preset threshold, the prediction result of the surrogate model is used as the basis for fitness correction. Based on the corrected fitness values, non-dominated sorting and selection are performed to drive the iterative evolution of the population, guiding the population to converge toward the Pareto optimal frontier to obtain the Pareto optimal solution set. The digital twin dynamics simulation engine is configured to perform UAV take-off, landing and cruise dynamics simulations on the take-off and landing point layouts corresponding to individuals in the multi-objective optimization solution module and the site selection schemes in the digital twin simulation verification module, extract energy consumption values ​​and airflow disturbances, and output the energy consumption values ​​and airflow disturbances to the multi-objective optimization solution module or the digital twin simulation verification module. The digital twin simulation verification module is configured to perform digital twin simulation verification on candidate location schemes in the Pareto optimal solution set and output the optimal location scheme.

8. The system as described in claim 7, characterized in that, The data acquisition and preprocessing module includes a data access submodule, a data cleaning submodule, and a standardization processing submodule. The data access submodule receives urban 3D model data, airspace restriction data, population density data, logistics demand data, and geographic information data from external data sources. The urban 3D model data includes urban building 3D models, terrain elevation models, and land cover data, sourced from urban CIM platforms or oblique photogrammetry data. The airspace restriction data includes no-fly zone boundaries, airport airspace protection zones, military management zones, high-voltage power line corridors, and temporary restricted airspace, sourced from civil aviation management departments and geographic information platforms. The population density data includes the distribution of permanent residents, the thermal distribution of transient populations, and the location of densely populated areas, sourced from statistical departments and mobile phone signaling data. The logistics demand data includes the spatiotemporal distribution of historical orders, delivery demand density, and user distribution characteristics, sourced from logistics company operational data. The geographic information data includes road networks, land use properties, and the distribution of transportation facilities. The data cleaning submodule performs outlier removal processing on the received data. The standardization processing submodule is used to perform coordinate system and format standardization on the data after outlier removal, and outputs the standardized data to the three-dimensional accessibility analysis module, the spatial constraint modeling module and the safety risk assessment module. The three-dimensional accessibility analysis module includes a spatial gridding submodule and a accessibility assessment submodule. The spatial gridding submodule receives urban three-dimensional model data output by the data acquisition and preprocessing module, establishes a three-dimensional gridded spatial model, wherein the grid resolution is dynamically adjusted according to the urban building density, with a resolution of 5m in the urban core area and 10m in the suburbs. The accessibility assessment submodule calculates the three-dimensional spatial accessibility index for each grid cell based on the three-dimensional gridded spatial model, comprehensively considers building obstruction, terrain undulation, and electromagnetic interference factors, generates a accessible three-dimensional spatial map, marks flight obstacle areas and recommended flight paths, and outputs the accessible three-dimensional spatial map to the multi-objective optimization modeling module. The airspace constraint modeling module includes an airspace layer construction submodule and a buffer zone setting submodule. The airspace layer construction submodule receives airspace restriction data output from the data acquisition and preprocessing module, performs spatial vectorization and rasterization processing on various airspace restriction areas, establishes an airspace restriction layer, and sets different restriction levels according to airspace type: no-fly zones are Level I (absolutely prohibited), restricted-fly zones are Level II (allowed under special conditions), and warning zones are Level III (risk assessment required). The buffer zone setting submodule establishes spatial buffer zones for airspace restrictions based on the airspace restriction layer. A transition buffer zone of no less than 100 meters is set around the no-fly zone, and a gradient buffer zone is set for the airport airspace clearance zone according to the airport level and runway direction. The airspace constraint conditions including the buffer zones are output to the multi-objective optimization modeling module. The safety risk assessment module includes a spatial overlay analysis submodule and a risk classification submodule. The spatial overlay analysis submodule receives population density data output from the data acquisition and preprocessing module, performs spatial overlay analysis on the population density data and candidate site selection areas, and calculates the population safety risk index R for each candidate location. The calculation formula is: R = w1 × D_pop + w2 × D_sensitive + w3 × D_activity, where D_pop is the normalized population density, D_sensitive is the proximity of sensitive locations, D_activity is the intensity of human activity, and w1, w2, and w3 are weighting coefficients, which are 0.4, 0.35, and 0.25, respectively; the risk classification submodule is used to divide the candidate area into high-risk, medium-risk, and low-risk areas based on the population safety risk index R, and output the risk classification results to the multi-objective optimization modeling module.

9. The system as described in claim 7, characterized in that, The multi-objective optimization solution module includes an encoding initialization submodule, a fitness evaluation submodule, a surrogate model management submodule, an evolutionary operation submodule, and an iteration termination judgment submodule. The encoding initialization submodule is used to design an encoding scheme based on the take-off and landing point layout, employing real-number encoding. Each chromosome represents a set of take-off and landing point location coordinates and quantities. Combining spatial constraints and population safety constraints, it generates an initial population within the feasible region. The fitness evaluation submodule receives the initial population output by the encoding initialization submodule, decodes each individual into a take-off and landing point layout scheme, and then calls the digital twin. The mechanical simulation engine obtains accurate energy consumption values ​​and airflow disturbance amounts, calculates the delivery efficiency objective function F1, the safety objective function F2, and the cost objective function F3, and constructs correction factors based on the prediction confidence output by the proxy model management submodule. Based on the correction factors, the fitness values ​​of the three objective functions are corrected. Specifically, the correction of F1 introduces both the energy consumption correction factor ΔE and the airflow disturbance correction factor ΔW, the correction of F2 introduces only the airflow disturbance correction factor ΔW, and the correction of F3 introduces only the energy consumption correction factor ΔE. The corrected fitness values ​​are then output to the evolutionary operation submodule. The proxy model management submodule is used to construct a proxy model based on Gaussian process to selectively approximate and replace the digital twin dynamics simulation engine. It determines whether to call the digital twin dynamics simulation engine based on the prediction confidence. When the prediction confidence is lower than a preset threshold, it calls the digital twin dynamics simulation engine for accurate evaluation and updates the proxy model with the evaluation result. When the prediction confidence is not lower than the preset threshold, it uses the prediction result of the proxy model as the basis for fitness correction. The surrogate model management submodule is used in the initial stage of iterative optimization, that is, from the first iteration until the number of training samples reaches the preset model initialization threshold, to input the take-off and landing point layout parameters corresponding to the individual into the digital twin dynamics simulation engine to perform simulation, obtain the energy consumption value and airflow disturbance amount of the simulation output, calculate the energy consumption correction factor and airflow disturbance correction factor based on the energy consumption value and airflow disturbance amount to correct the fitness value, and store the input parameters and the energy consumption value and airflow disturbance amount of the simulation output as training samples into the database, and train the surrogate model based on the Gaussian process based on the training samples; In subsequent iterations after the number of training samples reaches the model initialization threshold, the prediction confidence is determined based on the ratio of the prediction variance to the prediction mean output by the surrogate model. When the prediction confidence is lower than the preset threshold, the digital twin dynamics simulation engine is invoked for accurate evaluation and the surrogate model is updated with the evaluation result. When the prediction confidence is not lower than the preset threshold, the prediction result of the surrogate model is used as the basis for fitness correction. The evolutionary operation submodule is used to drive the iterative evolution of the population based on the modified fitness value using the non-dominated sorting mechanism of NSGA-III and the selection strategy based on the reference point. It adaptively adjusts the crossover probability and mutation probability according to the evolutionary stage of the population to avoid getting trapped in local optima. The iteration termination judgment submodule is used to monitor the current iteration number and the change in the Pareto front distribution. When the preset maximum iteration number is reached or the change in the Pareto front distribution for multiple consecutive generations is less than the preset convergence threshold, the iteration is terminated and the Pareto optimal solution set is output to the digital twin simulation verification module.

10. The system as claimed in claim 7, characterized in that, The digital twin simulation verification module includes a full-condition simulation submodule and a scheme ranking submodule. The full-condition simulation submodule receives the Pareto optimal solution set output by the multi-objective optimization solution module, calls the digital twin dynamics simulation engine to construct a digital twin model of the candidate site selection scheme for each scheme, simulates the entire process of UAV take-off, landing, flight, and delivery in a virtual environment, and simulates the operation performance under different weather conditions and different business loads to obtain the comprehensive performance index of each scheme. The scheme ranking submodule is used to comprehensively score each scheme based on the comprehensive performance index according to the scoring formula Score = α × E_normalized + β × S_normalized + γ × C_normalized, where E_normalized is the normalized delivery efficiency, S_normalized is the normalized safety score, C_normalized is the normalized cost score, and α, β, and γ are weight coefficients, which are 0.4, 0.35, and 0.25 respectively. The scheme with the highest comprehensive score is output as the final recommended scheme, and a site selection report is generated.

Images