A Smart Flight Departure Time Prediction Method Based on Multi-Model Machine Learning

Abstract

The present invention proposes a flight departure time intelligent prediction method based on multi-model machine learning, including: constructing a feature data set and selecting key features; establishing multiple flight departure time prediction models, training the flight departure time TOT as a label, and using the trained flight departure time prediction models to make predictions respectively; analyzing the trained multiple flight departure time prediction models to obtain the optimal flight departure time prediction model and key features; selecting the optimal flight departure time prediction model and key features to complete the flight departure time intelligent prediction based on multi-model machine learning. The present invention establishes the correlation between key influencing factors and flight departure time through data mining, extracts the key features of flight departure time prediction, realizes the accurate prediction of the departure time of airport flights, and provides technical support for field resource scheduling, flight schedule adjustment, etc. in a dynamic and real-time manner in combination with the actual operation status of the field.

Description

Technical Field

[0001] The present invention relates to a flight take-off time prediction method, in particular to a flight take-off time intelligent prediction method based on multi-model machine learning. Background Art

[0002] This section merely provides background information related to the present disclosure and is not necessarily prior art.

[0003] Typical studies on airport flight time prediction methods include, for example, in 2017, Feng Xia et al., aimed at the problems of poor traffic flow caused by the rough estimation of flight taxi time at large and busy airports. Based on KNN and SVR methods, they built a taxi time prediction model for departing flights. The model took into account factors such as taxi distance, the number of flights taxiing on the same runway during taxiing, and the number of flights pushed out within 15 minutes after wheel block removal. The number of take-off and landing flights using the same runway during taxiing was predicted. Based on the above prediction results, the taxi time of flights was further predicted. For example, in 2022, Chen Zhen et al. created a total of 550 candidate features including primary and secondary indicators from three dimensions of time, space and environment, and studied the taxi time prediction method of departing flights with the help of machine learning algorithms. The traditional airport flight time prediction method often focuses on the taxi time of the departure stage. The existing research on the prediction of airport flight take-off time currently has the following problems: the research methods mainly focus on predicting the taxi time of the departure stage, and there is relatively little research on the prediction of individual flight take-off time.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0005] Purpose of the invention: The technical problem to be solved by the present invention is to provide a method for intelligent prediction of flight departure time based on multi-model machine learning in view of the shortcomings of the prior art.

[0006] In order to solve the above technical problems, the present invention discloses a flight departure time intelligent prediction method based on multi-model machine learning, comprising the following steps:

[0007] Step S1, construct a feature data set and select key features, collect multi-source heterogeneous flight historical operation data, perform data cleaning, and construct new indicators for feature extraction to form a feature data set; perform data standardization on the feature data set, analyze the correlation factors between features and labels, and select key features based on the correlation;

[0008] Step S2, establishing multiple flight departure time prediction models, using the flight departure time TOT as a label for training, and using the trained flight departure time prediction models to perform predictions respectively;

[0009] Step S3, analyzing the trained multiple flight departure time prediction models to obtain the optimal flight departure time prediction model and the key features most relevant to the flight departure time;

[0010] Step S4, select the optimal flight departure time prediction model and key features selected in step S3 to complete the intelligent prediction of flight departure time based on multi-model machine learning.

[0011] Furthermore, the step S1 of constructing a feature data set and selecting key features specifically includes:

[0012] Step S1-1, data cleaning, including: processing missing values, outliers and duplicate data in the data;

[0013] Step S1-2, feature extraction, includes: extracting features from the flight historical operation data after data cleaning, and constructing quantitative indicators;

[0014] Step S1-3, key feature selection, includes: standardizing the extracted features, calculating the Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient between the features and the flight departure time TOT, and selecting key features in descending order of the correlation coefficients.

[0015] Furthermore, in step S1-1, the outliers in the data are processed, that is, a data statistics method based on a Z-core standardized score is used to identify and process outlier data points greater than a threshold.

[0016] Furthermore, the identification and processing of abnormal data points greater than a threshold value described in step S1-1 specifically includes:

[0017] Assume that the flight history operation data set is: F = {f1,f2,...,f n}, where f n represents the nth data point, whose mean is μ and standard deviation is σ, then the kth data point f in the data set F is k The Z-core standardized score Z k for:

[0018]

[0019] Data points with scores greater than the threshold are removed.

[0020] Furthermore, in step S1-2, the feature extraction includes:

[0021] Step S1-2-1, extracting scene flow characteristics, specifically as follows:

[0022] According to the overlapping relationship between the scheduled off-block time SOBT(j), scheduled take-off time STOT(j), scheduled landing time SLDT(j), scheduled on-block time SIBT(j) of other flights and the currently scheduled taxiing flight STOT(i), i represents the currently scheduled taxiing flight and j represents other flights; 8 surface flow characteristics are derived, among which:

[0023] The first scene flow characteristic D1 is expressed as follows:

[0024] D1(i)=count(j),t SOBT (j)<t SOBT (i)&t SOBT (i)<t STOT (j)<t STOT (i)

[0025] The second scene flow characteristic D2 is expressed as follows:

[0026] D2(i)=count(j),t SOBT (j) <t SOBT (i)&t STOT (j)>t STOT (i)

[0027] The third scene flow characteristic D3 is expressed as follows:

[0028] D3(i)=count(j),t SOBT (j)>t SOBT (i)&t SOBT (i) <t STOT (j) <t STOT (i)

[0029] The fourth scene flow characteristic D4 is expressed as follows:

[0030] D4(i)=count(j),t SOBT (i) <t SOBT (j) <t STOT (i)&t STOT (j)>t STOT (i)

[0031] The fifth scene flow characteristic D5 is expressed as follows:

[0032] D5(i)=count(j),t SLDT (j) <t SOBT (i)&t SOBT(i) <t SIBT (j) <t STOT (i)

[0033] The sixth scene flow characteristic D6 is expressed as follows:

[0034] D6(i)=count(j),t SLDT (j)<t SOBT (i)&t SIDT (j)>t STOT (i)

[0035] The seventh scene flow characteristic D7 is expressed as follows:

[0036] D7(i)=count(j),t SLDT (j)>t SOBT (i)&t SOBT (i) <t SIBT (j) <t STOT (i)

[0037] The eighth scene flow characteristic D8 is expressed as follows:

[0038] D8(i)=count(j),t SOBT (i) <t SLDT (j) <t STOT (i)&t SIBT (j)>t STOT (i)

[0039] Step S1-2-2, calculate the average sliding time t within the preset time slice d The features are as follows:

[0040]

[0041] in, represents the average taxi-out time of the i-th departing flight, t STOT(i) -t SOBT(i) is the planned taxi-out time of the i-th departing flight, and n is the time slice Δ l The number of departing flights that slide out, m represents the mth time slice;

[0042] Step S1-2-3, calculate the arrival and departure flight ratio α as a feature, as follows:

[0043] Calculate the ratio of the number of all scheduled departure flights to all scheduled arrival flights during the entire scheduled taxiing process of the current flight i, α i , the method is as follows:

[0044]

[0045] Among them, D 1,i It indicates the number of flights that meet the first scene flow characteristics during the planned taxi-out process of the current flight, D 2,i It indicates the number of flights that meet the second field traffic characteristics during the planned taxi-out process of the current flight, D 3,i It indicates the number of flights that meet the third field traffic characteristics during the planned taxi-out process of the current flight, D 4,i It indicates the number of flights that meet the fourth field traffic characteristics during the planned taxi-out process of the current flight, D 5,i It indicates the number of flights that meet the fifth field traffic characteristics during the planned taxi-out process of the current flight, D 6,i It indicates the number of flights that meet the sixth field traffic characteristics during the planned taxi-out process of the current flight, D 7,i It indicates the number of flights that meet the seventh field traffic characteristics during the planned taxi-out process of the current flight, D 8,i Indicates the number of flights that meet the eighth field traffic characteristics during the planned taxi-out process of the current flight;

[0046] Step S1-2-4, selecting the departure runway allocation R as a feature;

[0047] Step S1-2-5, dividing the airport operation time period T as a feature, that is, using the K-means method to cluster the features according to the planned departure time and the flight operation time period of the airport;

[0048] Step S1-2-6, selecting model M as a feature;

[0049] Step S1-2-7, select airline attribute A as the feature;

[0050] Step S1-2-8, selecting a runway and a stand combination as a feature;

[0051] Step S1-2-9, select the block time of the preceding flight as a feature.

[0052] Furthermore, the standardization process described in step S1-3 is to scale the feature data so that it falls within the range [0,1]. The specific method is as follows:

[0053]

[0054] Among them, x norm represents the feature data after standardization, x s Represents the sth feature data, max{} represents the maximum value, and min{} represents the minimum value.

[0055] Furthermore, the Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient between the calculated features described in step S1-3 and the flight departure time TOT.

[0056] Furthermore, the prediction model of the multiple flight departure times described in step S2 includes:

[0057] Take-off time prediction model based on BP neural network, take-off time prediction model based on multivariate linear regression and take-off time prediction model based on support vector machine.

[0058] Further, the analysis of the trained multiple flight departure time prediction models described in step S3 specifically includes:

[0059] Step S3-1, design analysis indicators, the analysis indicators include mean square error MSE, mean absolute error MAE, prediction accuracy R 2 and running time;

[0060] Step S3-2, inputting the test data set into the prediction model of all flight departure times, performing prediction, and calculating the analysis index according to the prediction results;

[0061] Step S3-3, selecting the model with the best analysis index as the optimal flight departure time prediction model.

[0062] Furthermore, the key features most relevant to the flight departure time obtained in step S3 specifically include:

[0063] In the process of prediction in step S3-2, the Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient of all features are calculated, and a preset number of features are selected as the final key features in the order of the three types of correlation coefficients from high to low.

[0064] Beneficial effects:

[0065] Based on the influencing factors related to flight departure time, the present invention constructs a multi-angle feature set, integrates the distribution characteristics of different data, calculates and screens data correlation from the perspective of linear and nonlinear, and selects the flight departure time prediction model with the best performance through training with three classic machine learning models. The beneficial effects are as follows:

[0066] 1. The present invention mines and analyzes the historical data of the tripartite collaboration through data mining technology, establishes the correlation between key influencing factors and flight departure time, extracts and screens the key features for flight departure time prediction, and builds and trains a multi-factor flight departure time prediction model based on multi-model machine learning on this basis to accurately predict the departure time of airport flights.

[0067] 2. The present invention accurately predicts the key time nodes of airport flights, combines the actual operation status of the airport, and provides technical support for airport resource scheduling, flight schedule adjustment, etc. in a dynamic and real-time manner. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments, and the above and / or other advantages of the present invention will become more clear.

[0069] Figure 1 It is a schematic diagram of the overall process of the present invention.

[0070] Figure 2 Schematic diagram for constructing a feature dataset and selecting key features. DETAILED DESCRIPTION

[0071] The present invention discloses a flight departure time intelligent prediction method based on multi-model machine learning, such as Figure 1 As shown, including:

[0072] Step S1, feature engineering, includes collecting multi-source heterogeneous flight historical operation data, cleaning the data, and constructing new indicators for feature extraction to form a high-quality, usable feature data set. The processed feature set is subjected to data standardization, and the correlation factors between features and between features and labels are analyzed, and key features are selected according to the degree of correlation;

[0073] Step S2, establishing a prediction model for flight departure time, using the departure time TOT of the departing flight as a label for training and accurate prediction;

[0074] Step S3, designing a prediction algorithm to efficiently and accurately predict the departure time of relevant flights.

[0075] Furthermore, in step S1, the feature engineering is as follows: Figure 2 As shown in the figure, it includes collecting multi-source heterogeneous flight historical operation data, cleaning the data, and constructing new indicators for feature extraction to form a high-quality, usable feature data set. The processed feature set is subjected to data standardization, and the correlation factors between features and between features and labels are analyzed, and key features are selected according to the degree of correlation. Among them,

[0076] Step S1-1, data cleaning includes three items: processing missing values, processing outliers, and processing duplicate data;

[0077] Step S1-2, feature extraction includes extracting new features from the original data and constructing new quantifiable indicators to improve the performance of the model. The features include TOT-related influencing factors.

[0078] Step S1-3, feature selection includes feature standardization calculation, calculation of the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall correlation coefficient between the feature and TOT, and selecting the most critical features in descending order of correlation.

[0079] Furthermore, in step S1-1, the data cleaning includes:

[0080] Step S1-1-1, processing missing values;

[0081] Step S1-1-2, processing abnormal values;

[0082] Step S1-1-3, processing duplicate data.

[0083] Furthermore, in step S1-1-1, the method for processing missing values ​​is to fill in the missing values ​​existing in the acquired original data or delete the data points containing missing values.

[0084] Furthermore, in step S1-1-2, the method for processing outliers is a data statistical method based on Z-core standardized scores to identify and process outlier data points greater than a certain threshold. Assume that the flight collaborative operation data set X = {x1, x2, ..., x n}, its mean is μ, and its standard deviation is σ, then the data point x i The calculation method of the Z-core standardized score is shown in formula (1).

[0085]

[0086] Furthermore, in step S1-1-3, the method for processing duplicate data is to detect and remove duplicate data records to ensure the uniqueness and consistency of the data.

[0087] Furthermore, in step S1-2, the TOT-related influencing factor characteristics include relevant quantifiable factors related to the taxi-out process of the ground flight, including:

[0088] (1) Scene flow characteristics

[0089] According to the relationship between SOBT(j), STOT(j), SLDT(j), SIBT(j) of other flights and ATOT(i) of sample flight, eight surface flow characteristics (D1 to D8) can be derived as shown in the following table.

[0090] Table 1 Surface flow characteristics of departing flights

[0091] Feature variables definition <![CDATA[D1]]> <![CDATA[D1(i)=count(j),t SOBT (j)<t SOBT (i)&t SOBT (i)<t STOT (j)<t STOT (i)]]> <![CDATA[D2]]> <![CDATA[D2(i)=count(j),t SOBT (j)<t SOBT (i)&t STOT (j)>t STOT (i)]]> <![CDATA[D3]]> <![CDATA[D3(i)=count(j),t SOBT (j)>t SOBT (i)&t SOBT (i)<t STOT (j)<t STOT (i)]]> <![CDATA[D4]]> <![CDATA[D4(i)=count(j),t SOBT (i)<t SOBT (j)<t STOT (i)&t STOT (j)>t STOT (i)]]> <![CDATA[D5]]> <![CDATA[D5(i)=count(j),t SLDT (j)<t SOBT (i)&t SOBT (i)<t SIBT (j)<t STOT (i)]]> <![CDATA[D6]]> <![CDATA[D6(i)=count(j),t SLDT (j)<t SOBT (i)&t SIDT (j)>t STOT (i)]]> <![CDATA[D7]]> <![CDATA[D7(i)=count(j),t SLDT (j)>t SOBT (i)&t SoBT (i)<t SIBT (j)<t STOT (i)]]> <![CDATA[D8]]> <![CDATA[D8(i)=count(j),t SOBT (i)<t SLDT (j)<t STOT (i)&t SIBT (j)>t STOT (i)]]>

[0092] In the above table, D1(i) indicates the number of other flights that meet the first type of surface flow characteristics during the planned taxi-out process of the current flight, count(j) indicates the count of flight j that meets the first type of surface flow characteristics, and t SOBT (j) represents the planned off-block time of other flight j, t SOBT (i) represents the planned off-block time of flight i currently scheduled to taxi, t STOT (j) represents the scheduled departure time of other flight j, t STOT (i) indicates the scheduled departure time of flight i currently scheduled to taxi; D2(i) indicates the number of other flights that meet the second type of surface flow characteristics during the current flight’s scheduled taxiing process; D3(i) indicates the number of other flights that meet the third type of surface flow characteristics during the current flight’s scheduled taxiing process; D4(i) indicates the number of other flights that meet the fourth type of surface flow characteristics during the current flight’s scheduled taxiing process; D5(i) indicates the number of other flights that meet the fifth type of surface flow characteristics during the current flight’s scheduled taxiing process; t SLDT (j) represents the planned landing time of other flight j, t SIBT D6(i) represents the number of other flights that meet the sixth type of surface flow characteristics during the planned taxi-out process of the current flight; D7(i) represents the number of other flights that meet the seventh type of surface flow characteristics during the planned taxi-out process of the current flight; D8(i) represents the number of other flights that meet the eighth type of surface flow characteristics during the planned taxi-out process of the current flight.

[0093] (2) Average slip-out time t in a 30-minute time slice d , unit is seconds.

[0094]

[0095] Among them, t ATOT(i) -t AOBT(i) is the taxi-out time of the i-th departing flight, and n is the time slice Δ l The number of departing flights that are taxied out requires that the actual off-block time of the departing flights must fall within t AOBT(i) Time slice Δ l Inside.

[0096] (3) Ratio of arriving and departing flights α

[0097] The ratio of all departing aircraft to all arriving aircraft during the entire taxiing process of flight i.

[0098]

[0099] (4) Departure runway allocation R

[0100] Since the time node of the flight taxiing process is related to the taxiing distance and flight flow, and the flight flow of different runways is different, it will reversely affect the time node of the flight assigned to take off from this runway.

[0101] According to the actual data of K Airport, it is a dual-runway airport with four operating directions, namely 03, 21, 04, and 22, numbered 1 to 4 respectively.

[0102] (5) Operation period T

[0103] The airport is in different operating periods, and its taxiing situation is different. Cluster analysis can be performed based on the scheduled departure time and the flight operation period of Airport K to obtain specific time period classification. The clustering method uses the most commonly used K-means method, and the clustering feature uses the total number of scheduled departures from Airport K within 1 to 24 hours throughout 2023.

[0104] By setting parameters, the clustering is divided into 5 categories, and the whole day is clustered into 5 time periods, which are numbered 1 to 5. Among them, [0:00-5:59] is time period 1, [6:00-6:59]∪[22:00-23:59] is time period 2, [7:00-8:59] is time period 3, [9:00-9:59]∪[12:00-13:59]∪[15:00-15:59] is time period 4, and [10:00-11:59]∪[14:00-14:59]∪[16:00-21:59] is time period 5.

[0105] Therefore, the operating time period mainly includes the time period to which the planned departure time belongs and the category to which the planned departure time belongs.

[0106] (6) Model M

[0107] Different aircraft types have different requirements for taxiways, parking spaces, etc. Therefore, taking the aircraft type into consideration, the aircraft types are divided into four categories: C, D, E, and F, corresponding to numbers 1 to 4 respectively.

[0108] (7) Airline Attribute A

[0109] There are some differences in the priority of airports for domestic and international airline operations, so the airline attributes are also critical to the taxi-out process of departing flights. International airlines are 1, and domestic airlines are 0.

[0110] (8) Runway-aircraft combination

[0111] Since the flight departure time is also affected by different taxi paths, the runway-aircraft stand combination is used as a feature input.

[0112] (9) Actual block time of the preceding flight

[0113] Furthermore, in step S1-3, the feature selection

[0114] Step S1-3-1, feature standardization processing;

[0115] Step S1-3-2, calculate the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall correlation coefficient between the feature and TOT.

[0116] Furthermore, in step S1-3-1, the data standardization processing method is to scale the data so that it falls within the range [0,1] so as to have a similar scale. The standardization formula is shown in formula (4).

[0117]

[0118] Furthermore, in step S1-3-2, the Pearson correlation coefficient, Spearman correlation coefficient, and Kendall correlation coefficient between the calculated features and TOT are as follows.

[0119] The most relevant or informative features to the predicted object are selected through the Pearson, Spearman, and Kendall correlation methods to reduce the dimension of the data set and improve the generalization ability of the model to prevent overfitting of the model. The specific implementation method is to evaluate the correlation between the extracted features and the predicted time nodes, sort them according to the correlation, extract the key features with greater correlation, and predict the time nodes.

[0120] The Pearson correlation coefficient has the characteristics of translation invariance and scale invariance, and can be used to calculate the correlation between two vectors. Thanks to this characteristic, the Pearson coefficient is superior to many correlation indicators and can generally be used as a universal standardized indicator. Its formula is:

[0121]

[0122] Considering that the relationship between the aforementioned characteristics is not obvious, the Spearman rank correlation coefficient and Kendall correlation coefficient are used to comprehensively reflect the relationship between the two variables.

[0123] The Spearman coefficient formula is as follows:

[0124]

[0125] Where n is the sample size. The Spearman rank correlation coefficient has much looser requirements on data conditions than the Pearson correlation coefficient. As long as the observations of the two variables appear in pairs, or are both converted from continuous values ​​at the same time, it can reflect the degree of correlation without considering the overall distribution characteristics and sample space capacity.

[0126] The Kendall rank coefficient formula is as follows:

[0127]

[0128] Among them, C represents the number of pairs of elements with consistency in samples X and Y (two elements form a pair); D represents the number of pairs of elements with inconsistency in samples X and Y. As a non-parametric hypothesis test, the data condition requirements are basically the same as those of the Spearman correlation coefficient. In the case of comprehensive consideration of the above three coefficients, the correlation of the features has basically met the requirements.

[0129] Furthermore, in step S2, the flight take-off time prediction model established is a take-off time prediction model based on BP neural network, a take-off time prediction model based on multivariate linear regression, and a take-off time prediction model based on support vector machine.

[0130] (1) Flight departure time prediction model based on BP neural network

[0131] The BP neural network algorithm is a multi-layer feedforward neural network. It performs reverse propagation training based on the error of each output. The process of back propagation is to derive the error to change the parameters. Therefore, the essence of the BP neural network is a process of training model parameters.

[0132] From the basic neuron structure, we can see that the basic unit is composed of input X i , weight w ij , output y j , threshold θ, etc. The threshold θ can be regarded as a weight w i0 The input of the neuron is f(), and f() represents the activation function. The integration function is used to synthesize the information of other neuron inputs, and the synthesis result is used as the input of this neuron. The relationship between input and output is

[0133]

[0134] Y i =f(net i ) (9)

[0135] If net i If the value is greater than 0, the neuron immediately jumps to an excited state and randomly outputs relevant information; if the net i If it is less than 0, the neuron is in an inhibited state and no relevant information is output. In the model, the activation function must be a differentiable function in nature. The more commonly used activation function is the logsig function with a value range of (0,1):

[0136]

[0137] And the tansig function with a range of (-1, 1)

[0138]

[0139] Usually, a training set is required to train the BP neural network. The training data set is usually divided into data for BP training and data for expected output.

[0140] (2) Flight departure time prediction model based on multivariate linear regression

[0141] The linear combination of model parameters with multiple independent variable regression coefficients is called multivariate linear regression. In linear regression, data is modeled using a linear prediction function, and all unknown parameters are estimated based on the data. Linear fitting is performed using least absolute error regression.

[0142] Let Y be the dependent variable, X1, X2, …, X n When is the independent variable and there is a linear relationship between the independent variable and the dependent variable, the multiple linear regression model is:

[0143] Y=a+b1X1+b2X2+…+b n X n +ε (12)

[0145] Among them, a is a constant term, b1, b2, …, b n is the regression coefficient, b1 is X1, X2,…, X n When x is fixed, the effect of each unit increase of X1 on Y is the partial regression coefficient of X1 on Y, and so on; ε is the error term.

[0146] (3) Flight departure time prediction model based on support vector machine

[0147] Support Vector Machine (SVM) can not only solve the classification problem of discrete variables, but also solve the prediction problem of continuous variables. Compared with traditional linear regression, it has the characteristics of less demanding data distribution constraints, the model is not affected by multicollinearity, and the model is less affected by outliers.

[0148] In the linear regression model, the loss function of the model is actually the difference between the predicted value and the actual value. When the two are equal, the loss is 0, and when the two are not equal, the loss is calculated. Support vector machine regression has a certain degree of fault tolerance, which uses the relaxation factor ζ (*) , ensuring the inequality

[0149] |Y i -f( X i )|-ζ(*) ≤ε (13)

[0150] In the above formula, Y i is the actual label of the i-th dataset, f(X i ) is X i The predicted value of |Y i -f(X i )| is the absolute value of the difference between the actual data and the predicted data, and ε is the error term formed randomly.

[0151] According to the objective function and constraints of the SVM regression model

[0152]

[0153] In the above model, f(X i )=w′X i +b is a hyperplane, and the parameters w and b are the normal vector and intercept of the hyperplane respectively. i -f(X i )|-ζ (*) ≤ε is written as two inequality constraints in the objective function; ζ i and Indicates that the sample points x outside the "division zone" i The cost required to pull it back into the two bands. By solving this model, we can get the predicted value that meets the requirements.

[0154] Furthermore, in step S3, the design prediction algorithm mainly selects three prediction algorithms for comparative analysis:

[0155] (1) Flight departure time prediction algorithm based on BP neural network

[0156] Using the BP neural network model to predict the departure time of inbound and outbound flights, the sample data set after feature reconstruction can be divided into a training set and a test set according to a certain weight distribution, and then the flight sample data of the training set is input into the neural network. The initial weight of each neuron contained in the layers of the neural network is randomly assigned. In the learning process, the feedforward algorithm is run to obtain the output result, and the error is obtained by subtracting the actual output result in the test set. The error function can reflect the effect of learning and training. Then the error is back-propagated from the output layer to the input layer. At this time, the gradient steepest descent method is used to minimize the error function. The weight and threshold are corrected according to the error term propagated back from the previous layer. Finally, the error term meets the parameter setting requirements, that is, the training process ends. The learning cycle of the neural network is repeated many times, mainly because there are many weight components. Neurons interact in the hidden layer, and the network performance depends on the number of hidden neurons and the selection of weights. Nonlinear conversion functions are often applied to the output layer.

[0157] ① Setting the number of neurons and hidden layers

[0158] The key points of implementing a good neural network include the number of neurons and the number of hidden layers. Generally speaking, for many features that are less relevant to the results, a neural network with many layers is required. However, too many neurons will lead to slow convergence or overfitting due to the high complexity of the model. For general problems, setting a 3-layer neural network is sufficient to train data sets with nonlinear relationships. For capacity prediction problems, a 3-layer BP neural network is selected: 1 input layer, 1 hidden layer, and 1 output layer.

[0159] The 17 features are used as the input layer of the neural network, so the number of nodes in the input layer is set to 17; the capacity evaluation value label obtained by simulation is used as the output layer, and the number of nodes in the output layer is set to 1; the number of layers and nodes in the hidden layer are set according to the rules specified by the empirical formula (15):

[0160]

[0161] Among them, N is the total number of nodes contained in the hidden layer, N in is the total number of nodes contained in the input layer, N out is the number of nodes in the output layer, l is an integer in the interval [1,10]. The number of hidden layer nodes is set to 15.

[0162] ②Use of activation function

[0163] In the BP neural network prediction process, activation functions are also needed. Different activation functions have different prediction accuracy. The ReLU function is not differentiable in the entire interval, but its slope is a constant in the positive interval, so it solves the problem of gradient vanishing and gradient descent. The Leaky ReLU function sets the slope to a variable value in the positive interval, not a constant 1. Therefore, compared with the ReLU function, the Leaky ReLU function inherits its characteristics and has more flexibility. Therefore, when predicting the time of key nodes of airport flights, the experimental environment is tensorflow, and the activation function uses the Leaky ReLU function.

[0164] ③Optimizer selection

[0165] Gradient Descent refers to finding the optimal value in the opposite direction of the gradient. The learning rate determines the step size of each step of the optimization. When approaching the optimal value, choosing too large a learning rate will cause the function to zigzag, and too small a learning rate will increase the number of iterations. Since SGD is prone to oscillation when encountering gullies, the momentum algorithm Momentum is introduced on this basis to suppress oscillation. The above algorithms all have fixed learning rates. In more cases, the learning rate is preferably able to change with the update frequency of the parameters to reduce the convergence time of the algorithm. Therefore, first-order momentum and second-order momentum are introduced to dynamically adjust the learning rate. The more famous one is the Adam optimization algorithm, which uses the first-order momentum to dynamically adjust the learning rate. In the present invention, the Adam optimizer is used.

[0166] Finally, the present invention sets the BP neural network model to have 17 input nodes, 1 hidden layer, 15 nodes in the hidden layer, and 1 output node; the optimizer is Adam; the dropout method is adopted to effectively avoid the occurrence of overfitting, and the dropout value is set to 0.2; the maximum number of iterations is set to 200, and the learning rate is set to 0.01; the LOSS function selects cross entropy, and divides the valid flight feature set into 0.8% for the training set and 0.2% for the test set;.

[0167] (2) Flight departure time prediction algorithm based on multivariate linear regression

[0168] ① Arrange the feature data set and divide it into a training set and a test set. Divide the valid flight feature set into a training set with a ratio of 0.8 and a test set with a ratio of 0.2;

[0169] ② Establish a linear regression model according to the linear regression formula, solve the partial regression coefficient of the model, and evaluate the prediction accuracy of the model;

[0170] (3) Flight departure time prediction algorithm based on support vector machine

[0171] ① Aiming at the nonlinear separability of flight feature data sets, the kernel function method is introduced to transform the high-dimensional sample data space into a low-dimensional space that can be linearly segmented;

[0172] ②Use the maximum interval method to obtain the segmentation line with the largest interval, and then find the sample data on the maximum separation hyperplane, which is the required support vector;

[0173] ③ Use the separation hyperplane and support vector obtained in the previous step to classify and predict the new samples.

[0174] The experimental environment is Python 3.11, and the library used is scikit-learn. The convergence condition of SVM is specified as 0.001, and the loss function is set to represent |γ q -f(X q )|-ζ (*) The “epsilon_insensitive” of ≤ε, the maximum number of iterations of the prediction model is specified as 200, the Gaussian kernel function is selected as the kernel function, and the valid flight feature set is divided into a training set with a ratio of 0.8 and a test set with a ratio of 0.2.

[0175] For the above three prediction methods, the prediction performance indicators are mean square error (MSE), mean absolute error (MAE) and prediction accuracy (R 2 ).

[0176] Taking Table 2 as an example, the performance of TOT-related influencing factor characteristics in different correlation evaluation methods is shown.

[0177] Table 2 Correlation between characteristics and TOT and significance test results

[0178]

[0179]

[0180] Note: The values ​​in brackets above are negative numbers, and the values ​​in bold are absolute values ​​not less than 0.1.

[0181] As shown in Table 5, for the same feature, the three correlation coefficients show different performances, but the most relevant features are the planned departure period, the actual block time of the preceding flight (Spearman correlation coefficient and Kendall correlation coefficient are large), the planned departure period category (all three correlation coefficients are greater than 0.40), the average taxi-out time of 30-minute time slices, D5-D8, and the number of departing / arriving flights during the taxi-out period. These nine features are selected for subsequent prediction work.

[0182] Taking Table 3 as an example, the prediction results of three methods, namely BP neural network, multivariate linear regression and SVM regression, are shown.

[0183] Table 3 Comparison of the results of three prediction methods of TOT

[0184] Prediction Methods Mean Square Error (MSE) Mean Absolute Error (MAE) <![CDATA[Prediction accuracy (R 2 )]]> Running time (min) BP Neural Network 7403.40 24.64 91.96％ 2 Multiple Linear Regression 10653.77 30.69 88.43％ 1 SVM Regression 11162.71 31.13 87.88％ 10

[0185] From the above table, we can see that BP neural network has good performance in MSE, MAE, R 2The three aspects performed best, showing good prediction performance, with an accuracy rate of up to 91.96%, and its running time was also short. Therefore, the BP neural network model is the best choice for flight departure time prediction. It can accurately predict the flight departure time in real time based on the actual operation situation of the field, and provide technical support for field resource scheduling, flight schedule adjustment, etc.

[0186] In a specific implementation, the present application provides a computer storage medium and a corresponding data processing unit, wherein the computer storage medium can store a computer program, and when the computer program is executed by the data processing unit, the invention content of the intelligent prediction method of flight departure time based on multi-model machine learning provided by the present invention and some or all steps in each embodiment can be executed. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM), etc.

[0187] Those skilled in the art can clearly understand that the technical solutions in the embodiments of the present invention can be implemented by means of computer programs and their corresponding general hardware platforms. Based on such an understanding, the technical solutions in the embodiments of the present invention can be essentially or partly contributed to the prior art in the form of computer programs, i.e., software products, which can be stored in a storage medium and include several instructions for enabling a device including a data processing unit (which can be a personal computer, a server, a single-chip microcomputer, an MCU or a network device, etc.) to execute the methods described in various embodiments of the present invention or certain parts of the embodiments.

[0188] The present invention provides a method and idea for intelligent prediction of flight departure time based on multi-model machine learning. There are many methods and approaches to implement the technical solution. The above is only a preferred implementation of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the protection scope of the present invention. All components not specified in this embodiment can be implemented by existing technologies.

Claims

1. A method for intelligent prediction of flight departure time based on multi-model machine learning, characterized in that: The following steps are involved: Step S1, construct a feature data set and select key features, collect multi-source heterogeneous flight historical operation data, perform data cleaning, and construct new indicators for feature extraction to form a feature data set; perform data standardization on the feature data set, analyze the correlation factors between features and labels, and select key features based on the correlation; Step S2, establishing multiple flight departure time prediction models, using the flight departure time TOT as a label for training, and using the trained flight departure time prediction models to perform predictions respectively; Step S3, analyzing the trained multiple flight departure time prediction models to obtain the optimal flight departure time prediction model and the key features most relevant to the flight departure time; Step S4, select the optimal flight departure time prediction model and key features selected in step S3 to complete the intelligent prediction of flight departure time based on multi-model machine learning.

2. The method for intelligent prediction of flight departure time based on multi-model machine learning according to claim 1, characterized in that: The step S1 of constructing a feature data set and selecting key features specifically includes: Step S1-1, data cleaning, including: processing missing values, outliers and duplicate data in the data; Step S1-2, feature extraction, includes: extracting features from the flight historical operation data after data cleaning, and constructing quantitative indicators; Step S1-3, key feature selection, includes: standardizing the extracted features, calculating the Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient between the features and the flight departure time TOT, and selecting key features in descending order of the correlation coefficients.

3. The method for intelligent prediction of flight departure time based on multi-model machine learning according to claim 2, characterized in that: In step S1-1, the outliers in the data are processed, that is, a data statistical method based on a Z-core standardized score is used to identify and process outlier data points that are greater than a threshold.

4. The method for intelligent prediction of flight departure time based on multi-model machine learning according to claim 3, characterized in that: The identification and processing of abnormal data points greater than a threshold value described in step S1-1 specifically includes: Assume that the flight history operation data set is: F = {f1,f2,...,f n }, where f n represents the nth data point, whose mean is μ and standard deviation is σ, then the kth data point f in the data set F is k The Z-core standardized score Z k for: Data points with scores greater than the threshold are removed.

5. The method for intelligent prediction of flight departure time based on multi-model machine learning according to claim 4, characterized in that: In step S1-2, the feature extraction includes: Step S1-2-1, extracting scene flow characteristics, specifically as follows: According to the overlapping relationship between the scheduled off-block time SOBT(j), scheduled take-off time STOT(j), scheduled landing time SLDT(j), scheduled on-block time SIBT(j) of other flights and the currently scheduled taxiing flight STOT(i), i represents the currently scheduled taxiing flight and j represents other flights; 8 surface flow characteristics are derived, among which: The first scene flow characteristic D1 is expressed as follows: D1(i)=count(j),t SOBT (j)<t SOBT (i)&t SOBT (i)<t STOT (j)<t STOT (i) The second scene flow characteristic D2 is expressed as follows: D2(i)=count(j),t SOBT (j)<t SOBT (i)&t STOT (j)>t STOT (i) The third scene flow characteristic D3 is expressed as follows: D3(i)=count(j),t SOBT (j)>t SOBT (i)&t SOBT (i)<t STOT (j)<t STOT (i) The fourth scene flow characteristic D4 is expressed as follows: D4(i)=count(j),t SOBT (i)<t SOBT (j)<t STOT (i)&t STOT (j)>t STOT (i) The fifth scene flow characteristic D5 is expressed as follows: D5(i)=count(j),t SLDT (j)<t SOBT (i)&t SOBT (i)<t SIBT (j)<t STOT (i) The sixth scene flow characteristic D6 is expressed as follows: D6(i)=count(j),t SLDT (j)<t SOBT (i)&t SIDT (j)>t STOT (i) The seventh scene flow characteristic D7 is expressed as follows: D7(i)=count(j),t SLDT (j)>t SOBT (i)&t SOBT (i)<t SIBT (j)<t STOT (i) The eighth scene flow characteristic D8 is expressed as follows: D8(i)=count(j),t SOBT (i)<t SLDT (j)<t STOT (i)&t SIBT (j)>t STOT (i) Step S1-2-2, calculate the average sliding time t within the preset time slice d The features are as follows: in, represents the average taxi-out time of the i-th departing flight, t STOT(i) -t SOBT(i) is the planned taxi-out time of the i-th departing flight, and n is the time slice Δ l The number of departing flights that slide out, m represents the mth time slice; Step S1-2-3, calculate the arrival and departure flight ratio α as a feature, as follows: Calculate the ratio of the number of all scheduled departure flights to all scheduled arrival flights during the entire scheduled taxiing process of the current flight i, α i , the method is as follows: Among them, D 1,i It indicates the number of flights that meet the first scene flow characteristics during the planned taxi-out process of the current flight, D 2,i It indicates the number of flights that meet the second field traffic characteristics during the planned taxi-out process of the current flight, D 3,i It indicates the number of flights that meet the third field traffic characteristics during the planned taxi-out process of the current flight, D 4,i It indicates the number of flights that meet the fourth field traffic characteristics during the planned taxi-out process of the current flight, D 5,i It indicates the number of flights that meet the fifth field traffic characteristics during the planned taxi-out process of the current flight, D 6,i It indicates the number of flights that meet the sixth field traffic characteristics during the planned taxi-out process of the current flight, D 7,i It indicates the number of flights that meet the seventh field traffic characteristics during the planned taxi-out process of the current flight, D 8,i Indicates the number of flights that meet the eighth field traffic characteristics during the planned taxi-out process of the current flight; Step S1-2-4, selecting the departure runway allocation R as a feature; Step S1-2-5, dividing the airport operation time period T as a feature, that is, using the K-means method to cluster the features according to the planned departure time and the flight operation time period of the airport; Step S1-2-6, selecting model M as a feature; Step S1-2-7, select airline attribute A as the feature; Step S1-2-8, selecting a runway and a stand combination as a feature; Step S1-2-9, select the block time of the preceding flight as a feature.

6. The method for intelligent prediction of flight departure time based on multi-model machine learning according to claim 5, characterized in that: The standardization process described in step S1-3 is to scale the feature data so that it falls within the range [0,1]. The specific method is as follows: Among them, x norm represents the feature data after standardization, x s Represents the sth feature data, max{} represents the maximum value, and min{} represents the minimum value.

7. The method for intelligent prediction of flight departure time based on multi-model machine learning according to claim 6, characterized in that: The Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient between the calculated features and the flight departure time TOT described in step S1-3.

8. The method for intelligent prediction of flight departure time based on multi-model machine learning according to claim 7, characterized in that: The prediction model of multiple flight departure times described in step S2 includes: Take-off time prediction model based on BP neural network, take-off time prediction model based on multivariate linear regression and take-off time prediction model based on support vector machine.

9. The method for intelligent prediction of flight departure time based on multi-model machine learning according to claim 8, characterized in that: The analysis of the trained multiple flight departure time prediction models in step S3 specifically includes: Step S3-1, design analysis indicators, the analysis indicators include mean square error MSE, mean absolute error MAE, prediction accuracy R 2 and running time; Step S3-2, inputting the test data set into the prediction model of all flight departure times, performing prediction, and calculating the analysis index according to the prediction results; Step S3-3, selecting the model with the best analysis index as the optimal flight departure time prediction model.

10. The method for intelligent prediction of flight departure time based on multi-model machine learning according to claim 9, characterized in that: The key features most relevant to the flight departure time obtained in step S3 specifically include: In the process of prediction in step S3-2, the Pearson correlation coefficient, Spearman correlation coefficient and Kendall correlation coefficient of all features are calculated, and a preset number of features are selected as the final key features in the order of the three types of correlation coefficients from high to low.

Images