Knowledge and data dual-drive-based flight delay prediction method and system
By combining expert knowledge and data-driven methods, an integrated learning model is built, which solves the problems of poor interpretability and insufficient robustness of the model in flight delay prediction, and achieves higher prediction accuracy and credibility.
Patent Information
- Application Number
- CN202411058755.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-02
- Publication Date
- 2024-11-08
AI Technical Summary
The existing flight delay prediction method only uses data models, lacking the use of expert experience and knowledge, which makes it difficult for the model to explain decision-making basis and performance needs to be improved; at the same time, single model prediction has problems such as overfitting, low accuracy and insufficient robustness. .
Using a dual-drive method based on knowledge and data, expert knowledge feature data is generated through the ATMAP algorithm, and an association rule library is built with the FP-Growth association analysis algorithm, and a preamble selection algorithm and SVM metamodel based on the difference degree of entropy E are selected to construct integrated learning. Models, fuse flight and meteorological data for preprocessing and training, and generate integrated learning models.
It improves the accuracy and credibility of flight delay prediction, enhances the interpretability and robustness of the model, effectively improves the prediction performance, and solves the problems of overfitting and low accuracy of a single model.
Smart Images

Figure CN118917485A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of traffic delay prediction, and in particular to a flight delay prediction method and system based on knowledge and data dual drive. Background Art
[0002] As the domestic air transport market gradually recovers and air traffic gradually increases, the problem of flight delays has become increasingly prominent.
[0003] In the early days of studying flight delays, scholars mainly used traditional techniques such as statistical analysis, but the prediction accuracy of these methods was not ideal. With the rapid development of artificial intelligence technology, especially deep learning, deep neural networks have begun to be used to predict flight delays. Compared with traditional methods, this data-driven method has significantly improved accuracy and is therefore widely used. However, this method ignores the use of human experience and knowledge. As a black box model, the data-driven deep learning model is uninterpretable and cannot well explain the decision-making basis of the model, making the output of the model transparent and credible. At the same time, the pure data-driven method may ignore the domain expert knowledge, such as industry rules, historical experience, etc., which makes the model lack sufficient accuracy and robustness in specific scenarios. In addition, in previous studies on flight delay prediction, most of them only used a single model for prediction, which has the disadvantages of being more prone to overfitting, lower accuracy, and insufficient robustness than the integrated learning model.
[0004] After research, the inventors found that the flight delay prediction method in the prior art still has at least the following defects:
[0005] Only using data models for prediction lacks the use of expert experience and knowledge, making it difficult for the model to explain its decision-making basis and its performance needs to be further improved; only using a single model for prediction is more prone to overfitting, has low accuracy and insufficient robustness.
[0006] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention
[0007] The purpose of the present invention is to improve the prediction accuracy and prediction performance of a flight delay prediction model.
[0008] The present invention provides a flight delay prediction method based on knowledge and data dual drive, comprising the steps of:
[0009] S11, obtaining METAR message data, flight data and AWOS weather data;
[0010] S12, generating new feature data into which expert knowledge is injected from the METAR message data according to the ATMAP algorithm; constructing an association rule base and a knowledge model according to the flight data and the AWOS meteorological data through the FP-Growth association analysis algorithm;
[0011] S13, selecting a base model according to a pre-order selection algorithm based on entropy E difference, and constructing a data model;
[0012] S14, selecting an optimal meta-model, and generating an integrated learning model according to the optimal meta-model, the data model and the knowledge model;
[0013] S15, fusing the new feature data, the flight data and the AWOS meteorological data to generate a fused data set;
[0014] S16, preprocessing the fused data set to obtain a new data set including a training set, a validation set and a test set;
[0015] S17. Setting parameters of each base model of the integrated learning model, training each base model with the new data set, and generating a flight delay prediction model based on knowledge and data dual drive.
[0016] Preferably, in an embodiment of the present invention, the construction of the association rule base includes:
[0017] Preprocessing the flight data and the AWOS meteorological data by data merging, data cleaning, feature selection, data fusion and data transformation;
[0018] Discretizing the flight data and the AWOS meteorological data that need to be subjected to association analysis to obtain a data set for association analysis;
[0019] The correlation between meteorological factors and flight delay conditions is obtained through the FP-Growth association analysis algorithm. Rainfall level, runway visibility, low cloud level and crosswind level are selected as the antecedents, and flight delay conditions are selected as the posterior items to obtain several association rules. The several association rules are summarized to generate an association rule library for flight delays.
[0020] Preferably, in an embodiment of the present invention, the construction of the knowledge model includes:
[0021] For flights that need delay prediction, the corresponding flight data and AWOS meteorological data are quantified and matched in the corresponding flight delay association rule library to obtain the confidence level of whether the flight will be delayed or not.
[0022] Preferably, in an embodiment of the present invention, the construction of the data model includes:
[0023] Using the selective integration strategy, a subset of learning models is selected through a preset mechanism to build the final integrated model.
[0024] Preferably, in an embodiment of the present invention, the selecting of the base model according to the pre-order selection algorithm based on the entropy E difference comprises:
[0025] The base model is measured by the entropy E difference metric, including the following formula:
[0026]
[0027] In the formula, L is the number of base models, N is the number of samples, and Z j is a sample, l(Z j ) is the number of samples correctly predicted in any base model.
[0028] Preferably, in an embodiment of the present invention, the optimal meta-model includes:
[0029] The SVM model is used as a meta-model.
[0030] Preferably, in an embodiment of the present invention, the fused data set is preprocessed to obtain a new data set including a training set, a validation set and a test set, including:
[0031] In data cleaning, for inconsistent date formats, use timestamps to make the dates consistent. For null values, choose to delete the corresponding features or delete the rows according to the number of null values.
[0032] In data labeling, for classification problems, the label column is marked as delay status. The delay time is calculated based on the difference between the scheduled take-off and landing time of the flight and the actual take-off and landing time. Among them, flights with a delay time less than the preset time are marked as not delayed, and the remaining flights are marked as delayed;
[0033] In data encoding, different encoding methods are used for continuous and discrete data. Min-Max encoding is used for continuous features, and CatBoost encoding is used for discrete features.
[0034] In data segmentation, the data is randomly divided into training set, validation set and test set according to a certain ratio.
[0035] Preferably, in an embodiment of the present invention, the construction of the flight delay prediction model based on knowledge and data dual drive includes:
[0036] Set the initial parameters of the base model, including the number of blocks, the number of fully connected layers in each block, and the feature dimension of the input data;
[0037] Set the initial parameters of the training method, including: learning rate during training, optimizer during training, and loss function;
[0038] Preferably, in an embodiment of the present invention, training each base model by using the new data set includes:
[0039] Forward propagation and back propagation, where
[0040] The forward propagation is to input the preprocessed data into the network model, extract features through the fully connected layer, attention module and non-linear activation layer, and calculate the predicted value and the true label value through the loss function to obtain the prediction error;
[0041] In the classification task, the cross entropy function is used as the loss function, including the following formula:
[0042]
[0043] In the formula, I is the number of delay levels, y(x i ) is the true label of the i-th category of the x-th sample, is the predicted probability of the i-th category of the x-th sample;
[0044] The back propagation is a process of chain derivation of the loss function obtained in the forward propagation.
[0045] In another aspect of the present invention, a flight delay integrated prediction system based on knowledge and data dual drive is also provided. The prediction system implements the flight delay prediction method based on knowledge and data dual drive. The prediction system includes:
[0046] Data acquisition module, used to obtain METAR message data, flight data and AWOS weather data;
[0047] A knowledge use module, used to generate new feature data injected with expert knowledge from the METAR message data according to the ATMAP algorithm; to construct an association rule base and a knowledge model according to the flight data and the AWOS meteorological data through the FP-Growth association analysis algorithm;
[0048] A fusion driving module, used for fusing the new feature data, the flight data and the AWOS meteorological data to generate a fused data set;
[0049] A data preprocessing module, used to preprocess the fused data set to obtain a new data set including a training set, a validation set and a test set;
[0050] An integrated learning module is used to select a base model based on an entropy E-prequer selection algorithm to construct a data model; select an optimal meta-model, and generate an integrated learning model based on the optimal meta-model, the data model and the knowledge model;
[0051] A flight delay prediction model construction module, used to set the parameters of each base model of the ensemble learning model, train each base model through the new data set, and generate a flight delay prediction model based on knowledge and data dual drive;
[0052] The result acquisition module is used to generate a flight delay prediction result through the flight delay prediction model.
[0053] Compared with the prior art, the present invention has the following beneficial effects:
[0054] In the present invention, METAR message data, flight data and AWOS meteorological data are firstly obtained; new feature data injected with expert knowledge is generated from the METAR message data according to the ATMAP algorithm; an association rule base is constructed according to the flight data and the AWOS meteorological data through the FP-Growth association analysis algorithm, and a knowledge model is constructed; a base model is selected according to a pre-order selection algorithm based on entropy E difference, and a data model is constructed; an optimal meta-model is selected, and an integrated learning model is generated according to the optimal meta-model, the data model and the knowledge model; the new feature data, flight data and AWOS meteorological data are fused to generate a fused data set; the fused data set is preprocessed to obtain a new data set including a training set, a validation set and a test set; the parameters of each base model of the integrated learning model are set, and each base model is trained through the new data set to generate a knowledge and data-based integrated learning model. A dual-driven flight delay prediction model; the present invention utilizes the ATMAP algorithm containing expert knowledge to process METAR message data, generates new features, and fuses them with flight data and AWOS meteorological data to construct a new data set, thereby realizing knowledge injection; the FP-Growth association analysis algorithm is used to mine hidden knowledge, and an association analysis database is constructed for flight delay prediction, which is used as a knowledge model to realize dual-drive of knowledge and data, thereby improving the accuracy and credibility of flight delay prediction; by using an integrated learning method, the data model and the knowledge model are integrated to achieve complementary advantages between models, and effectively improve the prediction performance of the flight delay prediction model, thereby solving the problem in the prior art that only the data model is used for prediction, which makes it difficult for the model to explain its decision basis and the performance needs to be improved; only a single model is used for prediction, which is more prone to overfitting, has a low accuracy rate and is not robust enough.
[0055] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention and implement it according to the contents of the specification, and to make the above and other purposes, technical features and advantages of the present invention easier to understand, one or more preferred embodiments are listed below and described in detail with the accompanying drawings as follows. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] In order to more clearly illustrate the technical solution of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0057] Figure 1 It is a step diagram of the flight delay prediction method based on knowledge and data dual drive described in the present invention;
[0058] Figure 2 is a flow chart of the flight delay prediction method based on knowledge and data dual drive described in the present invention;
[0059] Figure 3 It is a process diagram of data preprocessing when building an association rule base in the present invention;
[0060] Figure 4 is a flow chart of the pre-order selection algorithm based on entropy E difference described in the present invention;
[0061] Figure 5 It is the flow chart of the ATMAP algorithm described in the present invention;
[0062] Figure 6 It is a data processing diagram of the model training described in the present invention;
[0063] Figure 7 It is a schematic diagram of the structure of the knowledge model, the data-driven model and the integrated learning module described in the present invention;
[0064] Figure 8 is a confusion matrix diagram of the flight delay prediction result of the single data model described in the present invention;
[0065] Fig. 9 It is a confusion matrix diagram of the prediction results of the flight delay prediction model based on knowledge and data dual drive described in the present invention;
[0066] Fig.10 It is a schematic diagram of the structure of the flight delay prediction system based on knowledge and data dual drive described in the present invention; DETAILED DESCRIPTION
[0067] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation modes.
[0068] In this document, the terms “first”, “second”, etc. are used to distinguish two different elements or parts, but are not used to limit specific positions or relative relationships.
[0069] Embodiment 1
[0070] In order to improve the prediction performance of the flight delay prediction model, such as Figure 1 and Figure 2 As shown, in an embodiment of the present invention, a flight delay prediction method based on knowledge and data dual drive is provided, comprising the steps of:
[0071] S11, obtaining METAR message data, flight data and AWOS weather data;
[0072] The inventive ideas of the present invention include: the flight delay prediction model based on machine learning or deep learning in the prior art only uses a single model for prediction, which has the problems of easy overfitting, low accuracy and insufficient robustness. Therefore, the present invention proposes a flight delay prediction method based on knowledge and data dual drive combined with ensemble learning. Through ensemble learning, the base models can complement each other's advantages and effectively improve the model prediction performance.
[0073] In the embodiment of the present invention, firstly, METAR message data, flight data provided by the authority and AWOS meteorological data are obtained.
[0074] S12, generating new feature data into which expert knowledge is injected from the METAR message data according to the ATMAP algorithm; constructing an association rule base and a knowledge model according to the flight data and the AWOS meteorological data through the FP-Growth association analysis algorithm;
[0075] In the embodiment of the present invention, the Air Traffic Management Airport Airside Performance Weather Assessment Algorithm (ATMAP algorithm) is jointly developed by experts in multiple fields such as meteorology, air traffic control, airport operators, airlines, and mathematicians from the European Union. It is mainly used to (1) uniformly measure the weather conditions of European airports; (2) objectively and comprehensively measure the intensity and duration of weather phenomena that may make ANS and airport airside operations more difficult; and (3) divide operating days into two categories for higher-dimensional analysis. This algorithm helps air traffic managers make decisions by analyzing the impact of weather conditions on airport operating efficiency, thereby optimizing flight scheduling and ground operations and improving the overall operating efficiency of the airport.
[0076] The ATMAP algorithm quantifies airport meteorological data into five categories: visibility, wind, precipitation, icing conditions, and hazardous weather. Each category has a severity level and provides coefficients to reflect the different severity levels of different meteorological conditions. For each assigned severity level code, a value is given, called the summary score coefficient, which enables the description of the nonlinear performance of several weather events.
[0077] In an embodiment of the present invention, the visibility and cloud base height scores in the ATMAP score are shown in Table 1 below. The ANS performance (runway throughput, delay) of the airport changes continuously with the visibility level. When visibility is low, the airport performance depends on the instrument landing system or other approach assistance systems. The reduction in visibility affects both take-off and landing flights. Under the same meteorological conditions, the degradation of the airport's ANS performance when visibility is poor will be different, depending on the auxiliary systems used (ILS, MLS, satellite-aided systems, multilateral ground motion radar, etc.) and the approved air traffic control procedures used. In the visibility and cloud base height scoring example, the METAR message provides a visibility value (VIS), a runway visual range value (RVR), and three sets of cloud values (CLD_base, CLD_type, CLD_cover). Code1, Code2, Code3, and Code4 correspond to the limit values under non-precision approach runway conditions, Category I precision approach runway conditions, Category II precision approach runway conditions, and Category III precision approach runway conditions, respectively. The scores for each category are set based on the flight interval, operation complexity, and error probability under different landing procedures.
[0078] Table 1:
[0079]
[0080] like Figure 5 As shown, in an embodiment of the present invention, after using the pandas library in Python to read the CSV file containing the METAR message, the wind speed, visibility and cloud base height, precipitation, icing state, and dangerous weather phenomenon information in the message are parsed and extracted, and then the score is calculated according to different weather elements and different severity levels, and the result is stored in a new CSV file.
[0081] The construction of the association rule base includes:
[0082] The flight data and AWOS meteorological data are preprocessed by data merging, data cleaning, feature selection, data fusion and data transformation. The specific preprocessing process is as follows: Figure 3 As shown;
[0083] Discretize the flight data and AWOS meteorological data that need to be analyzed for association to obtain the data set for association analysis;
[0084] The association between meteorological factors and flight delays is obtained through the FP-Growth association analysis algorithm. Rainfall level, runway visual range, low cloud level and crosswind level are selected as the antecedents, and flight delays are selected as the posterior terms. Several association rules are obtained, and several association rules are summarized to generate an association rule base for flight delays. If a total of 212 association rules are obtained, the 212 association rules are summarized to form an association rule base for flight delays.
[0085] The construction of knowledge model includes:
[0086] For flights that need delay prediction, the corresponding flight data and AWOS meteorological data are quantified and matched in the corresponding flight delay association rule library to obtain the confidence level of whether the flight will be delayed or not, that is, to mine hidden knowledge from historical data.
[0087] S13, selecting a base model according to a pre-order selection algorithm based on entropy E difference, and constructing a data model;
[0088] In the embodiment of the present invention, the base model is selected according to the standard of "good but different". The selection of the base model in ensemble learning is one of the key factors affecting the overall performance of the model. The selection of the base model mainly considers diversity and accuracy, that is, to achieve "good but different". Diversity means that the errors of the base model on the training data should be as different as possible. This can be achieved by using heterogeneous learning models of different algorithm types, or by using subsets with the same algorithm but different model parameters or different training data. A combination of learning models with high diversity is more likely to complement each other in the ensemble, thereby improving the overall performance.
[0089] Furthermore, in the embodiment of the present invention, the construction of the data model includes:
[0090] Using the selective integration strategy, a subset of learning models is selected through a preset mechanism to build the final integrated model. Using the selective integration strategy, it does not simply combine all available base models, but selects a subset of learning models through a certain mechanism to build the final integrated model. The core idea of this method is that "a few elites" are better than "most mediocre", that is, in some cases, a carefully selected subset of learning models may be more effective than the combination of all learning models. The differences between different models are the key factors that need to be considered in selective integration.
[0091] To compare the differences between different base models, the base models can be measured by entropy E difference measurement to compare the classification effects of different base models, including the following formula:
[0092]
[0093] In the formula, L is the number of base models, N is the number of samples, and Z j is a sample, l(Z j ) is the number of samples correctly predicted in any base model. For two base models, l(Z j ) is 0 or 2, indicating that the sample is correctly or incorrectly classified by both base models. j ) and Ll(Z j ) is 0; if l(Z j ) is 1, then the minimum value is still 1. The summed value reflects the difference between the models. The value of E is between 0 and 1. The larger the value, the greater the difference between the two base models.
[0094] Preferably, in the embodiment of the present invention, in order to meet the selection criteria of "good and different" for selecting the base model, it is necessary to comprehensively consider the accuracy and diversity of the model, and for this purpose, the evaluation index C shown in the following formula is set for the pre-order selection algorithm.
[0095] C=(1-r)×Acc+r×E
[0096] Among them, Acc is the prediction accuracy of the base model, r is a constant ranging from 0 to 1, which is used as a weight to balance accuracy and variability, and E is the degree of variability. Since there is an order of magnitude difference between the degree of variability, the numerical value and the accuracy, the degree of variability is processed using a linear transformation.
[0097] In specific application scenarios, NR-DenseNet, LightGBM, XGBoost, RF, and DT models can be selected as base models in the Stacking-based flight delay prediction ensemble learning model. The prediction accuracy of each base model can be calculated, and the difference between the base model and the NR-DenseNet model can be calculated. Finally, the C value of each base model is obtained, and the obtained C values are arranged in descending order for use in the pre-order selection algorithm. The results are shown in Table 2 below.
[0098] Table 2:
[0099] Base model name Difference Accuracy C Value order LightGBM 0.4922 0.9603 0.7731 4 RF 0.5211 0.9517 0.7795 3 XGBoost 0.5174 0.9556 0.7803 2 DT 0.5279 0.9542 0.7837 1
[0100] In the embodiment of the present invention, the overall process of the preamble selection algorithm is as follows: Figure 4As shown in the figure, the order of adding the base models is determined according to the C value of each model, and the base models with the highest ranking are added to the ensemble learning model in turn. If the accuracy of the ensemble learning model is improved, it is added to the set S that constitutes the ensemble learning model, otherwise it is removed from the base model set T. This operation is repeated until the base model set T is empty. At this time, the model in the set S is the best combined base model selected by the pre-order selection algorithm. In addition, the dual drive of knowledge and data helps to improve the performance of the model and increase the interpretability. Therefore, the knowledge model constructed by the association analysis algorithm is added to the ensemble learning model as a special base model. For the classification prediction problem of flight delays, after testing, the best base model combination finally selected is NR-DenseNet, LightGBM, XGBoost, DT, and association analysis knowledge model.
[0101] S14, selecting an optimal meta-model, and generating an integrated learning model according to the optimal meta-model, the data model and the knowledge model;
[0102] In the embodiment of the present invention, the core idea of ensemble learning is to improve the overall prediction performance by combining the prediction results of multiple base models. Compared with a single model that directly outputs the results, the ensemble learning model processes the results according to a certain strategy based on different tasks and then makes a final judgment, which often achieves better results than a single model. In order to formally understand this, the following is an explanation through some simplified assumptions and mathematical reasoning.
[0103] Assume that for a simple binary classification problem, the result can only be -1 or 1. For a data sample, it can be expressed as {(x i ,y i )}, let the operation function be f, select M base models with error rate ε, where the i-th base model can be expressed as h i , the prediction result of the integrated model is based on the prediction result of the majority model. Assume that a simple majority voting mechanism is adopted here. If more than half of the models predict the result correctly, then the integrated learning model will also give a correct prediction, as shown in the following formula,
[0104]
[0105] In the formula, H(x) is the ensemble learning model, and T is the number of base models used. x represents a data sample, and the prediction result is calculated by the ensemble learning model H through the majority voting mechanism.
[0106] If the base models are independent of each other, the error rate of the integrated model composed of T base models can be expressed by the following formula:
[0107]
[0108] Let Q be an independent event with probability p, and G(n) be the number of occurrences after n tests. Then the probability of the event occurring at most k times is:
[0109]
[0110] According to the Hoeffding inequality, There exists k = n(p-δ), so:
[0111]
[0112] Among them, δ is the threshold value of deviation from the expected value, and e is a natural constant, which is approximately equal to 2.71828.
[0113] When p = 1-ε, n = T, hour:
[0114]
[0115] It can be seen that when the error rate is , the error rate of the integrated model will decrease with the increase of the number of base models T. The premise for this conclusion is that the base models have only low correlation and are independent of each other. Therefore, we can conclude that under certain conditions, the integrated learning algorithm can achieve better performance than a single algorithm.
[0116] Preferably, in an embodiment of the present invention, in constructing an integrated learning model, the selection of a metamodel is crucial because it directly affects the performance of the entire model. The task of the metamodel is to learn how to best combine the prediction results of the underlying base model to improve the accuracy and robustness of the overall prediction. The selection of an appropriate metamodel requires consideration of complementarity with the underlying model, the complexity and flexibility of the model, and the efficiency of training and prediction. Since the base model of the first layer of the model often uses complex nonlinear transformations and the output results are not complex, the metamodel often selects a simpler classifier. Based on this, four simple models, AdaBoost, LR, SVM, and NB, are selected as candidate metamodels. After comparative experimental testing, the SVM model with the best performance is selected as the metamodel.
[0117] S15, fusing the new feature data, the flight data and the AWOS meteorological data to generate a fused data set;
[0118] In the embodiment of the present invention, the airport's take-off and landing flight data, AWOS meteorological data, and new feature data obtained after knowledge injection are matched and fused according to time to generate a fused data set; data fusion is a commonly used method in feature engineering, and the purpose of enriching the feature items of the data set is achieved by merging different data sets. With the time key value as the key primary key, the flight data is matched and fused with the corresponding meteorological data. At the same time, the new features obtained after processing by the knowledge model are fused with the flight data set according to time; in order to fully consider the important impact of meteorological factors on flight delays, the flight data and meteorological data are fused in 10 minutes during the fusion process.
[0119] S16, preprocessing the fused data set to obtain a new data set including a training set, a validation set and a test set;
[0120] In the embodiment of the present invention, the fused data set is preprocessed, and the specific preprocessing is as follows: Figure 6 As shown, a new data set including training set, validation set and test set is obtained, including:
[0121] In data cleaning, for inconsistent date formats, use timestamps to make the dates consistent. For null values, select to delete the corresponding features or delete the rows according to the number of null values. Specifically, in the embodiment of the present invention, in data cleaning, first use the pandas library in python to query the abnormalities in the data. After inspection, the abnormal values mainly include inconsistent date formats and redundant null values. For the problem of inconsistent date formats, use timestamps to make the dates consistent. For the problem of null values, select to delete the corresponding features or delete the rows according to the number of null values.
[0122] In data labeling, for classification problems, the label column is marked as delay status, and the delay time is calculated based on the difference between the scheduled take-off and landing time of the flight and the actual take-off and landing time. Among them, flights with a delay time less than the preset time are marked as not delayed, and the remaining flights are marked as delayed; due to the use of supervised deep learning algorithms, the input of the model should include a label column in addition to the feature column. For classification problems, the label column should be the delay status, and the delay time is calculated based on the difference between the scheduled take-off and landing time of the flight and the actual take-off and landing time. Among them, flights with a delay time less than 15 minutes are marked as not delayed, and the remaining flights are marked as delayed.
[0123] In data encoding, different encoding methods are used for continuous and discrete data. Min-Max encoding is used for continuous features, and CatBoost encoding is used for discrete features. In data encoding, different encoding methods are used for continuous and discrete data. For continuous features, if the data magnitude difference of the input model is large, features with small values may be ignored, which has a negative impact on the feature extraction ability and prediction effect of the deep learning model. Using Min-Max encoding can map the input values within the range of [0,1], which can effectively eliminate the impact of numerical differences. For discrete features, CatBoost encoding can automatically process category features without manual label encoding or one-hot encoding. It is very suitable for situations with a large number of high-cardinality features, and can also reduce overfitting and improve model performance.
[0124] In data splitting, the data is randomly divided into training set, validation set and test set in a certain proportion.
[0125] The data is randomly divided into training set, validation set and test set in a ratio of 6:2:2. While ensuring the amount of model training data, the validation set and test set can be reasonably used to evaluate and improve model performance.
[0126] S17. Setting parameters of each base model of the integrated learning model, training each base model with the new data set, and generating a flight delay prediction model based on knowledge and data dual drive.
[0127] In an embodiment of the present invention, the construction of a flight delay prediction model based on knowledge and data dual drive includes:
[0128] Set the initial parameters of the base model, including the number of blocks, the number of fully connected layers in each block, and the feature dimension of the input data;
[0129] Set the initial parameters of the training method, including: learning rate during training, optimizer during training, and loss function;
[0130] Furthermore, in the embodiment of the present invention, each base model is trained using a new data set, including:
[0131] Forward propagation and back propagation, where
[0132] Forward propagation is to input the preprocessed data into the network model, extract features through the fully connected layer, attention module and nonlinear activation layer, and calculate the predicted value and the true label value through the loss function to obtain the prediction error;
[0133] In the classification task, the cross entropy function is used as the loss function, including the following formula:
[0134]
[0135] In the formula, I is the number of delay levels, which is 2. i ) is the true label of the i-th category of the x-th sample, is the predicted probability of the i-th category of the x-th sample;
[0136] Backpropagation is the process of chaining the derivation of the loss function obtained in forward propagation.
[0137] Specifically, the set base model parameters, such as the initial parameters of the NR-DenseNet network: the number of blocks, the number of fully connected layers of each block, and the feature dimension of the input data; the initial parameters of the training method include: the learning rate during training, the optimizer during training, the loss function, etc.; other base model hyperparameter settings are shown in Table 3 below:
[0138] Table 3:
[0139]
[0140]
[0141] Taking the NR-DenseNet model as an example, for model training, forward propagation is first performed to input the preprocessed data into the network model, then the data features are extracted, and the errors of the training set and the validation set are calculated through the loss function. The performance of the network is evaluated based on the error situation to determine whether it has converged or whether underfitting or overfitting has occurred. If the network converges, the training parameters of the model are saved for subsequent predictions; if the error between the training set and the validation set differs too much, it means that the current training hyperparameter settings may be unreasonable, and the hyperparameters need to be reset before training. Next is back propagation, which calculates the gradient value of the network by performing a chain derivation process on each network layer from back to front of the entire network, thereby updating the network parameters and accelerating the convergence of the training process.
[0142] Further, in the embodiment of the present invention, the fusion method of knowledge-driven and data-driven is diverse and can be mainly divided into four modes: parallel mode, guide mode, serial mode and feedback mode. In the parallel mode, knowledge-driven and data-driven operate independently and produce results, and then these results are combined to form the final prediction or decision. This method improves the overall performance by combining the advantages of the two methods. In the serial mode, the input data is first preprocessed or optimized using a knowledge-driven model, and then the processed data is sent to the data-driven model, which can improve the accuracy and effect of the data-driven model. The guide mode is based on an existing, interpretable mechanism model to guide or influence the construction process of the data-driven model, with the aim of creating a more reasonable and effective model so that the prediction results of the model have better logic and interpretability. In the feedback mode, a data-driven method is used to correct or replace certain parts or parameters in the mechanism model. In this way, the model can be continuously optimized based on the feedback of actual data to improve the adaptability and accuracy of the model.
[0143] Specifically, Figure 7 As shown in the figure, the knowledge model, data-driven model and integrated learning module use both serial and parallel modes. In the serial mode, the ATMAP algorithm incorporating expert knowledge is used to process the METER message information, generate new features with a strong correlation with flight delays and send them to the data model to improve the prediction accuracy and effect of the data model. In the parallel mode, the knowledge model built based on the FP-Growth association analysis algorithm and the data model built based on machine learning give delay predictions and probabilities respectively, and these results are integrated through learning to form the final prediction.
[0144] The confusion matrix diagram of the flight delay prediction results of the single data model is as follows: Figure 8 As shown in the figure, the output confusion matrix of the NR-DenseNet model with the best performance among the single data models is shown; the confusion matrix diagram of the prediction results of the flight delay prediction model based on knowledge and data dual drive is shown in the figure below. Fig. 9 As shown, the confusion matrix output by the knowledge-based integrated learning model is used. Each column in the confusion matrix represents the predicted flight delay status, and the total number of each column represents the number of flights predicted to be in this delay category. Each row in the matrix represents the actual flight delay category of the data, and the total number of each row represents the number of flights that are actually in this delay category. The numbers on the diagonal represent the number of correct predictions for each delay category. By comparison, it can be concluded that the integrated learning model using knowledge in the present invention reduces the number of prediction errors by 34.1% on the test set compared to the single data model, and the model accuracy is significantly improved.
[0145] It can be seen from the above embodiments that the flight delay prediction model of the present invention provides a more accurate and reliable flight delay prediction service.
[0146] In summary, the embodiment of the present invention first obtains METAR message data, flight data and AWOS meteorological data; generates new feature data injected with expert knowledge from the METAR message data according to the ATMAP algorithm; constructs an association rule base and a knowledge model according to the flight data and the AWOS meteorological data through the FP-Growth association analysis algorithm; selects a base model according to the pre-order selection algorithm based on the entropy E difference, and constructs a data model; selects the best meta-model, and generates an integrated learning model according to the best meta-model, the data model and the knowledge model; fuses the new feature data, flight data and AWOS meteorological data to generate a fused data set; pre-processes the fused data set to obtain a new data set including a training set, a validation set and a test set; sets the parameters of each base model of the integrated learning model, trains each base model with the new data set, and generates a base model. The invention relates to a knowledge and data dual-driven flight delay prediction model; the invention utilizes the ATMAP algorithm containing expert knowledge to process METAR message data, generates new features, and fuses them with flight data and AWOS meteorological data to construct a new data set, thereby realizing knowledge injection; the FP-Growth association analysis algorithm is used to mine hidden knowledge, and an association analysis database is constructed for flight delay prediction, which is used as a knowledge model to realize knowledge and data dual-drive, thereby improving the accuracy and credibility of flight delay prediction; by using an integrated learning method, the data model and the knowledge model are integrated to realize complementary advantages between models, and effectively improve the performance of the flight delay prediction model, thereby solving the problems in the prior art of only using the data model for prediction, making it difficult for the model to explain its decision basis and requiring performance improvement; only using a single model for prediction has the problems of being more prone to overfitting, low accuracy and insufficient robustness.
[0147] Embodiment 2
[0148] Corresponding to the method embodiment, in another aspect of the embodiment of the present invention, a flight delay integrated prediction system based on knowledge and data dual drive is also provided. Fig.10 The structure diagram of the flight delay integrated prediction system based on knowledge and data dual drive provided by the embodiment of the present invention is shown. Specifically, the flight delay integrated prediction system based on knowledge and data dual drive in the embodiment of the present invention includes:
[0149] Data acquisition module 01, used to obtain METAR message data, flight data and AWOS meteorological data;
[0150] In the embodiment of the present invention, firstly, METAR message data, flight data provided by the authority and AWOS meteorological data are obtained.
[0151] The knowledge use module 02 is used to generate new feature data injected with expert knowledge from the METAR message data according to the ATMAP algorithm; to construct an association rule base and a knowledge model according to the flight data and the AWOS meteorological data through the FP-Growth association analysis algorithm;
[0152] In the embodiment of the present invention, the ATMAP algorithm quantifies airport weather data into five categories: visibility, wind, precipitation, icing conditions, and hazardous weather. Each category has a severity level and provides coefficients to reflect the different severity levels under different meteorological conditions. For each specified severity level code, a value is given, called the summary score coefficient. This enables the description of the nonlinear performance of several weather events.
[0153] The construction of the association rule base includes:
[0154] The flight data and AWOS meteorological data are preprocessed by data merging, data cleaning, feature selection, data fusion and data transformation. The specific preprocessing is as follows: Figure 3 As shown;
[0155] Discretize the flight data and AWOS meteorological data that need to be analyzed for association to obtain the data set for association analysis;
[0156] The association between meteorological factors and flight delays is obtained through the FP-Growth association analysis algorithm. Rainfall level, runway visual range, low cloud level and crosswind level are selected as the antecedents, and flight delays are selected as the posterior terms. Several association rules are obtained, and several association rules are summarized to generate an association rule base for flight delays. If a total of 212 association rules are obtained, the 212 association rules are summarized to form an association rule base for flight delays.
[0157] The construction of knowledge model includes:
[0158] For flights that need delay prediction, the corresponding flight data and AWOS meteorological data are quantified and matched in the corresponding flight delay association rule library to obtain the confidence level of whether the flight will be delayed or not, that is, to mine hidden knowledge from historical data.
[0159] The data model construction module 03 is used to select a base model according to a pre-order selection algorithm based on entropy E difference and construct a data model;
[0160] In the embodiment of the present invention, the construction of the data model includes:
[0161] Using the selective integration strategy, a subset of learning models is selected through a preset mechanism to build the final integrated model. Using the selective integration strategy, instead of simply combining all available base models, a subset of learning models is selected through a certain mechanism to build the final integrated model. The core idea of this method is that "a few elites" are better than "most mediocre", that is, in some cases, a carefully selected subset of learning models may be more effective than the combination of all learning models. The differences between different models are the key factors that need to be considered in selective integration.
[0162] To compare the differences between different base models, the base models can be measured by entropy E difference measurement to compare the classification effects of different base models, including the following formula:
[0163]
[0164] In the formula, L is the number of base models, N is the number of samples, and Z j is a sample, l(Z j ) is the number of samples correctly predicted in any base model. For two base models, l(Z j ) is 0 or 2, indicating that the sample is correctly or incorrectly classified by both base models. j ) and Ll(Z j ) is 0; if l(Z j ) is 1, then the minimum value is still 1. The summed value reflects the difference between the models. The value of E is between 0 and 1. The larger the value, the greater the difference between the two base models.
[0165] Preferably, in the embodiment of the present invention, in order to meet the selection criteria of "good and different" for selecting the base model, it is necessary to comprehensively consider the accuracy and diversity of the model, and for this purpose, the evaluation index C shown in the following formula is set for the pre-order selection algorithm.
[0166] C=(1-r)×Acc+r×E
[0167] Among them, Acc is the prediction accuracy of the base model, r is a constant ranging from 0 to 1, which is used as a weight to balance accuracy and variability, and E is the degree of variability. Since there is an order of magnitude difference between the degree of variability, the numerical value and the accuracy, the degree of variability is processed using a linear transformation.
[0168] A fusion driving module 04, configured to fuse the new feature data, the flight data and the AWOS meteorological data to form a fusion data set;
[0169] In the embodiment of the present invention, the airport's take-off and landing flight data, AWOS meteorological data, and new feature data obtained after knowledge injection are matched and fused according to time to generate a fused data set; data fusion is a commonly used method in feature engineering, and the purpose of enriching the feature items of the data set is achieved by merging different data sets. With the time key value as the key primary key, the flight data is matched and fused with the corresponding meteorological data. At the same time, the new features obtained after processing by the knowledge model are fused with the flight data set according to time; in order to fully consider the important impact of meteorological factors on flight delays, the flight data and meteorological data are fused in 10 minutes during the fusion process.
[0170] A data preprocessing module 05 is used to preprocess the fused data set to obtain a new data set including a training set, a validation set and a test set;
[0171] In the embodiment of the present invention, the fused data set is preprocessed. The specific preprocessing process is as follows: Figure 6 As shown, a new data set including training set, validation set and test set is obtained, including:
[0172] In data cleaning, for inconsistent date formats, use timestamps to make the dates consistent. For null values, choose to delete the corresponding features or delete the rows according to the number of null values.
[0173] In data labeling, for classification problems, the label column is marked as delay status. The delay time is calculated based on the difference between the scheduled take-off and landing time of the flight and the actual take-off and landing time. Among them, flights with a delay time less than the preset time are marked as not delayed, and the remaining flights are marked as delayed;
[0174] In data encoding, different encoding methods are used for continuous and discrete data. Min-Max encoding is used for continuous features, and CatBoost encoding is used for discrete features.
[0175] In data splitting, the data is randomly divided into training set, validation set and test set in a certain proportion.
[0176] The data is randomly divided into training set, validation set and test set in a ratio of 6:2:2. While ensuring the amount of model training data, the validation set and test set can be reasonably used to evaluate and improve model performance.
[0177] An integrated learning module 06, configured to select an optimal meta-model and generate an integrated learning model according to the optimal meta-model, the data model and the knowledge model;
[0178] In the embodiment of the present invention, the core idea of ensemble learning is to improve the overall prediction performance by combining the prediction results of multiple learning base models. Compared with a single model that directly outputs the results, the ensemble learning model will process the results according to a certain strategy based on the different tasks, and then make a final judgment, which can often achieve better results than a single model.
[0179] Assume that for a simple binary classification problem, the result can only be -1 or 1. For a data sample, it can be expressed as {(x i ,y i )}, let the operation function be f, select M base models with error rate ε, where the i-th base model can be expressed as h i , the prediction result of the integrated model is based on the prediction result of the majority model. Assume that a simple majority voting mechanism is adopted here. If more than half of the models predict the result correctly, then the integrated learning model will also give a correct prediction, as shown in the following formula,
[0180]
[0181] In the formula, H(x) is the ensemble learning model, and T is the number of base models used. x represents a data sample, and the prediction result is calculated by the ensemble learning model H through the majority voting mechanism.
[0182] If the base models are independent of each other, the error rate of the integrated model composed of T base models can be expressed by the following formula:
[0183]
[0184] Let Q be an independent event with probability p, and G(n) be the number of occurrences after n tests. Then the probability of the event occurring at most k times is:
[0185]
[0186] According to the Hoeffding inequality, There exists k = n(p-δ), so:
[0187]
[0188] Among them, δ is the threshold value of deviation from the expected value, and e is a natural constant, which is approximately equal to 2.71828.
[0189] When p = 1-ε, n = T, hour:
[0190]
[0191] It can be seen that when the error rate is , the error rate of the integrated model will decrease with the increase of the number of base models T. The premise for this conclusion is that the base models have only low correlation and are independent of each other. Therefore, we can conclude that under certain conditions, the integrated learning algorithm can achieve better performance than a single algorithm.
[0192] Preferably, in an embodiment of the present invention, in constructing an integrated learning model, the selection of a metamodel is crucial because it directly affects the performance of the entire model. The task of the metamodel is to learn how to best combine the prediction results of the underlying base model to improve the accuracy and robustness of the overall prediction. The selection of an appropriate metamodel requires consideration of complementarity with the underlying model, the complexity and flexibility of the model, and the efficiency of training and prediction. Since the base model of the first layer of the model often uses complex nonlinear transformations and the output results are not complex, the metamodel often selects a simpler classifier. Based on this, four simple models, AdaBoost, LR, SVM, and NB, are selected as candidate metamodels. After comparative experimental testing, the SVM model with the best performance is selected as the metamodel.
[0193] The flight delay prediction model building module 07 is used to set the parameters of each base model of the ensemble learning model, train each base model through the new data set, and generate a flight delay prediction model based on knowledge and data dual drive;
[0194] In an embodiment of the present invention, the construction of a flight delay prediction model based on knowledge and data dual drive includes:
[0195] Set the initial parameters of the base model, including the number of blocks, the number of fully connected layers in each block, and the feature dimension of the input data;
[0196] Set the initial parameters of the training method, including: learning rate during training, optimizer during training, and loss function;
[0197] Furthermore, in the embodiment of the present invention, each base model is trained using a new data set, including:
[0198] Forward propagation and back propagation, where
[0199] Forward propagation is to input the preprocessed data into the network model, extract features through the fully connected layer, attention module and nonlinear activation layer, and calculate the predicted value and the true label value through the loss function to obtain the prediction error;
[0200] In the classification task, the cross entropy function is used as the loss function, including the following formula:
[0201]
[0202] In the formula, I is the number of delay levels, and its value is I=2. i ) is the true label of the i-th category of the x-th sample, is the predicted probability of the i-th category of the x-th sample;
[0203] Back propagation is the process of chain-deriving the loss function obtained in the forward propagation.
[0204] Further, in the embodiment of the present invention, the fusion method of knowledge-driven and data-driven is diverse and can be mainly divided into four modes: parallel mode, guide mode, serial mode and feedback mode. In the parallel mode, knowledge-driven and data-driven operate independently and produce results, and then these results are combined to form the final prediction or decision. This method improves the overall performance by combining the advantages of the two methods. In the serial mode, the input data is first preprocessed or optimized using the knowledge-driven model, and then the processed data is sent to the data-driven model, which can improve the accuracy and effect of the data-driven model. The guide mode is based on the existing, interpretable mechanism model to guide or influence the construction process of the data-driven model, with the aim of creating a more reasonable and effective model so that the prediction results of the model have better logic and interpretability. In the feedback mode, a data-driven method is used to modify or replace certain parts or parameters in the mechanism model, so that the model can be continuously optimized according to the feedback of actual data to improve the adaptability and accuracy of the model.
[0205] The result acquisition module 08 is used to generate a flight delay prediction result through the flight delay prediction model.
[0206] In an embodiment of the present invention, a flight delay prediction result is generated by a flight delay prediction model. The flight delay prediction model in the present invention has a high prediction performance for flight delay prediction. The binary delay prediction at Ningbo Lishe International Airport reaches an accuracy of 97.97%, the recall rate reaches 95.68%, and the F1 score reaches 0.953. Compared with the single data model in the base model, the accuracy is increased by 2.8% at most and 1.14% at least, the recall rate is increased by 10.35% at most and 4.13% at least, and the F1 score is increased by 8.37% at most and 2.73% at least.
[0207] In summary, the embodiment of the present invention first obtains METAR message data, flight data and AWOS meteorological data; generates new feature data injected with expert knowledge from the METAR message data according to the ATMAP algorithm; constructs an association rule base and a knowledge model according to the flight data and the AWOS meteorological data through the FP-Growth association analysis algorithm; selects a base model according to the pre-order selection algorithm based on the entropy E difference, and constructs a data model; selects the best meta-model, and generates an integrated learning model according to the best meta-model, the data model and the knowledge model; fuses the new feature data, flight data and AWOS meteorological data to generate a fused data set; pre-processes the fused data set to obtain a new data set including a training set, a validation set and a test set; sets the parameters of each base model of the integrated learning model, trains each base model through the new data set, and generates a knowledge model based on the knowledge set. The invention proposes a flight delay prediction model driven by knowledge and data. The invention uses the ATMAP algorithm containing expert knowledge to process METAR message data, generates new features, and fuses them with flight data and AWOS meteorological data to construct a new data set, thereby realizing knowledge injection. The FP-Growth association analysis algorithm is used to mine hidden knowledge, and an association analysis database is constructed for flight delay prediction. As a knowledge model, the knowledge and data dual drive is realized, thereby improving the accuracy and credibility of flight delay prediction. The integrated learning method is used to integrate the data model and the knowledge model, thereby realizing the complementary advantages between the models and effectively improving the prediction performance of the flight delay prediction model. Thus, the problem that only the data model is used for prediction in the prior art, which makes it difficult for the model to explain its decision basis and the performance needs to be improved is solved. The problem that only a single model is used for prediction is more prone to overfitting, has a low accuracy rate and is not robust enough is solved.
Claims
1. A flight delay prediction method based on knowledge and data dual drive, characterized in that: Includes steps: S11, obtaining METAR message data, flight data and AWOS weather data; S12, generating new feature data into which expert knowledge is injected from the METAR message data according to the ATMAP algorithm; constructing an association rule base and a knowledge model according to the flight data and the AWOS meteorological data through the FP-Growth association analysis algorithm; S13, selecting a base model according to a pre-order selection algorithm based on entropy E difference, and constructing a data model; S14, selecting an optimal meta-model, and generating an integrated learning model according to the optimal meta-model, the data model and the knowledge model; S15, fusing the new feature data, the flight data and the AWOS meteorological data to generate a fused data set; S16, preprocessing the fused data set to obtain a new data set including a training set, a validation set and a test set; S17. Setting parameters of each base model of the integrated learning model, training each base model with the new data set, and generating a flight delay prediction model based on knowledge and data dual drive.
2. The flight delay prediction method based on knowledge and data dual drive according to claim 1 is characterized in that: The construction of the association rule base includes: Preprocessing the flight data and the AWOS meteorological data by data merging, data cleaning, feature selection, data fusion and data transformation; Discretizing the flight data and the AWOS meteorological data that need to be subjected to association analysis to obtain a data set for association analysis; The correlation between meteorological factors and flight delay conditions is obtained through the FP-Growth association analysis algorithm. Rainfall level, runway visibility, low cloud level and crosswind level are selected as the antecedents, and flight delay conditions are selected as the posterior items to obtain several association rules. The several association rules are summarized to generate an association rule library for flight delays.
3. The flight delay prediction method based on knowledge and data dual drive according to claim 2 is characterized in that: The construction of the knowledge model includes: For flights that need delay prediction, the corresponding flight data and AWOS meteorological data are quantified and matched in the corresponding flight delay association rule library to obtain the confidence level of whether the flight will be delayed or not.
4. The flight delay prediction method based on knowledge and data dual drive according to claim 3 is characterized in that: The construction of the data model includes: Using the selective integration strategy, a subset of learning models is selected through a preset mechanism to build the final integrated model.
5. The flight delay prediction method based on knowledge and data dual drive according to claim 4 is characterized in that: The method of selecting a base model according to a pre-order selection algorithm based on entropy E difference includes: The base model is measured by the entropy E difference metric, including the following formula: In the formula, L is the number of base models, N is the number of samples, and Z j is a sample, l(Z j ) is the number of samples correctly predicted in any base model.
6. The flight delay prediction method based on knowledge and data dual drive according to claim 1 is characterized in that: The optimal meta-model comprises: The SVM model is used as a meta-model.
7. The flight delay prediction method based on knowledge and data dual drive according to claim 1 is characterized in that: The fused data set is preprocessed to obtain a new data set including a training set, a validation set and a test set, including: In data cleaning, for inconsistent date formats, use timestamps to make the dates consistent. For null values, choose to delete the corresponding features or delete the rows according to the number of null values. In data labeling, for classification problems, the label column is marked as delay status. The delay time is calculated based on the difference between the scheduled take-off and landing time of the flight and the actual take-off and landing time. Among them, flights with a delay time less than the preset time are marked as not delayed, and the remaining flights are marked as delayed; In data encoding, different encoding methods are used for continuous and discrete data. Min-Max encoding is used for continuous features, and CatBoost encoding is used for discrete features. In data splitting, the data is randomly divided into training set, validation set and test set in a certain proportion.
8. The flight delay prediction method based on knowledge and data dual drive according to claim 7 is characterized in that: The construction of the flight delay prediction model based on knowledge and data dual drive includes: Set the initial parameters of the base model, including the number of blocks, the number of fully connected layers in each block, and the feature dimension of the input data; Set the initial parameters of the training method, including the learning rate, optimizer and loss function during training.
9. The flight delay prediction method based on knowledge and data dual drive according to claim 8 is characterized in that: Training each of the base models using the new data set includes: Forward propagation and back propagation, where The forward propagation is a process of inputting the preprocessed data into the network model, extracting features through the fully connected layer, the attention module and the nonlinear activation layer, and calculating the predicted value and the true label value through the loss function to obtain the prediction error; In the classification task, the cross entropy function is used as the loss function, including the following formula: In the formula, I is the number of delay levels, y(x i ) is the true label of the i-th category of the x-th sample, is the predicted probability of the i-th category of the x-th sample; The back propagation is a process of chain derivation of the loss function obtained in the forward propagation.
10. A flight delay integrated prediction system based on knowledge and data dual drive, characterized in that: Implementing the flight delay prediction method based on knowledge and data dual drive as described in any one of claims 1 to 9, the prediction system comprises: Data acquisition module, used to obtain METAR message data, flight data and AWOS weather data; A knowledge use module, used to generate new feature data injected with expert knowledge from the METAR message data according to the ATMAP algorithm; to construct an association rule base and a knowledge model according to the flight data and the AWOS meteorological data through the FP-Growth association analysis algorithm; A data model building module, used to select a base model according to a pre-order selection algorithm based on entropy E difference and build a data model; A fusion driving module, used for fusing the new feature data, the flight data and the AWOS meteorological data to form a fusion data set; A data preprocessing module, used to preprocess the fused data set to obtain a new data set including a training set, a validation set and a test set; An integrated learning module, used for selecting an optimal meta-model and generating an integrated learning model according to the optimal meta-model, the data model and the knowledge model; A flight delay prediction model construction module, used to set the parameters of each base model of the ensemble learning model, train each base model through the new data set, and generate a flight delay prediction model based on knowledge and data dual drive; The result acquisition module is used to generate a flight delay prediction result through the flight delay prediction model.
Citation Information
Cited By
Flight delay knowledge base construction method and flight delay prediction method
CN121119302A
Airport flight time dynamic adjustment method, system, device, medium and product
CN121119651A
Methods, systems, equipment, media, and products for dynamic adjustment of airport flight schedules
CN121119651B
