Intelligent aviation flow prediction method based on MLP-LightGBM hybrid model

Abstract

The invention discloses an intelligent aviation flow prediction method based on an MLP-LightGBM hybrid model, and relates to the technical field of air transportation management, artificial intelligence and big data analysis. The problems of data isomerism and space-time alignment are solved by innovatively adopting periodic time-triangular transform coding and an aviation cloud influence index model. Then, constructing a'micro-macro 'two-stage prediction architecture: a micro layer processes unstructured meteorological data by using MLP, and predicts single flight delay; and the macroscopic layer aggregates the MLP result and combines the structured features, and the LightGBM performs hour-level flow accurate prediction and correction. According to the method, the three bottlenecks of insufficient critical mutation capture, cross-scale modeling fault and space-time decoupling in the high influence weather in the existing method are effectively solved, the prediction error is reduced by 7.2% compared with the optimal single model, the threshold accuracy is as high as 93.75%, and an advanced and reliable technical support is provided for improvement of airspace operation efficiency and prevention and control of large-area flight delay.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of air transportation management, artificial intelligence and big data analysis, and particularly relates to an air traffic intelligent prediction method based on a MLP-LightGBM hybrid model. BACKGROUND

[0002] As a key hub of modern social and economic operation, air transportation has shown explosive growth in recent years. As the second largest aviation market in the world, the daily takeoff and landing times of the core hub airport in China have exceeded 1100, and the airspace resources are under great pressure. This rapid growth directly leads to a sharp contradiction between the rigid constraints of airspace resources and the demand for operational efficiency, and the most prominent manifestation is the frequent occurrence of large-area flight delays.

[0003] According to the statistics of the Civil Aviation Administration of China, the national flight normal rate in 2023 is 87.8%, which means that more than 12% of flights are plagued by delays. The economic losses caused by delays are very significant, including direct costs such as passenger compensation and additional fuel consumption, as well as derivative social costs such as airport disorder and damage to the reputation of airlines. For example, the "Dusuoer" typhoon caused 2870 flights to be canceled within 72 hours at the East China airport group in 2023, resulting in direct economic losses of 920 million yuan.

[0004] Studies have shown that flight delays are a complex result of the dynamic coupling of multiple factors. Among them, weather conditions account for 62.3%, far higher than air traffic control restrictions (18.7%) and airline reasons (12.1%) and other factors. In particular, high-impact weather (HIW) such as thunderstorms and heavy fog has a significant nonlinear amplification effect, exhibiting three characteristics of low-frequency high-loss, runway capacity index decay, and network cascading effect.

[0005] However, in the high-dimensional multi-source heterogeneous big data environment generated by modern aviation operations, the current prediction methods face three bottlenecks in dealing with HIW scenarios: first, the lack of critical mutation capture, traditional feature engineering is difficult to effectively capture the dramatic change characteristics of meteorological parameters; second, the cross-scale modeling fault, there is a lack of effective transmission mechanism model between micro-meteorological changes and macro-traffic control; finally, the problem of spatio-temporal decoupling, the mainstream model simplifies the airport network into independent time series, ignoring its inherent topological network dynamics.

[0006] These defects are significant in practical applications. In extreme weather scenarios such as typhoons, the average absolute error of the existing CNN-LSTM model is still as high as 24.3 minutes, the prediction variance is 52.4, and the feasibility of the recovery scheme generated is less than 74%. This fully demonstrates that developing a new prediction framework that can integrate meteorological mutation response, cross-scale transmission mechanism and network topology dynamics has urgent theoretical value and practical significance. SUMMARY

[0007] The present application aims to overcome the three major bottlenecks of the existing aviation flow prediction methods under high-impact weather, namely, insufficient capture of critical mutations, cross-scale modeling faults, and spatio-temporal decoupling. The specific application purposes include: Significantly improve the prediction accuracy and robustness by constructing a MLP-LightGBM hybrid model and adopting a "micro-macro" two-level prediction mechanism to accurately depict the nonlinear aviation operation situation.

[0008] Break through the cross-scale modeling and spatio-temporal coupling problems by establishing an effective transmission mechanism between micro-meteorological changes, flight class delays, and airport-level flows, and integrating spatio-temporal features to accurately reflect the propagation law of delays in the airspace network.

[0009] Solve the multi-source heterogeneous data fusion dilemma by innovatively using periodic time triangle transformation coding and aviation cloud layer impact index model to achieve efficient fusion and feature extraction of flight operation data and meteorological data.

[0010] Optimize the model training efficiency by combining the efficient engineering characteristics of LightGBM and the early stopping method dynamic parameter tuning strategy to significantly improve the calculation efficiency while ensuring the prediction accuracy, providing practical technical support for aviation flow management.

[0011] To achieve the above application purposes, the present application provides the following technical solutions: In a first aspect, the present application provides an aviation flow intelligent prediction method based on a MLP-LightGBM hybrid model, comprising the following steps: Step S1: Obtain and preprocess multi-source heterogeneous aviation data, the aviation data at least including flight operation data and airport meteorological observation data; Step S2: Based on the preprocessed data, a "micro-macro" two-level hybrid prediction model is constructed and trained; wherein, The micro-prediction layer adopts a multi-layer perception machine MLP model, taking a single flight as a sample, inputting features including time, weather, and historical flow, and outputting the predicted delay time of the flight; The macro-prediction layer adopts a light gradient boosting machine LightGBM model, the input features of which include: the hourly prior flow features obtained by spatio-temporal aggregation of all flight prediction results output by the micro-prediction layer, and the original structured features from step S1; the output of the macro-prediction layer is the airport hourly flow prediction result; Step S3: Use the trained hybrid prediction model to predict the aviation flow of the target airport in the future period.

[0012] Further, the preprocessing of aviation data in step S1 specifically includes: The flight schedule timestamps are discretized, and the hours and minutes are periodically encoded using a trigonometric sine-cosine transform. The transform formula includes: ; Where h represents hours and m represents minutes.

[0013] Furthermore, the preprocessing of aviation data in step S1 also includes constructing an aviation cloud impact index: Analyze cloud information from meteorological observation data, and calculate the cloud impact index based on cloud cover and cloud base height using the following formula: ; Among them, C r C represents cloud coverage. h Where is the cloud base height, and k is a preset adjustment coefficient.

[0014] Furthermore, the preprocessing of aeronautical data in step S1 also includes a spatiotemporal data fusion step: A spatiotemporal matching algorithm based on a decay function is used to match the optimal meteorological observation record for each flight near its scheduled time. The formula for the decay function is: ; in, For the scheduled departure time of the flight, α represents the meteorological observation time, and α is the attenuation coefficient.

[0015] Furthermore, the input features of the macro-prediction layer LightGBM model also include dynamic engineering features, which include at least one of the following: Lag characteristics: historical traffic data for 1 hour, 2 hours, 3 hours, 24 hours, and 48 hours; Moving average characteristics: The average flow rate calculated based on 3-hour, 6-hour, and 12-hour sliding windows.

[0016] Furthermore, during the model training process in step S2, early stopping is used for dynamic parameter tuning, specifically including: During training, the loss function on the validation set is monitored in real time. When the loss function no longer decreases within N consecutive iterations, training is automatically terminated and the model parameters are rolled back to the optimal performance state, where N is a preset patience value.

[0017] Furthermore, the macro-prediction layer LightGBM model is trained using at least one of the following optimization strategies: Leaf growth strategy; One-sided gradient sampling method; Mutually exclusive feature binding method.

[0018] In a second aspect, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0019] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0020] Fourthly, the present invention provides an intelligent air traffic forecasting system for implementing the above-described method, the system comprising: The data preprocessing module is used to perform step S1 and is configured to clean, reconstruct and fuse multi-source heterogeneous aviation data. A hybrid model engine, used to perform step S2, is configured to carry and run the "micro-macro" two-level hybrid prediction model; The prediction and output module is used to perform step S3, and is configured to receive user requests, call the hybrid model engine to make predictions, and output a structured traffic prediction report.

[0021] Compared with the prior art, the present invention has the following advantages and beneficial effects: 1. Prediction accuracy has been significantly improved. Through an innovative MLP-LightGBM hybrid architecture, this invention achieves a qualitative leap in prediction accuracy. Experimental verification shows that the root mean square error (RMSE) of this invention is reduced by 7.2% compared to the single optimal model (LightGBM), and the threshold accuracy (error ≤15%) reaches as high as 93.75%. This significant improvement is mainly due to the complementary advantages of MLP's powerful nonlinear feature extraction capabilities and LightGBM's efficient structured feature processing capabilities. The micro-level MLP model can automatically learn complex nonlinear patterns from raw meteorological data and accurately predict single flight delays; the macro-level LightGBM then performs secondary corrections based on this, achieving a precise scale leap from micro to macro.

[0022] 2. Enhanced adaptability to complex scenarios This invention particularly enhances its adaptability to complex scenarios such as high-impact weather. Through an innovative aviation cloud impact index model and periodic time trigonometric transformation coding, the system effectively captures the critical abrupt changes in meteorological parameters. Simultaneously, the "micro-macro" two-level architecture establishes a complete multi-scale transmission mechanism, enabling the model to accurately reflect the dynamic process of the impact of micro-meteorological changes on macro-airflow. This characteristic allows the invention to maintain stable predictive performance even under high-impact weather scenarios such as typhoons and thunderstorms, providing an effective solution to the most challenging problems in airflow forecasting.

[0023] 3. Breakthrough in data processing capabilities This invention achieves significant breakthroughs in data processing. The proposed periodic time trigonometric transformation encoding innovatively solves the problem of abrupt changes in periodic boundaries inherent in traditional discrete encoding, perfectly preserving the temporal topological structure. The aviation cloud impact index model achieves, for the first time, a quantitative assessment of cloud impact, providing a scientific basis for meteorological impact analysis. The spatiotemporal fusion mechanism, through an attenuation matching algorithm, effectively solves the spatiotemporal misalignment problem between flight events and meteorological observations. These innovations collectively construct a complete and efficient data processing workflow.

[0024] 4. Significantly improved computational efficiency While maintaining accuracy, this invention achieves a significant improvement in computational efficiency through a multi-level optimization strategy. The synergistic effect of LightGBM's four main components—the Histogram algorithm, Leaf-wise growth strategy, unilateral gradient sampling, and mutually exclusive feature binding—significantly reduces computational and memory overhead. The introduction of early stopping enables intelligent control of the training process, avoiding ineffective iterations. Compared to a single deep learning model, this invention reduces training time by approximately 40% and memory usage by approximately 35% while maintaining accuracy, creating favorable conditions for practical deployment and application.

[0025] 5. Significant practical value This invention provides airspace management departments and airlines with a reliable decision support tool. Its accurate 24-hour traffic forecasting capability enables relevant departments to formulate traffic management strategies in advance, effectively preventing and mitigating large-scale flight delays. Statistics show that applying this system can increase the feasibility of flight recovery plans under extreme weather conditions from 74% using traditional methods to over 90%, demonstrating significant practical value in reducing operating costs and improving operational efficiency. Attached Figure Description

[0026] Figure 1 This is a flowchart of the multi-stage data preprocessing process of the present invention; Figure 2 This is a flowchart of the model training process of the present invention; Figure 3 This is a schematic diagram of the Level-wise and Leaf-wise strategies of the present invention; Figure 4 This is a flowchart of the model prediction process in this invention; Figure 5 This is a schematic diagram illustrating the working mechanism of the early stopping method in this invention; Figure 6 This is a schematic diagram showing the importance distribution of different features in this invention; Figure 7 This refers to the predicted average hourly traffic flow of Guangzhou Baiyun Airport over five random days in May 2025, as per this invention. Detailed Implementation

[0027] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0028] Firstly, this embodiment provides an intelligent air traffic forecasting method based on the MLP-LightGBM hybrid model, comprising the following steps: Step S1: Acquire and preprocess multi-source heterogeneous aviation data, wherein the aviation data includes at least flight operation data and airport meteorological observation data; Step S2: Based on the preprocessed data, construct and train a two-level hybrid prediction model of "micro-macro"; wherein, The micro-prediction layer adopts a multilayer perceptron (MLP) model. Taking a single flight as a sample, it inputs features including time, weather, and historical traffic flow, and outputs the predicted delay time of the flight. The macro-prediction layer adopts the Lightweight Gradient Boosting Machine (LightGBM) model. Its input features include: hourly prior traffic features obtained by spatiotemporal aggregation of all flight prediction results output by the micro-prediction layer, and the original structured features from step S1; the output of the macro-prediction layer is the hourly traffic prediction result of the airport. Step S3: Using the trained hybrid prediction model, predict the air traffic flow of the target airport in the future.

[0029] Furthermore, the preprocessing of aeronautical data in step S1 specifically includes: The flight schedule timestamps are discretized, and the hours and minutes are periodically encoded using a trigonometric sine-cosine transform. The transform formula includes: ; Where h represents hours and m represents minutes.

[0030] Furthermore, the preprocessing of aviation data in step S1 also includes constructing an aviation cloud impact index: Analyze cloud information from meteorological observation data, and calculate the cloud impact index based on cloud cover and cloud base height using the following formula: ; Among them, C r C represents cloud coverage. h Where is the cloud base height, and k is a preset adjustment coefficient.

[0031] Furthermore, the preprocessing of aeronautical data in step S1 also includes a spatiotemporal data fusion step: A spatiotemporal matching algorithm based on a decay function is used to match the optimal meteorological observation record for each flight near its scheduled time. The formula for the decay function is: ; in, For the scheduled departure time of the flight, α represents the meteorological observation time, and α is the attenuation coefficient.

[0032] Furthermore, the input features of the macro-prediction layer LightGBM model also include dynamic engineering features, which include at least one of the following: Lag characteristics: historical traffic data for 1 hour, 2 hours, 3 hours, 24 hours, and 48 hours; Moving average characteristics: The average flow rate calculated based on 3-hour, 6-hour, and 12-hour sliding windows.

[0033] Furthermore, during the model training process in step S2, early stopping is used for dynamic parameter tuning, specifically including: During training, the loss function on the validation set is monitored in real time. When the loss function no longer decreases within N consecutive iterations, training is automatically terminated and the model parameters are rolled back to the optimal performance state, where N is a preset patience value.

[0034] Furthermore, the macro-prediction layer LightGBM model is trained using at least one of the following optimization strategies: Leaf growth strategy; One-sided gradient sampling method; Mutually exclusive feature binding method.

[0035] The specific implementation process of this method is as follows: 1. Data Preprocessing The accuracy of aviation operation forecasting heavily relies on high-quality datasets. This study uses two core sets of raw data provided by China Southern Airlines: the first is "May Flight Operation Data (Anonymized)," containing over 70,000 historical flight operation records for the entire month of May, covering 25 full-process operation indicators such as flight number, aircraft type, planned / actual departure station four-letter code, and planned / actual departure time; the second is "May Airport Meteorological Observation Data," covering METAR format meteorological reports from various airports over 31 days (sample METARAGGH010000Z05004KT9999FEW020FEW021CB31 / 26Q1011RMCBTOE / SHTOS=), recording observation stations, dates, and multidimensional weather characteristics.

[0036] This study proposes a multi-stage preprocessing framework, such as... Figure 1As shown, this framework aims to integrate two heterogeneous data sources: flight operations and meteorological observations. Compared with traditional methods that employ shallow data analysis and coarse-grained time matching strategies, this framework focuses on addressing three core challenges: firstly, data heterogeneity, namely the structured differences between aviation operation parameters and meteorological elements; secondly, spatiotemporal misalignment, namely the spatiotemporal registration problem between discrete flight events and continuous meteorological observations; and thirdly, feature redundancy, namely noise interference caused by the dimensional expansion of multi-source data.

[0037] 1.1 Data Cleaning and Feature Reconstruction The core of flight operation data cleaning lies in the discretization of spatiotemporal features and the extraction of key dimensions. The original timestamp fields (such as "scheduled departure time," "scheduled takeoff time," and "scheduled arrival time," in the format "2023-05-01 08:30:00") are discretized into structured features. Given that this prediction model focuses on 24-hour traffic changes and flight operations exhibit monthly periodicity, the year-month dimension has a weak impact on short-term traffic fluctuations, so it is merged. The date, hour, and minute dimensions, however, have a significant impact on traffic distribution and are retained as independent input features. This processing converts the time field to the ISO 8601 standard, as shown in the following formula: ; Then Periodic feature transformation is performed, and time periodicity is captured through trigonometric transformation, as shown in the following formula: ; This periodic feature transformation innovatively solves the problems of abrupt changes and discontinuities in the periodic boundaries of traditional discrete coding, while using trigonometric transformation to perfectly preserve the topological structure of time and enhance the model's convergence efficiency.

[0038] The original airport code is a four-letter code, which is not a suitable data type for model input. We use LabelEncoding to map the airport code to numerical values: ; The above processing effectively captures the periodic patterns and traffic characteristics of flight operations, while eliminating redundant fields such as flight number and aircraft type. Data type conversion follows the tree model optimization principle, and airport codes are converted into categorical variables.

[0039] Meteorological data processing employs ICAO Annex 3 standards for message validity verification, filtering out non-METAR / TAF messages (such as SPECI). METAR messages are parsed using Python's built-in libraries. After parsing, 12 key aviation meteorological indicators are retained, with all distance quantities uniformly converted to metric units. Continuous meteorological features such as visibility are standardized using Z-scores, as shown in the following formula: ; This study innovatively proposes an aviation impact index model for cloud information (e.g., FEW020FEW021CB) in meteorological reports. Based on the WMO cloud classification standard, a formula for quantifying cloud impact is constructed: ; in Cloud coverage (FEW=0.25, SCT=0.5, BKN=0.75, OVC=1.0). Here, is the cloud base height, and k is a preset adjustment coefficient. This index comprehensively characterizes the inhibitory effect of clouds on flight.

[0040] Based on the above processing, the various weather data parsed from the METAR message are summarized, and the single weather phenomenon identifiers such as thunderstorms are converted into Boolean flags. The final summarized and processed effective weather information is shown in Table 1.

[0041] Table 1. Valid weather information extracted from METAR reports .

[0042] 1.2 Spatiotemporal Fusion Mechanism and Dynamic Feature Engineering Construction The fusion of flight and meteorological data employs a two-level spatiotemporal matching algorithm. Spatially, a mapping relationship between flight arrival and departure points and meteorological stations is established using airport ICAO codes (such as ZGGG). Temporally, an attenuation matching model is used. ; The most recent valid weather record before the scheduled time is selected for each flight, and a fused dataset containing both departure and arrival weather views is generated.

[0043] To capture the dynamic characteristics of the operating system, we constructed two types of derived features. The first is the delay feature, where the actual departure delay is calculated as a moving average of the historical average delay based on the flight number-aircraft type-airport combination, using the following formula: ; Second, airport hourly traffic volume, specifically the traffic volume of the previous hour. The formula reflecting the traffic flow transmission effect is as follows: ; Subsequently, a three-tiered data quality control strategy was applied to the integrated data, including outlier filtering, missing value imputation, and indicator validation. The final dataset contained 67,062 data points, achieving a meteorological matching success rate of 88.4% and a maximum missing value of <2.3%. The spatiotemporal panel data of the dataset is as follows: ; Where N is the number of airports, T is the time slice, and D is the feature dimension. The training, validation, and test sets are divided in a 7:2:1 ratio to ensure temporal continuity. The parameters of the dataset finally input into the model are shown in Tables 2 and 3, and the model training flowchart is as follows.Figure 2 As shown.

[0044] Table 2 Input Parameters of the First Layer MLP Model .

[0045] Table 3 Input Parameters for the Second Layer LightGBM Model .

[0046] 2. Flight Delay Prediction Model Based on MLP+LightGBM 2.1 Model Introduction 2.1.1 MLP Algorithm Principle A multilayer perceptron (MLP) is a typical feedforward neural network, consisting of an input layer, several hidden layers, and an output layer connected in series. Through layer-by-layer nonlinear mapping, the MLP elevates the low-dimensional original feature space to a high-dimensional representation space, thereby characterizing the complex coupling relationships in flight delays that are difficult to model explicitly. The forward propagation formula is as follows.

[0047] ; in, The model represents the first Predicted values ​​for the delay duration of a sample flight; Indicates the first The low-dimensional original feature input vector of the sample; This indicates the total number of layers in the multilayer perceptron (i.e., the output layer number). and They represent the first layer( =1,2..., The weight matrix and bias vector of (). This represents the Swish activation function, whose mathematical expression is: .

[0048] The HuberLoss loss function is used to enhance the model's robustness to abnormal delay data, as shown in the following formula: ; in, This indicates that the multilayer perceptron is in the current parameter set. The total loss value below; This represents the set of all learnable weights and bias parameters in the network. For sample index; Indicates the first The actual delay time of each flight sample; The corresponding predicted value; This is the core calculation term of the Huber loss function; Indicates the predicted residual; The preset threshold hyperparameter is used to control the critical point at which the loss function transitions from squared error to linear error. This occurs when the absolute value of the residual... Less than or equal to Squared loss is used when the time is right, and linear loss is used otherwise.

[0049] 2.1.2 LightGBM Algorithm Principle LightGBM is an efficient engineering implementation of Gradient Boosting Decision Tree (GBDT). Through four main components—Histogram algorithm, Leaf-wise growth, one-sided gradient sampling, and mutually exclusive feature binding—it significantly reduces computational and memory overhead while maintaining accuracy. Its core idea remains the same: using the negative gradient of the loss function as an approximation of the residual of the current tree, continuously adding weak learners to make the overall prediction approximate the true distribution.

[0050] 2.1.3 Principle of the MLP+LightGBM Algorithm Built in this Model This system employs a two-level cascaded framework of "micro-macro" to fully integrate the deep nonlinear capabilities of MLP with the efficient search advantages of LightGBM's tree structure: At the micro level, MLP single-flight modeling is performed, using each flight as a sample, inputting 20-dimensional time, weather, and traffic features, and outputting the delay in minutes. Subsequently, features are aggregated by airport, date, and hour to generate 2-dimensional hour-level prior features. (Number of normal departures predicted by MLP) (Number of normally executed arrivals predicted by MLP). At the macro level, LightGBM hourly panel correction is performed, using 72-dimensional features (including MLP priors, lags, moving averages, weather, etc.) of airport × date × hour as input, to fit the actual hourly flight volume again, achieving secondary error reduction.

[0051] Core Formula 1: Feature Aggregation from Micro to Macro 1. Predicted Traffic Baseline Features: Based on the single-flight predicted delay status output by the micro-level MLP model, the number of departing or arriving flights under a specific time and space is estimated, as shown in the following formula: ; in, and They represent at the airport The predicted number of departing and arriving flights with normal execution (or low delay) at the granularity of date d and hour h; This refers to the collection of all target flights within the specific airport, date, and hourly time period specified above. For flight indexing; For the MLP model to the first The predicted delay time for each flight; The preset time threshold for determining whether a flight is delayed (e.g., 30 minutes). This is an indicator function that takes the value 1 when the condition inside the parentheses is met, and 0 otherwise.

[0052] 2. Statistical characteristics of delay levels: The delay times predicted by the MLP are statistically aggregated to reflect the overall delay level for that period. The formula is as follows: ; in, At the airport The predicted average delay time at the granularity of date (d) and hour (h); Represents a set The total number of flights. Other symbols have the same meaning as above.

[0053] Core Formula Two: Macro-level Correction Stage: LightGBM Objective Function In the cascaded architecture of this invention, the macroscopic layer model employs the LightGBM algorithm. This model receives aggregated features generated in the previous stage (MLP microscopic layer) and other dynamically engineered features as input, and its training aims to minimize the prediction residuals. LightGBM is based on the gradient boosting decision tree (GBDT) principle, optimizing in the function space using Newton's method. The objective function of each iteration does not include the parameters of the MLP, but focuses on optimizing the current decision tree structure. Its second-order Taylor expansion is as follows: ; in, It is the value of the objective function in the k-th iteration. This represents the total number of samples at the macro level (airport-hour granularity). This represents the input feature vector of the i-th sample (containing prior features obtained by outputting and aggregating the MLP). This represents the prediction increment of the k-th tree for sample xj. These are the first-order gradient and second-order gradient (Hessian) of the loss function relative to the previous round of predictions. It is a regularization term that includes penalties for the number and weight of leaf nodes, used to prevent overfitting.

[0054] 2.1.4 Algorithm Advantages Traditional end-to-end deep learning models typically use a single forward propagation to complete all mappings for hourly forecasting, lacking explicit modeling of time-varying fluctuations and sudden events, leading to amplified errors during peak periods. The MLP+LightGBM architecture of this invention fully leverages the engineering advantages of LightGBM at the macroscopic level: it uses the Histogram algorithm to divide continuous features (including MLP aggregated priors, lagged flow, weather indices, etc.) into... A discrete interval, only need to be in this Selecting the optimal split point within each interval greatly improves training speed and memory utilization efficiency.

[0055] Furthermore, the macro-tree construction phase employs a leaf-wise growth strategy instead of a level-wise growth strategy, such as... Figure 3 As shown, a maximum depth limit is set to prevent overfitting while maintaining high efficiency. On the sample side, the macroscopic layer utilizes one-sided gradient sampling (GOSS): during training, samples with large gradients (i.e., samples that the current model struggles to fit) are retained, while samples with small gradients are randomly sampled, thus reducing the amount of data while ensuring the training accuracy of key samples. On the feature side, mutually exclusive feature bundling (EFB) is employed: mutually exclusive variables such as time period, weather, and traffic lag within a conflict threshold are merged to achieve dimensionality reduction without loss of information.

[0056] 2.2 Prediction Process This study proposes a cascaded hybrid prediction model based on multilayer perceptron (MLP) and lightweight gradient booster (LightGBM). Taking major airports in China as the research object, it integrates flight operation data and meteorological monitoring data to construct spatiotemporal fusion features.

[0057] First, data preprocessing is performed, including missing value imputation (0-filling), periodic feature encoding (sine and cosine transforms for hours and minutes), airport encoding (LabelEncoder encoding), and historical traffic statistics (1-hour arrival and departure flight counts). Then, a feature hierarchy construction strategy is adopted, including basic features (airport codes, hours, etc.), lag features (1-, 2-, 3-, 24-, and 48-hour traffic lag values), moving average features (3-, 6-, and 12-hour sliding window averages), and meteorological features (wind speed, visibility, and weather phenomenon markers).

[0058] The prediction task is divided into two stages: the first stage uses an MLP model to predict micro-level flight delay minutes; the second stage aggregates the MLP prediction results into hourly airport traffic features (such as predicted departure / arrival volume, average delay duration, etc.), and fuses them with the original structured features as input to the macro-level LightGBM model. In the feature fusion stage, a "prior enhancement" strategy is adopted, using the nonlinear aggregated features output by the MLP as high-order semantic information to assist LightGBM in residual correction and final traffic fitting. Model training employs an early-stop mechanism (150 rounds of patience) and validation set tuning, and finally evaluates the model on an independent test set using mean absolute error (MAE) and error rate as metrics. The algorithm prediction process is as follows: Figure 4 As shown.

[0059] 2.3 Dynamic parameter tuning based on early stopping method This model employs early stopping for dynamic parameter tuning, a highly efficient parameter optimization strategy. This method monitors the validation set performance in real time and automatically terminates training when model performance no longer improves, while simultaneously rolling back to the optimal weight state. The final results are shown in Table 4. Compared to traditional grid search, early stopping has three core advantages: first, it avoids ineffective iterations, significantly shortening training time; second, it adaptively selects the optimal training epochs, preventing overfitting; and third, it works in conjunction with the learning rate decay mechanism to achieve joint optimization of two parameters.

[0060] Table 4. Optimized parameter tuning range and results .

[0061] The dataset uses a time-series splitting method, with the last 3 days used as the test set (approximately 12%) and the remaining 88% as the training set. In the early stopping mechanism, if the validation loss fails to improve for 150 consecutive rounds, the system automatically rolls back to the optimal model state. The early stopping mechanism works as follows: Figure 5 As shown.

[0062] 3. Experimental Results and Analysis 3.1 Training Effects and Indicators This study conducts experimental validation based on real flight operation data from China Southern Airlines in May 2025. Three days were selected from the processed dataset as test data to prevent overfitting. A sliding window cross-validation strategy was employed, dividing the 67,062 processed data points into training, validation, and test sets in a 7:3:1 ratio. Features with the highest importance in the prediction process were extracted, such as... Figure 6As shown above, the key features can be divided into two categories: first, factors with a strong influence on actual operations from historical flight data, such as scheduled departure and arrival times and historical traffic flow; and second, weather factors generally considered to have a significant impact on flight takeoff and landing, such as visibility, wind speed, and cloud cover. Weather features account for approximately 24.4% of the total key features.

[0063] Table 5 shows the training results of the two-layer training of MLP and LightGBM during the model training process. All indicators such as MAE clearly meet the error requirements of the air traffic prediction model, and the model training effect is good.

[0064] Table 5 Analysis of Training Results .

[0065] 3.2 Model Results and Comparison This study proposes a hybrid model based on MLP and LightGBM, which demonstrates significant predictive performance for flight delay prediction. This model combines micro-level flight delay prediction with macro-level hourly traffic flow prediction, achieving accurate hourly traffic flow forecasting for airports. Figure 7 As shown, the visualization analysis of the model's prediction performance further verifies its prediction accuracy and reliability.

[0066] Through the Figure 7 Analysis shows that this model has high accuracy and stability in flight delay prediction. Based on the anonymized flight data and weather reports of Guangzhou Baiyun Airport in May 2025 provided by China Southern Airlines, five days were randomly selected for model validation. The hourly average traffic flow predicted by the model was calculated, compared with the actual traffic flow data, and a radar comparison chart was drawn. This demonstrates the accuracy of the model in hourly traffic flow prediction and reflects the model's good adaptability to daily flight traffic changes.

[0067] To further evaluate the overall performance of this fusion algorithm in time series prediction tasks, it is compared and analyzed with current mainstream single LightGBM, XGBoost, and single MLP models. All models use the same hyperparameter configuration during the training phase (maximum depth is uniformly set to 15 layers), and the performance metrics of the test set are shown in Table 6.

[0068] Multiple experimental results demonstrate that the fusion algorithm proposed in this study has a significant advantage in prediction accuracy, achieving the best performance in all four core evaluation metrics: RMSE, SMAPE, R², and threshold accuracy. Specifically, the fusion model significantly reduces the RMSE by 7.2% compared to the best single model, LightGBM; improves the relative performance of SMAPE by 9.3% compared to the second-best model; achieves the highest R² of 0.94; and maintains an absolute lead in threshold accuracy for key business indicators with a recognition success rate of 93.75%.

[0069] Furthermore, this fusion model demonstrates excellent computational efficiency. Compared to single LightGBM and XGBoost gradient boosting models, the fusion algorithm significantly reduces iterative computation requirements through a synergistic mechanism of automatic feature extraction from neural networks and efficient ensemble learning of tree models. Compared to computationally intensive single MLP models, the fusion strategy effectively avoids the parameter redundancy problem of fully connected neural networks, significantly reducing memory and computational overhead. These results validate that the fusion strategy achieves synergistic optimization of computational efficiency while maintaining prediction accuracy.

[0070] In summary, the MLP+LightGBM fusion model proposed in this study demonstrates significant advantages in both prediction accuracy and computational efficiency, validating its effectiveness in addressing the challenges of air traffic forecasting in high-impact weather (HIW) scenarios.

[0071] Table 6 Performance Comparison of Different Models on Test Sets .

[0072] 4. Conclusion This study aims to address the challenges of air traffic forecasting under high-impact weather (HIW) scenarios, overcoming the shortcomings of existing methods in capturing abrupt changes, cross-scale transmission, and spatiotemporal dependence. Faced with the complex nonlinear characteristics of weather-driven delays and the limitations of existing models under HIW, this study proposes and validates an intelligent forecasting framework based on deep fusion of MLP and LightGBM.

[0073] First, a preprocessing workflow integrating time-periodic trigonometric transformation and cloud impact index was designed to effectively improve data quality. Second, the model pioneered a two-level "micro-macro" architecture, combining the deep representation capabilities of MLP with the efficient feature processing of LightGBM to achieve accurate predictions ranging from single flight delays to hourly airport traffic. Simultaneously, the model's performance was optimized using LightGBM's efficient leaf-growth strategy and early-stop method.

[0074] Experimental results demonstrate that the proposed fusion model exhibits superior performance, with a 7.2% reduction in root mean square error (RMSE) compared to the optimal single model (LightGBM), and a threshold accuracy (error ≤15%) of 93.75%. This framework provides effective theoretical support and practical technical tools for improving airspace operational efficiency and preventing large-scale flight delays, and is of significant value in promoting the intelligent development of air transport.

[0075] Secondly, this embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the method described above.

[0076] Specifically, the electronic device includes: Processor: It uses an Intel Xeon Silver 4210 series CPU and is equipped with at least one NVIDIA Tesla T4 GPU for efficient parallel computing of deep neural networks (MLP) and gradient boosting decision trees (LightGBM).

[0077] Memory: Includes 128GB RAM for model runtime data caching, and 2TB NVMe SSD for storing trained model files, historical databases, and real-time data streams.

[0078] Communication interface: Equipped with a gigabit Ethernet port and a 5G communication module, it is used to receive multi-source data streams from air traffic management automation systems, airline operation control systems and civil aviation meteorological data centers in real time.

[0079] Input / output devices include high-resolution displays for visualizing airport traffic forecast heatmaps, error analysis reports, and early warning information for the next 24 hours.

[0080] Thirdly, this embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0081] Specifically, this embodiment provides a non-volatile computer-readable storage medium, such as an enterprise-grade solid-state drive (SSD) or an encrypted USB flash drive. This medium stores an executable computer program that, when installed on an electronic device as described in Embodiment 4 or any compatible server and executed by a processor, can fully implement the intelligent air traffic prediction method described in this invention.

[0082] The program stored in the storage medium contains the following core modules: The data preprocessing module includes Python-metar library wrapper code for parsing METAR messages, function library for implementing periodic trigonometric transformation, algorithm script for calculating cloud impact index, and implementation code for spatiotemporal decay matching algorithm.

[0083] Model Library Module: Contains MLP model weight files (.pth format) and LightGBM model files (.txt or .model format) that have been trained and optimized using massive amounts of historical data.

[0084] Prediction Engine Module: The main program, responsible for scheduling the data preprocessing module, loading the model library, executing the "micro-macro" two-level prediction process, and generating prediction results.

[0085] Configuration and Interface Module: Includes database connection configuration, API interface definition and log management functions, facilitating integration with external systems.

[0086] Fourthly, this embodiment provides an intelligent air traffic forecasting system for implementing the above-described method, the system comprising: The data preprocessing module is used to perform step S1 and is configured to clean, reconstruct and fuse multi-source heterogeneous aviation data. A hybrid model engine, used to perform step S2, is configured to carry and run the "micro-macro" two-level hybrid prediction model; The prediction and output module is used to perform step S3, and is configured to receive user requests, call the hybrid model engine to make predictions, and output a structured traffic prediction report.

[0087] Specifically, the system includes the following three physical modules connected via an internal network: Data acquisition and preprocessing server: Function: Dedicated to communicating with external data sources. It pulls flight data (FlightData) and meteorological observation data (METAR) from major airports across the country in real time via dedicated lines.

[0088] Processing: The built-in high-performance stream processing framework (such as Apache Flink) performs real-time parsing and cleaning of the raw packets, and performs the aforementioned time-periodic encoding, cloud index calculation and spatiotemporal fusion to form a standardized feature data stream.

[0089] Hybrid model prediction server (core computing node): Hardware: namely, the dedicated electronic device described in Example 4.

[0090] Software: Runs the complete program in the storage medium described in Example 5.

[0091] Task: Receives standardized data streams from the preprocessing server and invokes the deployed MLP-LightGBM hybrid model for real-time inference. This server automatically triggers a rolling forecast for the next 24 hours every hour and writes the forecast results to the central database.

[0092] Visualization and Decision Support Terminal: Equipment: Multiple workstations deployed at the traffic management console.

[0093] Function: The system reads the latest forecast results from the central database and provides intuitive decision-making support to air traffic controllers and traffic management engineers through self-developed visualization software, presenting data in the form of error heatmaps, traffic flow trend charts, and airport situation radar charts. When it is predicted that an airport's traffic will exceed capacity or be severely affected by severe weather in the next few hours, the system will automatically trigger audible and visual alarms.

[0094] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An aviation flow intelligent prediction method based on an MLP-LightGBM hybrid model, characterized in that, The method comprises the following steps: Step S1: obtaining and preprocessing multi-source heterogeneous aviation data, the aviation data comprising at least flight operation data and airport meteorological observation data; Step S2: based on the preprocessed data, constructing and training a "micro-macro" two-level hybrid prediction model; wherein, the micro prediction layer adopts a multi-layer perception MLP model, taking a single flight as a sample, inputting features containing time, weather and historical flow, and outputting the predicted delay time of the flight; the macro prediction layer adopts a light gradient boosting machine LightGBM model, the input features of which comprising: hour-level prior flow features obtained by spatio-temporal aggregation of all flight prediction results output by the micro prediction layer, and original structured features from step S1; the output of the macro prediction layer being an airport hour-level flow prediction result; Step S3: using the trained hybrid prediction model to predict the aviation flow of a target airport in a future period.

2. The method of claim 1, wherein, The preprocessing of aviation data in step S1 specifically comprises: discretizing flight plan timestamps and using a triangular sine-cosine transformation to periodically encode hours and minutes, the transformation formula comprising: ； wherein h is the hour and m is the minute.

3. The method according to claim 1 or 2, characterized in that, The preprocessing of aviation data in step S1 further comprises constructing an aviation cloud layer influence index: The cloud layer information in the meteorological observation data is analyzed, and a cloud layer influence index is calculated according to the cloud coverage and the cloud bottom height through the following formula: ; wherein, is the cloud coverage, is the cloud base height, k is a preset adjustment coefficient.

4. The method of claim 1, wherein, The preprocessing of aviation data in step S1 further comprises a spatio-temporal data fusion step: A spatio-temporal matching algorithm based on a decay function is used to match the optimal weather observation record for each flight around its scheduled time, the formula of which is: ; wherein is the planned departure time of the flight, is the time of the weather observation, and a is a decay coefficient.

5. The method of claim 1, wherein, The input features of the macro prediction layer LightGBM model further comprise at least one of the following dynamic engineering features: lag features: flow data of 1 hour, 2 hours, 3 hours, 24 hours and 48 hours in the past; moving average features: flow mean values calculated based on 3-hour, 6-hour and 12-hour sliding windows.

6. The method of claim 1, wherein, In the model training process of step S2, the early stopping method is used for dynamic parameter tuning, specifically comprising: In the training process, the loss function on the validation set is monitored in real time, and when the loss function no longer decreases in continuous N iterations, the training is automatically terminated and the model parameters are rolled back to the best performance, wherein N is a predetermined patience value.

7. The method of claim 1, wherein, The macro prediction layer LightGBM model is trained using at least one of the following optimization strategies: leaf growth strategy; one-sided gradient sampling method; exclusive feature bundling method.

8. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the method of any one of claims 1 to 7.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the method of any one of claims 1 to 7.

10. An intelligent air traffic flow prediction system, characterized by, The system for implementing the method of any one of claims 1 to 7 comprises: a data preprocessing module for performing step S1, configured to clean, reconstruct and fuse multi-source heterogeneous aviation data; a hybrid model engine for performing step S2, configured to carry and run the "micro-macro" two-level hybrid prediction model; a prediction and output module for performing step S3, configured to receive user requests, call the hybrid model engine for prediction, and output a structured flow prediction report.

Images