Low-altitude turbulence forecasting and route planning method based on machine learning
By using machine learning-based multi-source data processing and model building, the problem of insufficient turbulence forecasting accuracy for low-altitude aircraft has been solved, enabling dynamically optimized route planning and improving the safety and real-time early warning capabilities of low-altitude flights.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANTONG UNIV
- Filing Date
- 2026-01-14
- Publication Date
- 2026-04-24
AI Technical Summary
In existing technologies, low-altitude aircraft are susceptible to turbulence. Traditional turbulence prediction methods are not accurate enough, resulting in low safety in route planning and making it difficult to meet real-time safety requirements.
By employing machine learning-based methods, multi-source meteorological data is collected and preprocessed to construct a mid-term turbulence forecasting model based on the random forest regression algorithm and a short-term turbulence forecasting model based on the parameter inversion algorithm. Combined with the safety distance constraints between the path and the no-fly zone boundary, dynamic optimization of flight path planning is achieved.
It has achieved automated and quantitative mapping from low-resolution weather conditions to high-resolution aviation turbulence risks, established an autonomous short-term early warning system, and can proactively plan the safest flight trajectory to bypass future strong turbulence areas, thereby improving the safety assurance capability of low-altitude flight.
Smart Images
Figure CN121920226A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of route planning technology, and in particular to a method for low-altitude turbulence forecasting and route planning based on machine learning. Background Technology
[0002] Currently, the low-altitude economy has become a strategic emerging industry, the core of which is to carry out various aircraft flight activities in the airspace below 1,000 meters (and in some scenarios, up to 3,000 meters). However, low-altitude aircraft are small in size and fly at low altitudes, making them susceptible to turbulence, which can lead to safety incidents such as turbulence and yaw. The formation mechanism of turbulence is complex and is affected by a combination of factors such as heat and topography, making it difficult to monitor and forecast accurately.
[0003] In existing technologies, meteorological detection equipment (such as wind profiler radar and Doppler radar) can acquire data on wind, temperature, and humidity, but cannot directly measure turbulence intensity; it must be calculated indirectly. Traditional turbulence forecasting methods are mostly based on a single data source or physical model, which suffers from insufficient spatiotemporal resolution and limited prediction accuracy, making it difficult to meet the real-time safety requirements of low-altitude flight; for example... k-ω Due to their computational cost, sensitivity to initial values, and unsuitability as forward models, these models are difficult to directly apply to the operational monitoring and forecasting of low-altitude turbulence, which requires extremely high real-time performance, automation, and computational efficiency.
[0004] Meanwhile, existing route planning technologies rarely take into full account the spatiotemporal evolution characteristics of turbulence, and cannot provide aircraft with dynamically optimized safe paths; therefore, there is an urgent need for turbulence forecasting and route planning methods that integrate multi-source data and advanced models to improve the safety assurance capabilities for low-altitude flights. Summary of the Invention
[0005] To address the shortcomings of existing methods, this invention solves the problems of insufficient accuracy in turbulence forecasting and low safety in route planning.
[0006] The technical solution adopted in this invention is: a method for low-altitude turbulence forecasting and route planning based on machine learning, comprising the following steps: Step 1: Collect meteorological data from different sources; In a preferred embodiment of the present invention, the meteorological data includes: wind speed, temperature, altitude, atmospheric pressure, and humidity.
[0007] In a preferred embodiment of the present invention, meteorological data are preprocessed using natural neighbor interpolation, Z-score standardization, 3σ criterion, and Pearson correlation coefficient method.
[0008] Step 2: Construct a medium-term turbulence forecasting model based on the random forest regression algorithm and / or a short-term turbulence forecasting model based on the parameter inversion algorithm; As a preferred embodiment of the present invention, the formula for the mid-term turbulence prediction model is: ; in, B For the number of decision trees, T b Indicates the first b A decision tree, X The input feature vector.
[0009] As a preferred embodiment of the present invention, grid search and 5-fold cross-validation are used to optimize the hyperparameters of the mid-term turbulence prediction model.
[0010] As a preferred embodiment of the present invention, the construction of the short-term turbulence prediction model includes: Constructing an advection-attenuation physical framework, the formula is: ; in, For Lagrange operators; t 0 The starting time; Step size; Let be the parameters, representing the advection velocity in the east-west direction, the advection velocity in the north-south direction, and the decay time scale, respectively; collectively denoted as . ; Based on the three-dimensional turbulence intensity model; The loss function is constructed as follows: ; in, t i It is a historical point in time. N It is the number of valid sample pairs; Inversion by minimizing the loss function u e , v e and ,get: .
[0011] Step 3: Set safety distance constraints between the planned path and the no-fly zone boundary for the medium-term turbulence forecasting model and / or the short-term turbulence forecasting model; In a preferred embodiment of the present invention, the formula for the safety distance constraint between the planned path and the no-fly zone boundary is: (5) in, J The objective function is... t o The starting time; t fThe termination time; EDR ( x , y , z , t This refers to a medium-term or short-term turbulence forecasting model. .
[0012] As a preferred embodiment of the present invention, a machine learning-based low-altitude turbulence forecasting and route planning system includes: a memory for storing instructions executable by a processor; and a processor for executing the instructions to implement a machine learning-based low-altitude turbulence forecasting and route planning method.
[0013] As a preferred embodiment of the present invention, a computer-readable medium storing computer program code implements a machine learning-based method for low-altitude turbulence forecasting and route planning when executed by a processor.
[0014] The beneficial effects of this invention are: 1. Through supervised learning, a random forest regression model with core physical parameters of numerical weather prediction as features and benchmark data as labels was trained for mid-term turbulence forecasting. Its core is to realize the automated and quantitative direct mapping from low-resolution weather pattern forecasts to high-resolution aviation turbulence risk forecasts, solving the operational bottleneck of traditional methods that rely on manual interpretation and cannot output quantitative data. 2. A short-term turbulence forecasting model based on the advection-attenuation physical framework and historical data parameter inversion was designed. Its core contribution is to establish an autonomous short-term early warning system that does not rely on external numerical forecast input, effectively filling the "early warning blind zone" between operational numerical forecast updates and realizing minute-level high-frequency capture of the rapid evolution of turbulence in the next 0-1 hour. 3. The dynamically fused turbulence forecast field is transformed into a cost field for path optimization, and a spatiotemporal path search model with the goal of minimizing cumulative turbulence exposure is constructed. Its core is to realize the leap from avoiding static obstacles to avoiding dynamic meteorological risks in route planning, and to proactively and intelligently plan the safest flight trajectory to bypass future strong turbulence areas. Attached Figure Description
[0015] Figure 1 This is a flowchart of the machine learning-based low-altitude turbulence forecasting and route planning method of the present invention; Figure 2 This is a three-dimensional turbulence intensity distribution diagram of model C; Figure 3 It is the first k The training process of a decision tree; Figure 4 It is a random forest classification process; Figure 5Model D is a multi-level horizontal distribution; Figure 6 This is the turbulence intensity prediction diagram for model D at points 05-08; Figure 7 This is a three-dimensional distribution diagram of turbulence intensity of model E at points 05-06; Figure 8 This is a comparison of the three-dimensional turbulence intensity of model E over multiple time periods; Figure 9 This is the optimal flight path diagram for model D; Figure 10 This is the optimal route map for model E. Detailed Implementation
[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments. The drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.
[0017] like Figure 1 As shown, a machine learning-based method for low-altitude turbulence forecasting and route planning includes the following steps: Step 1: Collect meteorological data from different sources; By collecting multi-source meteorological data from ground-based automatic weather stations, wind profiler radar, and Doppler weather radar, the data includes: wind speed, temperature, altitude, atmospheric pressure, humidity, etc. The data were preprocessed and standardized. Missing values were filled in using natural neighbor interpolation. Dimensional differences were eliminated using the Z-score standardization formula. Outliers were removed using the 3σ criterion. The correlation between meteorological elements was analyzed using the Pearson correlation coefficient method to identify turbulence-prone areas. A turbulence intensity model was constructed using fused multi-source data, with a horizontal resolution of 100 meters, a vertical resolution of 50 meters, and a height range of 0-2000 meters, to simulate turbulence intensity. A basic three-dimensional turbulence intensity model C is constructed. This model can be based on the Doppler spectral width method to invert turbulence intensity; or it can be based on the turbulence inertial subregion theory, such as the Kolmogorov inertial subregion theory. The formula is as follows: (1) in, ε This represents the turbulent energy dissipation rate.
[0018] Step 2: Construct a mid-term turbulence prediction model D based on the random forest regression algorithm; Model D was trained using meteorological data from different sources; Establishing large-scale variables and high-resolution turbulence intensity in numerical weather prediction EDRThe nonlinear statistical relationship between them; the random forest regression algorithm in the ensemble learning method is selected, which has the advantage of being able to effectively handle the complex interactions between high-dimensional features and has strong robustness against overfitting. Random forest with K Decision Tree h X , k , k 1, 2,..., K The base classifier is used as a combined classifier obtained through ensemble learning; when a sample to be classified is input, the classification result output by the random forest is determined by a simple vote of the classification results of each decision tree. definition k , k 1, 2,..., K It is a sequence of random variables, determined by the two major randomization principles of random forest; Bagging Idea: From the original sample set X Random selection with replacement K A training sample set of the same size as the original sample set. Tk , k 1, 2,..., K Each training sample set Tk Construct the corresponding decision tree; Feature subspace concept: When splitting each node of the decision tree, a subset of attributes is randomly selected from all attributes with equal probability (usually taking...). Log 2 ( M ) 1 attribute, M (where the total number of features is 1), and then select an optimal attribute from this subset to split the node; Training a random forest is essentially training individual decision trees. Since the training of each decision tree is independent, random forest training can be implemented in parallel, significantly improving the efficiency of the generative model. k Decision Tree h ( X , kThe training process is as follows: Figure 3 As shown; will be trained in the same way to obtain K By combining several decision trees, a random forest can be obtained. When a sample to be classified is input, the classification result output by the random forest is determined by a simple vote (i.e., taking the mode) of the outputs of each decision tree. The random forest classification process is as follows: Figure 4 As shown; Random forests are more tolerant of noise and outliers, can process input samples with high-dimensional features without dimensionality reduction, and can evaluate the importance of each feature in classification problems. They also have good scalability and parallelism.
[0019] Random forest algorithms can perform regression analysis based on classification. By performing certain calculations on the results of sample classification, the importance of each feature can be obtained. The importance of a feature indicates the degree of influence of the feature on the prediction result. The greater the importance of a feature, the greater its influence on the prediction result, and the smaller the importance, the smaller its influence on the prediction result.
[0020] Random forests make predictions by constructing multiple decision trees and performing ensemble voting (or averaging for regression problems); a single decision tree divides the data space into multiple homogeneous sub-regions by recursively selecting the best features and split points. The mid-term turbulence prediction model D can be expressed as: (2) in, B For the number of decision trees, T b Indicates the first b A decision tree, X The input feature vector; Feature selection and training data construction: Model D uses data from 02:00 to 05:00 as the training period; feature variables are selected from physical quantities output by the NWP model, including but not limited to: near-surface meteorological elements such as wind speed, temperature, and humidity of each standard isobaric surface, as well as dynamic and thermodynamic parameters such as vertical wind shear VWS and Richardson number Ri derived from them; the target variable is the high-resolution EDR field calculated by model C during the same period.
[0021] After interpolating the NWP data to a 3D grid consistent with EDR, each grid point constitutes a sample at each time step. X i , y i ;in, X i For feature vectors,y i The corresponding true EDR values were used to construct a training dataset.
[0022] A random forest model was implemented using the Scikit-learn library; key hyperparameters were optimized using grid search and 5-fold cross-validation; the optimized parameters included: the number of decision trees n and the maximum tree depth depth. max The optimization objective is to minimize the mean squared error (MSE) on the cross-validation set. After optimization, the final hyperparameter combination is determined as follows: n=300, depth... max =25, and the remaining parameters use the library default values; train the final model using all training data and the optimal hyperparameters.
[0023] Step 3: Construct a short-term turbulence forecasting model E based on the parameter inversion algorithm; Step 31: The prediction framework of Model E is based on two physical concepts: first, turbulence intensity, as an atmospheric property, exhibits spatial distribution variations that follow scalar advection laws; second, turbulence energy decays in an approximately exponential manner without external forcing. Therefore, these two processes are decoupled, forming the theoretical basis for the prediction equations. Constructing an advection-attenuation physical framework, the prediction equations are as follows: (3) in, For Lagrange operators; t 0 The starting time; Step size; Let be the parameters, representing the advection velocity in the east-west direction, the advection velocity in the north-south direction, and the decay time scale, respectively; collectively denoted as . .
[0024] Step 32: Learning algorithm based on historical sequences; The parameters in formula (3) can be determined by solving a nonlinear optimization problem; for each grid point, the predicted start time is used. t The loss function is constructed using historical data from the three hours prior to 0, and is used to measure the difference between the extrapolated predicted values and the actual observed values. The formula for the loss function is: (4) in, t i It is a historical point in time. N It represents the number of valid sample pairs.
[0025] Step 33: Invert the advection velocity by minimizing the loss function. u e ,v e and attenuation coefficient After obtaining the core parameters, we get: (5) The constructed nonlinear time-dimensional model E can be applied to calculate the turbulence intensity in three-dimensional space between points 05 and 06 by following these steps: 1. In t 0 At 05:00, the optimal parameters for each grid point across the entire field are obtained. Finally, the prediction process begins; 2. Analyze the EDR field of model C at 05:00. C This is the initial data; 3. Determine the prediction step size t 5 minutes, calculate future time. EDR E t0 t ; 4. Using the forecast results from the previous step as new initial data, repeat step 2 until the target forecast lead time of 06:00 is reached.
[0026] Using the output of the baseline model C as the initial field, short-term turbulence prediction is completed through iterative calculations at 5-minute intervals; RMSE, MAE, and R... 2 The reliability of the model is verified by indicators, requiring RMSE ≤ 0.025, MAE ≤ 0.015, and R0.025. 2 ≥0.9.
[0027] Step 4: Optimal route planning based on turbulence forecast results; Optimal route planning aims to find a route that minimizes the total turbulence intensity experienced by the aircraft throughout its flight. This objective function minimizes the spatiotemporal integral of the turbulence intensity along the entire flight path, and its discrete form can be expressed as: (5) in, J The objective function is... t o The starting time; t f The termination time; EDR ( x , y , z , t() refers to either the medium-term turbulence forecasting model D or the short-term turbulence forecasting model E; the short-term usually refers to 0-1 hours, and the medium-term usually refers to 3-6 hours.
[0028]
[0029] min J Clearly define the safety distance constraints between the planned route and the no-fly zone boundary; By integrating forecast data from the mid-term turbulence forecasting model and the short-term turbulence prediction module, the turbulence distribution during the planning period can be obtained. The optimal path is solved using a linear programming algorithm, and recommended flight altitude and route are output. In this invention, the time-period turbulence prediction model D requires the turbulence intensity model C as the verification standard, and uses the given data to build a model to predict the future turbulence intensity; the short-period turbulence prediction model E requires only three consecutive hours of data to build a model to predict the turbulence intensity in the next hour and verify its reliability; the difficulty lies in the fact that the medium-period turbulence prediction model D needs to extract turbulence correlation features from low-resolution numerical forecast data and optimize it using the turbulence intensity model C as the standard; the short-period turbulence prediction model E relies only on 3 hours of observation for nonlinear extrapolation, making it difficult to capture the randomness of turbulence; the route planning needs to quantify the objective function under the goal of minimizing turbulence.
[0030] Random forest regression algorithm and parameter inversion algorithm were used to establish a mid-term turbulence prediction model D and a short-term turbulence prediction model E, respectively. Based on the results, the turbulence intensity from 05:00 to 08:00 and from 05:00 to 06:00 was predicted. RMSE, MAE, and R were calculated. 2 The reliability of the short-term turbulence prediction model E was verified. Finally, a linear regression equation was established to obtain the optimal route based on the medium-term turbulence prediction model D and the short-term turbulence prediction model E.
[0031] Experimental results: Based on model C, calculate the turbulence intensity at different heights from point 02 to point 05, such as... Figure 2 As shown, Figure 2 The left figure visually reveals the non-uniform distribution characteristics of turbulence in three-dimensional space; the shape of the turbulence is an irregular clump or layered structure, which indicates that the intensity of turbulence changes with spatial location; the turbulence in the low-altitude boundary layer is often more active, while the structure of turbulence gradually becomes sparser with increasing altitude. Figure 2 As shown in the right figure, the distribution trend of turbulence in the horizontal direction is presented by horizontal slices, which are characterized by patchy or banded patterns. At low altitudes, due to the influence of surface friction and thermal disturbance, the turbulence intensity is usually high and the distribution is more chaotic. However, as the altitude increases, the turbulence will be more controlled by the large-scale wind field and will show a more orderly gradient change.
[0032] Model Forecast: The trained model D is applied to the NWP forecast data from 05 to 08; for each grid point within the forecast period, its feature vector is extracted. X forecast By inputting model D, the predicted EDR value for that point can be obtained. The results of calculating the turbulence intensity in three-dimensional space from point 05 to 08 using the established model D are as follows: Figure 5 , Figure 5 At a height of 100 meters near the ground, turbulence intensity is at its highest and exhibits significant spatial variability, primarily due to the influence of surface friction and thermal convection. As the altitude increases to 500 meters, turbulence gradually begins to detach from the surface, with increased wind shear and weather system influences, leading to a zonal distribution. At the lower atmosphere at 1000 meters, turbulence intensity further weakens but sudden bursts of strong turbulence occur, resulting in a more uniform spatial distribution. At 1500 meters, near the boundary layer top, boundary layer top shear occurs, intensifying the turbulence. This vertical variation trend from strong to weak and from localized to systemic fully reveals the three-dimensional structural characteristics of turbulence transitioning from surface forcing to atmospheric dynamic processes.
[0033] Figure 6 The forecast indicates that turbulence intensity gradually increases from 05:00, reaching its peak between 06:30 and 07:30, followed by a slight decline before 08:00. This diurnal variation pattern is mainly due to the development of thermal instability caused by enhanced solar radiation. Surface heating begins after 05:00, while boundary layer convection intensifies. The combined effect of wind shear and thermal instability leads to a continuous increase in turbulence intensity, reaching its strongest point in the early afternoon, when turbulence activity is also most vigorous. The upward trend in the forecast map indicates that the morning period is a high-incidence period for low-level turbulence, which poses a significant threat to aviation safety. Therefore, special attention needs to be paid to the critical period between 06:00 and 08:00.
[0034] Model validation: Verifying the reliability of the established model E by evaluating its RMSE, MAE, and R-squared values. 2 The calculations are performed separately, and the results are as follows; Table 1. RMSE, MAE, and R of Model E 2
[0035] The calculation results show that the RMSE value of model E is small, indicating that the overall bias control is good; MAE reflects the average absolute deviation between the predicted value and the true value, and this value is also at a small level, while the model is not greatly affected by extreme errors. R 2 A model with a value close to 0.9 fits the data well; therefore, this demonstrates the reliability of model E.
[0036] After the Figure 7 and Figure 8 The analysis revealed that the intensity of three-dimensional turbulence during the period from 5 to 6 o'clock showed a clear spatiotemporal evolution pattern. The intensity of turbulence was relatively low at 5:00, then reached its peak at 5:30, and although it weakened at 6:00, it still maintained a moderate intensity. In terms of spatial distribution, the high intensity of turbulence was mainly concentrated in the vortex center and areas with complex topography, and slowly moved eastward over time.
[0037] Vertically, the turbulence is strongest and most significant in the lower layers, but relatively stable in the middle layers, and gradually decreasing in the upper layers. This spatiotemporal distribution is mainly controlled by wind shear, thermal instability, and topographic effects, reflecting the typical characteristics of atmospheric boundary layer evolution during the morning transition period.
[0038] Planning Results and Analysis: Route planning was performed based on the EDRs provided by Model D and Model E respectively, resulting in two optimal routes, such as... Figure 9 and 10 As shown, by Figure 9 It can be seen that the route is at an altitude of 500 meters, an optimal flight path from west to east. The generated route successfully avoids three rectangular no-fly zones, effectively avoids three strong turbulence zones in the northwest, central and southeast, and utilizes the relatively calm zone in the southwest, keeping the overall turbulence intensity at a low level. Although the route is slightly longer than the straight-line distance due to the detour, it achieves the best balance between safety and comfort. The path is smooth and feasible, and can provide a reliable reference for actual flights.
[0039] like Figure 10 This is the optimal flight path for model E, inferred based on limited observational data. The turbulence field exhibits a non-uniform distribution, with highly turbulent areas mainly concentrated on the right side of the grid. The optimal flight path successfully bypasses these no-fly zones, selecting a path with acceptable turbulence intensity and high reliability of observational data. The path from the start to the end point is not a straight line but rather a circuitous route through multiple waypoints. Furthermore, the flight path maintains a safe distance around no-fly zones, ensuring flight compliance.
[0040] The comparison shows that the route planned by Model D can more effectively avoid strong turbulence areas in the medium and long term, while the route of Model E is similar to that of Model D in the short term, but lacks the ability to avoid long-term systems; this confirms the importance of accurate medium-term forecasts for long-term route planning.
[0041] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.
Claims
1. A method for low-altitude turbulence forecasting and flight path planning based on machine learning, characterized in that, Includes the following steps: Step 1: Collect meteorological data from different sources; Step 2: Construct a medium-term turbulence forecasting model based on the random forest regression algorithm and / or a short-term turbulence forecasting model based on the parameter inversion algorithm; Step 3: Set safety distance constraints between the planned path and the no-fly zone boundary for the medium-term turbulence forecasting model and / or the short-term turbulence forecasting model.
2. The method for low-altitude turbulence forecasting and flight path planning based on machine learning according to claim 1, characterized in that, The formula for the mid-term turbulence prediction model is: ； in, B For the number of decision trees, T b Indicates the first b A decision tree, X The input feature vector.
3. The method for low-altitude turbulence forecasting and flight path planning based on machine learning according to claim 1, characterized in that, Short-term turbulence prediction models include: Constructing an advection-attenuation physical framework, the formula is: ； in, For Lagrange operators; t 0 The starting time; Step size; Let represent the advection velocity in the east-west direction, the advection velocity in the north-south direction, and the decay time scale, respectively; collectively denoted as . ; Based on the three-dimensional turbulence intensity model; The loss function is constructed as follows: ； in, t i It is a historical point in time. N It is the number of valid sample pairs; Inversion by minimizing the loss function u e , v e and ,get: .
4. The machine learning-based low-altitude turbulence forecasting and route planning method according to any one of claims 2 or 3, characterized in that, The formula for the safe distance constraint between the planned path and the no-fly zone boundary is: ； in, t o The starting time; t f The termination time; EDR ( x , y , z , t This refers to a medium-term or short-term turbulence forecasting model. .
5. The low-altitude turbulence forecasting and route planning method based on machine learning according to claim 2, characterized in that, Hyperparameter optimization of mid-term turbulence prediction models was achieved using grid search and 5-fold cross-validation.
6. The method for low-altitude turbulence forecasting and flight path planning based on machine learning according to claim 1, characterized in that, Meteorological data includes: wind speed, temperature, altitude, atmospheric pressure, and humidity.
7. The method for low-altitude turbulence forecasting and flight path planning based on machine learning according to claim 1, characterized in that, Meteorological data were preprocessed using natural neighbor interpolation, Z-score standardization, 3σ criterion, and Pearson correlation coefficient method.
8. A machine learning-based low-altitude turbulence forecasting and flight path planning system, characterized in that, include: Memory is used to store instructions that can be executed by the processor; A processor for executing instructions to implement the machine learning-based low-altitude turbulence forecasting and route planning method as described in any one of claims 1-7.
9. A computer-readable medium storing computer program code, characterized in that, The computer program code, when executed by a processor, implements the machine learning-based low-altitude turbulence forecasting and route planning method as described in any one of claims 1-7.