Airport flight quantity and throughput prediction method and system

Abstract

The invention discloses an airport flight quantity and throughput prediction method and system, and relates to the technical field of intelligent airport management. The method comprises the following steps: firstly, determining a future target time period to be predicted, and obtaining related multi-source data according to the future target time period; preprocessing the data and extracting time sequence and non-time sequence features; processing time sequence features by using the trained long-short-term memory neural network model, processing non-time sequence features by using the extreme gradient boosting tree model, and respectively obtaining prediction results; and integrating the outputs of the two types of models through a weighted fusion strategy to obtain an accurate prediction value of the flight sortie and the single-machine passenger capacity, and further calculating to obtain a passenger throughput prediction result. According to the method and the system provided by the invention, through the architecture of multi-source data fusion and hybrid intelligent model cooperation, the prediction precision and reliability are effectively improved, and scientific decision support is provided for airport operation management.

Description

Technical Field

[0001] This invention relates to the field of intelligent airport management technology, and in particular to a method and system for predicting airport flight volume and throughput. Background Technology

[0002] As the core hub of the air transport network, the operational efficiency of airports directly impacts passenger experience and airline profitability. Accurately predicting future flight volume and passenger throughput is crucial for airports to optimize resource allocation (such as ground staff, boarding gates, and security checkpoints), for airlines to arrange capacity, and for improving passenger service quality.

[0003] Currently, existing forecasting methods suffer from the following limitations: First, they often rely on single types of historical time-series data, failing to fully consider the influence of multiple external factors such as flight schedules, ticket prices, search popularity, holidays, and major events, leading to insufficient forecast accuracy, especially with significant deviations in the face of unforeseen circumstances. Second, traditional statistical models or single machine learning models struggle to simultaneously capture the complex temporal dependencies and nonlinear feature interactions within the data. Third, existing methods typically lack a systematic and interpretable framework for data processing and model fusion, resulting in a need to improve the stability and reliability of the forecast results.

[0004] Therefore, there is an urgent need for a comprehensive solution that can integrate multi-source data, combine the advantages of models, and provide high-precision predictions. Summary of the Invention

[0005] To address the technical problems existing in the background art, this invention proposes a method and system for predicting airport flight volume and throughput.

[0006] This invention proposes a method for predicting airport flight volume and throughput, comprising:

[0007] Obtain the target airport's future target time period to be predicted, and obtain multi-source data related to the target airport's future target time period;

[0008] The multi-source data is preprocessed to obtain a standardized feature dataset containing both temporal and non-temporal features;

[0009] The temporal features are input into a trained long short-term memory neural network model to obtain a first prediction result, while the non-temporal features are input into a trained extreme gradient boosting tree model to obtain a second prediction result.

[0010] The first prediction result and the second prediction result are weighted and fused according to the preset fusion weights to obtain the flight number prediction value and the passenger capacity prediction value per aircraft.

[0011] Based on the predicted flight frequency and the predicted passenger capacity per aircraft, the predicted passenger throughput of the target airport in the future target time period is calculated.

[0012] Preferably, the first prediction result includes a first-time prediction value of flight frequency and a first-time prediction value of passenger capacity per aircraft, and the second prediction result includes a second-time prediction value of flight frequency and a second-time prediction value of passenger capacity per aircraft; the step of weighting and fusing the first prediction result and the second prediction result according to a preset fusion weight to obtain the predicted value of flight frequency and the predicted value of passenger capacity per aircraft specifically involves:

[0013] The first time prediction value of the flight number and the second time prediction value of the flight number are fused using the first fusion weight and the second fusion weight to obtain the flight number prediction value, and the first time prediction value of the single-aircraft passenger volume and the second time prediction value of the single-aircraft passenger volume are fused to obtain the single-aircraft passenger volume prediction value.

[0014] Preferably, the long short-term memory neural network model and the extreme gradient boosting tree model are trained through the following process:

[0015] The historical feature dataset is divided into training set, validation set and test set according to a preset ratio;

[0016] The training set is used to iteratively train the long short-term memory neural network model and the extreme gradient boosting tree model, and the model loss is calculated using the validation set after each iteration.

[0017] Training of the corresponding model is stopped when the loss of the validation set fails to decrease for a preset number of consecutive iterations.

[0018] Preferably, the historical feature dataset is obtained and constructed in the following manner:

[0019] Obtain actual airport operation data within a historical time period;

[0020] Obtain historical data for the same period corresponding to the historical time period, aviation industry weekday effect coefficient, flight schedule data, online travel agency search index data, airfare data, number of holiday days, and quantitative values ​​of the influence of popular events;

[0021] The acquired data undergoes preprocessing, including data cleaning, data standardization, and data feature extraction, to generate a historical feature dataset.

[0022] Preferably, the training process of the long short-term memory neural network model includes:

[0023] Configure the input layer structure of the long short-term memory neural network model so that the number of its neurons matches the dimension of the temporal features;

[0024] Construct a network architecture that includes at least one hidden layer, wherein the hidden layer is composed of long short-term memory units;

[0025] The time-series feature data of the training set are input into the model for forward and backward propagation to optimize the model parameters;

[0026] The training process is monitored using the validation set. When the validation set loss fails to decrease for multiple consecutive iterations, an early stopping mechanism is triggered to terminate the training.

[0027] Preferably, the training process of the extreme gradient boosting tree model includes:

[0028] Configure the input dimension of the extreme gradient boosting tree model to match the dimension of the non-temporal features;

[0029] An ensemble learning model based on gradient boosting decision trees is constructed, which generates multiple regression trees iteratively, with the complexity of each tree limited by a preset depth.

[0030] The non-temporal feature data of the training set is input into the ensemble learning model, and the forward stepwise algorithm is used for training. In each iteration, the negative gradient direction of the current model is fitted.

[0031] The model complexity is controlled by introducing a regularization term, and the model's generalization ability is improved by adopting sample subsampling and feature sampling strategies.

[0032] The newly generated tree model is superimposed on the current model at a predetermined learning rate, and the objective function is continuously optimized until the predetermined number of trees is reached or the performance on the validation set no longer improves.

[0033] Finally, a well-trained extreme gradient boosting tree model is obtained that can output prediction results based on input non-temporal features.

[0034] Preferably, the multi-source data includes recent airport operation data, historical data for the same period of the airport, aviation industry weekday effect coefficient, future flight schedule data for the target airport, online travel agency search index data related to the target airport, airfare data for the target airport, number of holidays, and quantitative values ​​of the influence of popular events; the time-series features include at least preprocessed recent airport operation data and historical data for the same period of the airport, and the non-time-series features include at least preprocessed aviation industry weekday effect coefficient, number of holidays, and quantitative values ​​of the influence of popular events.

[0035] Preferably, the preprocessing includes data cleaning to handle missing and outlier values, data standardization to eliminate dimensional differences, and data feature extraction. During the data cleaning process, when the proportion of missing values ​​is less than 5%, the mean of the same type of data is used for filling; when the proportion of missing values ​​is between 5% and 15%, linear interpolation is used for filling; when the proportion of missing values ​​is greater than 15%, a data re-collection process is triggered.

[0036] Preferably, the method for determining the first fusion weight and the second fusion weight is as follows: the prediction accuracy of the long short-term memory neural network model on the validation set for flight numbers is divided by the sum of the prediction accuracy of the long short-term memory neural network model and the extreme gradient boosting tree model on the validation set for flight numbers, to obtain the value of the first fusion weight, and the value of the second fusion weight is equal to one minus the value of the first fusion weight.

[0037] This invention proposes an airport flight volume and throughput prediction system, comprising:

[0038] The data acquisition module is used to acquire the target airport's future target time period to be predicted, and to acquire multi-source data related to the future target time period to be predicted;

[0039] The preprocessing module is used to process the multi-source data to obtain a standardized feature dataset containing time-series and non-time-series features;

[0040] The prediction calculation module includes a trained long short-term memory neural network model and a trained extreme gradient boosting tree model, which is used to receive the temporal features and non-temporal features and output the first prediction result and the second prediction result accordingly.

[0041] The fusion calculation module is used to perform weighted fusion of the first prediction result and the second prediction result according to preset weights, and output the flight number prediction value and the single aircraft passenger capacity prediction value.

[0042] The throughput calculation module is used to calculate the passenger throughput forecast value based on the flight number forecast value and the single aircraft passenger capacity forecast value.

[0043] The proposed airport flight volume and throughput prediction method and system overcomes the limitations of traditional prediction methods, such as single data sources, neglect of external factors, and poor model adaptability, by integrating multi-source data and constructing a hybrid machine learning model. This significantly improves prediction accuracy and robustness. The method can provide accurate and forward-looking decision support for airport human and equipment resource allocation and business operation planning. It also helps to formulate passenger flow management plans in advance, optimize passenger travel experience, and strengthen operational safety assurance capabilities. Ultimately, it achieves scientific and reliable comprehensive prediction of airport flight volume and throughput, providing an efficient data-driven solution for intelligent airport operation and management. Attached Figure Description

[0044] Figure 1 This is a schematic diagram illustrating the workflow of an airport flight volume and throughput prediction method proposed in this invention.

[0045] Figure 2 This is a schematic diagram of the system architecture of an airport flight volume and throughput prediction system proposed in this invention. Detailed Implementation

[0046] Reference Figure 1 and Figure 2 The present invention proposes a method for predicting airport flight volume and throughput, comprising the following steps:

[0047] S1. Obtain the target airport's future target time period to be predicted, and obtain multi-source data related to the target airport's future target time period to be predicted.

[0048] In this embodiment, the multi-source data includes recent airport operation data, historical data of the same period of the airport, aviation industry weekday effect coefficient, future flight schedule data of the target airport, online travel agency search index data related to the target airport, airfare data of the target airport, number of holidays and quantitative value of influence of popular events; the time-series features include at least the preprocessed recent airport operation data and historical data of the same period of the airport, and the non-time-series features include at least the preprocessed aviation industry weekday effect coefficient, number of holidays and quantitative value of influence of popular events.

[0049] S2. Preprocess the multi-source data to obtain a standardized feature dataset containing both temporal and non-temporal features.

[0050] In this embodiment, preprocessing includes data cleaning to handle missing and outlier values, data standardization to eliminate dimensional differences, and data feature extraction. During the data cleaning process, when the proportion of missing values ​​is less than 5%, the mean of the same type of data is used for filling; when the proportion of missing values ​​is between 5% and 15%, linear interpolation is used for filling; when the proportion of missing values ​​is greater than 15%, the data re-collection process is triggered.

[0051] S3. Input the temporal features into the trained long short-term memory neural network model to obtain the first prediction result, and simultaneously input the non-temporal features into the trained extreme gradient boosting tree model to obtain the second prediction result.

[0052] In this embodiment, the Long Short-Term Memory Neural Network model and the Extreme Gradient Boosting Tree model are trained through the following process:

[0053] The historical feature dataset is divided into training set, validation set and test set according to a preset ratio;

[0054] The long short-term memory neural network model and the extreme gradient boosting tree model are iteratively trained using the training set, and the model loss is calculated using the validation set after each iteration.

[0055] Training of the corresponding model is stopped when the validation set loss fails to decrease for a preset number of consecutive iterations.

[0056] In this embodiment, the historical feature dataset is obtained and constructed in the following manner:

[0057] Obtain actual airport operation data within a historical time period;

[0058] Obtain historical data for the same period corresponding to the historical time period, aviation industry weekday effect coefficient, flight schedule data, online travel agency search index data, airfare data, number of holiday days, and quantitative values ​​of the influence of popular events;

[0059] The acquired data undergoes preprocessing, including data cleaning, data standardization, and data feature extraction, to generate a historical feature dataset.

[0060] Specifically, the ratio of the training set, validation set, and test set is 7:1.5:1.5.

[0061] In this embodiment, the training process of the Long Short-Term Memory (LSTM) neural network model includes:

[0062] Configure the input layer structure of the long short-term memory neural network model so that the number of neurons matches the dimension of the temporal features;

[0063] Construct a network architecture that includes at least one hidden layer, which is composed of long short-term memory units;

[0064] The time-series feature data of the training set is input into the model for forward and backward propagation in order to optimize the model parameters;

[0065] The training process is monitored using a validation set. If the validation set loss does not decrease for several consecutive iterations, an early stopping mechanism is triggered to terminate the training.

[0066] In this embodiment, the training process of the extreme gradient boosting tree model includes:

[0067] Configure the input dimension of the extreme gradient boosting tree model to match the dimension of the non-temporal features;

[0068] An ensemble learning model based on gradient boosting decision trees is constructed, which generates multiple regression trees iteratively, with the complexity of each tree limited by a preset depth.

[0069] The non-temporal feature data of the training set is input into the ensemble learning model, and the forward stepwise algorithm is used for training. In each iteration, the negative gradient direction of the current model is fitted.

[0070] The model complexity is controlled by introducing a regularization term, and the model's generalization ability is improved by adopting sample subsampling and feature sampling strategies.

[0071] The newly generated tree model is superimposed on the current model at a predetermined learning rate, and the objective function is continuously optimized until the predetermined number of trees is reached or the performance on the validation set no longer improves.

[0072] Finally, a well-trained extreme gradient boosting tree model is obtained that can output prediction results based on input non-temporal features.

[0073] Specifically, the Long Short-Term Memory (LSTM) neural network model is specifically the LSTM neural network model, and the Extreme Gradient Boosting Tree (XGBoost) neural network model is specifically the XGBoost neural network model. The LSTM neural network model is used to process data with time-series characteristics, such as recent airport operation data, historical data from the same period, and OTA search index data, capturing long-term time-series dependencies in the data. The XGBoost neural network model is used to process non-time-series data, such as Dow effect coefficients, quantified values ​​of holidays and popular events, and ticket price characteristics, uncovering non-linear relationships in the data.

[0074] It should be noted that during model training, the preprocessed feature data is divided into training, validation, and test sets. First, the LSTM and XGBoost neural network models are trained separately using the training set. The training parameters for the LSTM model are set as follows: 128 hidden layer neurons, a learning rate of 0.001, 100 iterations, and a batch size of 32 (batch size: the number of samples used in one model parameter update). The training parameters for the XGBoost neural network model are set as follows: 100 trees, a learning rate of 0.01, a maximum tree depth of 6, a subsample of 0.8 (subsample: the proportion of samples randomly drawn from all training samples), and a colsample_bytree of 0.8 (colsample_bytree: the proportion of features randomly drawn from all features). Then, the training process of both sub-models is monitored using the validation set, employing an early stopping strategy (training stops if the loss function of the validation set does not decrease within 10 consecutive iterations) to prevent overfitting. Finally, the outputs of the trained LSTM neural network model and XGBoost neural network model are fused according to weights (the weights are determined by the prediction accuracy of the validation set; if the prediction accuracy of the LSTM neural network model on the validation set is 85% and the prediction accuracy of the XGBoost neural network model is 82%, then the weight of the LSTM neural network model is 0.51 and the weight of the XGBoost neural network model is 0.49).

[0075] Furthermore, the preprocessed feature data corresponding to the prediction time period is input into the trained fusion prediction model. The model first processes the temporal feature data using an LSTM neural network model, outputting preliminary temporal prediction results; simultaneously, it processes the non-temporal feature data using an XGBoost neural network model, outputting preliminary non-temporal prediction results. Then, the two preliminary prediction results are fused and calculated according to preset weights to obtain the final predicted flight frequency, passenger capacity per aircraft, and passenger throughput data for the airport in the future. Specifically, the passenger throughput prediction data = flight frequency prediction data × passenger capacity per aircraft prediction data.

[0076] S4. The first prediction result and the second prediction result are weighted and fused according to the preset fusion weights to obtain the predicted value of flight number and the predicted value of passenger volume per aircraft.

[0077] In this embodiment, the first prediction result includes the first time prediction value of flight number and the first time prediction value of passenger capacity per aircraft; the second prediction result includes the second time prediction value of flight number and the second time prediction value of passenger capacity per aircraft; step S4 specifically includes:

[0078] The first and second time-based predicted values ​​of flight number are fused using the first and second fusion weights to obtain the predicted value of flight number. The first and second time-based predicted values ​​of passenger volume per aircraft are fused to obtain the predicted value of passenger volume per aircraft.

[0079] In this embodiment, the preset fusion weights include a first fusion weight and a second fusion weight. The method for determining the first fusion weight and the second fusion weight is as follows: the prediction accuracy of the long short-term memory neural network model on the validation set for flight numbers is divided by the sum of the prediction accuracy of the long short-term memory neural network model and the extreme gradient boosting tree model on the validation set for flight numbers, to obtain the value of the first fusion weight. The value of the second fusion weight is equal to one minus the value of the first fusion weight.

[0080] S5. Based on the predicted flight frequency and passenger capacity per aircraft, calculate the predicted passenger throughput of the target airport in the future target time period.

[0081] In this embodiment, three indicators—mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE)—are used to evaluate the prediction results. Their calculation formulas are as follows:

[0082]

[0083] Where n is the number of samples in the prediction data, y i These are actual data values. For predicted data values.

[0084] If the MAPE in the evaluation result is less than 10%, the prediction result is considered qualified and can be output directly. If the MAPE is between 10% and 20%, the weights of the prediction model are fine-tuned based on the prediction deviation of the same period in history and the special characteristics of the current data (such as whether there are popular events that have not been fully considered), and the prediction result is recalculated. If the MAPE is greater than 20%, the data collection and preprocessing process is re-examined, data anomalies are eliminated, the model is retrained and prediction is performed until the prediction result meets the requirement of MAPE less than 10%.

[0085] Example 1:

[0086] Taking the prediction of flight frequency, passenger capacity per aircraft, and passenger throughput data for a major hub airport from October 1st to October 7th, 2024 (National Day Golden Week) as an example, the specific implementation steps are as follows:

[0087] 1. Data Collection (July 1, 2024 - September 20, 2024)

[0088] Collect recent airport operation data: Collect daily data on actual flight counts, actual passenger throughput, actual passenger capacity per aircraft, and flight punctuality rate for the airport from June 1, 2024 to August 31, 2024, totaling 92 records.

[0089] Collect historical data from the same period of the airport: obtain flight numbers, passenger throughput, and passenger capacity per aircraft from October 1st to October 7th of each year from 2019 to 2023, as well as holiday arrangements and records of major events during the same period, totaling 15 records.

[0090] Data collection on the Dow effect in the industry: Daily flight counts and passenger throughput data for all airports in the industry from 2019 to 2023 were obtained, totaling 1825 records. The average growth rate of flight counts and passenger throughput on different days within a week were calculated to obtain the Dow effect coefficient. For example, the Dow effect coefficient for flight counts on Monday is 0.95 (a 5% decrease compared to the weekly average), while the Dow effect coefficient for flight counts on Sunday is 1.12 (a 12% increase compared to the weekly average).

[0091] Collect data on future available flight schedules: Obtain information on available flight schedules for the airport from October 1st to October 7th, 2024, including daily planned flight number, planned flight route, planned flight time, and planned number of seats, for a total of 7 records.

[0092] OTA search index data: Obtain search volume data for routes related to this airport from September 1, 2024 to September 20, 2024, including the number of searches and the number of people searching for each route each day. Calculate the daily OTA search index; for example, the search index for September 1 is 105 (a 5% increase compared to the average number of searches over the past 30 days), and the search index for September 10 is 130 (a 30% increase compared to the average number of searches over the past 30 days), totaling 20 records.

[0093] Collect ticket price data: Obtain the available ticket prices for all routes at this airport from October 1st to October 7th, 2024, including prices and discount rates for economy class, business class, and first class. Calculate the average ticket price and price fluctuation range for each route daily. For example, the average ticket price for the Beijing-Shanghai route on October 1st is 1200 yuan, with a price fluctuation range of 15% (highest price 1400 yuan, lowest price 1100 yuan), totaling 7 records.

[0094] Data collection on holidays and trending events: October 1-7, 2024, was designated as the National Day Golden Week, a total of 7 days. Information on trending events in and around the airport during October 1-7, 2024, was collected from news media platforms, Weibo, WeChat official accounts, etc. It was found that a large-scale international tourism expo will be held near the airport from October 3-5, and the influence level of this trending event was determined to be significant.

[0095] 2. Data Preprocessing (September 21, 2024 - September 23, 2024)

[0096] Data cleaning: Upon inspection, the single-aircraft passenger volume data for August 15, 2024, was missing in the recent airport operation data (missing percentage of approximately 1.1%, less than 5%). This missing data was filled in using the average single-aircraft passenger volume for August 14, 16, and 17 of that month (152, 148, and 150 passengers respectively, with an average of 150 passengers). There were no missing values ​​in the historical data for the same period. The data for September 8, 2024, in the OTA search index data deviated from the mean by 4.2 times the standard deviation (exceeding the 3σ principle) and was identified as an outlier. It was replaced with the median (118) of the OTA search index series.

[0097] Data Standardization: Z-score standardization was performed on all cleaned data using Python's NumPy library. Taking daily flight counts from recent airport operation data as an example, the original data had a mean μ = 780 flights and a standard deviation σ = 65 flights. If the original flight count X for a certain date was 850 flights, then the standardized data Z = (850-780) / 65 ≈ 1.08. This process uniformly converts data on flight counts, passenger throughput, OTA search indexes, ticket prices, etc., of different dimensions into standardized data with a mean of 0 and a standard deviation of 1.

[0098] Data Feature Extraction: Based on the standardized data, various features were extracted using Python's Pandas library. For example, in the recent airport operation trend features, the weekly growth rate of flight frequency over the past month (August 1st - August 31st, 2024) was calculated as: (This week's average daily flights - Last week's average daily flights) / Last week's average daily flights × 100%. This yielded a growth rate of 3.2% for the second week of August relative to the first week, and 2.8% for the third week relative to the second week. In the holiday and popular event features, there were 7 holiday days from October 1st to October 7th, 2024; the popular event quantification value was 3 for October 3rd to October 5th; and there were no popular events on other dates, with a quantification value of 0.

[0099] 3. Prediction Model Construction and Training (September 24, 2024 - September 26, 2024)

[0100] Data partitioning: The extracted feature data (a total of 120 data points, including recent running features, historical period features, Dow effect features, and other types of feature data) were divided into training set (84 data points), validation set (18 data points) and test set (18 data points) in a ratio of 7:1.5:1.5.

[0101] LSTM Neural Network Model Training: An LSTM neural network model was built using TensorFlow 2.8. The input layer dimension was set to the dimension of the feature data (15 dimensions in total, covering 15 different feature classes), the hidden layer had 128 neurons, and the output layer had 1 neuron (corresponding to a single prediction metric, such as flight count). The learning rate was set to 0.001, the number of iterations to 100, the batch size to 32, the mean squared error (MSE) loss function, and the Adam optimizer. The training set data was input into the model for training. After each iteration, the model's loss value was calculated using the validation set data. When the validation set loss value did not decrease for 10 consecutive iterations (e.g., after the 45th iteration, the validation set loss value stabilized at 0.08 without further decrease), an early stopping strategy was triggered, and training was stopped. At this point, the LSTM neural network model achieved a flight count prediction accuracy of 86.2% on the validation set.

[0102] XGBoost Neural Network Model Training: An XGBoost neural network model was built using XGBoost 1.6.1, with 100 trees, a learning rate of 0.01, a maximum tree depth of 6, a subsample rate of 0.8, and a colsample_bytree rate of 0.8. The objective function was a regression task (predicting continuous values ​​such as flight count and passenger capacity per aircraft). The training set data was input into the model for training, and the training process was monitored using a validation set. Training was stopped when the validation set loss value did not decrease for 10 consecutive iterations. Ultimately, the XGBoost neural network model achieved a flight count prediction accuracy of 83.5% on the validation set.

[0103] Model fusion: The first fusion weight is calculated based on the prediction accuracy of the XGBoost neural network model and the LSTM neural network model on the validation set. The formula for the first fusion weight is: First fusion weight = LSTM neural network model accuracy / sum of the prediction accuracies of the XGBoost neural network model and the LSTM neural network model on the validation set. In this case, the first fusion weight = 86.2% / (86.2% + 83.5%) ≈ 0.51, and the second fusion weight = 1 - 0.51 = 0.49.

[0104] 4. Implementation Schedule (September 27, 2024)

[0105] The preprocessed feature data (7 data points in total, 1 data point per day) for October 1st to October 7th, 2024, were input into the XGBoost neural network model and the LSTM neural network model, respectively. Taking October 1st as an example, the input features included: weekly growth rate of flight frequency in recent operating trends of 3.0%, historical growth rate of flight frequency in the same period of 5.1%, Dow effect coefficient of 1.08 (October 1st is a Tuesday, corresponding to the industry Dow effect coefficient), planned flights of 850, OTA search index of 135, average ticket price of 1200 yuan, number of holiday days of 1 (the day was a holiday), and quantitative value of hot events of 0, etc.

[0106] After processing the input time-series features (such as recent flight growth rate, historical year-on-year growth rate, OTA search index changes, etc.), the LSTM neural network model outputs a preliminary prediction of 842 flights and 153 passengers per aircraft for October 1st. After processing the input non-time-series features (such as Dow effect coefficient, number of holiday days, average ticket price, etc.), the XGBoost neural network model outputs a preliminary prediction of 838 flights and 151 passengers per aircraft for October 1st.

[0107] The preliminary predictions were then merged using weights of 0.51 and 0.49:

[0108] Forecast for October 1st: Flight frequency = 842 × 0.51 + 838 × 0.49 ≈ 840 flights; Passenger capacity per aircraft = 153 × 0.51 + 151 × 0.49 ≈ 152 passengers; Passenger throughput = 840 × 152 = 127,680 passengers. Similarly, calculate the forecast data for October 2nd to October 7th. For example, for October 3rd (a peak event day), the predicted flight frequency is 865 flights, the predicted passenger capacity per aircraft is 158 passengers, and the predicted passenger throughput is 136,670 passengers.

[0109] 5. Evaluation and Revision of Forecast Results (September 28, 2024)

[0110] Results Evaluation: The model's predictive performance was evaluated using test set data (18 historical feature data points). The test set feature data was input into the fusion model to obtain prediction results. The MAE, RMSE, and MAPE were then calculated using the corresponding actual data from the test set. The calculated results were as follows: MAE for flight frequency prediction = 12.5 flights, RMSE = 15.8 flights, MAPE = 8.2%; MAE for passenger capacity prediction = 3.2 passengers, RMSE = 4.5 passengers, MAPE = 7.5%; and MAE for passenger throughput prediction = 2850 passengers, RMSE = 3520 passengers, MAPE = 9.1%. All three MAPE values ​​were less than 10%, indicating that the model's predictive performance was satisfactory.

[0111] Reference Figure 1 and Figure 2 The present invention proposes an airport flight volume and throughput prediction system, comprising:

[0112] The data acquisition module is used to acquire the target airport's future target time period to be predicted, and to acquire multi-source data related to the future target time period to be predicted;

[0113] The preprocessing module is used to process multi-source data to obtain a standardized feature dataset containing both time-series and non-time-series features;

[0114] The prediction calculation module includes a trained long short-term memory neural network model and a trained extreme gradient boosting tree model, which are used to receive temporal features and non-temporal features and output the first prediction result and the second prediction result accordingly.

[0115] The fusion calculation module is used to weight and fuse the first prediction result and the second prediction result according to the preset weights, and output the flight number prediction value and the single aircraft passenger capacity prediction value.

[0116] The throughput calculation module is used to calculate the passenger throughput forecast based on the flight number forecast and the single aircraft passenger capacity forecast.

[0117] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An airport flight volume and throughput prediction method, characterized by, The method comprises the following steps: acquire a target airport to be predicted in a future target period, and acquire multi-source data related to the future target period to be predicted; preprocess the multi-source data to obtain a standardized feature dataset containing time series features and non-time series features; input the time series features into a trained long short-term memory neural network model to obtain a first prediction result, and input the non-time series features into a trained extreme gradient boosting tree model to obtain a second prediction result; weight and fuse the first prediction result and the second prediction result according to a preset fusion weight to obtain a flight frequency prediction value and a single-aircraft passenger capacity prediction value; calculate a passenger throughput prediction value of the target airport in the future target period according to the flight frequency prediction value and the single-aircraft passenger capacity prediction value.

2. The method of claim 1, wherein, The first prediction result includes a flight frequency first-time prediction value and a single-aircraft passenger capacity first-time prediction value, and the second prediction result includes a flight frequency second-time prediction value and a single-aircraft passenger capacity second-time prediction value; the weight and fusion of the first prediction result and the second prediction result according to the preset fusion weight to obtain the flight frequency prediction value and the single-aircraft passenger capacity prediction value are specifically as follows: fuse the flight frequency first-time prediction value and the flight frequency second-time prediction value using a first fusion weight and a second fusion weight to obtain the flight frequency prediction value, and fuse the single-aircraft passenger capacity first-time prediction value and the single-aircraft passenger capacity second-time prediction value to obtain the single-aircraft passenger capacity prediction value.

3. The method of claim 1 or 2, wherein, The long short-term memory neural network model and the extreme gradient boosting tree model are trained through the following process: divide a historical feature dataset into a training set, a validation set and a test set according to a preset ratio; use the training set to iteratively train the long short-term memory neural network model and the extreme gradient boosting tree model, and calculate the model loss using the validation set after each iteration; stop training the corresponding model when the validation set loss does not decrease for a preset number of consecutive iterations.

4. The method of claim 3, wherein, The historical feature dataset is obtained and constructed in the following way: acquire airport operation actual data in a historical period; acquire historical contemporaneous data corresponding to the historical period, aviation industry day-of-week effect coefficients, flight plan data, online travel agency search index data, route ticket price data, holiday days and popular event influence power values; perform data preprocessing such as data cleaning, data standardization and data feature extraction on the acquired data to generate a historical feature dataset.

5. The method of claim 3, wherein, The training process of the long short-term memory neural network model comprises: configure the input layer structure of the long short-term memory neural network model so that the number of neurons matches the dimension of the time series features; build a network architecture containing at least one hidden layer composed of long short-term memory units; input the time series feature data of the training set into the model for forward propagation and back propagation to optimize the model parameters; use the validation set to monitor the training process, and trigger an early stop mechanism when the validation set loss does not decrease for a plurality of consecutive iterations to terminate the training.

6. The method of claim 3, wherein, The training process of the extreme gradient boosting tree model comprises: The input dimension of the extreme gradient boosting tree model is matched with the dimension of the non-time sequence feature; An ensemble learning model based on gradient boosting decision tree is constructed, and multiple regression trees are generated by iteration, wherein the complexity of each tree is limited by a preset depth; The non-time sequence feature data of the training set is input into the ensemble learning model, and a forward step-by-step algorithm is used for training, and in each iteration, the negative gradient direction of the current model is fitted; The model complexity is controlled by introducing a regularization term, and the model generalization ability is improved by using sample subsampling and feature sampling strategies; The newly generated tree model is superimposed into the current model with a predetermined learning rate, and the target function is continuously optimized until the predetermined number of trees is reached or the performance of the validation set no longer improves; Finally, a trained extreme gradient boosting tree model is obtained, which can output prediction results according to input non-time sequence features.

7. The method of claim 1, wherein, The multi-source data includes airport recent operation data, airport historical same period data, aviation industry week effect coefficient, future on-sale flight plan data of the target airport, online travel agency search index data related to the target airport, route ticket price data of the target airport, holiday day number and popular event influence power value; the time sequence features at least include preprocessed airport recent operation data and airport historical same period data, and the non-time sequence features at least include preprocessed aviation industry week effect coefficient, holiday day number and popular event influence power value.

8. The method of claim 1, wherein, The preprocessing includes data cleaning to handle missing values and outliers, data standardization to eliminate dimension differences, and data feature extraction; in the data cleaning process, when the proportion of missing values is less than 5%, the mean value of the same type of data is used for filling; when the proportion of missing values is between 5% and 15%, a linear interpolation method is used for filling; when the proportion of missing values is greater than 15%, a data reacquisition process is triggered.

9. The method of claim 2, wherein, The determination method of the first fusion weight and the second fusion weight is: the prediction accuracy of the long short-term memory neural network model on the validation set is divided by the sum of the prediction accuracy of the long short-term memory neural network model and the extreme gradient boosting tree model on the validation set, to obtain the value of the first fusion weight, and the value of the second fusion weight is equal to one minus the value of the first fusion weight.

10. An airport flight volume and throughput prediction system, characterized by, It comprises: a data acquisition module for acquiring a target airport to be predicted in a future target period and acquiring multi-source data related to the to-be-predicted future target period; a preprocessing module for processing the multi-source data to obtain a standardized feature data set containing time sequence features and non-time sequence features; a prediction calculation module including a trained long short-term memory neural network model and the trained extreme gradient boosting tree model, for receiving the time sequence features and non-time sequence features and outputting first prediction results and second prediction results accordingly; a fusion calculation module for weighting and fusing the first prediction results and the second prediction results according to a preset weight, and outputting flight cycle prediction values and single-machine passenger capacity prediction values; a throughput calculation module for calculating passenger throughput prediction values based on the flight cycle prediction values and the single-machine passenger capacity prediction values.

Images