ATFM delay prediction method and device based on discrete fusion of time and space feature matrixes

Abstract

The embodiment of the invention discloses an ATFM delay prediction method and device based on discrete fusion of time and space feature matrixes, and relates to the technical field of air traffic control. The server obtains flight plan data in a specified area according to a model verification request sent by the client, then generates an ATFM delay prediction data set and establishes an ATFM delay prediction network; using the ATFM delay prediction data set to generate a feature matrix of time and space scales; constructing a total feature matrix and a total target matrix through an Optuna framework based on Bayesian optimization, and taking the total feature matrix and the total target matrix as the input of a prediction model; and verifying the operation effect of each sub-model and the combination mode of the sub-models under the ATFM delay prediction network, and feeding back the operation effect to the client. Therefore, the ATFM delay prediction model can be further optimized, and the prediction accuracy is improved.

Description

Technical Field

[0001] The present invention relates to the field of air traffic control technology, and in particular to an ATFM delay prediction method and device based on the discrete fusion of time and space feature matrices. Background Art

[0002] Air traffic flow management delay (ATFM delay) is the difference between the calculated take-off time (CTOT) of the traffic flow management system and the planned take-off time (TTOT) of the aircraft operator. In the operation of the aviation network, with the dynamic changes in airspace capacity, some waypoints and routes will be congested due to the mismatch between capacity and demand. When the aircraft operator enters the planned take-off time (TTOT) into the system, the air traffic flow management system will determine the congested nodes that the aircraft route will pass through, and queue up the aircraft that exceed the capacity limit, thereby allocating the calculated take-off time (CTOT) to the aircraft. Therefore, ATFM delay reflects the degree of match between air transport demand and key resources of the aviation network, and can be used to evaluate the operational capabilities of the ATFM system in a targeted manner.

[0003] Furthermore, ATFM delays can have complex downstream effects. When a flight experiences a significant ATFM delay, the delay can propagate to subsequent flights. Even if the ATFM delay for a flight subject to air traffic flow management is minimal or even zero, it can still prevent subsequent flights from departing within their allocated time slots, necessitating a CTOT reallocation and resulting in even more significant delays. Therefore, controlling ATFM delays is crucial to the operation of route networks and airlines' operations.

[0004] However, current traffic management still faces many challenges. For example, existing implementations can already accurately classify the causes of ATFM delays, which provides a theoretical basis for subsequent ATFM delay prediction implementations. In terms of predictive model construction, ATFM delay prediction has undergone a transition from traditional machine learning to advanced deep learning, from focusing on structural modeling of air traffic network characteristics to data-driven approaches that capture the temporal and spatial relationships present in the data. Previous researchers have attempted various prediction models, including the k-nearest neighbor algorithm (KNN), support vector machine (SVM), naive Bayes, and others, but the actual prediction accuracy of these existing solutions remains unsatisfactory.

[0005] Therefore, how to further optimize the ATFM delay prediction model to improve prediction accuracy has become a topic that requires further in-depth research. Summary of the Invention

[0006] The embodiments of the present invention provide an ATFM delay prediction method and device based on the separate fusion of temporal and spatial feature matrices, which can further optimize the ATFM delay prediction model and thus improve the prediction accuracy.

[0007] To achieve the above objectives, the embodiments of the present invention adopt the following technical solutions:

[0008] In a first aspect, an embodiment of the present invention provides a method comprising:

[0009] S1. The server obtains flight schedule data within the specified area based on the model verification request sent by the client, then generates an ATFM delay prediction dataset and establishes an ATFM delay prediction network.

[0010] S2. generating a feature matrix of time and space scales using the ATFM delay prediction dataset;

[0011] S3. Construct the total feature matrix and the total target matrix through the Optuna framework based on Bayesian optimization and use them as inputs of the prediction model;

[0012] S4. Verify the operating effect of each sub-model and the combination of sub-models under the ATFM delay prediction network, and feed back the operating effect to the client.

[0013] In the ATFM delay prediction network, airports and waypoints are used as nodes, and routes are used as edges, wherein waypoints are points that a planned route will pass through, and there is a directed edge between two airports that a route passes through, the starting point of the directed edge represents the departure airport, and the end point of the edge represents the destination airport; the ATFM delay prediction network is represented as G, G = (V, E), V represents the node set, E represents the edge set, the number of nodes is n = |V|, the number of edges is m = |E|, and the directed edge from node V1 to node V2 is represented as (V1, V2).

[0014] In S2, generating a time-scale feature matrix includes: breaking down the flight's TTOT according to the time scale, and then mapping it into time features according to date types, where the date types include: holidays / non-holidays, weekends / weekdays; and filling the true values ​​of the time features into the time-scale feature matrix.

[0015] Generating a spatial-scale feature matrix involves: using the XGBoost regression model to capture the nonlinear relationship between spatial features and targets, and fitting it to obtain ATFM delay data; determining the contribution of different spatial features to the prediction of ATFM delays, thereby obtaining the weight value corresponding to each spatial feature; using the obtained weight value to calculate the eigenvalue of each flight data separately, and filling in the spatial-scale feature matrix.

[0016] The prediction model includes at least 6 sub-models based on decision trees, and the types of sub-models include: decision tree DT, random forest RF, gradient boosting tree GBDT, XGBoost, LightGBM and CatBoost; 6. The method according to claim 5 is characterized in that it also includes: combining random forest RF with XGBoost to obtain a first combined sub-model; and combining random forest RF, XGBoost, gradient boosting tree GBDT and LightGBM to obtain a second combined sub-model.

[0017] Specifically, S3 includes: for each sub-model, tuning the key parameters of the sub-model through the Optuna framework based on Bayesian optimization and genetic algorithm, and then evaluating the prediction accuracy of the sub-model through cross-validation.

[0018] In a second aspect, an embodiment of the present invention provides a device, which is deployed on a server. The device includes:

[0019] The data extraction module is used to obtain flight schedule data within a specified area based on the model verification request sent by the client, and then generate an ATFM delay prediction dataset and establish an ATFM delay prediction network;

[0020] A preprocessing module, configured to generate a feature matrix of time and space scales using the ATFM delay prediction dataset;

[0021] The data screening module is used to construct the total feature matrix and the total target matrix through the Optuna framework based on Bayesian optimization, and use them as inputs for the prediction model;

[0022] The verification feedback module is used to verify the operating effect of each sub-model and the combination of sub-models under the ATFM delay prediction network, and feed back the operating effect to the client.

[0023] The present invention provides an ATFM delay prediction method and device based on the separate fusion of temporal and spatial feature matrices. The method divides the flight schedule dataset into training, validation, and test sets. The ATFM delay duration for each flight is calculated and matched, and airport and route information is integrated to form the original ATFM delay prediction dataset. The ATFM delay prediction network scope is determined, and the latitude and longitude information of airports and the route network topology are integrated to construct the ATFM delay prediction network. Based on the results of previous theoretical analysis, several factors that significantly influence ATFM delay performance are selected as data features. The true values ​​of temporal features are retained, and Bayesian optimization is used to find the optimal weights for spatial features, thereby assigning eigenvalues. Feature matrices and target matrices are formed for each set, respectively, at the temporal and spatial scales. Methods for fusing the temporal and spatial matrices at the feature and model levels are explored. The Bayesian optimization-based Optuna framework is used to construct the total feature matrix and target matrix, which serve as input for the subsequent prediction model. Six decision tree-based models are used to investigate prediction accuracy, identifying the input method and prediction model that yield the best prediction results. During the prediction process, we again used the Optuna framework and genetic algorithm based on Bayesian optimization to tune key model parameters. We then added cross-validation to evaluate and compare the accuracy of each prediction model. Based on the ATFM delay prediction results, we selected the model with the highest prediction accuracy from the previous step for the final experimental verification. This allowed us to further optimize the ATFM delay prediction model and improve its accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the embodiments. 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 any creative work.

[0025] Figure 1 Schematic diagram of the ATFM delay prediction network model provided by an embodiment of the present invention;

[0026] Figure 2 Two feature fusion method prediction flow charts provided in embodiments of the present invention;

[0027] Figure 3 Schematic diagram of the Optuna optimization process of the prediction model provided in an embodiment of the present invention;

[0028] Figure 4 、 5 A schematic diagram of time / space feature weights provided by an embodiment of the present invention;

[0029] Figure 6 A schematic diagram of the weights of each feature in the total feature matrix provided by an embodiment of the present invention;

[0030] Figure 7 MAE convergence curve diagram of each model test set provided by the embodiment of the present invention;

[0031] Figure 8 A comparison chart of predicted values ​​and actual values ​​provided by an embodiment of the present invention (high flight volume airport);

[0032] Figure 9 A comparison chart of predicted values ​​and actual values ​​provided by an embodiment of the present invention (low flight volume airport);

[0033] Figure 10 A comparison chart of predicted values ​​and actual values ​​provided by an embodiment of the present invention (high flight volume routes);

[0034] Figure 11 A comparison chart of predicted values ​​and actual values ​​provided by an embodiment of the present invention (low flight volume routes). DETAILED DESCRIPTION

[0035] To enable those skilled in the art to better understand the technical solutions of the present invention, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention will be described in detail below, with examples of the embodiments illustrated in the accompanying drawings. Throughout, identical or similar reference numerals represent identical or similar elements or elements having identical or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and intended only to explain the present invention and are not to be construed as limiting the present invention. Those skilled in the art will appreciate that, unless otherwise stated, the singular forms "a," "an," "said," and "the" used herein may also include the plural forms. It should be further understood that the term "comprising" as used in the description of the present invention refers to the presence of the stated features, integers, steps, operations, elements, and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when an element is referred to as being "connected" or "coupled" to another element, it may be directly connected or coupled to the other element, or intervening elements may be present. Furthermore, "connected" or "coupled" as used herein may include wireless connections or couplings. The term "and / or" as used herein includes any and all combinations of one or more associated listed items. It will be understood by those skilled in the art that, unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those skilled in the art in the art to which the present invention belongs. It should also be understood that terms such as those defined in common dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and will not be interpreted in an idealized or overly formal sense unless defined as such herein.

[0036] An embodiment of the present invention provides an ATFM delay prediction method based on separate fusion of temporal and spatial feature matrices, including:

[0037] Step (a) Data collection and preprocessing: The flight schedule dataset is divided into training, validation, and test sets. The ATFM delay duration of each flight is calculated and matched, and airport and route information is integrated to form the original ATFM delay prediction dataset. The scope of the ATFM delay prediction network is determined, and the latitude and longitude information of the airports and the route network topology are integrated to construct the ATFM delay prediction network.

[0038] Step (b) Feature screening and weight calculation: Based on the results of the previous theoretical analysis, several factors that have a significant impact on ATFM delay performance were selected as data features. The true value of the time feature was retained, and Bayesian optimization was used to find the optimal weight for the spatial feature to assign a feature value. The feature matrix and target matrix of the time and space scales were formed on each set respectively.

[0039] Step (c) Feature Matrix Fusion: We explored the fusion methods of the time matrix and the spatial matrix at the feature level and the model level, and used the Optuna framework based on Bayesian optimization to construct the total feature matrix and the total target matrix as the input of the subsequent prediction model.

[0040] Step (d) Prediction model construction: Input 6 decision tree-based models to explore the prediction accuracy and find the input method and prediction model with the best prediction effect. In the prediction process, the Optuna framework based on Bayesian optimization and genetic algorithm are used again to tune the key parameters of the model, and cross-validation is added for evaluation to compare the prediction accuracy of each prediction model.

[0041] Step (e) Evaluation of prediction results: Based on the ATFM delay prediction results, the model with the highest prediction accuracy in the previous step is selected for final experimental verification. Several pairs of airports and routes are selected for further examples to analyze the causes of ATFM delays.

[0042] Specifically:

[0043] In step (a), the flight plan dataset is divided into training, validation, and test sets. The ATFM delay duration of each flight is calculated and matched, and the airport and route information is integrated to form the original ATFM delay prediction dataset. The ATFM delay prediction network range is determined, and the latitude and longitude information of the airports and the route network topology are integrated to construct the ATFM delay prediction network. The specific design steps are as follows:

[0044] Step a-1. This embodiment is experimentally verified based on the actual flight operation data from May 1 to June 15, 2023 in China. Taking into account the positive correlation between the prediction effect and the degree of data adequacy, in order to achieve the best prediction accuracy, the embodiment of this article uses all valid data within this interval. In order to eliminate the impact of data integrity on the accuracy of the experiment, the original flight plan data set used in this experiment was screened based on all flight plan data from May 1 to June 15, 2023 in China. Only flights whose take-off and landing airports are both located in China were retained, with a total of 589,842 valid data. Among them, 193,440 flights were subject to air traffic flow control, accounting for approximately 32.80% of the total.

[0045] Step a-2: construct an ATFM delay prediction network in the form of a directed network graph, as shown in the attached figure. Figure 1 As shown. Airports and waypoints are regarded as nodes, and routes are regarded as edges. Waypoints are points that the planned route will pass through. If a route passes between two airports, there is a directed edge between the two airports. The starting point of the edge represents the departure airport, and the end point of the edge represents the destination airport. The ATFM delay prediction network model is represented by G, G = (V, E). V represents the node set, and E represents the edge set. The number of nodes in the network is n = |V|, and the number of edges is m = |E|. (V1, V2) represents the directed edge from node V1 to node V2. A brief schematic diagram of the ATFM delay prediction network model is shown below. Figure 2 As shown, the ATFM delay prediction objects are AC edge, BD edge, CA edge and DB edge.

[0046] In step (b), based on the results of previous theoretical analysis, several factors that significantly affect ATFM delay performance were selected as data features. The true value of the temporal feature was retained, and Bayesian optimization was used to find the optimal weights for the spatial features to assign eigenvalues. Feature matrices and target matrices at the temporal and spatial scales were formed for each set. The specific design steps are as follows:

[0047] Step b-1. Consider the time-related feature variables as time features. Break down the flight's TTOT into hours and minutes, and map them into independent features according to information such as holidays / non-holidays, weekends / weekdays, etc. In particular, considering that the congestion situation on a certain route or region will generally have a lasting impact on the current and subsequent flights, the delay performance of the previous flights is of reference significance to a certain extent for delay prediction. This article will use the average ATFM delay duration of other flights passing through the same take-off and landing airport as the predicted flight within 4 hours before departure as a separate time feature. Feature variables related to geographic location and their combination are considered spatial features, including take-off / landing airport information, airport geographic information, airport type (international / domestic), cumulative traffic of airports and routes, cumulative ATFM delay duration, and other information. Specific time and space characteristics are shown in Table 1.

[0048] Table 1 Characteristics affecting ATFM delay prediction

[0049]

[0050]

[0051] Step b-2: Since there are relatively few time features and no significant interactions between them, the feature matrix is ​​populated directly with real values. Information such as time delay (TTOT), holiday information, weekend information, and the average historical delay within a 4-hour time window is extracted from the fused dataset. TTOT is broken down into hours and minutes. For ease of calculation, the hourly portion is converted to sine and cosine values, enabling the model to better understand the cyclical nature of time. Minutes are used directly as feature values. Holidays and non-holidays are assigned values ​​of 1 and 0, respectively. Similarly, weekends (Friday to Sunday) and weekdays are assigned values ​​of 1 and 0, respectively. A sliding window is used to calculate the average ATFM delay value for the historical data of the target flight and use it as a feature value.

[0052] Step b-3: Considering the complexity of spatial features and the inclusion relationship between features, it is necessary to use the XGBoost regression model to capture the nonlinear relationship between features and targets to fit the ATFM delay data, learn the contribution of different spatial features to the prediction of ATFM delays, and thus obtain the optimal weight for each spatial feature; then use this weight combination to calculate the eigenvalue of each flight data separately to obtain the eigenvalue of each spatial feature.

[0053] In step (c), we explored how to fuse the temporal matrix and the spatial matrix at the feature and model levels. We used the Optuna framework based on Bayesian optimization to construct the total feature matrix and the total target matrix as inputs for the subsequent prediction model. The specific design steps are as follows:

[0054] Step c-1, for the selection of prediction models, related embodiments have previously tried methods such as support vector machines. This article mainly explores the prediction effect of genetic algorithm decision tree prediction models, including random forest, GBDT, XGBoost, etc. The input form accepted by this type of prediction model is mainly a structured numerical feature matrix. Each row of the matrix corresponds to a data sample, and each column corresponds to a numerical feature. In the previous section, the features within the time feature matrix have been fused together according to the optimal weight. Similarly, the spatial feature matrix is ​​the same. Therefore, before the two sets of matrices obtained from the feature engineering in the previous section are put into the prediction model as input, it is necessary to discuss the matrix fusion mode. This section mainly focuses on the different performances of the fusion methods of the two feature matrices at the feature level and the model level in terms of prediction accuracy.

[0055] Step c-2, feature-level fusion, involves fusing the temporal and spatial matrices before inputting them into the model, forming a single feature matrix encompassing all data features. This matrix is ​​then fed into the model for a single training step. This predictive model can exploit correlations between features through unified learning. This simplifies the model structure and streamlines the training process, improving prediction efficiency to a certain extent. However, when the data features are high in dimensionality, overfitting is more likely to occur. Furthermore, if the data features vary significantly, this can interfere with the model's judgment.

[0056] Model-level fusion, in contrast, builds and trains models for the temporal and spatial feature matrices separately, then fuses the training results to produce the final prediction. This approach has the advantage of allowing each model to optimize its training process based on the characteristics of different input matrices, thereby improving overall performance. It also makes it easier to control the relevant parameters and complexity of each model. However, this requires separating the training processes, which increases complexity and does not necessarily guarantee the stability of the combined prediction results.

[0057] In step c-3, whether the time and space matrices are combined first and then trained, or trained separately and then combined, the time and space matrices must be concatenated according to certain weights. This combination is achieved using Bayesian optimization. Bayesian optimization searches through different weight combinations to find the fusion method that maximizes performance. In the feature-level combination method, Bayesian optimization is used to find the optimal weights for combining the time and space matrices to form the total feature matrix. In the model-machine combination method, Bayesian optimization is used to find the optimal weights for fusing the two prediction results.

[0058] In step (d), we input six decision tree-based models to explore their prediction accuracy and identify the input method and prediction model that yield the best prediction results. During the prediction process, we again used the Bayesian optimization-based Optuna framework and genetic algorithm to tune key model parameters. We also added cross-validation to evaluate and compare the prediction accuracy of each prediction model, as follows:

[0059] In step d-1, selecting a prediction model, decision tree-based ensemble learning methods excel in improving model robustness and accuracy. The most prominent examples, as shown in Table 3, are the random forest and XGBoost models. Random forest constructs independent decision trees by sampling training examples and then averaging the predictions from the different trees. XGBoost, on the other hand, fits the residuals of the previous tree structure at each step and incorporates a regularization term to prevent overfitting. Given the complex data required for ATFM delay prediction, this paper incorporates an ensemble of random forest and XGBoost to combine the high robustness of random forest with the high prediction accuracy of XGBoost. Furthermore, this combination of the two approaches balances variance and bias, improving model diversity and thus reducing the risk of overfitting.

[0060] Table 2 Prediction model based on decision tree

[0061]

[0062]

[0063] In step d-2, based on conventional decision tree-based models, this paper proposes using the Optuna optimization method to tune hyperparameters of the prediction model. Combining early stopping and parallel acceleration, this method significantly improves parameter tuning efficiency while preserving model accuracy. By incorporating Optuna optimization into conventional prediction models, the model can set an optimization direction after setting a search space, finding the optimal solution among a large number of parameter combinations. For both ensemble models, after completing parameter tuning for all base models, a genetic algorithm is used to optimize the parameters of the final fusion process.

[0064] Step d-3, taking the prediction process of the random forest + XGBoost ensemble model using model-level fusion as an example, the optimization process of Optuna + genetic algorithm is shown in the attached figure. Figure 3 The specific steps are as follows:

[0065] (1) Data preprocessing: The ATFM delay prediction dataset is used to generate feature matrices and target matrices according to time and space, and then divided into training, validation, and test sets, which are then put into the initial random forest and XGBoost models for training respectively;

[0066] (2) Optuna optimization and parameter adjustment: The Optuna framework is used to perform Bayesian optimization and parameter adjustment on the random forest model and the XGBoost model respectively, and the sub-prediction models of the two models under the optimal hyperparameter configuration are obtained;

[0067] (3) Genetic algorithm optimization of model fusion weight: Individuals represent the model fusion weight, generate n individuals to form the initial population, establish a fitness function, and find the optimal individual through selection, crossover and mutation. This process is iterated until the termination condition is reached, and the optimal fusion weight of the random forest and XGBoost models is obtained;

[0068] (4) The trained random forest and XGBoost models are weightedly fused according to the optimal fusion weights found in the previous step, and the prediction results of the fusion model are evaluated on the test set.

[0069] In step (e), based on the ATFM delay prediction results, the model with the highest prediction accuracy in the previous step is selected for final experimental verification. Several pairs of airports and routes are selected for further application in analyzing the causes of ATFM delays.

[0070] The following simulation verification is carried out on the ATFM delay prediction method of the present invention which integrates the time and space feature matrices:

[0071] Step e-1: To ensure model training quality, the original dataset needs to be partitioned and feature engineering performed on each partition. Considering that the dataset partitioning should be comprehensive in terms of temporal and spatial characteristics, the following time-series partitioning method is used: flight data from May 1st to May 31st is used as the training set; flight data from June 1st to June 8th is used as the validation set; and flight data from June 9th to June 15th is used as the test set. The data volume ratios of the training set, validation set, and test set are: 68%, 17%, and 15%.

[0072] The 236 domestic airports included in the flight data set were selected as nodes in the ATFM delay prediction network, including 94 international airports and 142 domestic airports. All 7,140 routes involved in flight operations in the data set were selected as edges in the network. Data items included each flight's departure / destination airports, time to takeoff / landing times, and other information. After preprocessing, the flight schedule data was fused with other ATFM delay-related information to form the original ATFM delay prediction dataset.

[0073] The ATFM delay prediction network graph is a weighted directed network graph. Node and edge weights represent the influence of airports and routes in the route network on ATFM delay prediction. We use the SHAP optimization method to assess the importance of each feature and combine it with XGBoost to find the optimal weight combination among the features, thereby deriving the weights for each airport and route.

[0074] Step e-2: Constructing the temporal feature matrix: Decompose the TTOT into month, day, hour, minute, and second. Incorporate some implicitly relevant information as features, including holiday and weekend information, and the average historical delay of preceding flights. An XGBoost regression model is used to learn the relationship between each feature and the predicted target ATFM delay duration, thereby calculating the weight of each feature. The feature values ​​are multiplied by their respective weights to construct a weighted temporal feature matrix. The weights of each temporal feature are shown in the figure. To prevent data noise from negatively impacting the experimental results, the final temporal feature matrix was removed from the model due to its low weights: month (1.78%) and second (0%). The remaining six features were retained. The average historical delay within the four hours before departure accounted for 56.86%, playing a dominant role in the temporal features; day and hour also played a significant role. The spatial feature matrix was constructed similarly. Because the fusion of airport and route weights as underlying features performed poorly in the model, the weights and eigenvalues ​​of each spatial feature were calculated separately when constructing the spatial feature matrix, for a total of 18 spatial features. In the final spatial feature matrix, 10 features with weights below 1% were removed, retaining the remaining eight features. The dominant spatial features were the destination airport's latitude, the number of flights on the route, and the route's cumulative ATFM delays. In the model-level combination approach, the temporal and spatial feature matrices, along with their target matrix, are directly fed into the prediction model as inputs, and after completing the predictions separately, they are fused.

[0075] Step e-3: Next, construct the total feature matrix and target matrix for each data set. Use Bayesian optimization to search for the optimal weight of each feature, and reassign each feature according to the optimal weight. The weight of each feature in the final total feature matrix is ​​as follows: Figure 4 、 5 As shown in the figure, holiday information and the destination airport's latitude and longitude account for the largest proportions of the total feature matrix, potentially indicating significant geographic variation in the distribution of ATFM delays. The historical average delay within the four hours before departure also exhibits a significant impact, validating the inclusion of this time feature. The factors that contribute least to delay prediction are the destination airport's cumulative delays and the number of flights at the departure airport. In feature-level combination prediction, the total feature matrix and its target matrix serve as inputs to the prediction model, and prediction results are obtained after training.

[0076] To verify the effectiveness of feature engineering, a random forest model with default parameters was added for a control experiment, with the original dataset and the total feature matrix input, respectively. The resulting prediction results are shown in the table. It can be seen that using the feature matrix as input resulted in a certain degree of improvement in various prediction indicators: the average absolute prediction error per flight decreased by 0.26 minutes, indicating an improvement in overall prediction accuracy; the mean squared error (MSE) decreased significantly, indicating a reduction in large deviations in the prediction results and a more robust prediction; the coefficient of determination (R2) increased from 0.06 to 0.26, increasing the model's interpretability by more than four times its original value, and significantly improving the regression fit, as shown in Table 3.

[0077] Table 3 Comparison of two input methods

[0078]

[0079]

[0080] Step e-4: First, for the model-level fusion prediction method, the temporal and spatial feature matrices are placed in independent models for training, and then the prediction results are fused using the Optuna framework based on Bayesian optimization. The prediction results of each model are shown in Table 4.

[0081] Table 4 Model-level fusion prediction results (default parameters)

[0082]

[0083] For the feature-level fusion prediction method, the time and space feature matrices are optimally weighted in the Optuna framework based on Bayesian optimization to form a total feature matrix, which is then put into 6 prediction models for training and prediction. The prediction results are shown in Table 5.

[0084] Table 5 Feature-level fusion prediction results (default parameters)

[0085]

[0086] The prediction results show that, with the exception of Random Forest, the other five models all achieved lower MAE, MSE, and higher R² when using feature-level fusion prediction, demonstrating superior prediction performance. Although Random Forest achieved a lower MAE when using model-level fusion prediction, its overall performance was significantly inferior to the other models. Therefore, the next phase of optimization will focus solely on feature-level fusion prediction.

[0087] Next, the total feature matrix and total target matrix on each dataset were used as input, MAE was set as the target loss function, and hyperparameter optimization was performed on each model. After parameter adjustment, the performance of each model on the test set is shown in the table. The mean absolute error (MAE) of GBDT and LightGBM is significantly lower than that of the other models, and the mean square error (MSE) and coefficient of determination (R2) also perform very well. Under the optimal parameter configuration, GBDT trains 129 decision trees, each with a maximum depth of 7 layers, the proportion of samples used in the training process is 65.6%, and the model learning rate is approximately 0.10; LightGBM trains 830 trees, each with a maximum depth of 8 layers, the proportion of samples used in training is 93.7%, and the model learning rate is approximately 0.16. Although the MSE and R2 performance of the integrated model RF+XGBoost+GBDT+LightGBM are very good, its MAE is large, which is somewhat different from the previous two, as shown in Table 6.

[0088] Table 6 Prediction results of each model after parameter adjustment

[0089]

[0090] After determining the optimal parameters for each model, each model was rebuilt using the optimal parameters, and the training set, validation set, and test set matrices were input into the prediction model. The MAE convergence process of each model on the test set was recorded and summarized into a loss function curve, as shown in the figure. The horizontal axis represents the number of decision trees established, and the vertical axis represents the loss value. Each model achieved the optimal prediction accuracy at 150, 1000, 129, 830, 177, and 50 trees, respectively. For ease of observation, only the loss function of the first 150 trees is shown in the figure. The loss function MAE performance and model convergence speed of GBDT and LightGBM are significantly better than other models. Among them, LightGBM is slightly better than GBDT, and they were selected for subsequent experiments.

[0091] Step e-5: To further analyze the prediction performance of the GBDT model and the LightGBM model at different airports, we selected four airports with high flight volume and four airports with low flight volume. Consistent with the conclusions drawn in Section 4.3, the LightGBM model outperformed the GBDT model at all airports. As shown in the table, the LightGBM model's predicted absolute error (MAE) was relatively stable, with only slightly higher values ​​at ZGSZ and ZYBS. Relatively accurate predictions were obtained at the remaining six airports. The model's R2 performance fluctuated greatly, with airports with low flight volume performing significantly better than airports with high flight volume. The model achieved the best R2 at ZYJZ airport, reaching 0.81, while the R2 at ZGGG airport was only 0.05.

[0092] Table 10 Airport prediction results

[0093]

[0094] Compare the LightGBM model's prediction results for the above airports from June 9 to June 15, 2023 with the actual ATFM delay values, as shown below: Figure 8 and Figure 9 As shown in the figure, the model's predictions are generally lower than the actual ATFM delay values, with the model performing better at airports with low flight volume. Furthermore, the model's predictions are significantly correlated with the actual ATFM values. When the actual ATFM values ​​are high, the predictions are less accurate, while when the actual ATFM values ​​are low, the model's predictions are closer to the true values. It can be seen that the predictions for ZBDH Airport are almost identical to the true values, demonstrating the best prediction performance.

[0095] Next, we focus on the model's performance in route prediction. We also selected four high-volume routes for comparison with four low-volume routes, with the results shown in Table 7. At high-volume airports, although the LightGBM model's MAE and R² still outperform the GBDT model, the gap is small. At low-volume airports, the GBDT model's MAE performance tends to surpass the LightGBM model, while the two models have different R² performances.

[0096] Table 7 Route prediction results

[0097]

[0098]

[0099] Similarly, the LightGBM model's prediction results for the above routes on the test set are compared with the actual ATFM delay values. Figure 10 and Figure 11 As shown in the figure, similar to the airport prediction results, the model's predictions are still lower than the actual ATFM delay values ​​in most cases. Furthermore, the model performs better when ATFM delays are low. When the actual ATFM delays of a flight are high, the model struggles to provide accurate predictions.

[0100] It can be seen that this solution can realize ATFM delay prediction for a single flight. After fusing the time feature matrix and the spatial feature matrix and inputting them into the prediction model, the MAE can be reduced to 6.33 minutes per flight. It can also realize ATFM delay prediction for airports. After fusing the time feature matrix and the spatial feature matrix and inputting them into the prediction model, the MAE at specific airports can be reduced to 2.73 minutes per flight. It can also realize ATFM delay prediction for routes. After fusing the time feature matrix and the spatial feature matrix and inputting them into the prediction model, the MAE on specific routes can be reduced to 1.24 minutes per flight.

[0101] This embodiment also provides an ATFM delay prediction device based on the separate fusion of time and space feature matrices, wherein the device is deployed on a server;

[0102] The device comprises:

[0103] The data extraction module is used to obtain flight schedule data within a specified area based on the model verification request sent by the client, and then generate an ATFM delay prediction dataset and establish an ATFM delay prediction network;

[0104] A preprocessing module, configured to generate a feature matrix of time and space scales using the ATFM delay prediction dataset;

[0105] The data screening module is used to construct the total feature matrix and the total target matrix through the Optuna framework based on Bayesian optimization, and use them as inputs for the prediction model;

[0106] The verification feedback module is used to verify the operating effect of each sub-model and the combination of sub-models under the ATFM delay prediction network, and feed back the operating effect to the client.

[0107] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited to this. Any changes or replacements that can be easily thought of by any technician familiar with this technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention. Therefore, the protection scope of the present invention should be based on the protection scope of the claims.

Claims

1. An ATFM delay prediction method based on discrete fusion of temporal and spatial feature matrices, characterized by: include: Based on the model verification request sent by the client, the server obtains the flight schedule data in the specified area, generates an ATFM delay prediction dataset, and establishes an ATFM delay prediction network. Using the ATFM delay prediction dataset, a feature matrix of time and space scales is generated; The total feature matrix and the total target matrix are constructed through the Optuna framework based on Bayesian optimization and used as inputs for the prediction model; Verify the operating effect of each sub-model and the combination of sub-models under the ATFM delay prediction network, and feed back the operating effect to the client.

2. The method according to claim 1, characterized in that In the ATFM delay prediction network, airports and waypoints are used as nodes, and routes are used as edges. Waypoints are points that a planned route will pass through. There is a directed edge between two airports that a route passes through. The starting point of the directed edge represents the departure airport, and the end point of the edge represents the destination airport. The ATFM delay prediction network is represented as G, G = (V, E), V represents the node set, E represents the edge set, the number of nodes is n = |V|, the number of edges is m = |E|, and the directed edge from node V1 to node V2 is represented as (V1, V2).

3. The method according to claim 1, characterized in that The feature matrix of the generated time scale includes: The flight's TTOT is broken down into time scales and then mapped to time features based on date types, including holidays / non-holidays and weekends / weekdays. Fill the true value of the time feature into the feature matrix of the time scale.

4. The method according to claim 3, characterized in that Generating the feature matrix of spatial scale includes: The XGBoost regression model is used to capture the nonlinear relationship between spatial features and targets, and fit the ATFM delay data; Determine the contribution of different spatial features to the prediction of ATFM delays, and thus obtain the weight value corresponding to each spatial feature; The obtained weight values ​​are used to calculate the eigenvalues ​​of each flight data and fill in the characteristic matrix of the spatial scale.

5. The method according to claim 1 or 4, characterized in that The prediction model includes at least 6 sub-models based on decision trees, and the types of sub-models include: decision tree DT, random forest RF, gradient boosting tree GBDT, XGBoost, LightGBM and CatBoost.

6. The method according to claim 5, characterized in that Also includes: Combine Random Forest RF with XGBoost to get the first combined sub-model; And, the random forest RF, XGBoost, gradient boosting tree GBDT and LightGBM are combined to obtain the second combination sub-model.

7. The method according to claim 5, characterized in that The construction of the total feature matrix and the total target matrix by the Optuna framework based on Bayesian optimization includes: For each sub-model, the key parameters of the sub-model are tuned using the Optuna framework based on Bayesian optimization and genetic algorithm, and the prediction accuracy of the sub-model is evaluated through cross-validation.

8. An ATFM delay prediction device based on the discrete fusion of time and space feature matrices, characterized in that: The device is deployed on a server; The device comprises: The data extraction module is used to obtain flight schedule data within a specified area based on the model verification request sent by the client, and then generate an ATFM delay prediction dataset and establish an ATFM delay prediction network; A preprocessing module, configured to generate a feature matrix of time and space scales using the ATFM delay prediction dataset; The data screening module is used to construct the total feature matrix and the total target matrix through the Optuna framework based on Bayesian optimization, and use them as inputs for the prediction model; The verification feedback module is used to verify the operating effect of each sub-model and the combination of sub-models under the ATFM delay prediction network, and feed back the operating effect to the client.

Images