Energy consumption prediction and path planning method for heavy-load unmanned aerial vehicle hoisting operation

Abstract

This invention relates to the field of unmanned aerial vehicle (UAV) application technology, specifically a method for energy consumption prediction and path planning in heavy-payload UAV hoisting operations. The method includes the following steps: loading the original flight data and mission target point data of the heavy-payload UAV; constructing a hybrid energy consumption model coupled with physical mechanisms and residual correction; introducing a heavy-load unbalanced weighted training mechanism to train the hybrid energy consumption model, assigning significantly higher weights to samples exceeding a preset heavy-payload threshold; using an improved K-Means clustering algorithm to divide the spatial domain of discrete operation points, introducing the hybrid energy consumption model as an evaluation function into the clustering iterative closed loop to achieve region clustering based on real operation point corrections; and calling the trained hybrid energy consumption model to perform energy simulation, performing path planning and search based on the limit efficiency of a safe power threshold. This invention effectively solves the safety hazard of inaccurate endurance estimation in heavy-payload operations and maximizes hoisting efficiency within a single takeoff and landing cycle.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) application technology, and in particular to a method for energy consumption prediction and path planning for UAV hoisting operations with large payloads. Background Technology

[0002] With the development of smart agriculture and the low-altitude economy, agricultural drones have expanded from simple spraying operations to heavy-payload (80kg-149.9kg) lifting operations. The demand for heavy-payload drones is surging in scenarios such as agricultural product harvesting and transfer, transportation of field equipment and consumables, and centralized lifting and distribution of agricultural supplies. Heavy-payload drone lifting operations differ from conventional light logistics, facing unique challenges:

[0003] First, current energy consumption models fall short when handling heavy loads. Most existing energy consumption calculation methods are based on a simple linear relationship P=Fv, which, while applicable under light loads, shows significantly increased errors under heavy loads, easily leading to flight accidents due to sudden battery depletion. Energy consumption under heavy loads is extremely sensitive to load conditions, especially during high-load operation, where motor efficiency is low and nonlinear heat loss rises rapidly with increasing battery discharge resistance. These factors make it difficult for existing models to accurately predict energy consumption under full load and low battery conditions, resulting in significant deviations in endurance estimations. For example, near the maximum takeoff weight of the drone, the rotor must accelerate to maintain necessary lift, which not only increases induced drag and parasitic drag (which are powers of 1.5 or higher with mass) but also exacerbates energy consumption.

[0004] Furthermore, during sustained high-load tasks, such as transporting a bag of fertilizer from the foot of a mountain to a work site at the summit, or relocating equipment, drones typically maintain maximum load throughout the entire process, unable to gradually reduce the burden as in package delivery. This further limits the operational flexibility and efficiency of drones. For agricultural work sites distributed in complex terrain, the long distances between each lifting point, coupled with limited battery life under heavy loads (typically only lasting 10 to 20 minutes), make efficient use of a single charge a key challenge. Current technology is insufficient in maximizing battery power utilization, making it difficult to determine the maximum number of work sites that can be covered in a single flight while ensuring a safe return, leading to frequent takeoffs and landings for battery swaps, which significantly reduces work efficiency. Traditional path planning algorithms such as TSP (Traveling Salesman Problem) or VRP (Vehicle Routing Problem) often focus on the shortest path, ignoring the energy consumption differences during ascent and level flight under different load conditions, failing to consider energy consumption fluctuations caused by load changes during the "pickup-transportation-return" process, and the nonlinear impact of load on energy consumption. In heavy cargo transportation, the energy consumption of fully loaded flight is much higher than that of empty flight. Traditional algorithms tend to plan long-distance fully loaded routes, reducing the amount of tasks that a drone can complete in each sortie, and may even lead to forced landings due to battery depletion. Traditional pure data-driven models lack physical constraints, and in cases with a small number of extremely heavy load samples, they are prone to predicting results that violate physical principles, making them unsuitable for direct application in heavy-load agricultural scenarios.

[0005] Therefore, developing a planning method that can accurately describe the nonlinear power consumption model under heavy loads and can also "squeeze out" the maximum number of operation points within a limited power supply based on the power consumption model is of great significance for the large-scale application of heavy-load UAVs. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of existing technologies, such as linear models not being applicable to heavy-load conditions and path planning not considering the actual heavy-load energy consumption limitations. It provides an energy consumption prediction and path planning method for heavy-load UAV lifting operations. Through accurate energy consumption modeling and multi-stage clustering optimization, it maximizes the operation efficiency under a single battery charge. It aims to solve the range anxiety and safety hazards caused by inaccurate energy consumption prediction under heavy-load conditions, and achieves the effects of reducing operational energy consumption and improving operational safety.

[0007] The present invention employs the following technical solution in its embodiments: an energy consumption prediction and path planning method for heavy-load UAV hoisting operations, comprising the following steps:

[0008] S1. Initialization: Load the raw flight data and mission target point data of the high-payload UAV.

[0009] S2. Construct a hybrid energy consumption model that couples physical mechanisms with residual correction as the energy calculation kernel for path planning;

[0010] S3. Introduce a heavy-load unbalanced weighted training mechanism to train the constructed hybrid energy consumption model. Give significantly higher weights to samples that exceed the preset heavy-load threshold, forcing the model to prioritize fitting the residual fluctuations under heavy load during the training process.

[0011] S4. An improved K-Means clustering algorithm is used to divide the spatial domain of discrete work points. The actual work point closest to the centroid in each cluster is selected as the collection point or supply center of the corresponding region. The hybrid energy consumption model is introduced into the clustering iteration closed loop as the evaluation function to realize regional clustering based on real work points.

[0012] S5. Call the trained hybrid energy consumption model to perform energy simulation. Use a greedy truncation strategy based on real energy consumption simulation, a greedy search strategy based on forward energy simulation and safe truncation, and perform path planning and search based on the limit efficiency of the safe power threshold.

[0013] Compared with the prior art, the technical effects achieved by the present invention include:

[0014] 1. This invention maximizes operational efficiency under a single battery charge by using precise energy consumption modeling and multi-stage clustering optimization. It aims to solve the range anxiety and safety hazards caused by inaccurate energy consumption prediction under heavy load conditions, and achieves the effects of reducing operational energy consumption and improving operational safety.

[0015] 2. In the model optimization process, the heavy-load unbalanced weighted training mechanism of this invention forces the loss function to amplify the prediction error of heavy-load conditions by many times, thereby ensuring that the model has extremely high prediction confidence in the high-risk operation range of heavy loads above 80kg.

[0016] 3. This invention adopts an improved clustering algorithm, which clusters the work area based on the actual work point correction. It can automatically find a balance point, which avoids the division being too fine, causing the drone to frequently travel back and forth to the base, and avoids the division being too coarse, causing the workload of a single area to exceed the battery endurance limit. This makes the workload of each work area naturally match the maximum single-run work capacity of the drone. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the energy consumption prediction and path planning method for large-payload UAV hoisting operations in this embodiment of the invention;

[0018] Figure 2 This is a schematic diagram of the structure of the hybrid energy consumption model constructed in the embodiments of the present invention;

[0019] Figure 3This is a flowchart of the training process for the hybrid energy consumption model using the unbalanced load weighted training mechanism introduced in this embodiment of the invention.

[0020] Figure 4 This is a schematic diagram of the region clustering structure based on real work points in an embodiment of the present invention;

[0021] Figure 5 This is a logic diagram for planning the efficiency of a single extreme hoisting operation in an embodiment of the present invention;

[0022] Figure 6 This is a schematic diagram of the overall efficient hoisting process in an embodiment of the present invention. Detailed Implementation

[0023] Taking the lifting operation of a large-payload quadcopter UAV in a complex mountainous environment as an example, the specific implementation of the present invention will be described in detail with reference to the accompanying drawings and embodiments.

[0024] Example

[0025] This embodiment provides a method for energy consumption prediction and path planning for heavy-load UAV hoisting operations, such as... Figure 1 As shown, it includes the following steps:

[0026] S1. Initialization: Load the raw flight data and mission target point data of the high-payload UAV.

[0027] The raw flight data includes state parameters such as timestamps, voltage, current, horizontal velocity, vertical velocity, flight altitude, flight phase, and total aircraft weight; the mission target point data includes a set of three-dimensional coordinates of N discrete target points within the area to be operated. And the cargo loading requirements for each target point.

[0028] After initialization, the system process is divided into parallel model building branches and path planning branches for processing.

[0029] S2. Construct a hybrid energy consumption model that couples physical mechanisms with residual correction.

[0030] This step aims to build and train a large-load hybrid energy consumption model as the energy calculation kernel for subsequent path planning.

[0031] First, the raw flight data undergoes data preprocessing and feature engineering to extract features including total mass. 1.5th power and horizontal flight speed Key physical feature vectors including the cube of . :

[0032] ;

[0033] in, Characterizing the induced resistance factor, Characterizing parasitic resistance factor, This is the acceleration due to gravity.

[0034] To address the nonlinear relationship between aerodynamic drag and power under heavy load conditions, this embodiment constructs a two-layer parallel prediction architecture of "physical baseline + residual correction". Specifically, the constructed hybrid energy consumption model includes a physical baseline model and a weighted residual neural network. In the physical baseline model, linear regression is used to fit physical characteristics, outputting the physical baseline power. In a weighted residual neural network, the nonlinear bias and nonlinear error of the unbalanced weighted training mechanism are fitted using the neural network, and the residual power is output. ,like Figure 2 As shown.

[0035] More specifically, the physical baseline model uses linear regression to fit the basic lift and drag power consumption. Input features include total mass, flight speed, and climb power; the output is the physical baseline power. The calculation formula is as follows:

[0036] ;

[0037] Where m is the total mass (i.e. the load mass), and the exponent 1.5 indicates the nonlinear growth of induced drag under heavy loads; For horizontal flight speed, Indicates the parasitic resistance factor; Vertical flight speed, It is the acceleration due to gravity. Characterizes the power output during vertical climb; All are regression coefficients. The physical baseline model ensures that the prediction results conform to basic physical laws: the greater the load, the greater the power.

[0038] The weighted residual neural network employs a multilayer perceptron (MLP) neural network, specifically a fully connected neural network with three hidden layers (128-64-32 neurons), to predict nonlinear residual power that cannot be explained by the physical baseline model. Its input vector include:

[0039] ;

[0040] in This indicates the drone's current relative flight altitude. This indicates the current remaining percentage of battery power. A status label describing the current flight status.

[0041] The total power prediction output by the hybrid energy consumption model is .

[0042] S3. Introduce a heavy-load unbalanced weighted training mechanism to train the constructed hybrid energy consumption model.

[0043] Since there are far more light-load samples than heavy-load samples in actual flight data for heavy-load UAVs, conventional training can easily mask the prediction bias in the heavy-load range due to the overall average error. To address the problem that the scarcity of heavy-load samples in the training data causes the model to shift towards the light-load range, resulting in insufficient prediction accuracy under key operating conditions, this step designs a load-sensitive sample weighting strategy during the model training phase, namely, a cost-sensitive learning-based unbalanced weighted training mechanism for heavy-load systems (Weighted Heavy-Load Training).

[0044] like Figure 3 As shown, the unbalanced weighted training mechanism assigns significantly higher weights (e.g., 5 times) to samples exceeding a preset load threshold (e.g., 80 kg), forcing the model to prioritize fitting residual fluctuations under high loads during training, thereby improving prediction accuracy in high-load scenarios.

[0045] This embodiment embeds the overloaded unbalanced weighted training mechanism into the training closed loop of the hybrid energy consumption model. The specific implementation steps are as follows:

[0046] S31. Based on the heavy-load unbalanced weighted training mechanism, set a high-load judgment threshold for the load and define sample weights. The details are as follows:

[0047] ;

[0048] in The high load threshold for determining the load capacity is set to 80.0 kg in this embodiment; The penalty weight coefficient for high-load samples, i.e., high-multiplier weight, is set to 5.0 in this embodiment; For the first The total payload of the training samples, or simply the sample payload.

[0049] S32. Read the training samples from the original flight dataset, each training sample... It includes two key types of information: physical state characteristics and actual energy consumption labels.

[0050] Among them, the core of the physical state characteristics is the sample load. The actual energy consumption label is the actual power consumption. .

[0051] S33. Before calculating the loss function, an unbalanced weighted discriminator is introduced to scan the training samples in the original flight dataset one by one. Logical branch judgment is performed based on the sample load and the high load judgment threshold. Based on the judgment result, the training samples are assigned corresponding sample weights and the sample weight vector is output.

[0052] Specifically, the logical branch judgment process includes: loading the sample High load determination threshold with load capacity If a comparison is made, If the training sample is deemed a high-risk sample, and the corresponding data point belongs to the scarce operating condition of the motor under high load, then it is forcibly assigned a high weight. ;like If the training sample is determined to be a regular sample, and the corresponding data point belongs to a regular cruise condition, then it is assigned a standard weight. .

[0053] After evaluating each training sample in the original flight dataset, the output sample weight vector is: .

[0054] S34. Substitute the sample weight vector into the loss function of the hybrid energy consumption model and perform backpropagation training to optimize the weighted loss function and obtain the trained hybrid energy consumption model.

[0055] This embodiment uses the weighted mean squared error (MSE) as the optimization objective function. :

[0056] ;

[0057] in, The total predicted power is calculated using the physical baseline model and the weighted residual neural network. These are the model parameters.

[0058] By minimizing the weighted loss function, the final high-load hybrid model is trained, which serves as a validated high-precision hybrid energy consumption model and is encapsulated as a core computing component for real-time use in subsequent path planning.

[0059] The weighted training mechanism in this step amplifies the gradient contribution of heavy-load training samples by a factor of 5, forcing the model to prioritize fitting the nonlinear characteristics of the high-load range. This ensures that the model has extremely high prediction confidence in the high-risk operation range above 80kg. In other words, this mechanism forces the neural network to sacrifice some accuracy in the light-load range to fully optimize the prediction accuracy in the heavy-load range above 80kg, thereby ensuring the safety of hoisting operations.

[0060] Therefore, during model training, high-risk heavy-load samples can be automatically identified and assigned high loss weights, thus forcing the model to prioritize fitting the nonlinear characteristics under heavy-load conditions. The weights for high-load samples follow the cost-sensitive learning principle. In actual flight datasets, high-load samples account for approximately 20% of the total sample size, exhibiting a significant long-tail distribution problem. To eliminate the impact of data imbalance on the gradient and reflect the high safety cost of prediction errors under heavy-load conditions, this embodiment uses an inverse frequency weighting method combined with engineering experience risk factors. After multiple cross-validation experiments, the loss weight coefficient, i.e., the penalty weight coefficient, for high-load samples was finally set to 5.0.

[0061] S4. Introduce the hybrid energy consumption model as an evaluation function into the clustering iteration closed loop to achieve regional clustering based on real work points.

[0062] In the face of the problem of highly dispersed operation points and complex terrain in agricultural scenarios, simply relying on geometric clustering may result in planning points falling in inaccessible locations such as cliffs and water surfaces. This embodiment adopts the K-Means algorithm to divide the spatial domain of discrete operation points, but it is different from conventional clustering algorithms. Instead, it deeply couples the clustering results with the nonlinear energy consumption model by changing the evaluation criteria and decision logic of the algorithm.

[0063] The improved K-Means clustering algorithm uses a different approach: instead of using theoretical centroids, it forces the selection of the nearest existing work point within each cluster as the collection point or supply center for that region (the Medoid). This ensures that every planned transfer point is practically feasible, avoiding the awkward situation where the calculation is possible but the execution is not.

[0064] The region clustering structure corrected based on actual work points is as follows: Figure 4As shown. Unlike traditional static methods that rely solely on spatial geometric distance for clustering, this embodiment does not simply depend on geographical location for partitioning. Instead, it innovatively introduces a hybrid energy consumption model as an evaluation function into the clustering iterative closed loop, achieving region clustering based on real work points. This embodiment employs a region partitioning method based on full-process energy consumption simulation feedback. By simulating the entire "pickup-delivery" process under different partitioning dimensions, it maps the geometric Euclidean distance of physical space to energy consumption topological distance, thereby ensuring that the partitioned work areas are optimal in the energy dimension, rather than merely compact in the spatial dimension. After region partitioning, a Medoid optimization strategy based on real work points is adopted for each work cluster, forcibly selecting real work points within the cluster as collection centers, avoiding the problem of collection points being located in inaccessible areas and unable to land. Simultaneously, the Medoid optimization strategy based on real work points, combined with the hybrid energy consumption model, prioritizes the work point that minimizes the total weighted load distance within the work cluster as the core, minimizing the high-energy-consumption full-load flight range.

[0065] The regions divided using the improved clustering algorithm will automatically find a balance point, avoiding both overly fine divisions that cause drones to frequently travel to and from the base and overly coarse divisions that cause the workload of a single region to exceed the battery's maximum single-use capacity. This ensures that the workload of each work area is naturally adapted to the drone's maximum single-use capacity.

[0066] In other words, this step, by simulating the entire pickup-transportation process under different partitioning dimensions, constructs a mapping and optimization mechanism from physical Euclidean space to energy consumption topology space, specifically including the following steps:

[0067] S41. Define the energy consumption topology distance, reconstruct the spatial distance metric, nonlinearly map the physical geometric space to the energy consumption topology space, and establish the energy consumption topology graph.

[0068] Before performing cluster analysis, this step first reconstructs the spatial distance metric, completing the nonlinear mapping from physical geometric space to energy consumption topology space. Unlike traditional path planning which only calculates the geometric distance between two nodes, this embodiment defines energy consumption topology distance and utilizes a pre-trained high-load hybrid energy consumption model to transform the geometric Euclidean distance into energy consumption cost through mapping. This mapping is non-linear and depends on the current payload capacity of the drone. Strong correlation, the mapping formula is:

[0069] ;

[0070] in, For nodes With nodes The geometric Euclidean distance between them; This refers to the drone's cruising speed. This is the total power prediction value output by the hybrid energy consumption model.

[0071] Using this formula, the system can quantify the huge energy consumption difference when flying the same physical distance with a loaded load (e.g., 80kg) versus when unloaded, thus establishing an energy consumption topology map.

[0072] S42. Based on the improved K-Means clustering algorithm and energy consumption topology map, read the task point data for cluster analysis, and perform regional division and corresponding regional aerial collection point site selection correction.

[0073] First, in order to determine the optimal granularity of the work area division, the system reads the task point set and initiates clustering. The traversal loop, in which , The maximum allowed value for traversal is given; the discrete set of task points is divided into P geometric clusters using the improved K-Means clustering algorithm.

[0074] Then, to address the issue that the theoretical centroids generated by the K-Means clustering algorithm might be located in inaccessible areas (such as cliffs or water surfaces), real job point correction is performed. The geometric Euclidean distance from all real job points within a cluster to the theoretical centroid is calculated; the real job point with the smallest geometric Euclidean distance is selected as the aerial collection point (Medoid) for the corresponding region, ensuring the physical reachability of the relay station.

[0075] Finally, a closed-loop feedback mechanism is constructed to determine the optimal number of clusters.

[0076] For each candidate cluster number and its corresponding region partitioning scheme, the complete cargo pickup and delivery operation process of the UAV within each cluster is simulated, and the total global energy consumption is calculated. Total sorties Construct the objective function Find the optimal energy efficiency ratio solution while satisfying battery constraints:

[0077] ;

[0078] in, , These are the weighting coefficients. The system automatically compares the simulation results under different numbers of clusters and outputs the objective function. The regional division scheme and the list of collection points corresponding to the minimum value.

[0079] Furthermore, the number of clusters mentioned above In the traversal loop, when the number of clusters The loop condition is "Yes", meaning the current number of clusters has not yet been traversed. For each specific... The value is used to divide the N task points into groups using an improved K-Means clustering algorithm. A job cluster C1, C2,... C k ..., C p And the Medoid correction strategy is used to map the theoretical centroid to physically reachable real job points; for each job cluster C k Calculate the distance from each point within the cluster to the theoretical centroid. The geometric Euclidean distance is used to select the nearest real work point as the collection point for that work area, thereby generating the current data. The list of aerial collection points under the specified value ensures the physical reachability of relay stations. When the number of clusters... When the traversal loop condition is "No", meaning that all possible partitioning schemes have been calculated, the system enters the optimal decision-making stage.

[0080] S5. Call the trained hybrid energy consumption model to perform energy simulation, and perform path planning and search based on the limit efficiency of the safe power threshold.

[0081] In the path planning phase, traditional distance constraints are abandoned, and a greedy truncation strategy based on real energy consumption simulation, forward energy simulation, and a greedy search strategy based on safe truncation are adopted. Throughout the entire UAV sortie planning, it is assumed that the UAV always carries the maximum lifting load (including its own weight and cargo weight), without considering unloading and weight reduction mid-flight, which better reflects typical scenarios such as fertilizer lifting and equipment relocation. Starting from the current point, it sequentially attempts to reach the nearest unvisited work point, and calls the previously trained hybrid energy consumption model to accurately calculate the energy consumption of the entire process, including takeoff, hovering, level flight, and landing.

[0082] This step, based on step S4, proceeds to the path planning step based on actual energy consumption. This is the embodiment. Figure 1 The key to generating synergistic effects between the left and right branches. Figure 1 When the right branch performs path planning based on actual energy consumption, the system calls in real time. Figure 1 Energy simulation is performed on the high-load hybrid energy consumption model generated in the left branch. The simulation consists of three stages: parameter input, realistic energy consumption derivation, and energy hard constraint verification. In the parameter input stage, the UAV is set to maintain a fixed high-load state during the operational flight segment, and this parameter is input as a constant into the high-load hybrid energy consumption model. In the realistic energy consumption derivation stage, the system calculates the energy consumption from the current node at each step of the greedy path search. fly to candidate node Flight segment energy consumption At this point, the system no longer uses geometric Euclidean distance, but instead calculates it through model integration:

[0083] ;

[0084] in This represents the parameter set of the hybrid energy consumption model. Indicates the total weight of the drone. Indicates the instantaneous velocity at the current moment; This represents the predicted energy consumption for flying from the current location to the next target point. This energy consumption is calculated by the aforementioned large load hybrid energy consumption model based on fixed large load parameters.

[0085] Figure 5 The diagram illustrates the planning logic for single-segment maximum lifting efficiency. After determining the work area and aerial collection point, the system employs a dynamic truncation strategy under energy constraints to plan the flight path for a single sortie. This logic aims to maximize the workload of a single takeoff and landing by pushing battery performance to its limits. The specific implementation steps are as follows:

[0086] S51. First, initialize and set the status of the UAV, and set the workload according to the mission attributes.

[0087] Set the total battery capacity of the drone to be (Value is 1064.0 Wh, single battery capacity is 20Ah, nominal voltage is 53.2V); Retained capacity percentage is The value is 0.15 (can be modified according to the mission flight distance); defines the available battery capacity for effective safe operation. for:

[0088] ;

[0089] It also sets the drone to be located at the base or the previous collection point, with its battery fully charged (SOC=100%). The workload is set according to the mission attributes, and the drone is configured to maintain a fixed high payload state during the operational phase (e.g., ...). This parameter will be directly input into the hybrid energy consumption model.

[0090] S52. Next, a greedy strategy and truth simulation are used to search for candidate targets in the set of remaining unvisited task points, and a nearest neighbor search is used to find the point to be visited that is closest to the current position. Invoke the hybrid energy consumption model to calculate the flight path from the current location to the destination point. Single-segment actual power consumption .

[0091] S53. Perform energy constraint verification on the point to be visited to determine whether to use the point to be visited as the next target point.

[0092] Before deciding whether to visit a target point, the system performs a rigorous safety boundary check. Specifically, when planning a single hoisting sortie, the following energy inequality check is performed in real time for the next target point to be visited:

[0093] ;

[0094] in, This indicates the total amount of electricity consumed in the current flight. This is expressed as the predicted return energy consumption from the next target point back to the air collection point or base.

[0095] The next target point to be visited is added to the current flight sequence only when the energy consumption predicted by the hybrid energy model satisfies the constraints of the above energy inequality; otherwise, the system triggers flight segmentation logic, terminates the path expansion of the current flight, and triggers a return flight. Through this mechanism, the number of work points that can be covered by a single flight is maximized while ensuring the safety of the return flight, significantly reducing invalid takeoffs and landings and battery swapping frequency, and improving overall operational efficiency.

[0096] Specifically, the process of verifying energy inequalities includes:

[0097] (1) If the inequality holds, it means that the battery power is sufficient to support the visit to the next target point and the safe withdrawal. Add the next target point to the current voyage sequence and update the accumulated battery power. And use this next target point as a new starting point to continue the next round of search;

[0098] (2) If the inequality does not hold, it means that visiting the next target point will cause the battery to run out and the return trip will be impossible. In this case, the current flight will be immediately terminated, a path will be generated to return to the collection point from the current location, the current flight mission will be terminated, and the planning of the next flight will be triggered.

[0099] By "squeezing" battery performance in this way, the system can calculate the maximum number of points that the drone can lift in a single take-off and landing cycle, thereby maximizing operational efficiency.

[0100] The system will iterate through different numbers of clusters. (For example, from 1 to 8), repeat the above process and summarize all. The corresponding path planning results. Compare the total global energy consumption of each path planning scheme. Total sorties By constructing an objective function, the clustering dimension with the optimal energy efficiency ratio is selected. The process then generates an optimal sling-out trajectory containing a detailed waypoint sequence and flight instructions, concluding the entire process. The resulting overall sling-out trajectory is shown below. Figure 6 As shown.

[0101] In summary, addressing the challenges of dispersed work points, heavy and fixed loads, and the predictive failure of existing linear models in agricultural hoisting scenarios under heavy load conditions, this invention first constructs a hybrid energy consumption model coupling physical mechanisms and residual correction. During the training phase, a load imbalance weighting mechanism is introduced to force the model to fit the nonlinear power consumption characteristics under heavy load conditions. Second, a greedy traversal strategy based on real energy consumption simulation is proposed. Within work point clusters, segmental energy extrapolation is performed under a fixed full load state. Under strict battery safety threshold constraints, the maximum number of work points that can be covered in a single deployment is calculated. This invention effectively solves the safety hazard of inaccurate endurance estimation during heavy load operations and maximizes hoisting efficiency within a single take-off and landing cycle. It significantly improves the efficiency and safety of agricultural hoisting operations.

[0102] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A method for energy consumption prediction and path planning in the lifting operation of large-payload unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: S1. Initialization: Load the raw flight data and mission target point data of the high-payload UAV. S2. Construct a hybrid energy consumption model that couples physical mechanisms with residual correction as the energy calculation kernel for path planning; S3. Introduce a heavy-load unbalanced weighted training mechanism to train the constructed hybrid energy consumption model. Give significantly higher weights to samples that exceed the preset heavy-load threshold, forcing the model to prioritize fitting the residual fluctuations under heavy load during the training process. S4. An improved K-Means clustering algorithm is used to divide the spatial domain of discrete work points. The actual work point closest to the centroid in each cluster is selected as the collection point or supply center of the corresponding region. The hybrid energy consumption model is introduced into the clustering iteration closed loop as the evaluation function to realize regional clustering based on real work points. S5. Call the trained hybrid energy consumption model to perform energy simulation. Use a greedy truncation strategy based on real energy consumption simulation, a greedy search strategy based on forward energy simulation and safe truncation, and perform path planning and search based on the limit efficiency of the safe power threshold.

2. The energy consumption prediction and route planning method according to claim 1, characterized in that, Step S2 first performs data preprocessing and feature engineering on the raw flight data, extracting features including total mass. 1.5th power and horizontal flight speed Key physical feature vectors including the cube of . : ； in, Characterizing the induced resistance factor, Characterizing parasitic resistance factor, It is the acceleration due to gravity; To address the nonlinear relationship between aerodynamic drag and power under heavy load conditions, a hybrid energy consumption model was constructed, comprising a physical baseline model and a weighted residual neural network. The physical baseline model utilizes linear regression to fit physical characteristics and outputs the physical baseline power. In a weighted residual neural network, the nonlinear bias and nonlinear error of the unbalanced weighted training mechanism are fitted using the neural network, and the residual power is output. ; The total power prediction output by the hybrid energy consumption model is .

3. The energy consumption prediction and route planning method according to claim 2, characterized in that, Physical reference power output by the physical reference model The calculation formula is: ； Where m is the total mass, and the exponent 1.5 indicates the nonlinear growth of induced drag under heavy loads; For horizontal flight speed, Indicates the parasitic resistance factor; Vertical flight speed, It is the acceleration due to gravity. Characterizes the power output during vertical climb; All are regression coefficients.

4. The energy consumption prediction and route planning method according to claim 1, characterized in that, Step S3 embeds the overloaded unbalanced weighted training mechanism into the training closed loop of the hybrid energy consumption model, including the following steps: S31. Based on the heavy-load unbalanced weighted training mechanism, set a high-load judgment threshold for the load and define sample weights. : ； in The threshold for determining high load capacity; This is the penalty weighting coefficient for high-load samples, i.e., high-multiplier weighting; For the first The total payload of the training samples, i.e., the sample payload; S32. Read the training samples from the original flight dataset, each training sample... This includes physical state characteristics and actual energy consumption labels, with the core of the physical state characteristics being the sample load. The actual energy consumption label is the actual power consumption. ; S33. Introduce an unbalanced weighted discriminator to scan the training samples in the original flight dataset one by one, make logical branch judgments based on the sample load and the high load judgment threshold, assign corresponding sample weights to the training samples based on the judgment results, and output the sample weight vector. S34. Substitute the sample weight vector into the loss function of the hybrid energy consumption model and perform backpropagation training to optimize the weighted loss function and obtain the trained hybrid energy consumption model.

5. The energy consumption prediction and route planning method according to claim 4, characterized in that, The logical branch judgment process in step S33 includes: loading the sample High load determination threshold with load capacity If a comparison is made, If the training sample is deemed a high-risk sample, and the corresponding data point belongs to the scarce operating condition of the motor under high load, it will be assigned a high weight. ;like If the training sample is determined to be a regular sample, and the corresponding data point belongs to a regular cruise condition, then it is assigned a standard weight. ; After evaluating each training sample in the original flight dataset, the sample weight vector is output.

6. The energy consumption prediction and route planning method according to claim 1, characterized in that, Step S4 adopts a region division method based on full-process energy consumption simulation feedback. By simulating the entire process of picking up and transporting goods under different division dimensions, the geometric Euclidean distance of physical space is mapped to the energy consumption topological distance. After region division, a Medoid optimization strategy based on real work points is adopted for each work cluster to select the real work points in the work cluster as the collection center. At the same time, the Medoid optimization strategy based on real work points is combined with the hybrid energy consumption model to prioritize the work point that can make the total weighted load path within the work cluster the shortest as the core, so as to minimize the high-energy-consumption full-load flight range.

7. The energy consumption prediction and route planning method according to claim 6, characterized in that, Step S4 includes: S41. Define energy consumption topology distance, reconstruct the metric for spatial distance, nonlinearly map the physical geometric space to the energy consumption topology space, and establish an energy consumption topology graph. S42. Based on the improved K-Means clustering algorithm and energy consumption topology map, read the task point data for cluster analysis, and perform regional division and corresponding regional aerial collection point site selection correction.

8. The energy consumption prediction and route planning method according to claim 7, characterized in that, Step S41 utilizes a pre-trained hybrid energy consumption model to transform the geometric Euclidean distance into an energy consumption cost through mapping. The mapping formula is: ； in, For nodes With nodes The geometric Euclidean distance between them; This refers to the drone's cruising speed. This is the total power prediction value output by the hybrid energy consumption model.

9. The energy consumption prediction and route planning method according to claim 7, characterized in that, Step S42 includes: Read the task point set and start clustering. The traversal loop, in which , The maximum allowed value for traversal is given; the discrete task point set is divided into P geometric clusters using an improved K-Means clustering algorithm; Perform real job point correction, calculate the geometric Euclidean distance from all real job points within the cluster to the theoretical centroid generated by the K-Means clustering algorithm; select the real job point with the smallest geometric Euclidean distance as the aerial collection point for the corresponding region; Construct a closed-loop feedback mechanism to determine the optimal number of clusters.

10. The energy consumption prediction and route planning method according to claim 1, characterized in that, Step S5 includes: S51. Initialize and set the status of the UAV, and set the operational payload according to the mission attributes; S52. Using a greedy strategy and truth simulation, search for candidate targets in the set of remaining unvisited task points, and use nearest neighbor search to find the point to be visited closest to the current position. Invoke the hybrid energy consumption model to calculate the flight path from the current location to the destination point. Single-segment actual power consumption ; S53. Perform energy constraint verification on the point to be visited to determine whether to use the point to be visited as the next target point. Define the available battery capacity for effective and safe operation. for: ； in The total capacity of the human-machine interface battery. The percentage of electricity retained; The energy constraint verification uses the following energy inequality: ； in, This indicates the total amount of electricity consumed in the current flight. This is expressed as the predicted return energy consumption from the next target point back to the air collection point or base; The next target point to be visited will only be added to the current flight sequence if the energy consumption predicted by the hybrid energy model satisfies the energy inequality constraint; otherwise, the system will trigger the flight splitting logic, terminate the path extension of the current flight, and trigger the return flight.

Images