An air route EDR intelligent prediction method based on airborne QAR data

Abstract

The application discloses a kind of route EDR intelligent prediction methods based on airborne QAR data, belong to aviation meteorology field.The method first collects national flight QAR data and airborne EDR sample and is matched;Dynamic stratification weight sampling strategy is used to construct balanced training dataset;Through sliding window mechanism, construct the spatiotemporal correlation of block prediction dataset;Using multiscale ConvLSTM2D, extreme learning machine and the heterogeneous model of chaos gate cycle unit extracted spatiotemporal characteristics and realized turbulence prediction;Introduce cosine annealing weight mechanism optimization training, focus on high EDR sample;Finally, according to EDR value, divide low, middle, high risk interval, respectively using different evaluation criteria carries out risk assessment.The method can accurately predict route EDR value, with good fitting ability and generalization performance, enhance the interpretability of model, applicable to complex atmospheric turbulence scene under the whole route risk prediction.

Description

Technical Field

[0001] This invention belongs to the field of aviation meteorology, specifically relating to a method for intelligent prediction of flight path EDR based on airborne QAR data. Background Technology

[0002] With the intensification of global warming and the rapid expansion of civil aviation transport, aircraft are encountering extreme turbulence more frequently along their routes. This weather disturbance not only significantly reduces the quality of the passenger experience but also poses potential risks to the controllability of the navigation system. Severe atmospheric turbulence can also lead to serious consequences such as structural damage to the aircraft and accidental injuries to passengers. Therefore, effective prediction of turbulence is of paramount importance.

[0003] Current turbulence prediction methods mainly include: first, important weather information in the flight plan, which is usually communicated by the dispatcher to the flight crew 2-3 hours before flight, informing them of the turbulence conditions along the route; and second, pilot reporting, where pilots proactively report turbulence information and broadcast it to following aircraft after the preceding aircraft encounters turbulence. However, due to the current difficulty in transmitting long-term sequential data between aircraft, once the preceding aircraft observes or encounters actual turbulence, its own flight data cannot be transmitted to following aircraft in real time or downloaded in real time, making it difficult to accurately quantify and utilize turbulence information in real time. Summary of the Invention

[0004] To address the aforementioned problems, the present invention aims to provide a method for intelligent prediction of flight route EDR based on airborne QAR data.

[0005] To achieve the above objectives, the intelligent EDR prediction method based on airborne QAR data provided by the present invention includes the following steps performed in sequence:

[0006] In step 1), airborne QAR data and airborne EDR samples from various types of flights nationwide are collected. Location information (call, time, tail number) is used to collect airborne QAR data involving parameters such as vertical overload, radio altitude, indicated airspeed, angle of attack, flaps, pitch rate, roll angle, vertical speed, total weight, supersonic airspeed, Mach number, geographical location, wind speed, wind direction, and static air temperature. The airborne EDR samples are then matched with the airborne QAR data.

[0007] In step 2), to ensure balanced learning of the prediction model, a dynamic hierarchical weight sampling strategy is adopted when constructing the training dataset: firstly, based on the EDR values ​​(low / medium / high intensity corresponding to 0-0.1, 0.1-0.3, 0.3+m), a dynamic hierarchical weight sampling strategy is used. 2 / s 3The data is partitioned by physical meaning (intervals), and sampling probabilities for each turbulence intensity level are constructed—by statistically analyzing the sample size of each bin and assigning it an inversely proportional weight. These weight parameters are adaptively updated with each epoch iteration. Finally, a normalized probability density function is used to achieve systematic, equally probable sampling of samples from each turbulence layer, thereby eliminating the interference of EDR distribution skewness on parameter estimation.

[0008] In step 3), data blocks are extracted using a sliding window mechanism, and temporal mapping highlights the future EDR evolution characteristics, constructing a block prediction dataset with complete spatiotemporal correlation characteristics. This architecture is specifically designed for heterogeneous architectures such as multi-scale ConvLSTM2D, Extreme Learning Machine (ELM), and Chaotic Gated Recurrent Unit (ChaoticGRU) to realize turbulence prediction and dynamic risk assessment models.

[0009] In step 4), multi-scale ConvLSTM2D is used to extract the spatiotemporal features of the window data. This method combines a convolutional neural network (CNN) and a long short-term memory network (LSTM) to process spatiotemporal sequence data. Its core formula is as follows:

[0010] Input Gate:

[0011]

[0012] Forgotten Gate:

[0013]

[0014] Cell status update:

[0015]

[0016] Output gate:

[0017]

[0018] The attention mechanism generates feature weights through a softmax function, strengthening core features relevant to EDR prediction and weakening the interference of irrelevant features. In the model, this mechanism is applied to the ELM output, optimizing the feature selection process and improving the fitting accuracy for complex EDR patterns. The weighted summation in the formula achieves dynamic feature integration, ensuring the model focuses on high-impact features crucial for turbulence prediction.

[0019] ELM is a single-hidden-layer feedforward neural network, and its core formula is:

[0020] Hidden layer output:

[0021] H = g(W·X + b) (17)

[0022] Output weight calculation:

[0023]

[0024] ELM optimizes only the output weights by randomly initializing input weights and biases, avoiding the tedious iterations of traditional gradient descent and thus exhibiting high computational efficiency. In EDR prediction, ELM is used to quickly map spatiotemporal features extracted from ConvLSTM2D to EDR values, shortening training time while maintaining excellent nonlinear mapping capabilities. (The pseudo-inverse calculation in the formula is mentioned.) It ensures the stability of the least squares solution and is suitable for processing high-dimensional spatiotemporal data.

[0025] ChaoticGRU is a variant of GRU that adds Lyapunov stability constraints. Its core formulas include:

[0026] Update Gate:

[0027] z t =σ(W z ·[h t-1 x t ]+b z (19)

[0028] Reset Door:

[0029] r t =σ(W r ·[h t-1 x t ]+b r (20)

[0030] Lyapunov stability constraint regularization loss:

[0031]

[0032] GRU Part: GRU controls the retention and updating of information through update and reset gates, making it suitable for processing time series data (such as predicting EDR values ​​n steps ahead). The update gate determines how much historical information is retained, while the reset gate controls the influence of the current input on the candidate state. Lyapunov Constraints: The Lyapunov factor limits the dynamic range of the hidden state through exponential decay (exp(-|λ|)), preventing the model from exhibiting chaotic or unstable behavior in long-term predictions. Lyapunov Loss (L... lyapunov As a regularization term, it encourages the model to learn stable dynamic systems. In EDR prediction, ChaoticGRU, combined with Lyapunov constraints, ensures that the model maintains the stability and consistency of the time series when predicting multiple future steps, thus handling the nonlinear dynamics in turbulent data.

[0033] In step 5), the airborne QAR data and EDR are adapted to fit the model. A cosine annealing weighting mechanism is applied to high-value EDR samples, focusing on high EDR values ​​early on and balancing all samples later. The core formula of the cosine annealing weighting mechanism is as follows:

[0034]

[0035] Furthermore, the weights of the samples are dynamically adjusted based on their EDR values, with samples having higher EDR values ​​receiving greater weight. The corresponding mathematical formula is as follows:

[0036]

[0037] Where t represents the number of iterations (epochs), cosine annealing smoothly adjusts the learning rate using a cosine function, from the maximum value (η). max ) gradually decreases to the minimum value (η) min The simulation annealing process is used. The characteristics of the cosine function cause the learning rate to change slowly at the beginning of the cycle and decrease rapidly near the end of the cycle. The model explores a large parameter space in the early stage and makes adjustments in the later stage.

[0038] In step 6), the output EDR value is adapted to risk assessment, divided into three risk assessment intervals based on the EDR value, and different assessment methods are used for risk assessment. The low-risk interval (0 ≤ EDR < 0.1) uses an absolute difference between the actual and predicted values ​​of less than 0.02 as the criterion for accurate prediction. The medium-risk interval (0.1 ≤ EDR < 0.2) assesses the prediction effect through the mean squared error (MSE); the smaller the MSE value, the higher the prediction accuracy. The overall mathematical formula is:

[0039]

[0040] The medium-risk range (0.1 ≤ EDR < 0.2) corresponds to moderate turbulence intensity. MSE amplifies larger deviations through squared errors and is used to evaluate the model's overall prediction performance in medium-risk scenarios. The high-risk range (EDR ≥ 0.2) uses an absolute difference of less than 0.1 between the actual and predicted values ​​as the criterion for accurate prediction.

[0041] In step 7), based on airborne QAR data from real flight routes, data blocks are segmented using a sliding time window to extract spatiotemporal features. Future EDR is predicted using a time mapping and then input into the trained multi-scale turbulence_pre model. A cosine annealing weighting mechanism is employed to optimize model training and predict future turbulence risks. The prediction results are used to divide risk assessment intervals based on EDR values, adhering to relevant risk assessment standards to ensure high-precision turbulence prediction and full-route risk assessment.

[0042] The intelligent EDR prediction method for flight routes based on airborne QAR data provided by this invention has the following beneficial effects:

[0043] 1. This method is based on fitting eddy current dissipation rate (EDR) data transmitted from a large number of airborne algorithms for various flight routes. Compared with traditional EDR prediction methods, it can achieve more accurate and long-term predictions of EDR values ​​for a single flight route. Its prediction results show a significant correlation with parameters characterizing flight turbulence, such as vertical overload, and can effectively depict the dynamic changes in atmospheric turbulence.

[0044] 2. This model exhibits excellent fitting ability and maintains robust performance in detections independent of the training data. Advanced feature extraction processing effectively suppresses overfitting in the validation data, thereby ensuring the model's generalization ability.

[0045] 3. Through advanced feature extraction and importance ranking, high-frequency components of the airborne QAR parameters associated with flight turbulence were identified as key features. When these feature values ​​were input into a neural network to achieve EDR prediction fitting, the model exhibited strong physical interpretability, partially alleviating the inherent "black box" limitations of machine learning.

[0046] 4. This method is based on iterative neural network calculations. By mapping airborne QAR data with existing airborne calculated EDR data, accurate EDR prediction can be achieved without manually setting calculation rules. Research shows that this data-driven method of automatically capturing the correlation between the two demonstrates broad engineering application potential in the field of turbulence monitoring.

[0047] 5. By fitting EDR window QAR data from a large number of flights nationwide, the prediction of EDR for a single route was achieved, demonstrating strong applicability and broad application potential. Attached Figure Description

[0048] Figure 1 A schematic diagram of the intelligent EDR prediction method for airborne QAR data provided by this invention;

[0049] Figure 2 This is a schematic diagram illustrating the matching of airborne QAR parameters with vortex energy dissipation rate (EDR) and the division of the sliding time window in this invention.

[0050] Figure 3 This is a flowchart illustrating the EDR prediction using a hybrid deep learning model in this invention. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0052] like Figure 1 As shown, the intelligent EDR prediction method based on airborne QAR data provided by this invention includes the following steps performed in sequence:

[0053] 1) Collect airborne QAR data for all types of flights nationwide. Specific steps include collecting airborne QAR data and EDR samples transmitted from the aircraft, covering parameters such as vertical overload, radio altitude, indicated airspeed, angle of attack, flaps, pitch rate, roll angle, vertical speed, total weight, supersonic airspeed, Mach number, geographical location, wind speed, wind direction, and static air temperature. Matching is performed based on the aircraft flight number, flight time, and tail number. The extracted data contains 18 flight parameters; details are shown in Table 1.

[0054] Table 1

[0055]

[0056] 2) To ensure balanced learning of the prediction model, a dynamic hierarchical weight sampling strategy is adopted when constructing the training dataset: First, based on the EDR value (low / medium / high intensity corresponding to 0-0.1, 0.1-0.3, 0.3+m), a dynamic hierarchical weight sampling strategy is used. 2 / s 3 The data is partitioned by physical meaning (intervals), and sampling probabilities for each turbulence intensity level are constructed—by statistically analyzing the sample size of each bin and assigning it an inversely proportional weight. These weight parameters are adaptively updated with each epoch iteration. Finally, a normalized probability density function is used to achieve systematic, equally probable sampling of samples from each turbulence layer, thereby eliminating the interference of EDR distribution skewness on parameter estimation.

[0057] 3) By extracting data blocks through a sliding window mechanism and using temporal mapping to highlight the future EDR evolution characteristics, a block prediction dataset with complete spatiotemporal correlation characteristics is constructed. This architecture is specifically designed for heterogeneous architectures such as multi-scale ConvLSTM2D, Extreme Learning Machine (ELM), and Chaotic Gated Recurrent Unit (ChaoticGRU) to realize turbulence prediction and dynamic risk assessment models.

[0058] 4) Multi-scale ConvLSTM2D is used to extract the spatiotemporal features of the window data. It combines a convolutional neural network (CNN) and a long short-term memory network (LSTM) to process spatiotemporal sequence data. Its core formula is as follows:

[0059] Input Gate:

[0060]

[0061] Forgotten Gate:

[0062]

[0063] Cell status update:

[0064]

[0065] Output gate:

[0066]

[0067] The attention mechanism generates feature weights through a softmax function, strengthening core features relevant to EDR prediction and weakening the interference of irrelevant features. In the model, this mechanism is applied to the ELM output, optimizing the feature selection process and improving the fitting accuracy for complex EDR patterns. The weighted summation in the formula achieves dynamic feature integration, ensuring the model focuses on high-impact features crucial for turbulence prediction.

[0068] ELM is a single-hidden-layer feedforward neural network, and its core formula is:

[0069] Hidden layer output:

[0070] H=g(W·X+b)#(29)

[0071] Output weight calculation:

[0072]

[0073] ELM optimizes only the output weights by randomly initializing input weights and biases, avoiding the tedious iterations of traditional gradient descent and thus exhibiting high computational efficiency. In EDR prediction, ELM is used to quickly map spatiotemporal features extracted from ConvLSTM2D to EDR values, shortening training time while maintaining excellent nonlinear mapping capabilities. (The pseudo-inverse calculation in the formula is mentioned.) It ensures the stability of the least squares solution and is suitable for processing high-dimensional spatiotemporal data.

[0074] ChaoticGRU is a variant of GRU that adds Lyapunov stability constraints. Its core formulas include:

[0075] Update Gate:

[0076] z t =σ(W z ·[h t-1 x t ]+b z )#(31)

[0077] Reset Door:

[0078] r t =σ(W r ·[h t-1 x t ]+b r )#(32)

[0079] Lyapunov stability constraint regularization loss:

[0080]

[0081] GRU Part: GRU controls the retention and updating of information through update and reset gates, making it suitable for processing time series data (such as predicting EDR values ​​n steps ahead). The update gate determines how much historical information is retained, while the reset gate controls the influence of the current input on the candidate state. Lyapunov Constraints: The Lyapunov factor limits the dynamic range of the hidden state through exponential decay (exp(-|λ|)), preventing the model from exhibiting chaotic or unstable behavior in long-term predictions. Lyapunov Loss (L... lyapunov As a regularization term, it encourages the model to learn stable dynamic systems. In EDR prediction, ChaoticGRU, combined with Lyapunov constraints, ensures that the model maintains the stability and consistency of the time series when predicting multiple future steps, thus handling the nonlinear dynamics in turbulent data.

[0082] 5) Adapt the airborne QAR and EDR data to fit the model. Apply a cosine annealing weighting mechanism to high-value EDR samples, focusing on high EDR values ​​early on and balancing all samples later. The core formula of the cosine annealing weighting mechanism is as follows:

[0083]

[0084] Where t represents the number of iterations (epochs), cosine annealing smoothly adjusts the learning rate using a cosine function, from the maximum value (η). max ) gradually decreases to the minimum value (η) min The simulation annealing process is used. The characteristics of the cosine function cause the learning rate to change slowly at the beginning of the cycle and decrease rapidly near the end of the cycle. The model explores a large parameter space in the early stage and makes adjustments in the later stage.

[0085] 6) The output EDR value is adapted to risk assessment, divided into three risk assessment intervals based on the EDR value, and different assessment methods are used for risk assessment. In the low-risk interval (0≤EDR<0.1), the absolute difference between the actual value and the predicted value is less than 0.02 as the criterion for accurate prediction. In the medium-risk interval (0.1≤EDR<0.2), the prediction effect is evaluated by the magnitude of the mean squared error (MSE); the smaller the MSE value, the higher the prediction accuracy. The overall mathematical formula is:

[0086]

[0087] The medium-risk range (0.1 ≤ EDR < 0.2) corresponds to moderate turbulence intensity. MSE amplifies larger deviations through squared errors and is used to evaluate the model's overall prediction performance in medium-risk scenarios. The high-risk range (EDR ≥ 0.2) uses an absolute difference of less than 0.1 between the actual and predicted values ​​as the criterion for accurate prediction.

[0088] 7) Based on airborne QAR data from real flight routes, data blocks are segmented using a sliding time window to extract spatiotemporal features. Future EDR (Electronic Deflection Rate) is predicted using time mapping and then input into a trained multi-scale turbulence_pre model. A cosine annealing weighting mechanism is employed to optimize model training and predict future turbulence risks. The prediction results are used to divide risk assessment intervals based on EDR values, adhering to relevant risk assessment standards to ensure high-precision turbulence prediction and full-route risk assessment.

[0089] The above description is a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications and substitutions made by those skilled in the art within the scope of the technology disclosed in the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for intelligent prediction of flight path EDR based on airborne QAR data, characterized in that: The intelligent EDR prediction method based on airborne QAR data includes the following steps performed in sequence: S1 collects airborne QAR data and eddy current dissipation rate (EDR) samples from various types of flights nationwide, utilizing location information (call, time, tail number). The EDR samples are then matched with the airborne QAR data. S2 interval sampling balances the EDR sample by dividing the EDR values ​​into multiple intervals according to their statistical distribution. The number of samples in each interval is counted, and the sampling weight for each interval is calculated. Based on these weights, samples are randomly drawn from each interval, ensuring that the sampling probability matches the weight, ultimately generating a turbulence dataset with a balanced sample size. The S3 design uses a sliding window for prediction, a 750-second feature window (QAR), and a 5-minute label window (EDR). This window is applied to a large number of flight paths to create a turbulence prediction dataset. S4 constructs a Turb-Pre model and uses ConvLSTM to extract spatiotemporal data of the window; an ELM model with attention mechanism is used to quickly fit the future turbulence pattern; ChaoticGRU is used to predict the EDR for the next 5 minutes. S5 employs a dynamic cosine annealing weighting mechanism, initially focusing on training high EDR samples, and gradually balancing the weights of all samples in the later stages. It combines an early stopping mechanism to optimize model convergence and uses the Huber loss function to enhance the model's robustness to abnormal EDR values, balancing the advantages of mean squared error (MSE) and mean absolute error (MAE). The EDR value output by the S6 model is adapted to the turbulence prediction and assessment method designed for flight risk. It is divided into three risk assessment intervals according to the magnitude of the EDR value, and risk assessment is carried out according to different assessment methods. S7 uses airborne QAR data from real flight routes, divides it into time windows, and inputs the feature values ​​into a trained model. It then uses the model weights to predict future turbulence risks.

2. The intelligent EDR prediction method for flight routes based on airborne QAR data according to claim 1, characterized in that: In S1, airborne QAR data for all types of flights across the country is collected. The specific steps include collecting airborne QAR data and EDR samples transmitted from the airborne system, which involve parameters such as vertical overload, radio altitude, indicated airspeed, angle of attack, flaps, pitch rate, roll angle, vertical speed, total weight, vacuum speed, Mach number, geographical location, wind speed, wind direction, and static air temperature. The data is then matched based on the aircraft flight number, flight time, and tail number.

3. The intelligent EDR prediction method for flight routes based on airborne QAR data according to claim 1, characterized in that: In S2, the weighted interval sampling process is as follows: the EDR value is divided into multiple intervals according to the statistical distribution (e.g., low, medium, and high turbulence intensities: 0-0.1, 0.1-0.3, 0.3+m). 2 / s 3 The process involves counting the number of samples in each interval and calculating the sampling weight for each interval. The weight is inversely proportional to the sample size (intervals with fewer samples have higher weights). During sampling, samples are randomly drawn from each interval based on these weights, ensuring that the sampling probability matches the weight, ultimately generating a dataset with a balanced sample size. This entire process can be completed before training or dynamically executed in each training batch to adapt to the model training requirements.

4. The intelligent EDR prediction method for flight routes based on airborne QAR data according to claim 1, characterized in that: In S3, data blocks are extracted through a sliding window, and the future EDR trend is highlighted by combining time mapping. A multi-block prediction dataset is constructed, which retains complete spatiotemporal features and is adapted to multi-scale ConvLSTM2D, ELM and ChaoticGRU models to achieve high-precision turbulence prediction and risk assessment.

5. The intelligent EDR prediction method for flight routes based on airborne QAR data according to claim 1, characterized in that: In S4, multi-scale ConvLSTM2D is used to extract spatiotemporal features from window data. It combines a convolutional neural network (CNN) and a long short-term memory network (LSTM) to process spatiotemporal sequence data. Its core formula is as follows: Input Gate: Forgotten Gate: Cell status update: Output gate: The attention mechanism calculates feature weights using a softmax function, enhancing focus on key features for EDR prediction and suppressing the influence of irrelevant features. In the model, the attention mechanism acts on the ELM output, optimizing feature selection and improving the fitting ability to complex EDR patterns. The weighted summation in the formula achieves dynamic weighted combination of features, ensuring the model focuses on high-impact features relevant to turbulence prediction. ELM is a single-hidden-layer feedforward neural network, and its core formula is: Hidden layer output: H=g(W·X+b) (5) Output weight calculation: β=H T ·T (6) ELM avoids the complex iterations of traditional gradient descent by randomly initializing input weights and biases and training only the output weights, resulting in high computational efficiency. In EDR prediction, ELM is used to quickly fit the spatiotemporal features extracted from ConvLSTM2D to the EDR value, reducing training time while maintaining high nonlinear mapping capability. The pseudo-inverse calculation in the formula ensures the stability of the least squares solution, making it suitable for handling high-dimensional spatiotemporal data. ChaoticGRU is a variant of GRU that adds Lyapunov stability constraints. Its core formulas include: Update Gate: z t =σ(W z ·[h t-1 ,x t ]+b z ) (7) Reset Door: r t =σ(W r ·[h t-1 ,x t ]+b r ) (8) Lyapunov stability constraint regularization loss: GRU Component: GRU controls the retention and updating of information through update and reset gates, making it suitable for processing time series data (such as predicting EDR values ​​for the next n steps). The update gate determines how much historical information is retained, while the reset gate controls the impact of the current input on the candidate state. Lyapunov Constraints: The Lyapunov factor limits the dynamic range of the hidden state through exponential decay (exp(-|λ|)), preventing the model from exhibiting chaotic or unstable behavior in long-term predictions. The Lyapunov loss, using Lyapunov as a regularization term, encourages the model to learn stable dynamic systems. In EDR prediction, ChaoticGRU, combined with Lyapunov constraints, ensures that the model maintains the stability and consistency of the time series when predicting multiple future steps, handling the nonlinear dynamics in turbulent data.

6. The intelligent EDR prediction method for flight routes based on airborne QAR data according to claim 1, characterized in that: In S5, airborne QAR data and EDR are adapted to fit the model. A cosine annealing weighting mechanism is applied to high-value EDR samples, focusing on high EDR values ​​early on and balancing all samples later. The core formula of the cosine annealing weighting mechanism is as follows: Furthermore, the weights of the samples are dynamically adjusted based on their EDR values, with samples having higher EDR values ​​receiving greater weight. The corresponding mathematical formula is as follows: Here, t represents the number of iterations (epoch). Cosine annealing smoothly adjusts the learning rate using a cosine function, gradually decreasing it from its maximum value to its minimum value, simulating the annealing process. The characteristics of the cosine function cause the learning rate to change slowly at the beginning of the cycle and decrease rapidly near the end of the cycle. The model explores a large parameter space in the early stages and makes adjustments in the later stages.

7. The intelligent EDR prediction method based on airborne QAR data according to claim 1, characterized in that: The EDR value output by the S6 model is used for risk assessment, divided into three risk assessment intervals based on the EDR value, each assessed using a different method. The low-risk interval (0 ≤ EDR < 0.1) uses an absolute difference between the actual and predicted values ​​of less than 0.02 as the criterion for accuracy. The medium-risk interval (0.1 ≤ EDR < 0.2) assesses the prediction effect through the mean squared error (MSE); a smaller MSE indicates a more accurate prediction. The overall mathematical formula is: The medium-risk range (0.1 ≤ EDR < 0.2) represents moderate turbulence intensity. MSE penalizes larger deviations through squared errors, making it suitable for evaluating the overall predictive performance of the model in medium-risk scenarios. The high-risk range (EDR ≥ 0.2) uses an absolute difference between the actual and predicted values ​​of less than 0.1 as the criterion for accuracy.

8. The intelligent EDR prediction method based on airborne QAR data according to claim 1, characterized in that: Based on airborne QAR data from real flight routes, S7 divides the data into blocks using a sliding time window and extracts spatiotemporal features. These features are then combined with a time-mapped future EDR trend and input into a trained tur_pre model to predict future turbulence risk. The prediction results are assessed according to EDR value intervals to ensure high-precision turbulence prediction and comprehensive flight route risk assessment.

Images