A machine learning-based method for predicting strategic flight schedule delay distribution

Abstract

The present invention discloses a method for predicting the delay distribution of a strategic flight schedule based on machine learning. The method comprises fitting the delay distribution of each flight using a normal distribution based on historical flight operation data, and splicing the obtained delay distribution into the strategic flight schedule to obtain a data set. Features used in the strategic flight schedule for predicting flight delay distribution are selected, and different types of features are encoded. A machine learning algorithm is used to learn and predict flight delay distribution information in the strategic flight schedule. For the prediction results, evaluation indicators are proposed to evaluate the prediction results of different machine learning algorithms. Finally, a machine learning algorithm with better performance is selected to predict the delay distribution of the strategic flight schedule. The method provides flight schedule coordinators with potential delay distribution information of the strategic schedule. When bottlenecks in airport on-time performance are discovered, actions are taken to reduce flight delays, thereby supporting strategic flight schedule performance evaluation and strategic flight schedule optimization.

Description

Technical Field

[0001] The present invention relates to a machine learning method, and in particular to a method for predicting strategic flight schedule delay distribution based on machine learning. Background Art

[0002] In recent years, air transport demand has continued to grow, and the quality of air transport services has gradually improved. However, due to the limited growth rate of airport and airspace capacity, the air transport system has gradually exposed capacity saturation and low operational efficiency, and flight delays have become increasingly serious. In particular, at some hub airports, air transport demand far exceeds airport capacity. Therefore, demand and capacity management is necessary to improve capacity utilization efficiency and ensure flight service quality.

[0003] Currently available strategic flight schedules do not provide information on potential flight delays. The actual impact of strategic flight schedules on airport on-time performance is unknown. Furthermore, strategic flights are often operated multiple times throughout a future season, making their delays uncertain. To address this issue, a method is needed to predict the potential distribution of flight delays in strategic flight schedules. Summary of the Invention

[0004] Purpose of the invention: The purpose of the present invention is to provide a method for predicting strategic flight schedule delay distribution based on machine learning.

[0005] Technical solution: The machine learning-based strategic flight schedule delay distribution prediction method includes the following steps:

[0006] Step 1: Obtain strategic flight schedules and historical flight operation data;

[0007] Step 2: Fit the flight delay distribution of historical flight operation data with a normal distribution;

[0008] Step 3: Select features for flight delay distribution prediction and encode them accordingly;

[0009] Step 4: Use machine learning algorithms to train and predict the distribution of flight delays in the strategic flight schedule;

[0010] Step 5: Propose evaluation indicators to evaluate the prediction results of different machine learning algorithms;

[0011] Step 6: Select the best performing machine learning algorithm to predict the delay distribution of the strategic flight schedule;

[0012] Furthermore, the strategic flight schedule and historical flight operation data in step 1 include flight number, aircraft type, departure airport, destination airport, flight number, expected / actual departure time and expected / actual arrival time.

[0013] Furthermore, the flight delay distribution fitting in step 2 uses normal distribution fitting to obtain the mean μ and variance σ of each flight.

[0014] Furthermore, in step 3, the features selected for flight delay distribution prediction include aircraft type, airport, year, week, airline, strategic departure time, strategic arrival time, flight number, base airline, flight time and number of flights. Beta target encoding is used for categorical variables, minimum maximum encoding is used for numerical variables, and triangular transformation encoding is performed on triangular transformation variables.

[0015] Furthermore, the machine learning algorithms used in step 4 include a multilayer perceptron (MLP) with a loss function of mean absolute error (MAE), an MLP with a loss function of quantile, a light gradient boosting machine (LightGBM), and a random forest (RF).

[0016] Furthermore, the evaluation indicators selected in step 5 are prediction accuracy at different confidence levels, regional area at different interval levels, and distribution similarity. These three indicators are used to evaluate the prediction results of the strategic flight schedule delay distribution.

[0017] Beneficial Effects: This invention can provide flight schedule coordinators with information on the potential delay distribution of strategic schedules. When bottlenecks in airport on-time performance are identified, necessary actions can be taken to reduce flight delays. It can also evaluate strategic flight schedule performance and support strategic flight schedule optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the method of the present invention;

[0019] Figure 2 An example of a flight delay distribution histogram, kernel density curve, and normal distribution fit curve;

[0020] Figure 3 Example of a correct prediction at a 70% confidence level;

[0021] Figure 4 Example of prediction at 20-minute interval level;

[0022] Figure 5 Example of the Wasserstein distance between two distributions. DETAILED DESCRIPTION

[0023] like Figure 1As shown, the method for predicting the strategic flight schedule delay distribution based on machine learning in this embodiment has the following specific steps:

[0024] Step 1: Obtain strategic flight schedules and historical flight operation data;

[0025] The following data were obtained: ① Strategic flight schedule data (see Table 1), including flight number, aircraft type, departure airport, destination airport, flight number, estimated departure time, and estimated arrival time. ② Historical flight operation data, including actual departure time and actual arrival time.

[0026] Table 1 Strategic flight schedule

[0027]

[0028] Step 2: Fit the flight delay distribution of historical flight operation data with a normal distribution;

[0029] The flight delay distribution fitting uses normal distribution fitting to obtain the mean μ and variance σ of each flight. The fitting process is as follows: Figure 2 .

[0030] Step 3: Select features for flight delay distribution prediction and encode them accordingly;

[0031] The features selected for flight delay distribution prediction in step 3 include aircraft type, airport, year, week, airline, strategic departure time, strategic arrival time, flight frequency, base airline, flight duration, and number of flights. A description of each feature is shown in Table 2. Beta objective encoding is used for categorical variables, min-max encoding is used for numerical variables, and triangular transformation encoding is performed on triangular transformation variables.

[0032] Table 2. Feature description of the flight delay distribution prediction algorithm

[0033]

[0034]

[0035] λ (j) =τ / (N (j) +τ) (2)

[0036]

[0037] In the above constraints, expressions (1)-(3) represent the process of Beta target encoding, where μ prior represents the mean of the prior distribution, i represents the number of samples, N represents the number of samples, y irepresents the sample value, j represents the y value of a certain level (the level value here represents how many categorical variable categories there are in the feature), τ represents the effective sample size of the prior distribution, and λ (j) is a hyperparameter that controls the amount of smoothing in the estimate, x (j) Indicates the sample at this level;

[0038] Expression (4) represents the minimum-maximum process, x represents the value of the variable, min represents the minimum value of the variable, and max represents the maximum value of the variable;

[0039] Expression (5) represents the trigonometric transformation encoding process. Taking a day t as an example, sin and cos are used for encoding conversion to ensure the periodicity of time t.

[0040] Step 4: Use machine learning algorithms to train and predict the distribution of flight delays in the strategic flight schedule;

[0041] The data was divided into a training set and a test set in a 5:1 ratio. A regression prediction algorithm was used to predict the mean μ and variance σ of the normal distribution for each flight in the strategic flight schedule. Machine learning algorithms included RF, LightGBM, and MLP. Random search was used to tune hyperparameters during training. Six-fold cross-validation was used to obtain regression prediction results for different algorithms.

[0042]

[0043] Expression (6) indicates that the loss function is MAE, y i Represents the original data, Represents the predicted data, and m represents the number of samples in the original data.

[0044] Expression (7) indicates that the loss function is Quantile, y i Represents the original data, Represents the predicted data, m represents the number of samples in the original data, and r represents the quantile setting value, which ranges from 0 to 1.

[0045] Step 5: Propose evaluation indicators to evaluate the prediction results of different machine learning algorithms;

[0046] The evaluation metrics for delay distribution predictions for the strategic flight schedule include the following: ① Prediction accuracy at different confidence levels; ② Area at different interval levels; and ③ Distribution similarity. The prediction accuracy at different confidence levels is defined as follows: the algorithm predicts the mean and standard deviation of the delay distribution for a single flight and generates a corresponding normal distribution curve. If the mean of the actual delay distribution falls within this confidence interval at different confidence levels, the prediction is considered correct. Figure 3A correct prediction for a flight at a 70% confidence level is given. The prediction area at different interval levels is defined as follows. The algorithm predicts the mean and standard deviation of the delay distribution of a single flight and generates a corresponding normal distribution curve. The mean of the actual delay distribution corresponds to two points on the x-axis at different interval levels. The larger the area, the more accurate the prediction. Figure 4 An example is given where the mean of the true delay distribution is 30 minutes and the delay interval level is 20 minutes. The difference between the true delay distribution and the predicted delay distribution is measured using the Wasserstein distance. The smaller the value, the more accurate the prediction. Figure 5 An example of the Wasserstein distance between two distributions is given.

[0047] W(P,Q)=inf γ∈Π(P,Q) E (x,y)～γ [||xy||] (8)

[0048] Expression (8) represents the calculation method of Wasserstein distance, where P and Q represent the two distributions for which the Wasserstein distance is to be calculated, x represents the starting point, y represents the target point to be pushed, r represents the cost of the operation, Π(P,Q) represents all possible joint distributions of distributions P and Q, and the total cost is minimized using the Expectation Maximum (EM) method.

[0049] Based on the evaluation indicators for individual flight prediction results, three corresponding indicators are proposed to analyze the prediction results from a holistic perspective. ① Prediction accuracy at different confidence levels. Accuracy is defined as the proportion of flights in the strategic flight schedule that are correctly predicted at this confidence level. ② Average region area at different interval levels. Region area is defined as the average region area at this interval level in the strategic flight schedule. ③ Distribution similarity frequency distribution. A frequency distribution plot of the Wasserstein distance is used; taller frequency distributions indicate excellent prediction performance.

[0050] Step 6: Select the best performing machine learning algorithm to predict the delay distribution of the strategic flight schedule;

[0051] Select the best performing machine learning algorithm to predict the delay distribution of strategic flight schedules.

Claims

1. A machine learning-based method for predicting the distribution of strategic flight schedule delays, characterized by: The steps include: Step 1: Obtain strategic flight schedules and historical flight operation data; Step 2: Fit the flight delay distribution of historical flight operation data with a normal distribution; Step 3: Select features for flight delay distribution prediction, including aircraft type, airport, year, week, airline, strategic departure time, strategic arrival time, flight frequency, base airline, flight time, and number of flights. Use Beta target encoding for categorical variables, min-max encoding for numerical variables, and triangular transformation encoding for triangular transformation variables. Step 4: Use machine learning algorithms to train and predict the distribution of flight delays in the strategic flight schedule; Step 5: Propose evaluation indicators to evaluate the prediction results of different machine learning algorithms. The evaluation indicators include prediction accuracy at different confidence levels, regional area at different interval levels, and distribution similarity. These three indicators are used to evaluate the prediction results of strategic flight schedule delay distribution. Step 6: Select the best performing machine learning algorithm to predict the delay distribution of strategic flight schedules.

2. The method for predicting strategic flight schedule delay distribution based on machine learning according to claim 1, characterized in that: The strategic flight schedule and historical flight operation data in step 1 include flight number, aircraft type, departure airport, destination airport, flight number, expected / actual departure time and expected / actual arrival time.

3. The method for predicting strategic flight schedule delay distribution based on machine learning according to claim 1, characterized in that: In step 2, the flight delay distribution is fitted using a normal distribution to obtain the mean μ and variance σ of each flight.

4. The method for predicting strategic flight schedule delay distribution based on machine learning according to claim 1, characterized in that: The machine learning algorithms used in step 4 include a multilayer perceptron (MLP) with a loss function of mean absolute error (MAE), an MLP with a loss function of quantile, a light gradient boosting machine (LightGBM), and a random forest (RF).

Images