An eVTOL trajectory planning method based on wind field prediction and collaborative optimization

Abstract

This invention relates to the field of intelligent path planning for urban low-altitude aerial vehicles (eVTOLs), proposing an eVTOL trajectory planning method based on wind field prediction and collaborative optimization. The method includes: determining eVTOL platform parameters, urban building cluster mission scenario, and origin / endpoint; acquiring or simulating urban three-dimensional wind field velocity components and turbulent kinetic energy data under multiple incoming flow conditions; training an adaptive surrogate model to quickly predict urban canopy wind fields; establishing rotor aerodynamic power consumption, battery-end power, and segment energy consumption models; constructing an energy-time collaborative optimization model including turbulent kinetic energy threshold, maximum tilt angle, power, effective crosswind, velocity envelope, and energy budget constraints; performing edge optimization ensemble search on a safe KNN graph, dynamically solving for the optimal ground velocity for each candidate edge, and outputting a three-dimensional trajectory, velocity profile, energy consumption, and mission time that satisfy safety constraints. This method can significantly reduce turbulence exposure with relatively low energy consumption and achieve collaborative planning that is safe, energy-efficient, and time-saving.

Description

Technical Field

[0001] This invention relates to the field of intelligent path planning for urban low-altitude aircraft, and in particular to an eVTOL trajectory planning method based on wind field prediction and collaborative optimization. Background Technology

[0002] With the development of the low-altitude economy and urban air traffic, electric vertical take-off and landing (eVTOL) aircraft have advantages such as high mobility, flexible deployment and no restrictions from ground traffic congestion in scenarios such as last-mile delivery, emergency medical services and warehouse material handling. Therefore, a path planning method that can take into account both operational energy consumption and delivery timeliness is needed.

[0003] Traditional path planning methods often employ the shortest path approach or plan paths based on empirical models. However, when entering urban low-altitude operation scenarios, the coupling between path geometry, aircraft dynamics, and complex three-dimensional wind fields is significantly enhanced, making it difficult for the simple geometric shortest path to accurately reflect real flight safety and cost.

[0004] eVTOL is affected by multiple physical constraints such as battery capacity, power limit, maximum tilt angle, airspeed and ground speed envelope. Its energy consumption is highly sensitive to flight status and environmental disturbances. Under strong headwinds, crosswinds or building wakes, maintaining a given ground speed requires adjusting attitude and rotor speed, resulting in significant nonlinear changes in battery power.

[0005] Existing methods still have the following shortcomings: First, urban wind fields are often simplified to uniform wind or steady wind models, making it difficult to capture the three-dimensional flow characteristics of building canyons, flow around, wakes, and high gradient regions near walls; Second, minimum energy consumption, minimum time, or minimum distance are often treated in isolation, lacking an engineering scheme that unifies and coordinates the modeling of turbulence safety, energy consumption, and flight duration. Summary of the Invention

[0006] The purpose of this invention is to address the problems of three-dimensional wind field disturbance, nonlinear power consumption, and multi-objective coupling faced by eVTOL delivery tasks in complex urban wind fields, and to propose an eVTOL trajectory planning method based on wind field prediction and collaborative optimization.

[0007] The specific technical solution is as follows:

[0008] An eVTOL trajectory planning method based on wind field prediction and collaborative optimization includes the following steps:

[0009] Step 1: Determine the eVTOL platform parameters and urban delivery mission scenario, including: aircraft mass, rotor radius, propulsion efficiency, auxiliary power, battery capacity, safety margin, maximum tilt angle, maximum airspeed, ground speed upper and lower limits, takeoff and landing hovering time, as well as the geometry of urban building clusters, 3D planning mesh, starting point and ending point;

[0010] Step 2: Construct a three-dimensional wind field dataset of the urban building complex under multiple sets of incoming flow conditions, including incoming flow velocity, turbulence intensity and turbulence length, as well as the three-dimensional velocity components and turbulent kinetic energy under these conditions;

[0011] Step 3: Train and build a wind field proxy model using a 3D wind field dataset, so that it can predict the city's 3D wind speed vector field and turbulent kinetic energy field based on the incoming flow parameters and spatial coordinates on the 3D planning grid;

[0012] Step 4: Construct eVTOL flight kinematics model, rotor aerodynamic power consumption model and battery energy consumption model, and establish the mapping relationship from local wind vector, ground speed, airspeed to battery-end propulsion power and segment energy consumption;

[0013] Step 5: Construct an energy and time co-optimization model based on flight safety constraints, with the weighted sum of normalized flight time and normalized energy consumption as the objective function, and adjust the preference between time priority and energy consumption priority through task weights;

[0014] Step 6: Construct a safe K-nearest neighbor search graph; map the predicted 3D wind speed vector field and turbulent kinetic energy field to the city's 3D planning grid, generate candidate edges based on fixed connectivity, and prune candidate edges that intersect with building obstacles or whose turbulent kinetic energy exceeds the threshold to obtain a safe K-nearest neighbor search graph;

[0015] Step 7: Perform edge-level dynamic ground velocity optimization. For each candidate edge in the safe K-nearest neighbor search graph, use ground velocity as a continuous decision variable and solve for the optimal ground velocity and edge cost within the upper and lower bounds of the energy-time joint cost. If there is no feasible velocity, then remove the edge.

[0016] Step 8: Execute The search uses a safe K-nearest neighbor search graph as the search space and calls the optimal ground velocity and edge cost obtained in step 7 during the search process, so that path geometry selection and ground velocity optimization are completed within the same search framework.

[0017] Step 9: Output planning results; output three-dimensional waypoint sequence, optimal ground speed for each segment, total energy consumption, mission time, energy consumption per unit distance, maximum turbulent kinetic energy along the route, and feasibility assessment.

[0018] The beneficial effects of this invention are as follows:

[0019] By constructing a PIPT-AS wind field proxy model based on physical information constraints, rapid and high-precision prediction of three-dimensional wind speed and turbulent kinetic energy in complex urban building environments is achieved, providing reliable wind field information for trajectory planning. By combining wind field prediction, flight safety constraints, energy-time co-optimization, and edge-level dynamic ground speed optimization, the planning results can comprehensively consider path length, flight energy consumption, mission time, and turbulence exposure level, improving the overall performance of trajectory planning under complex wind field conditions. In areas with strong winds, headwinds, and high turbulence, the method can proactively avoid high-risk flight segments and obtain feasible trajectories that meet flight envelope, energy budget, and turbulence safety thresholds through trajectory adjustment and ground speed optimization, thereby improving the safety, economy, adaptability, and operational efficiency of eVTOL urban delivery missions. Attached Figure Description

[0020] Figure 1 The flowchart shows the eVTOL trajectory planning method based on wind field prediction and collaborative optimization of the present invention.

[0021] Figure 2 Flowchart of the PIPT-AS wind farm proxy model;

[0022] Figure 3 This is a 3D geometric schematic diagram of the wind field;

[0023] Figure 4 The geometric diagram of the wind field's starting and ending points;

[0024] Figure 5 Example of cross-sectional comparison between the actual CFD values ​​of a wind field and the surrogate model;

[0025] Figure 6 Examples of comparing 3D flight paths and built environments for different planning strategies. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other. To achieve the above objectives, this invention adopts the following technical solution.

[0027] like Figure 1 , Figure 2 As shown, this invention provides an eVTOL trajectory planning method based on wind field prediction and collaborative optimization, comprising the following steps:

[0028] Step 1: Determine the eVTOL platform parameters and the urban delivery mission scenario. Based on mission requirements, determine the aircraft mass, rotor radius, propulsion efficiency, auxiliary power, battery capacity, safety margin, maximum tilt angle, maximum airspeed, ground speed upper and lower limits, takeoff and landing hovering time, as well as the geometry of the urban building complex, the 3D planning mesh, and the starting and ending points.

[0029] Step 2: Construct a three-dimensional wind field dataset for the city. Using sensor data acquisition or computational fluid dynamics simulation, obtain three-dimensional velocity components and turbulent kinetic energy under multiple conditions of incoming flow velocity, turbulence intensity, and turbulence length, and divide the dataset into training, validation, and test sets.

[0030] Step 3: Construct a wind field surrogate model based on the Physics-Informed Point Transformer Adaptive Surrogate Model (PIPT-AS). Using five-dimensional incoming flow parameters and spatial coordinates as input, the model undergoes parameter encoding, Fourier position encoding, point-level feature fusion, global modeling of the Transformer, enhancement of mainstream velocity components, and component output heads to predict the u, v, and w velocity components and the turbulent kinetic energy k. Here, u, v, and w represent the velocity components along the x, y, and z directions at the spatial coordinate point, respectively, and k represents the turbulent kinetic energy, as detailed below:

[0031] Given five-dimensional incoming flow parameters And spatial coordinates on a fixed computational fluid dynamics (CFD) grid: ,in, For points in the three-dimensional wind field space, Let N be the coordinates of a point in the 3D wind field space, and N be the total number of points in the grid within the wind field space. A mapping is learned based on the wind field proxy model PIPT-AS:

[0032] ;

[0033] To find an approximate CFD solution .in, This indicates that it contains learnable parameters. The proxy model, This indicates the learnable parameters of the model. Represents the five-dimensional incoming flow parameters. Represents the coordinates of the i-th spatial sampling point. y represents the model prediction result. i This represents the corresponding CFD numerical solution. Five-dimensional incoming flow parameters. First, the data is encoded into 192-dimensional global features using two layers of MLP, and then shared as conditional input across all sampling points. For the incoming flow velocity component, For turbulence intensity, The turbulence length scale is used; the three-dimensional coordinates are encoded using multi-scale Fourier features to enhance the representation of high-frequency structures in the flow around structures and the wake. After point-level concatenation of the two types of features, global modeling is performed using a single layer of fused MLP and four transformer blocks with layer scaling, enabling each grid point to perceive the nonlocal effects of the entire flow field through self-attention. The mainstream velocity components are considered. The most critical factor for flight aerodynamic characteristics is the output of the main engine. An additional self-attention-gated enhancement module was designed to finely reweight the features. Finally, the network outputs through fully connected heads for each of the four components. ,in, , , These represent the velocity components in the x, y, and z directions predicted by the model at all grid sampling points, respectively. This represents the turbulent kinetic energy field predicted by the model at all grid sampling points; for the i-th spatial sampling point, its corresponding output can be denoted as... , , , , representing the predicted values ​​of velocity in the x-direction, y-direction, z-direction, and turbulent kinetic energy at that point, respectively. and A deeper prediction head is used to improve the fitting accuracy of these two key physical quantities. The number of parameters in the entire model remains in the millions, significantly smaller than that of conventional large-scale transformers, saving GPU memory.

[0034] During the training phase, considering component variance, local wind speed, and the training stage itself, this invention designs an adaptive weighted loss and a two-stage training strategy for high-wind-speed areas and... Components are given greater weight. In the normalized space, [the following is used]. and The weighted combined loss serves as the base term, and multiple weighting is introduced to highlight physically more important components and regions. That is, the four output components are adaptively weighted according to their standard deviation in the dataset. At the same time, for the mainstream velocity components and turbulent kinetic energy Apply additional amplification factor To enhance the fitting of these two types of quantities; each sampling point is also assigned different velocity weights based on the low / medium / high speed range to which the local velocity modulus belongs. This results in a more severe penalty for errors in high-wind-speed areas; furthermore, a first-order difference smoothing regularization is added to the point cloud sequence to suppress non-physical oscillations. Let the predicted velocity of adjacent grid points be denoted as... Difference regularization is and using hyperparameters By controlling its weight in the total loss, the overall loss can be written as:

[0035] ;

[0036] in, Based on the base loss, the optimizer uses AdamW, and the learning rate follows cosine annealing scheduling. Simultaneously, to balance convergence speed and fine-tuning, this invention employs a two-stage training strategy: Stage 1 involves balanced training of the entire model to gradually reduce the validation set RMSE; then, in Stage 2, the parametric encoder, position encoder, and some transformer modules are frozen, with only the latter half of the backbone and output head fine-tuned, while significantly reducing the learning rate and improving performance. The loss weights are adjusted to further improve the accuracy of key components without increasing training costs.

[0037] During the inference phase, given any new five-dimensional incoming flow parameter μ, a single forward propagation can generate a complete three-dimensional wind field (u,v,w,k) at all grid points.

[0038] Step 4: Establish eVTOL power consumption model. Construct eVTOL flight kinematics, rotor aerodynamic power consumption, and battery energy consumption models, and establish the mapping relationship from local wind vector, ground speed, airspeed, rate of climb, and load state to battery-end power and segment energy consumption; map ground speed to airspeed based on local wind vector, and calculate segment battery-end power and energy consumption by combining induced power, shape drag power, parasitic power, propulsion link efficiency, auxiliary equipment power consumption, and service hovering energy.

[0039] The flight path of an unmanned aerial vehicle (UAV) can be represented as a series of discrete waypoints or within a time interval. Continuous space curves on The trajectory must satisfy basic kinematic constraints, including limits on maximum horizontal velocity, maximum rate of ascent / descent, and maximum tilt angle. The three-dimensional wind field and wind speed at each arc length position s along the path are given as follows:

[0040] ;

[0041] in, This represents the wind speed in a three-dimensional wind field.

[0042] The ground speed of the drone is The corresponding airspeed is This explicitly reflects the impact of tailwind, headwind, and crosswind on the motion state. Energy consumption is obtained by integrating flight power over time. Power is modeled as a function of variables such as airspeed, vertical velocity, attitude, and load, so that each candidate path can be mapped to corresponding energy consumption and feasibility constraints under a given wind field, providing a foundation for subsequent optimization solutions.

[0043] According to the Galilean transformation, the airspeed of the drone relative to the air... With three-dimensional wind field and wind speed The relationship is:

[0044] ;

[0045] Will Considered as the only continuous decision variable on the candidate edge. Given the local wind field obtained by sampling along the edge, the airspeed, required thrust, and power consumption can be expressed as: The function, and in Internally solve a one-dimensional optimization problem to find a feasible speed that minimizes energy or energy consumption per unit distance while satisfying flight envelope constraints such as airspeed limit, thrust / bank angle, vertical velocity, and propulsion power; if no feasible speed exists... If the edge is not found, then that edge is removed from the graph.

[0046] The rotor aerodynamic power consumption and battery energy consumption models are as follows:

[0047] For multi-rotor platforms, a classic profile-induced-parasitic triple superposition model is used to characterize propulsion aerodynamic power. It is airspeed and quality Functions:

[0048] ;

[0049] in, This refers to the profile power during hovering. The linear velocity of the propeller tip; The induced power is related to the thrust T and the induced velocity. Related, This is the equivalent organism parasitic drag coefficient. For rotor realism, The total disk area of ​​the multi-rotor This refers to air density.

[0050] The battery energy consumption model is as follows:

[0051] ;

[0052] in, For battery output power, To increase aerodynamic power, To improve link efficiency, To add power to airborne equipment.

[0053] For candidate edge e, let its edge length be... The flight time and battery power consumption are as follows:

[0054] ;

[0055] ;

[0056] in, Let e ​​be the flight time of candidate edge e. For the battery energy consumption of candidate edge e, Indicates the speed of the ground The airspeed, determined together with the local wind speed vector, is 3600, which is used to convert watt-seconds to watt-hours.

[0057] In addition to cruising, the mission at the terminal also includes a hovering alignment and unloading service process. Let the terminal hovering time be... Then service energy modeling for:

[0058] ;

[0059] in, To improve link efficiency; For constant power of airborne equipment, constant power parameters are given; Here, represents the aerodynamic power during hovering, and represents the profile power during hovering. and induced power sum:

[0060] ;

[0061] ;

[0062] ;

[0063] in, The induced velocity is the velocity in the hovering state, which represents the downwash velocity of the airflow required for the rotor to generate a balance force. Hovering induced power represents the aerodynamic power required by the rotor to generate lift and overcome induced losses while hovering. This is the induced power correction factor, used to account for non-ideal induced losses.

[0064] A given path consists of several sets of graph edges. Composition, total battery energy for the mission for:

[0065] ;

[0066] in, Let e ​​be the energy consumed by the candidate edge.

[0067] Consider improving link efficiency and additional loads such as airborne equipment Battery-side propulsion power for:

[0068] ;

[0069] This power consumption model unifies multirotor aerodynamics, climb motion, and battery characteristics within the same framework, providing comparable energy assessments across different paths and velocity profiles.

[0070] Step 5: Construct an energy-time co-optimization model based on flight safety constraints. Using the weighted sum of normalized flight time and normalized energy consumption as the objective function, the preference between time priority and energy priority is adjusted through task weight λ. Turbulent kinetic energy threshold, maximum tilt angle, power, effective crosswind, velocity envelope, and energy budget constraints are explicitly introduced. The model is as follows:

[0071] Find the optimal path given a starting point and an ending point. This makes the total cost function Minimum:

[0072] ;

[0073] in, These represent preference weights. and For dimensionless time and energy, defined as follows.

[0074] Since time (seconds) and energy (watt-hours) typically differ in magnitude by 1-2 orders of magnitude, direct weighting would skew the optimization objective towards the larger value. Therefore, on a global scale, the set of edges corresponding to a path from the starting point to the ending point... To explicitly characterize the trade-off between time and energy consumption, dimensionless time and energy are introduced:

[0075] ;

[0076] in, For candidate edge energy consumption, The flight time of the candidate edge. and Typical flight time and typical energy consumption per unit path length were set respectively.

[0077] During the search process, when the algorithm examines an edge e, its cost is not a pre-calculated fixed value, but is obtained in real time by solving the following sub-optimization problem:

[0078] ;

[0079] in, This indicates the preference weight between time and energy consumption. The larger the value, the more important the task time is. The smaller the value, the more emphasis is placed on energy consumption; For the one-way flight distance, if the typical flight time per unit path length is used... and typical energy consumption After normalization, we have , , For the normalized flight time, This represents the normalized energy consumption.

[0080] To ensure the feasibility and safety of flight missions, the planning process must strictly meet the following hard physical constraints:

[0081] (1) Maximum tilt angle constraint: , The flight tilt angle during flight. The maximum flight tilt angle limits the aircraft's maneuverability in strong crosswinds;

[0082] (2) Power constraint: , Maximum propulsion power to prevent battery over-discharge or motor overheating;

[0083] (3) Through effective crosswinds Constrain crosswind accessibility to ensure that, given crosswind and safety margin There is still a feasible airspeed component that can offset the crosswind, therefore the effective crosswind is less than the maximum ground speed. : ;

[0084] (4) Speed ​​envelope: The airspeed must be less than the maximum airspeed. : Ground speed must meet ;

[0085] (5) Energy Budget: Total energy consumption must meet ; Energy required for the return journey.

[0086] Step 6: Construct a K-Nearest Neighbors graph (KNN) under safety constraints, where K represents the number of nearest neighbor candidate nodes connected to each discrete programming node. Then, use the wind speed vector field predicted by PIPT-AS... and turbulent kinetic energy field Mapped to a discrete urban planning grid, candidate edges are generated based on a fixed connectivity K, and candidate edges that intersect with building obstacles or whose turbulent kinetic energy exceeds a threshold are pruned.

[0087] Step 7: Perform edge-level dynamic ground velocity optimization. For each candidate edge in the search, ground speed is used as a continuous decision variable, and the optimal ground speed with the minimum energy-time joint cost is calculated within the upper and lower bounds of ground speed. If no feasible speed is found, the edge is discarded. For each candidate edge in the KNN graph generated under the safety constraints in step 6, the flight ground speed is used as a continuous decision variable, and the optimal ground speed with the minimum energy-time joint cost is calculated within the given upper and lower bounds of ground speed. For candidate edge e ij Its starting point is planning node i, its ending point is planning node j, and its side length is d. ij Local wind speeds are obtained from the wind speed vector field predicted by PIPT-AS. , , Interpolation is used to calculate the airspeed, flight time, and energy consumption of the edge. If no feasible speed satisfying the flight constraints exists within the allowable ground speed range for the candidate edge, the candidate edge is deleted; if a feasible speed exists, the optimal ground speed of the edge is recorded. Flight time Energy consumption and the corresponding edge cost c ij This step transforms each candidate edge from a simple geometric connection into a weighted candidate edge with optimal ground speed, flight time, energy consumption, and overall cost, providing edge-level optimization results for subsequent path search.

[0088] Step 8: Execute Search. Embed edge-level optimization costs. The graph search is used, and a physical information heuristic function is employed to guide the search, enabling path geometry selection and ground velocity profile optimization to be completed within the same framework. Energy-time optimization for edge integration The algorithm, short for Edge Optimization Integrated Energy-Time A-star search Algorithm, where EOI stands for Edge Optimization Integrated and ET stands for Energy-Time. This refers to the A-star search algorithm. After constructing safe candidate edges and optimizing edge-level ground velocities, EOI-ET-A uses the safe KNN graph obtained in step 6 as the search space and directly calls the optimal edge-level ground velocity obtained in step 7 during the search process. and edge cost This allows path geometry selection and ground velocity optimization to be completed within the same search framework.

[0089] Specifically, When expanding from the current node n to the adjacent node m, instead of calculating the edge cost using only a fixed velocity or geometric distance, the optimal ground velocity corresponding to the candidate edge is read. Flight time (t), energy consumption And add it to the path cumulative cost g(n), where The cumulative cost when the neighboring node is m:

[0090] ;

[0091] in, The algorithm calculates the optimal edge cost. Simultaneously, it uses the heuristic function h(m) to estimate the remaining energy-time cost from the current node m to the target point, constructs a node evaluation function f(m) = g(m) + h(m), and prioritizes expanding nodes with smaller evaluation functions. By integrating the edge-level ground velocity optimization results into the edge expansion process of A-search, h(m) represents the estimated remaining energy-time cost from node m to the target point, and f(m) represents the comprehensive evaluation function value of node m. It can simultaneously optimize path morphology, flight speed, total energy consumption, and mission time while meeting obstacle avoidance and turbulent kinetic energy safety constraints.

[0092] Step 9: Output planning results. Output the three-dimensional waypoint sequence, optimal ground speed for each segment, total energy consumption, mission time, energy consumption per unit distance, maximum turbulent kinetic energy along the route, and feasibility assessment.

[0093] This invention provides a specific embodiment of an eVTOL trajectory planning method based on wind field prediction and collaborative optimization. To make the technical solution and beneficial effects of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings, taking a single eVTOL delivery task in an urban building complex as an example:

[0094] Step 1: Determine the basic parameters of the eVTOL platform and the urban mission scenario. This example uses eVTOL parameters based on the publicly available technical specifications of the DJI Matrice300 RTK, setting the payload (including battery) to 8.3 kg, rotor radius to 0.105 m, nominal battery capacity to 274 Wh, safety margin to 27.4 Wh, usable energy to 246.6 Wh, maximum tilt angle to 25°, maximum airspeed to 23 m / s, ground speed range to 3 m / s to 15 m / s, and takeoff / landing hover time to 60 s.

[0095] The task scenario is a 3D computational domain of urban building complexes, such as... Figure 3 As shown, the starting point coordinates are [50, 400, 60] m, and the ending point coordinates are [850, 500, 80] m. The starting and ending points differ in both planar location and height, covering the area affected by street valley bypass and building wake, and are used to compare path topology, energy consumption, and safety constraint responses under different incoming flow conditions. Figure 4 As shown in (a) and (b). The main parameters of the implementation case are shown in Table 1.

[0096] Table 1. Main parameters for path planning

[0097]

[0098] Step 2: Construct a 3D wind field dataset for the city. This example uses the turbulent flow interface in COMSOL Multiphysics 6.2 to perform a 3D steady numerical simulation of the wind field around a fixed urban building complex; no-slip conditions are applied to the ground and building surfaces, symmetrical or slip conditions are applied to the top and side boundaries, and isobaric or zero-normal-stress conditions are applied to the outlet.

[0099] The inlet velocity is jointly specified by the incoming flow velocity, turbulence intensity, and turbulence length scale. Each sample set contains the velocity component and turbulent kinetic energy at a uniform grid point. The samples are divided into training, validation, and test sets and used as wind field inputs for training the PIPT-AS surrogate model and subsequent routing experiments.

[0100] Step 3: Train the PIPT-AS wind field proxy model. The five-dimensional incoming flow parameters are first encoded into global features through two layers of MLP, and the three-dimensional coordinates are encoded into positions through multi-scale Fourier features. The two are then concatenated at the point level and globally modeled using a fused MLP and a Transformer block with LayerScale.

[0101] Considering that the mainstream velocity component is the most critical to the flight aerodynamic characteristics, a self-attention-gated enhancement module is set after the main output to reweight the characteristics of this component; the network finally outputs u, v, w and k through component-specific output heads, among which the mainstream velocity component and turbulent kinetic energy use deeper prediction heads to improve accuracy.

[0102] During the training phase, an adaptive weighted loss and a two-stage training strategy are adopted, which combines L1 / L2 loss, component variance weights, local velocity weights, and first-order difference smoothing regularization terms in the normalized space; the optimizer adopts AdamW and is combined with cosine annealing learning rate scheduling.

[0103] During the inference phase, given new incoming flow parameters, PIPT-AS can generate a complete 3D wind field (u,v,w,k) in a single forward propagation, which is then used for safety graph construction and wind-aware planning algorithm evaluation. The wind field prediction results are as follows: Figure 5 As shown in Table 2, the comparison results between PIPT-AS and the benchmark model indicators are presented.

[0104] Table 2 Comparison of PIPT-AS and Benchmark Model Indicators

[0105]

[0106] Step 4: Establish the eVTOL power consumption model. The airspeed vector of the UAV relative to the air is determined by the difference between the ground speed vector and the local wind speed vector. The airspeed, climb status, load, and rotor parameters together determine the propulsion power.

[0107] The rotor aerodynamic power consumption adopts a superposition model of induced power, shape drag power and parasitic power. The battery power is corrected by the propulsion link efficiency and auxiliary equipment power consumption. The total mission energy is obtained by the sum of the energy consumption of each map edge segment and the terminal hovering service energy.

[0108] Step 5: Construct the energy-time co-optimization function. For a single flight segment, the cost function is: ,in λ represents the task preference weight; as λ increases from 0 to 1, the optimization strategy gradually shifts from energy-priority to time-priority bias.

[0109] The safety constraints include turbulent kinetic energy threshold constraints, maximum tilt angle constraints, power constraints, effective crosswind constraints, airspeed and ground speed envelope constraints, and battery energy budget constraints; if any constraint is not met, the corresponding flight segment will not be included in the feasible search graph.

[0110] Step 6: Construct a safe KNN search graph. Discretize the 3D city space into fixed grid nodes, establish a candidate edge set according to the KNN criterion, and remove obstacles from edges that intersect with building geometry; then remove high-risk edges according to the turbulence threshold rule to obtain a safe graph.

[0111] Step 7: Implement edge-level dynamic ground velocity optimization. In each... When expanding nodes, the local wind vector and turbulent kinetic energy are read for candidate edges, the feasible interval of ground speed that satisfies the flight envelope is calculated, and the ground speed with the minimum energy-time joint cost is solved within this interval.

[0112] Compared to a fixed cruise speed strategy, the edge-level dynamic ground speed optimization can increase ground speed in the downwind area and take advantage of local favorable wind fields, and adjust speed in segments in the upwind or high turbulence areas, thereby avoiding manual speed adjustment and parameter mismatch.

[0113] Step 8: Execute Search. The heuristic function integrates the spatial distance from the current node to the target node, the reference velocity, the minimum power, and the favorable wind speed assumption to form a physical information heuristic that satisfies the acceptability; the search process uses g(n)+h(n) as the sorting basis, gradually expands the safety graph, and records the edge-level optimal ground velocity.

[0114] Step 9: Output the final planning result, such as Figure 6 As shown. The output results include a safe and feasible three-dimensional trajectory, optimal ground speed for each segment, total energy consumption, flight time, energy consumption per unit distance, maximum turbulent kinetic energy exposure along the route, whether the threshold is exceeded, and parameter sensitivity analysis results.

[0115] In the embodiment, PIPT-AS achieved a MAE of 0.1699, an RMSE of 0.3536, an R² of 0.9883, and an NMAE of 2.95% on the same test set; the NMAEs of velocity components u, v, and w were 2.05%, 2.86%, and 0.24%, respectively, and the NMAE of turbulent kinetic energy k was 2.42%.

[0116] In the comparison of statistical path planning and The average total energy consumption is approximately 136Wh, and the average task time is approximately 130s. At λ=0.5, the average total energy consumption is 77.44Wh, the average task time is 53.0s, and the energy consumption per unit distance is 77.38Wh / km.

[0117] Table 3 Comparison of Statistical Performance of Different Programming Algorithms

[0118]

[0119] In Case 146, risk neutrality Although the energy consumption was reduced to 85.60Wh and the time was compressed to 46.3s, the maximum turbulent kinetic energy of 4.481 exceeded the threshold of 3.270. The method of the present invention achieves 82.47Wh and 44.1s under the condition that the maximum turbulent kinetic energy of 3.242 does not exceed the threshold.

[0120] In Case 1900, with an inflow velocity of 10.0 m / s and a wind direction of 4°, the method of this invention controls the maximum turbulent kinetic energy to 1.925 while maintaining feasibility. Compared with the risk-neutral approach, it only increases energy consumption by 0.15 Wh and reduces peak turbulent exposure by approximately 29.6%.

[0121] In Case 572, with an inflow velocity of 8.5 m / s and a wind direction of 184°, the method of this invention reduces the maximum turbulent kinetic energy from 1.082 in the risk-neutral scheme to 0.939, a reduction of 13.2%, while compared to It shortens the flight time by 23.2 seconds and the path by approximately 95 meters.

[0122] Table 4 Safety Comparison of Typical Wind Farm Cases

[0123]

[0124] Parameter sensitivity analysis shows that λ can stably adjust energy and time preferences across the entire range. When the graph connectivity K increases from 8 to 16, the path quality improves significantly. However, when K exceeds 16, the energy efficiency gains tend to saturate and the computational cost continues to increase. Therefore, K=16 is the configuration that balances solution quality and computational efficiency.

[0125] Ground speed strategy analysis shows that traditional geometric shortest path and fixed speed baseline are highly sensitive to manually set cruise speeds; the method of this invention does not require manual traversal of speeds, and can automatically solve the optimal ground speed for each edge based on local wind vectors and achieve stable energy efficiency performance under strict safety constraints.

[0126] The above description is merely a preferred embodiment of the present invention, based on the PIPT-AS wind field prediction, eVTOL power consumption modeling, security graph construction, edge-level dynamic ground velocity optimization, and other methods described in the present invention. All equivalent technical solutions found through the search fall within the protection scope of this invention.

Claims

1. An eVTOL trajectory planning method based on wind field prediction and collaborative optimization, characterized in that, Includes the following steps: Step 1: Determine the eVTOL platform parameters and urban delivery mission scenario, including: aircraft mass, rotor radius, propulsion efficiency, auxiliary power, battery capacity, safety margin, maximum tilt angle, maximum airspeed, ground speed upper and lower limits, takeoff and landing hovering time, as well as the geometry of urban building clusters, 3D planning mesh, starting point and ending point; Step 2: Construct a three-dimensional wind field dataset of the urban building complex under multiple sets of incoming flow conditions, including incoming flow velocity, turbulence intensity and turbulence length, as well as the three-dimensional velocity components and turbulent kinetic energy under these conditions; Step 3: Train and build a wind field proxy model using a 3D wind field dataset, so that it can predict the city's 3D wind speed vector field and turbulent kinetic energy field based on the incoming flow parameters and spatial coordinates on the 3D planning grid; Step 4: Construct eVTOL flight kinematics model, rotor aerodynamic power consumption model and battery energy consumption model, and establish the mapping relationship from local wind vector, ground speed, airspeed to battery-end propulsion power and segment energy consumption; Step 5: Construct an energy and time co-optimization model based on flight safety constraints, with the weighted sum of normalized flight time and normalized energy consumption as the objective function, and adjust the preference between time priority and energy consumption priority through task weights; Step 6: Construct a safe K-nearest neighbor search graph; map the predicted 3D wind speed vector field and turbulent kinetic energy field to the city's 3D planning grid, generate candidate edges based on fixed connectivity, and prune candidate edges that intersect with building obstacles or whose turbulent kinetic energy exceeds the threshold to obtain a safe K-nearest neighbor search graph; Step 7: Perform edge-level dynamic ground velocity optimization. For each candidate edge in the safe K-nearest neighbor search graph, use ground velocity as a continuous decision variable and solve for the optimal ground velocity and edge cost within the upper and lower bounds of the energy-time joint cost. If there is no feasible velocity, then remove the edge. Step 8: Execute The search uses a safe K-nearest neighbor search graph as the search space and calls the optimal ground velocity and edge cost obtained in step 7 during the search process, so that path geometry selection and ground velocity optimization are completed within the same search framework. Step 9: Output planning results; output three-dimensional waypoint sequence, optimal ground speed for each segment, total energy consumption, mission time, energy consumption per unit distance, maximum turbulent kinetic energy along the route, and feasibility assessment.

2. The eVTOL trajectory planning method based on wind field prediction and collaborative optimization according to claim 1, characterized in that, In step 3, the five-dimensional incoming flow parameters and spatial coordinates are used as inputs to the wind field surrogate model to predict the u, v, w velocity components and turbulent kinetic energy k, where u, v, and w represent the velocity components along the x, y, and z directions at the spatial coordinate points, as detailed below: Given five-dimensional incoming flow parameters And spatial coordinates on a fixed computational fluid dynamics grid: ,in, For points in the three-dimensional wind field space, Let N be the coordinates of a point in the 3D wind field space, and N be the total number of points in the grid within the wind field space. A mapping is learned based on the wind field proxy model: ； To obtain an approximate numerical solution for computational fluid dynamics. ;in, This indicates that it contains learnable parameters. The proxy model, This indicates the learnable parameters of the model. This represents the model prediction results; five-dimensional incoming flow parameters. First, the global features are encoded using two layers of MLP. Then, the 3D coordinates are encoded into positional features through multi-scale Fourier transform. After concatenating the two types of features at the point level, global modeling is performed, and finally, the output is... ,in, , , These represent the velocity components in the x, y, and z directions predicted by the model at all grid sampling points, respectively. This represents the turbulent kinetic energy field predicted by the model at all grid sampling points; for the i-th spatial sampling point, its corresponding output is denoted as... , , , .

3. The eVTOL trajectory planning method based on wind field prediction and collaborative optimization according to claim 2, characterized in that, Five-dimensional incoming flow parameters First, it is encoded into 192-dimensional global features through two layers of MLP, where, These are the inflow velocity components in the x, y, and z directions. For turbulence intensity, It is the length scale of turbulence.

4. The eVTOL trajectory planning method based on wind field prediction and collaborative optimization according to claim 2, characterized in that, Four output components Weights are adaptively assigned based on their standard deviation in the dataset. At the same time, the mainstream velocity x-direction component and turbulent kinetic energy Apply additional amplification factor Each sampling point is assigned a different velocity weight based on its local velocity magnitude, which falls into one of the three categories: low, medium, or high speed. Let the difference regularity of the predicted velocity between adjacent grid points be denoted as . and using hyperparameters By controlling its weight in the total loss, the overall loss... Written as: ； in, The basic loss.

5. The eVTOL trajectory planning method based on wind field prediction and collaborative optimization according to claim 4, characterized in that, Step 4 specifically involves: The flight trajectory of a drone is represented as a continuous spatial curve over a time interval. Given the three-dimensional wind field and wind speed at each arc length position s along the path: ； in, Represents the wind speed in a three-dimensional wind field; airspeed of a drone relative to the air With three-dimensional wind field and wind speed The relationship is: ； in, Let eVTOL be the ground velocity vector relative to the ground. Considered as the only continuous decision variable on the candidate edge; given the local wind field obtained by sampling along the edge, the airspeed, required thrust, and power consumption are expressed as... The function, and in Internally solve a one-dimensional optimization problem to find a feasible speed that minimizes energy or energy consumption per unit distance while satisfying constraints on airspeed, thrust / bank angle, vertical velocity, and propulsion power flight envelope. If no feasible speed exists... If the edge is not found, then that edge is removed from the graph.

6. The eVTOL trajectory planning method based on wind field prediction and collaborative optimization according to claim 5, characterized in that, The rotor aerodynamic power consumption model is as follows: Propulsion aerodynamic power It is airspeed and aircraft mass Functions: ； in, This refers to the profile power during hovering. The linear velocity of the propeller tip; The induced power is related to the thrust T and the induced velocity. Related, This is the equivalent organism parasitic drag coefficient. For rotor realism, The total disk area of ​​the multi-rotor air density; The battery energy consumption model is as follows: ； in, For battery output power, To increase aerodynamic power, To improve link efficiency, Add power to airborne equipment; For candidate edge e, let its edge length be... The flight time and battery power consumption are as follows: ； ； in, Let e ​​be the flight time of candidate edge e. For the battery energy consumption of candidate edge e, Indicates the speed of the ground The airspeed magnitude is determined jointly by the local wind speed vector; In addition to cruising, the aircraft also includes a hovering process; let the hovering time of the aircraft be... Then service energy modeling for: ； in, To improve link efficiency; Add power to airborne equipment; Here, represents the aerodynamic power during hovering, and represents the profile power during hovering. and induced power sum: ； ； ； in, The induced velocity is the velocity in the hovering state, which represents the downwash velocity of the airflow required for the rotor to generate a balance force. Hover induced power represents the aerodynamic power required by the rotor to generate lift and overcome induced losses while hovering. This is the induced power correction factor; A given path consists of several sets of graph edges. Composition, total battery energy for the mission for: ； in, Let e ​​be the energy consumed by the candidate edge.

7. The eVTOL trajectory planning method based on wind field prediction and collaborative optimization according to claim 6, characterized in that, In step 5, the energy and time co-optimization model is as follows: Find the optimal path given a starting point and an ending point. This makes the total cost function Minimum: ； in, For preference weights; where, and For dimensionless time and energy: ； in, For candidate edge energy consumption, For candidate edge flight time; and Typical flight time and typical energy consumption per unit path length were set respectively; During the search process, when considering an edge e, its cost is obtained in real time by solving the following sub-optimization problem: ； in, This indicates the preference weight between time and energy consumption. The larger the value, the more important the task time is. The smaller the value, the more emphasis is placed on energy consumption; For the one-way flight distance, if the typical flight time per unit path length is used... and typical energy consumption After normalization, we have , , For the normalized flight time, This represents the normalized energy consumption.

8. The eVTOL trajectory planning method based on wind field prediction and collaborative optimization according to claim 7, characterized in that, To ensure the feasibility and safety of flight missions, the planning process must meet the following hard physical constraints: Maximum tilt angle constraint: , The flight tilt angle during flight. This is the maximum flight tilt angle; Power constraints: , Maximum propulsion power; Through effective crosswinds Constraining crosswind accessibility ensures that, given crosswind and safety margin There is still a feasible airspeed component that can offset the crosswind, therefore the effective crosswind is less than the maximum ground speed. : ; Speed ​​envelope: Airspeed must be less than maximum airspeed : Ground speed must meet ; Energy Budget: Total Energy Consumption Meets Standards ; Energy required for the return journey.

9. The eVTOL trajectory planning method based on wind field prediction and collaborative optimization according to claim 7, characterized in that, Step 7 specifically involves: for candidate edge e ij Its starting point is planning node i, its ending point is planning node j, and its side length is d. ij Local wind speeds are obtained from the wind speed vector field predicted by PIPT-AS. , , Interpolation is used to calculate the airspeed, flight time, and energy consumption of the candidate edge. If no feasible speed satisfying the flight constraints exists within the allowable ground speed range, the candidate edge is deleted. If a feasible speed exists, the optimal ground speed of the candidate edge is recorded. Flight time Energy consumption and the corresponding edge cost .

10. The eVTOL trajectory planning method based on wind field prediction and collaborative optimization according to claim 9, characterized in that, Step 8 specifically involves: when expanding from the current node n to the adjacent node m, reading the optimal ground speed, flight time, and energy consumption corresponding to the candidate edge, and adding them to the cumulative path cost g(n), where... The cumulative cost when the neighboring node is m: ； in, The optimal edge cost is determined by the heuristic function h(m). Simultaneously, the remaining energy-time cost from the current node m to the target point is estimated using the heuristic function h(m). The node evaluation function f(m) = g(m) + h(m) is constructed, and nodes with smaller evaluation functions are expanded first. h(m) represents the estimated value of the remaining energy-time cost from node m to the target point, and f(m) represents the comprehensive evaluation function value of node m.

Images