Unmanned aerial vehicle path intelligent planning method based on wind environment forecast
By constructing a wind environment forecast model and a rolling window path planning algorithm, the problem of insufficient utilization of stroke environment data for drone path planning is solved, and efficient and stable flight of drones is achieved.
Patent Information
- Application Number
- CN202510409552.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-02
- Publication Date
- 2025-07-11
AI Technical Summary
The existing UAV path planning methods lack the utilization of wind environment data, resulting in low UAV flight efficiency, safety and mission completion rates.
The intelligent path planning method of drone based on wind environment forecasting, by obtaining historical wind environment data for preprocessing, building a wind environment information prediction model, and combining the path planning algorithm of the rolling window, using the Autoformer algorithm and genetic algorithm for path optimization to realize real-time path planning of drones.
Significantly reduce the energy consumption of drones, improve the working environment of drones, and improve the stability and efficiency of drone flight.
Smart Images
Figure CN120293141A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of UAV operation safety, and particularly relates to an intelligent UAV path planning method based on wind environment prediction. Background Art
[0002] With the rapid development and wide application of UAV technology, UAVs are increasingly widely used in various industries, including agriculture, logistics, security monitoring and other fields. During the flight of a UAV, the wind environment is an important consideration factor, which has an important impact on the flight trajectory and flight efficiency of the UAV.
[0003] The wind field has a direct impact on the flight trajectory and route planning of UAVs. Changes in wind speed and direction will affect the flight speed, energy consumption and arrival time of UAVs. Therefore, it is essential to consider wind field data in path planning.
[0004] For general logistics UAVs, due to their low flight altitude, they are easily affected by the environmental wind field when flying in the wind field. When the wind speed is greater than the wind resistance level of the UAV, the UAV is very likely to get out of control. And when flying against the wind in the wind field, when the wind speed is relatively large, it will cause the energy of the UAV to be consumed rapidly, reducing the working efficiency of the UAV.
[0005] The current UAV path planning methods lack the utilization of wind environment data, resulting in great room for improvement in the flight efficiency, safety and task completion rate of UAVs. Summary of the Invention
[0006] The purpose of the present invention is to provide an intelligent UAV path planning method based on wind environment prediction, which makes full use of wind field data and combines advanced path planning algorithms to provide a more accurate and efficient path planning scheme for the flight of UAVs.
[0007] The present invention provides an intelligent UAV path planning method based on wind environment prediction, including the following steps:
[0008] S1. Obtain historical wind environment data, perform preprocessing, and perform annotation to obtain a training data set;
[0009] S2. Based on the Autoformer algorithm, construct an initial wind environment information prediction model;
[0010] S3. Use the training data set obtained in step S1 to train the initial wind environment information prediction model obtained in step S2 to obtain a wind environment information prediction model;
[0011] S4. According to the wind environment information prediction model obtained in step S3, based on the rolling window path planning method, complete the real-time path planning of the UAV.
[0012] Step S4 is specifically as follows:
[0013] First, divide the flight area of the drone into grids of a preset size according to the grid resolution. During the flight in the same grid area, and the drone maintains a constant speed. If the stable flight distance of the drone in the i-th period of time is d i , then the energy consumption W of the drone during this period of time i is calculated using the following formula:
[0014]
[0015] where m p is the weight of the cargo on the drone; m v is the weight of the drone when it is empty; r is the lift-to-drag ratio; η is the power transmission efficiency of the motor and propeller; p is the power consumption of the electronic equipment on the drone; α is the ratio of the wind speed to the flight speed, and α < 1 when flying against the wind; after simplifying the formula, the energy consumption W of the drone flying in the k-th grid k is expressed using the following formula:
[0016]
[0017] where d k is the flight distance of the drone in the k-th grid; α k is the ratio of the wind speed to the flight speed of the drone in the k-th grid; σ is a constant and Therefore, the total energy consumption W of the drone for the whole journey is expressed using the following formula:
[0018]
[0019] To achieve the minimum cost objective of the drone path planning, from the perspective of energy consumption, set the drone path planning objective function, which is expressed using the following formula:
[0020]
[0021] where v k is the wind speed of the wind field of the drone in the k-th grid; v s is the flight speed of the drone;
[0022] Reconstruct the wind environment prediction path planning based on a rolling window, specifically as follows: Obtain the real-time wind environment information of the current window area of the UAV, and use the wind environment information prediction model obtained in step S3 to obtain the predicted wind environment information of the current window area of the UAV; within each window area, the UAV uses a heuristic method to generate optimized sub-goals according to the predicted wind environment information; according to the predicted wind environment information of the current window area of the UAV, perform local path planning in the current rolling window area; move to the next window area in the manner of the rolling window algorithm; repeat the above steps to achieve the results of optimization and feedback until the UAV flies to the destination;
[0023] The specific heuristic method is as follows: During the flight of the UAV, after obtaining the predicted wind environment information of the current window each time, use a heuristic function to determine the sub-goal, which is expressed by the following formula:
[0024] minf(P) = g(P) + h(P)
[0025] where P is a grid point in the window network; g(P) is the cost for the UAV to travel from the current position to point P; h(P) is the cost from point P to the end point; since g(P) depends on the position of point P and the predicted wind environment information of the current window, and h(P) depends on the energy consumption from point P to the end point, the window boundary point P obtained with the minimum value of the above function can balance the global optimization requirements and local limited information.
[0026] Perform local path planning within the rolling window area based on the genetic algorithm, including the following steps:
[0027] a. Population initialization: Obtain the start and end nodes of the path in the current window area, and randomly generate intermediate nodes, and traverse whether the two points between the nodes are continuous; if the absolute value of the difference in the horizontal and vertical coordinates between the two nodes is less than or equal to the preset value, it is judged that the two nodes are continuous, otherwise it is judged as discontinuous; if the two nodes are judged to be discontinuous, insert a node between the two nodes; the calculation method of the inserted node is the average of the horizontal and vertical coordinates of the two nodes; after inserting the node, update the latter node, and repeat the above operation until all nodes are continuous;
[0028] b. Crossover: Randomly select two flight paths, and judge whether the two paths have a crossover point. If there is no crossover point, reselect two; if there is one crossover point, exchange the paths after the crossover point of the two paths; if there are two or more crossover points, select one of the crossover points and exchange the paths after the crossover point of the two paths; add the newly generated path to the population;
[0029] c. Mutation: Randomly select a path and two nodes on it. Using the method in step a, randomly generate new nodes between these two nodes again, and ensure that the newly generated path nodes are continuous. Add the newly generated path to the population;
[0030] d. After generating a path each time, calculate the energy consumption corresponding to each path, keep the path with the lowest energy consumption, and form a new population with the newly generated paths;
[0031] e. Continuously repeat the above steps until the algorithm converges.
[0032] The present invention discloses an intelligent path planning method for unmanned aerial vehicles based on wind environment prediction, which can significantly reduce the energy consumption of unmanned aerial vehicles, improve the working environment of unmanned aerial vehicles, and enhance the flight stability of unmanned aerial vehicles. Description of the Drawings
[0033] Figure 1 is a schematic flow chart of the method of the present invention;
[0034] Figure 2 is an example diagram of the flight path of the unmanned aerial vehicle of the method of the present invention. Detailed Embodiments
[0035] The present invention provides an intelligent path planning method for unmanned aerial vehicles based on wind environment prediction. Its schematic flow chart is as Figure 1 shown, and it includes the following steps:
[0036] S1. Obtain historical wind environment data, perform preprocessing, and perform annotation to obtain a training data set;
[0037] The historical wind environment data consists of several parameters such as temperature, air pressure, and wind speed; the historical wind environment data is obtained through a wind speed sensor in a ground monitoring point;
[0038] Perform preprocessing on the historical wind environment data, and construct the wind environment data detected at N stations after preprocessing into a format of 1*L*W, where 1 is the number of channels to be observed in the modeling; L is the record length of the wind environment data, and W is the type of data. Annotate the station where each wind environment data is obtained according to the station longitude and latitude relationship, and fill the place without data with the value "0" to obtain a wind environment training data set; construct labels for the data in the wind environment training data set. The labels are simulation values with a preset resolution obtained using a meteorological simulation model, and perform interpolation, anomaly processing, and standardization processing on the labels to obtain a high-frequency simulation training data set;
[0039] The preprocessing includes data missing value completion, data anomaly handling, data standardization, and data structure handling. In terms of data missing values, according to the characteristic of discontinuous data missing, the forward interpolation method is used to complete the missing data. In terms of data anomalies, a method based on a probabilistic autoencoder is selected to process the data. In terms of data standardization, considering the inconsistent dimensions of wind speed and wind vector and the requirements of the training model, Z-score standardization is used to process the data. The data structure handling uses a sliding window method to process the time series data of wind environment data into the data structure required for neural network input.
[0040] S2. Based on the Autoformer algorithm, construct an initial wind environment information prediction model;
[0041] S3. Use the training data set obtained in step S1 to train the initial wind environment information prediction model obtained in step S2 to obtain a wind environment information prediction model;
[0042] In a specific embodiment, the initial wind environment information prediction model can also be:
[0043] The initial wind environment information prediction model includes a wind environment prediction model, a wind environment data conversion model A, and a wind environment data conversion model B; the wind environment prediction model processes the input wind environment data through a general algorithm to obtain predicted wind environment data; the wind environment data conversion model A constructs a convolutional neural network based on SRResNet; the wind environment data conversion model A and the wind environment data conversion model B have the same structure; the wind environment data conversion model A and the wind environment data conversion model B process the input predicted wind environment data to improve the resolution of the wind environment data and obtain the converted wind environment data;
[0044] In the wind environment data conversion model A and the wind environment data conversion model B, partial convolution operations combined with a mask mechanism are used to replace convolution operations; in partial convolution operations, the area where there is no numerical value is called an invalid area, and the area where there is a numerical value is called a valid area. Given a binary mask, the convolution result of partial convolution is only related to the valid area of each layer; partial convolution makes the convolution operation of the model unaffected by the invalid area and is represented by the following formula:
[0045]
[0046] Among them, W is the weight of the convolution kernel; X is the eigenvalue in the current sliding window; M is the corresponding binary mask; ⊙ is the element product; M1 has the same shape as M, but all elements are 1; x′ is the eigenvalue after convolution;
[0047] After each partial convolution operation, the binary mask is updated, and the update strategy is represented by the following formula:
[0048]
[0049] Among them, m' is the value of the binary mask.
[0050] The specific training process is as follows:
[0051] Use the wind environment training data set obtained in step S1 to train the wind environment prediction model in the initial wind environment information prediction model obtained in step S2 to obtain a preliminary trained wind environment prediction model;
[0052] Use the high-frequency simulation training data set obtained in step S1 to train the wind environment data conversion model A and the wind environment data conversion model B in the initial wind environment information prediction model obtained in step S2 respectively, adopting a two-stage training mechanism; the two-stage training mechanism is to map the input information graph into a high-frequency information graph through the wind field data conversion model A, and then use the output high-frequency information graph as the input and input it into the initial wind field data conversion model B, and still train with the same label to obtain a preliminary trained wind field data conversion model A and a preliminary trained wind field data conversion model B;
[0053] Then use the high-frequency simulation training data set obtained in step S1 to train the preliminary trained wind environment prediction model and the preliminary trained wind environment data conversion model A. When training, first input the data in the training data set into the preliminary trained wind environment prediction model, and use the obtained result as the input and input it into the preliminary trained wind field data conversion model A to obtain a wind environment prediction model and a wind field data conversion model A. The total loss function of the training is expressed by the following formula:
[0054] L = MSE_1 + γMSE_2
[0055] Among them, γ is the balance coefficient; MSE_1 is the loss of the initial wind environment information prediction model; MSE_2 is the loss of the wind field data conversion model A, which is expressed by the following formula:
[0056]
[0057]
[0058] Among them, Y1 is the output of the wind environment information prediction model; Y2 is the output of the wind field data conversion model A; y t is the label value in the high-frequency simulation training data set;
[0059] Input the high-frequency simulation training dataset obtained in step S1 into the wind environment prediction model and the wind field data conversion model A to obtain a high-frequency prediction simulation training dataset. Train the initially trained wind field data conversion model B with the same labels to obtain the wind field data conversion model B. Finally, obtain a wind environment information prediction model including a wind environment prediction model, a wind field data conversion model A, and a wind field data conversion model B connected in series in sequence.
[0060] S4. Based on the wind environment information prediction model obtained in step S3, use the path planning method based on a rolling window to complete the real-time path planning of the UAV. Specifically:
[0061] First, divide the UAV flight area into grids of a preset size according to the grid resolution. During the flight in the same grid area and with the UAV flying at a constant speed, if the stable flight distance of the UAV in the i-th time period is d i , then the energy consumption W of the UAV during this time period i is calculated using the following formula:
[0062]
[0063] where, m p is the weight of the cargo on the UAV; m v is the weight of the UAV when it is empty; r is the lift-to-drag ratio; η is the power transmission efficiency of the motor and propeller; p is the power consumption of the electronic equipment on the UAV; α is the ratio of the wind speed to the flight speed, and α < 1 when flying against the wind; after simplifying the formula, the energy consumption W of the UAV flying in the k-th grid is k expressed using the following formula:
[0064]
[0065] where, d k is the flight distance of the UAV in the k-th grid; α k is the ratio of the wind speed to the flight speed of the UAV in the k-th grid; σ is a constant and Therefore, the total energy consumption W of the UAV for the whole journey is expressed using the following formula:
[0066]
[0067] To achieve the minimum cost objective of UAV path planning, from the perspective of energy consumption, set the UAV path planning objective function, which is expressed using the following formula:
[0068]
[0069] where, v k is the wind speed of the wind field of the UAV in the k-th grid; v s is the flight speed of the UAV;
[0070] Reconstruct the wind environment prediction path planning based on a rolling window, specifically: obtain the real-time wind environment information of the current window area of the UAV, and use the wind environment information prediction model obtained in step S3 to obtain the predicted wind environment information of the current window area of the UAV; within each window area, the UAV uses a heuristic method to generate optimized sub-goals according to the predicted wind environment information; according to the predicted wind environment information of the current window area of the UAV, perform local path planning in the current rolling window area; move to the next window area in the manner of a rolling window algorithm; repeat the above steps to achieve the result of optimization and feedback until the UAV flies to the destination;
[0071] The heuristic method is specifically: during the flight of the UAV, after obtaining the predicted wind environment information of the current window each time, use a heuristic function to determine the sub-goal, which is expressed by the following formula:
[0072] minf(P) = g(P) + h(P)
[0073] where P is a grid point in the window network; g(P) is the cost for the UAV to travel from the current position to point P; h(P) is the cost from point P to the end point; since g(P) depends on the position of point P and the predicted wind environment information of the current window, and h(P) depends on the energy consumption from point P to the end point, the window boundary point P obtained with the minimum value of the above function can balance the global optimization requirements and local limited information.
[0074] Perform local path planning within the rolling window area based on the genetic algorithm, including the following steps:
[0075] a. Population initialization: Obtain the start and end nodes of the path in the current window area, and randomly generate intermediate nodes, and traverse whether the two points of the nodes are continuous; if the absolute value of the difference in the horizontal and vertical coordinates between the two nodes is less than or equal to the preset value, it is judged that the two nodes are continuous, otherwise it is judged that they are not continuous; if the two nodes are judged to be not continuous, insert a node between the two nodes; the calculation method of the inserted node is the average of the horizontal and vertical coordinates of the two nodes; after inserting the node, update the latter node, and repeat the above operation until all nodes are continuous;
[0076] b. Crossover: Randomly select two flight paths, and judge whether the two paths have intersection points. If there are no intersection points, reselect two; if there is one intersection point, exchange the paths after the intersection point of the two paths; if there are two or more intersection points, select one of the intersection points and exchange the paths after the intersection point of the two paths; add the newly generated path to the population;
[0077] c. Mutation: Randomly select a path and two nodes on it. Using the method in step a, randomly generate new nodes between these two nodes again, and ensure that the nodes on the newly generated path are continuous. Then add the newly generated path to the population;
[0078] d. After generating each path, calculate the energy consumption corresponding to each path, keep the path with the lowest energy consumption, and form a new population with the newly generated paths;
[0079] e. Continuously repeat the above steps until the algorithm converges.
Claims
1. An intelligent path planning method for an unmanned aerial vehicle based on wind environment prediction, characterized in that, It includes the following steps: S1. Obtain historical wind speed data, preprocess it, and annotate it to obtain a training dataset; S2. Based on the Autoformer algorithm, construct an initial wind environment information prediction model; S3. Use the training dataset obtained in step S1 to train the initial wind environment information prediction model obtained in step S2 to obtain a wind environment information prediction model; S4. According to the wind environment information prediction model obtained in step S3, based on the rolling window path planning method, complete the real-time path planning of the drone.
2. The intelligent path planning method for an unmanned aerial vehicle based on wind environment prediction according to claim 1, wherein Step S4 is specifically as follows: First, aiming at minimizing the cost of drone path planning, construct a drone path planning objective function from the perspective of energy consumption; Then construct a wind environment prediction path planning based on a rolling window, specifically: obtain the real-time wind environment information of the current window area of the drone, and use the wind environment information prediction model obtained in step S3 to obtain the predicted wind environment information of the current window area of the drone; within each window area, the drone uses a heuristic method to generate an optimized sub-goal according to the predicted wind environment information; according to the predicted wind environment information of the current window area of the drone, perform local path planning in the current rolling window area; move to the next window area in the manner of a rolling window algorithm; repeat the above steps to achieve the result of optimization and feedback until the drone flies to the destination.
3. The intelligent path planning method for an unmanned aerial vehicle based on wind environment prediction according to claim 2, wherein, The construction of the drone path planning objective function includes the following steps: First, divide the UAV flight area into grids of a preset size according to the grid resolution. During the flight in the same grid area and with the UAV flying at a constant speed, if the stable flight distance of the UAV in the $i$-th period of time is $d$ i , then the energy consumption $W$ of the UAV in this period of time i is calculated using the following formula: Among them, m p is the weight of the cargo on the UAV; m v is the weight of the UAV when it is unloaded; r is the lift-to-drag ratio; η is the power transmission efficiency of the motor and the propeller; p is the power consumption of the electronic equipment on the UAV; α is the ratio of the wind speed to the flight speed, and α < 1 when flying against the wind; after simplifying the formula, the energy consumption W k of the UAV flying in the k-th grid is expressed by the following formula: where d k is the distance the drone flies in the k-th grid; α k is the ratio of the wind speed to the flight speed of the drone in the k-th grid; σ is a constant and therefore, the total energy consumption W of the drone for the whole journey is expressed by the following formula: To achieve the goal of minimizing the cost of drone path planning, set the drone path planning objective function from the perspective of energy consumption, and use the following formula to represent it: Among them, v k is the wind speed of the UAV in the k-th grid; v s is the flight speed of the UAV.
4. The intelligent path planning method for unmanned aerial vehicles based on wind environment prediction according to claim 3, wherein The heuristic method is specifically as follows: during the flight of the drone, after obtaining the current window wind environment prediction information each time, use the heuristic function to determine the sub-goal, and use the following formula to represent it: minf(P) = g(P) + h(P) where P is a grid point in the window network; g(P) is the cost for the drone to travel from the current position to point P; h(P) is the cost from point P to the end point; since g(P) depends on the position of point P and the predicted wind environment information of the current window, and h(P) depends on the energy consumption from point P to the end point, the window boundary point P obtained with the minimum value of the above function as the sub-goal can balance the global optimization requirements and local limited information.
5. The intelligent path planning method for unmanned aerial vehicles based on wind environment prediction according to claim 4, wherein Based on the genetic algorithm, perform local path planning within the rolling window area, including the following steps: a. Population initialization: Obtain the start and end nodes of the path in the current window area, and randomly generate intermediate nodes, and traverse whether the two points of the nodes are continuous; if the absolute value of the difference in the horizontal and vertical coordinates between the two nodes is less than or equal to the preset value, it is judged that the two nodes are continuous, otherwise it is judged as discontinuous; if the two nodes are judged to be discontinuous, insert a node between the two nodes; the calculation method of the inserted node is the average of the horizontal and vertical coordinates of the two nodes; after inserting the node, update the latter node, and repeat the above operation until all nodes are continuous; b. Crossover: Randomly select two flight paths and determine whether there is an intersection point between the two paths. If there is no intersection point, select two new paths; if there is one intersection point, exchange the paths after the intersection point of the two paths; if there are two or more intersection points, select one of the intersection points and exchange the paths after the intersection point of the two paths; add the newly generated paths to the population; c. Mutation: Randomly select a path and randomly select two nodes in it. Using the method in step a, randomly generate new nodes between these two nodes again, and ensure that the nodes in the newly generated path are continuous. Add the newly generated path to the population; d. After each path is generated, calculate the energy consumption corresponding to each path, keep the path with the lowest energy consumption among them, and form a new population with the newly generated paths; e. Continuously repeat the above steps until the algorithm converges.