A flight guarantee time prediction method, device, medium and equipment

Abstract

The present application relates to the technical field of data prediction, and more particularly to a flight guarantee time prediction method, device, medium and equipment, which constructs a multi-dimensional initial feature set and fuses a channel attention mechanism, and captures long time sequence dependency, so that the capture ability of the delay conduction effect of the previous guarantee node is significantly improved; through the adaptive selection of the double module based on the node attribute, the prediction needs of different types of guarantee nodes are accurately matched, and the overall accuracy of the whole process node prediction is improved; by splicing the prediction value of the i-th guarantee node with the time sequence context feature to form the dynamic enhancement feature of the subsequent guarantee node, the transfer and fusion of the previous prediction information to the subsequent node are realized, so that the subsequent guarantee node prediction can adapt to the running state change of the previous guarantee node in real time, effectively reduces the subsequent prediction deviation caused by the delay of the previous guarantee node, and improves the dynamic adaptability and timeliness of the prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data prediction, in particular to a flight support time prediction method, device, medium and equipment. BACKGROUND

[0002] Flight ground support time prediction is a core technical support for improving airport operation efficiency and optimizing resource scheduling, and is of great significance to realize precise flight push control and reduce flight delay rate. With the promotion of digital transformation of civil aviation industry, a large amount of data covering flight operation, node execution, environmental state and other multi-dimensional data are generated in the flight support link. However, the existing node time prediction method mainly depends on ground personnel to record and judge by experience through handheld terminal, and more than ten items of nodes such as aircraft entering and leaving the airport, cabin door opening and closing, and various support vehicle scheduling need to be manually tracked. Not only the labor cost is high and the workload is large, but also the problems of missing and false filling are easy to occur due to human operation, which leads to poor data accuracy and low reference value. Especially during the peak period of flights, the support efficiency cannot meet the scheduling demand. Some methods use Markov Monte Carlo and other dynamic simulation methods to model the support process, but the fusion ability of multi-dimensional characteristics is limited, and it is difficult to fully capture the correlation between static attributes and dynamic time sequence. Another method relies on a single deep learning model to process all nodes, ignoring the attribute differences of different support nodes. Some nodes such as boarding start and end have strong time sequence dependence, while some nodes such as wheel block removal and actual take-off are more affected by static characteristics such as aircraft type and stand, and a single model cannot meet the prediction needs of both types of nodes, resulting in large overall prediction error.

[0003] Therefore, how to provide a prediction method capable of modeling the time sequence dynamics and multi-source feature interaction of the support process, so as to realize high-precision and integrated dynamic prediction of the whole process key node has become a problem to be solved. SUMMARY

[0004] In view of the above technical problems, the technical solution adopted by the present application is a flight support time prediction method, which comprises the following steps:

[0005] S1, performing local space feature extraction and weight calibration based on a channel attention mechanism on an initial feature set of a target flight by a feature extraction model to obtain a space feature sequence, wherein the initial feature set is a structured numerical representation constructed for joint modeling of space-time dynamics and resource constraints, which is used to describe the time sequence stage, static attribute and running context of each support node in the target flight.

[0006] S2, capturing the long time sequence dependence of the space feature sequence by a long short-term memory network model to obtain a time sequence context feature.

[0007] S3, according to the characteristic attribute of the i-th guarantee node, selecting a long short-term memory network model or a light gradient boosting machine model to process the dynamic enhanced features of the i-th guarantee node, and obtaining a time prediction value of the i-th guarantee node, wherein i=1, 2, …, N, N is the total number of guarantee nodes, when i=1, the long short-term memory network model is selected to process the time sequence context features, and the characteristic attribute is used to represent the bias of the completion time of the guarantee node to the sequential time sequence mode or to the static feature interaction.

[0008] S4, splicing the time prediction value of the i-th guarantee node and the time sequence context features to obtain dynamic enhanced features of the i+1-th guarantee node.

[0009] S5, if the time prediction values of all guarantee nodes are obtained, a guarantee time prediction sequence of the target flight is formed, otherwise, i=i+1 is updated and the step S3 is returned.

[0010] The application also provides a flight guarantee time prediction device, which comprises:

[0011] The first feature extraction module is used for performing local space feature extraction and weight calibration based on a channel attention mechanism on an initial feature set of the target flight through a feature extraction model, and obtaining a space feature sequence, wherein the initial feature set is a structured numerical representation for describing the time sequence stage, static attribute and running context of each guarantee node in the target flight, and is constructed for joint modeling of space-time dynamics and resource constraints.

[0012] The second feature extraction module is used for capturing long time sequence dependency of the space feature sequence through a long short-term memory network model, and obtaining time sequence context features.

[0013] The time prediction module is used for selecting a long short-term memory network model or a light gradient boosting machine model to process the dynamic enhanced features of the i-th guarantee node according to the characteristic attribute of the i-th guarantee node, and obtaining a time prediction value of the i-th guarantee node, wherein i=1, 2, …, N, N is the total number of guarantee nodes, when i=1, the long short-term memory network model is selected to process the time sequence context features, and the characteristic attribute is used to represent the bias of the completion time of the guarantee node to the sequential time sequence mode or to the static feature interaction.

[0014] The feature splicing module is used for splicing the time prediction value of the i-th guarantee node and the time sequence context features to obtain dynamic enhanced features of the i+1-th guarantee node.

[0015] The iterative updating module is used for forming a guarantee time prediction sequence of the target flight if the time prediction values of all guarantee nodes are obtained, otherwise, i=i+1 is updated and the time prediction module is returned.

[0016] The application further provides a non-transitory computer-readable storage medium, which stores at least one instruction or at least one program, and the at least one instruction or at least one program is loaded and executed by a processor to implement the flight guarantee time prediction method.

[0017] The application further provides an electronic device, which comprises a processor and the non-transitory computer-readable storage medium.

[0018] The application has at least the following beneficial effects: by constructing a multi-dimensional initial feature set and fusing a channel attention mechanism, the feature representation is more accurate and the key information is highlighted; by using a long short-term memory network model to capture long-time sequence dependency, the value of time sequence related features is fully mined, and the ability to capture the delay propagation effect of previous nodes is significantly improved; by using a double-module adaptive selection based on node attributes, the prediction demand of different types of guarantee nodes is accurately matched, the advantages of the long short-term memory network model in capturing time sequence dynamics are utilized, and the ability of the light gradient boosting machine in processing structured feature interaction is also utilized, so that the overall accuracy of the whole process node prediction is greatly improved; by splicing the prediction value of the i-th guarantee node with the time sequence context feature to form a dynamic enhanced feature of the subsequent guarantee node, the transfer and fusion of the previous prediction information to the subsequent node are realized, so that the subsequent guarantee node prediction can adapt to the running state change of the previous guarantee node in real time, the prediction deviation of the subsequent guarantee node caused by the delay of the previous guarantee node is effectively reduced, and the dynamic adaptability and timeliness of the prediction result are improved; by using a cascaded iterative whole node prediction process, the flight guarantee time prediction sequence is more complete and conforms to the actual process, which provides a more valuable decision basis for airport apron monitoring and resource scheduling, and helps to improve the flight operation guarantee efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0019] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0020] Figure 1 A flowchart of a flight guarantee time prediction method provided for the first embodiment of the present application;

[0021] Figure 2 A structural schematic diagram of a flight guarantee time prediction device provided for the second embodiment of the present application. DETAILED DESCRIPTION

[0022] With reference to the drawings and the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of the present application.

[0023] It should be noted that the terms "first", "second" and the like in the specification and claims of the present application and the above-described drawings are used to distinguish similar objects, and do not necessarily indicate a specific order or a chronological sequence. It can be understood that the above-described terms for distinguishing similar objects can be interchanged under appropriate circumstances, so that the present application can also be implemented in other embodiments in addition to the above-described illustrated embodiments or described embodiments. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not necessarily have to include only those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to the process, method, product or device.

[0024] Embodiment one

[0025] The flight guarantee time prediction method provided in the embodiment one includes the following steps, as shown in the following: Figure 1

[0026] S1, performing local space feature extraction and weight calibration based on a channel attention mechanism on an initial feature set of a target flight by a feature extraction model to obtain a space feature sequence, wherein the initial feature set is a structured numerical representation constructed for joint modeling of space-time dynamics and resource constraints, which is used to describe the time sequence stage, static attributes and running context of each guarantee node in the target flight.

[0027] The target flight is a specific flight to be predicted for guarantee time, and its basic identification information such as flight number, departure and arrival airport, and planned departure and arrival time can be determined, which is used to limit the prediction object, ensure the pertinence of subsequent data collection and feature extraction, and avoid irrelevant flight data interference. The guarantee node is a key operation link or state transition point with a clear starting point and completion end point in the ground guarantee whole process of the flight from actual landing to takeoff again, which is divided according to operation logic and scheduling requirements, and is the core object of airport A-CDM (Airport Collaborative Decision Making) system data collection, progress monitoring and resource scheduling, and specifically includes aircraft parking, wheel chock, cabin door opening, boarding start, boarding end, cabin door closing, wheel chock removal, aircraft push-out, actual takeoff, cargo door opening, luggage unloading completion, luggage loading completion, refueling completion, cleaning completion, meal serving completion, etc. ​

[0028] The initial feature set is a structured numerical set for describing the timing phase, static attribute, and running context of each support node of the target flight, which can integrate multi-dimensional information such as time dependence, fixed attribute, and real-time environment, solve the problem of single feature, and provide data support for accurate prediction.

[0029] The timing phase feature is a feature describing the time dimension of the support node, including previous flight delay time, arrival and departure time period, expected transit time, and minimum transit time, which is used to capture the time dynamics of flight operation, such as the transmission effect of previous delay on the current support, and provide time dimension basis for timing modeling. The static attribute feature is a fixed and semi-fixed attribute feature of the flight or airport, including aircraft type based on seat classification, airport size grade, etc., which is used to reflect the inherent differences in support demand, such as the need for longer support time for large aircraft, which is the basis for judging support resource consumption. The running context feature is a feature describing the real-time running environment of the support node, including the state of the pier, flight density, and stand type, which is used to adapt to the dynamic running scenario, such as high flight density during peak hours leading to tight support resources, and improve the adaptability of the model to the actual running environment.

[0030] The feature extraction model is a combined model containing convolution layer units and channel attention units, which is responsible for extracting local spatial features from the initial feature set and calibrating weights, realizing the integration of feature extraction and key enhancement, and improving the automation and accuracy of feature processing. Those skilled in the art know that any feature extraction model in the prior art falls within the protection scope of the present application, for example, a feature extraction model based on Convolutional Neural Network (CNN), which extracts the association pattern between adjacent features (such as the combined effect of aircraft type + stand type) from the initial feature set based on the local perception ability of CNN, thereby capturing the local interaction between features (such as the higher support efficiency of near stand + small aircraft), and improving the representation of feature interaction. The channel attention mechanism is a feature weight calibration mechanism based on the compression excitation module, which dynamically allocates weights by calculating the importance of each feature channel, and automatically identifies key feature channels (such as the previous flight delay time channel), thereby improving its contribution to the prediction result, while suppressing the interference of redundant channels.

[0031] The spatial feature sequence is the feature form output after local spatial feature extraction and weight calibration, which retains the spatial correlation information and key weight distribution of the feature, and provides a feature basis with spatial and weight dual-dimensional enhancement for capturing long time sequence dependence.

[0032] The above-mentioned scattered flight-related data is converted into a unified feature representation through structured processing, solving the problem of fragmentation of multi-source heterogeneous data; then, the channel attention mechanism is used to simulate the cognitive logic of focusing on key information, dynamically calibrating the feature channel weight, highlighting the features that play a core role in guaranteeing time prediction, and suppressing redundant feature interference, providing a feature basis with enhanced space and weight dimensions for subsequent time series modeling.

[0033] In a specific embodiment, S1 includes the following steps:

[0034] S11, obtaining multi-source data of the target flight, wherein the multi-source data includes flight operation basic data, guarantee node execution records, and environmental influence data.

[0035] S12, preprocessing the multi-source data to obtain reference data.

[0036] S13, extracting time sequence stage features, static attribute features, and operation context features from the reference data, and splicing them according to the feature dimension to form an initial feature set.

[0037] Among them, the multi-source data of the target flight is obtained within the time window of "M hours before the planned arrival time" to "M hours after the planned departure time", ensuring coverage of the complete guarantee process. The specific value of M can be set by the implementer according to the actual situation, for example, M = 1 or 2.

[0038] Specifically, the flight operation basic data is data describing the inherent attributes and basic scheduling information of the flight, serving as the static basis for guarantee time prediction, and can reflect the inherent demand differences of flight guarantee, such as different aircraft models leading to different guarantee times, and specifically includes flight number, aircraft model, registration number, estimated arrival time, actual arrival time, terminal and stand allocation information, etc., which can be obtained from the airport departure system and airline scheduling system.

[0039] The guarantee node execution record is data describing the completion time of each key node of the flight ground guarantee, serving as a dynamic core for capturing time sequence dependencies, and can provide time benchmarks for the guarantee process, such as calculating the relative duration of subsequent nodes through the last barrier time, and specifically includes time records of the last barrier, opening the cabin door, boarding start, boarding end, closing the cabin door, removing the barrier, and actual takeoff, etc., which can be obtained from the airport A-CDM system and the hand-held terminal records of ground personnel.

[0040] The environmental influence data is real-time environmental factor data that affects the efficiency of ground guarantee, serving as a key supplement to adapt to dynamic scenarios, and can explain external causes of guarantee time fluctuations, such as increased time consumption for passenger bus docking in strong wind weather, and specifically includes wind speed, wind direction, and temperature corresponding to the target flight arrival period, which can be obtained from the airport weather station real-time database and the air traffic control weather system.

[0041] The multi-source data has problems such as inconsistent format, abnormal values, and large dimension differences, such as time being a string, wind speed being a continuous value, and aircraft model being text, which cannot be directly input into the model for modeling. Through preprocessing operations such as format conversion, feature encoding and normalization, the original data is converted into numerical data recognizable by the model, eliminating the interference of data noise on subsequent prediction. Specifically, time format unification is to convert string type time data such as "08:48" into timestamp format, and calculate the time difference from "protection process start (such as aircraft in position)" to "node completion", such as cabin door node completion time = cabin door time - aircraft in position time. Feature encoding and normalization are to use "target encoding" for discrete classification features such as aircraft model and terminal, and use the average protection time of this category in historical data as the encoding value, and use Z-score normalization for continuous features such as wind speed and temperature to eliminate dimension differences.

[0042] The influencing factors of flight protection time can be divided into three categories: time dynamic factors (time sequence stage), inherent attribute factors (static attributes), and real-time environmental factors (running context). The three factors correspond to different dimensions of prediction logic, specifically, the time sequence stage feature describes the dynamic state of the time dimension of the protection node, which is used to reflect the time dependence relationship and capture the time dynamic influence (such as previous flight delay leading to compression of current protection time), which can include previous flight delay time, arrival and departure time period, expected transit time, minimum transit time, etc. Static attribute features describe the inherent attributes of flights and airports, which do not change dynamically with the protection process, reflecting inherent demand differences (such as large aircraft requiring longer protection time), which can include airport size, departure terminal, arrival terminal, departure stand, and arrival stand. Running context features describe the real-time running environment of the protection node, which changes dynamically with the scene, capturing real-time environmental influences (such as higher efficiency for near-stand protection than for far-stand protection), which can include jetty parking status, flight density, and stand type.

[0043] By classifying and extracting time sequence stage features, static attribute features and running context features from the preprocessed reference data, and concatenating them according to feature dimensions (such as 4-dimensional time sequence features, 2-dimensional static attribute features, and 3-dimensional running context features), a structured feature set with full-dimensional coverage, clear logic, and fixed dimensions is formed, ensuring that the model can capture the influence of different types of factors and avoid the one-sidedness of prediction caused by single-dimensional features.

[0044] The above, through multi-source data collection, makes the data cover flight attributes, progress guarantee, and environmental state full dimensions, solves the problem of single data source and missing key information, and provides complete data support for subsequent feature extraction; through standardized preprocessing and classification extraction of three types of features and splicing, the initial feature set has full-dimensional representation ability of "time sequence dynamic + static inherent + context real-time", solves the problem of one-sided feature dimension and logical confusion, and cooperates to provide comprehensive and accurate initial input for subsequent modeling.

[0045] In a specific embodiment, the feature extraction model includes a convolution layer unit and a channel attention unit connected in sequence, and S1 includes the following steps:

[0046] S14, inputting the initial feature set to the convolution layer unit for local spatial feature extraction to obtain a target channel feature map.

[0047] S15, inputting the target channel feature map to the channel attention unit to obtain weight coefficients of each channel.

[0048] S16, element-wise multiplying the weight coefficients of each channel and the target channel feature map according to the channel dimension to complete feature weight calibration and obtain a spatial feature sequence.

[0049] Among them, part of the influencing factors of flight guarantee need to be combined with features to reflect the actual role (such as the guarantee efficiency of the combination of "near gate + small aircraft" is much higher than that of "far gate + large aircraft"). And the convolution layer unit is the core component of the feature extraction model responsible for local spatial feature extraction, which is composed of multiple layers of convolution kernel, activation function and pooling layer, which is used to capture the interaction relationship of adjacent dimensions in the initial feature (such as the combination of "previous flight delay time + flight density"), and generate a target channel feature map containing deep interaction information, wherein each channel corresponds to a local spatial feature mode (such as channel 1 corresponds to "previous flight delay + near gate" mode, and channel 2 corresponds to "flight density + small aircraft" mode), as the input of the channel attention unit, providing multi-mode and multi-dimensional feature basis for weight calibration.

[0050] In this embodiment, the convolution layer unit includes two convolution structures, specifically, the first layer of convolution is convolved with 16 convolution kernels with a size of (3, 1) on the initial feature set, and outputs a 16-channel feature map through a ReLU activation function. The second layer of convolution is convolved with 32 convolution kernels with a size of (3, 1) on the 16-channel feature map, and outputs a 32-channel feature map through a ReLU activation function, that is, a target channel feature map.

[0051] Among all the channels of the target channel feature map, the contribution of different channels to the guarantee time prediction is significantly different (for example, the influence of the "previous flight delay time + aircraft model" channel is much greater than that of the "temperature + wind direction" channel). The channel attention unit is a core component responsible for feature weight calibration in the feature extraction model, which is constructed based on a Squeeze-and-Excitation (SE) module, including a global average pooling layer, a fully connected layer, and an activation function, for dynamically calculating the importance of each channel, generating weight coefficients, and realizing key channel strengthening and redundant channel suppression. Specifically, the target channel feature map is first globally averaged to obtain channel-dimension feature statistics, and then the feature statistics are compressed and excited through the fully connected layer to generate weight coefficients for each channel, thereby identifying the channels that play a key role in prediction and providing a quantitative basis for subsequent weight calibration to avoid interference from redundant channels. The greater the weight coefficient, the more significant the influence of the corresponding channel on the prediction result.

[0052] The weight coefficients output by the channel attention unit have quantified the importance of each channel. By element-wise multiplication (for example, the cth coefficient of the weight vector is multiplied by all elements of the cth channel of the feature map), the weight is assigned to the corresponding channel of the target channel feature map, which can achieve the effect of strengthening key channel features and weakening redundant channel features. The finally generated spatial feature sequence not only retains local interaction information but also highlights key influencing factors, providing high-quality input for the subsequent long short-term memory network model.

[0053] As described above, the local interaction patterns between features are captured through the convolution layer unit, solving the problem that isolated features cannot reflect combined effects. Then, the importance of each feature channel is calculated through the channel attention unit, solving the defect of equalization of feature weights. Finally, key feature strengthening is achieved through weight assignment, and a high-quality feature sequence with spatial interaction information and key weight distribution is output, providing accurate input for subsequent capture of temporal dependency.

[0054] S2, the long short-term memory network model is used to capture the long temporal sequence dependency of the spatial feature sequence, and the temporal context features are obtained.

[0055] Among all the channels of the target channel feature map, the contribution of different channels to the guarantee time prediction is significantly different (for example, the influence of the "previous flight delay time + aircraft model" channel is much greater than that of the "temperature + wind direction" channel). The channel attention unit is a core component responsible for feature weight calibration in the feature extraction model, which is constructed based on a Squeeze-and-Excitation (SE) module, including a global average pooling layer, a fully connected layer, and an activation function, for dynamically calculating the importance of each channel, generating weight coefficients, and realizing key channel strengthening and redundant channel suppression. Specifically, the target channel feature map is first globally averaged to obtain channel-dimension feature statistics, and then the feature statistics are compressed and excited through the fully connected layer to generate weight coefficients for each channel, thereby identifying the channels that play a key role in prediction and providing a quantitative basis for subsequent weight calibration to avoid interference from redundant channels. The greater the weight coefficient, the more significant the influence of the corresponding channel on the prediction result.

[0056] The Long Short-Term Memory Network (LSTM) is a core component responsible for capturing time-dependent relationships, which is composed of multiple LSTM units and a fully connected layer. Through the three gate control structures of the "input gate, forget gate, and output gate", the updating and forgetting of the cell state can be dynamically adjusted, which can not only retain key historical dependent information (such as previous flight delays), but also eliminate irrelevant temporal noise (such as temporary wind speed fluctuations), and accurately model long time sequence dependencies.

[0057] The skilled person in the art knows that the LSTM model in the prior art falls within the protection scope of the present application. For example, in the present embodiment, the LSTM model is configured as a two-layer hidden unit structure, the first layer captures short-term local time sequence association (such as feature interaction between adjacent time steps), and the second layer captures long-term global time sequence association (such as delay transmission across multiple time steps). Each layer contains 64 LSTM units, and the activation function uses tanh (the output value is mapped to [-1, 1], which adapts to the fluctuation range of time sequence features). Details are not repeated here.

[0058] As described above, the LSTM model is used to accurately capture the dependency relationship across multiple time steps in the spatial feature sequence, and the dispersed spatial features and time association are fused into unified time sequence context features, providing spatio-temporal collaborative feature support for subsequent security node prediction.

[0059] S3, according to the feature attribute of the i-th security node, selecting a Long Short-Term Memory Network model or a Light Gradient Boosting Machine model to process the dynamic enhancement features of the i-th security node, and obtaining a time prediction value of the i-th security node, wherein i = 1, 2, …, N, N is the total number of security nodes, when i = 1, the Long Short-Term Memory Network model is selected to process the time sequence context features, and the feature attribute is used to represent the bias of the completion time of the security node to the sequential time sequence pattern or to the static feature interaction.

[0060] Among them, the N security nodes of the flight ground security have significant attribute differences: the completion time of some security nodes (such as boarding start, boarding end, and cabin door closing) is highly dependent on the time sequence transmission of the previous security nodes (such as cabin door opening delay will inevitably lead to boarding start delay), and the time sequence dependency needs to be captured; the completion time of another part of the security nodes (such as wheel blocking removal and actual take-off) is more affected by static feature interactions such as aircraft type, aircraft position, and flight density (such as large aircraft + remote position requires longer wheel blocking removal preparation time), and the feature combination rule needs to be modeled.

[0061] If all the guarantee nodes are processed by a single model, it will lead to insufficient modeling of time sequence nodes or missing of static node interaction capture. Therefore, model selection is performed through attribute judgment to match the optimal model for different guarantee nodes, and dynamic enhancement features are used to deliver the previous sequence prediction information to achieve accurate prediction.

[0062] Specifically, the i-th guarantee node is the i-th key node arranged in the operation sequence in the flight ground guarantee process, N is the total number of nodes, and in this embodiment, N=7, i=1 for the upper chock, i=2 for the cabin door opening, i=3 for the boarding start, i=4 for the boarding end, i=5 for the cabin door closing, i=6 for the chock removal, and i=7 for the actual takeoff.

[0063] The long short-term memory network model (LSTM model) is good at capturing long time sequence dependency and has a gating mechanism, which can process the dynamic conduction demand of time sequence type nodes. It is used to predict time sequence mode biased nodes (such as boarding start and boarding end) and accurately capture the conduction influence of previous node delay on the current node. The Light Gradient Boosting Machine model (LightGBM) is a traditional machine learning model based on decision tree integration, which is good at processing high-dimensional structured feature interaction and is used to predict static feature interaction biased nodes (such as chock removal and actual takeoff) and accurately model the combined influence of features such as aircraft type and stand.

[0064] Any LightGBM model in the prior art falls within the protection scope of the present application, for example, the LightGBM model with parameters learning rate=0.05, leaf number=31, and tree number=200 in this embodiment, which will not be described here.

[0065] The feature attribute is an attribute label used to distinguish the guarantee node “time sequence dependency bias” or “static feature interaction bias”, which is determined by the dominant type of the influencing factors of the guarantee node completion time, and is used to ensure that the LSTM model is adapted to the time sequence type node and the LightGBM model is adapted to the static interaction type node.

[0066] The dynamic enhancement feature is an input feature set constructed for the i-th guarantee node prediction, which fuses “previous node prediction information + basic features”, wherein i=1 is the time sequence context feature, and i≥2 includes the previous node prediction value and the time sequence context feature, which is used to deliver the running state information of the previous node to avoid the deviation caused by isolated prediction and improve the dynamic adaptability of prediction.

[0067] The time prediction value is a prediction result of the completion time of the i-th guarantee node, in units of min, and is usually a relative time difference from the start of guarantee to the completion of the node, which is a core component of the dynamic enhancement feature of the subsequent node and provides an accurate reference for the completion time of the node for airport scheduling.

[0068] In a specific embodiment, the feature attribute is a time sequence pattern bias type or a static feature interaction bias type, and S3 includes the following steps:

[0069] When i≥2, if the feature attribute of the i-th guarantee node is a time sequence pattern bias type, a long short-term memory network model is selected.

[0070] If the feature attribute of the i-th guarantee node is a static feature interaction bias type, a light gradient boosting machine model is selected.

[0071] Before performing prediction, based on historical guarantee data and industry experience, the feature attribute determination rule of each guarantee node can be preset to form a "node-attribute-model" mapping relationship, avoiding subjective judgment deviation in the prediction process.

[0072] Specifically, the completion time of the upper wheel block is highly dependent on the aircraft parking time, the completion time of opening the cabin door is dependent on the completion time of the upper wheel block, the completion time of starting boarding is dependent on the completion time of opening the cabin door, the completion time of ending boarding is dependent on the starting boarding time and the time sequence change of passenger flow, and the completion time of closing the cabin door is dependent on the completion time of ending boarding, so the attribute types of the upper wheel block, the opening of the cabin door, the starting boarding, the ending boarding, and the closing of the cabin door are time sequence pattern bias types, and the adaptive model is LSTM. The completion time of removing the wheel block is influenced by the interaction of the aircraft type (long preparation for large aircraft), the stand type (more vehicles need to be dispatched for remote stands), and the flight density (resource shortage during peak hours), and the actual takeoff completion time is influenced by the interaction of the aircraft type (difference in takeoff sliding distance), the flight density (air traffic control), and the wind direction (adjustment of takeoff parameters for headwinds), so the attribute types of removing the wheel block and actual takeoff are static feature interaction bias types, and the adaptive model is LightGBM.

[0073] If the i-th guarantee node is a time sequence pattern bias node, the dynamic enhancement features of the i-th guarantee node are input into the fully connected output layer of the LSTM model, and the time prediction value of the i-th guarantee node is calculated by forward propagation.

[0074] If the i-th guarantee node is a static feature interaction bias node, the dynamic enhancement features of the i-th guarantee node are input into the LightGBM model, the model first performs histogram binning on the dynamic enhancement features, and then obtains the time prediction value of the i-th guarantee node by weighted summation of the split gains of multiple decision trees in the pre-trained decision tree ensemble model.

[0075] The above, through the adaptive selection of the double model based on the feature attribute, the differentiated prediction needs of the time sequence type node and the static interaction type node are accurately matched, the defects that a single model cannot consider the characteristics of the two types of nodes are solved, the LSTM model fully plays the time sequence dependence capture advantage, the LightGBM model fully plays the feature interaction modeling advantage, and the prediction accuracy of all nodes is improved; the dynamic enhancement features are constructed by fusing the prediction values of the previous nodes when i≥2, so that the subsequent node prediction can adapt to the state change of the previous node in real time, the problems of isolated prediction and inability to transmit previous information are solved, and the prediction deviation caused by delay transmission is reduced.

[0076] In a specific embodiment, S3 includes the following steps:

[0077] S31, when i≥2, calculating the time correlation degree according to the historical completion time sequence of the i-th guarantee node and the historical completion time sequence of the i-1-th guarantee node.

[0078] S32, if the time correlation degree is greater than a preset correlation degree threshold, determining that the feature attribute of the i-th guarantee node is time sequence mode biased.

[0079] S33, obtaining the contribution degree of the completion time of the i-th guarantee node affected by the static feature interaction, wherein the contribution degree is used to quantify the influence strength of the interaction of the static feature on the completion time of the current node.

[0080] S34, if the contribution degree is greater than a preset contribution degree threshold, determining that the feature attribute of the i-th guarantee node is static feature interaction biased.

[0081] The core feature of the time sequence mode biased node is that the completion time of the current node highly depends on the completion time of the previous node, and the time sequence dependence relationship can be quantified by the correlation of the completion time sequences of the two guarantee nodes in the historical data.

[0082] The historical completion time sequence is extracted from the historical database of the airport A-CDM system, and contains the complete guarantee records of the i-th and i-1-th nodes in the past year, forming two equal-length time sequences, providing data basis for time correlation degree calculation, and ensuring that the result is based on real guarantee rules rather than subjective assumptions.

[0083] The time correlation degree is an index for quantifying the linear correlation strength of the completion time of the i-th guarantee node and the i-1-th guarantee node, for example, calculating the Pearson correlation coefficient of the two historical completion time sequences, the larger the value, the closer the time sequence dependence of the two nodes, the more significant the time sequence conduction influence of the previous guarantee node on the current guarantee node, and thus the current guarantee node needs to capture the time sequence dependence.

[0084] The core feature of the static feature interaction bias type node is that the completion time is significantly affected by the combination of static features such as machine type, machine position type, and flight density. The contribution degree is quantified by the feature importance and interaction gain of the LightGBM model, reflecting the total explanatory power of static features and combinations on the node completion time. Correspondingly, the higher the contribution degree, the more dominant the influence of static feature interaction.

[0085] Specifically, in the embodiment, the historical completion time T i of the ith guarantee node is taken as the prediction label, and the input feature set X (such as the aircraft model code 1-5, the machine position type code 0 / 1, and the aircraft model x machine position type code 10 / 11) is taken as the “basic static feature + feature interaction term”; the model parameters are set as learning rate = 0.05, leaf number = 31, and tree number = 200, the feature set X is divided into training set and validation set according to 8:2, and the LightGBM model is trained to convergence; the split gain importance is extracted from the trained LightGBM model, including the gain Gain f of the basic static feature (such as the gain of the aircraft model is 25, and the gain of the machine position type is 18) and the gain Gain int of the feature interaction term (such as the gain of the aircraft model x machine position type is 12); the influence of the static feature and the interaction term is integrated by the formula to obtain the contribution degree C = ((Σ f∈FGainf )+ Gain int ) / (Σ all Gain all ), wherein F is the static feature set, Σ f∈FGainf is the total gain of the basic static feature, Gain int is the gain of the interaction term, Σ all Gain all is the total gain of all features in the model; finally, the contribution degree C is normalized to the interval [0, 1] to obtain the final contribution degree.

[0086] The specific values of the preset correlation degree threshold and the preset contribution degree threshold can be determined based on industry experience and experimental verification. For example, in the embodiment, the preset correlation degree threshold is 0.7, and the preset contribution degree threshold is 0.6.

[0087] S4, splicing the time prediction value of the ith guarantee node and the time sequence context feature to obtain the dynamic enhancement feature of the (i+1)th guarantee node.

[0088] The flight support nodes are in series-parallel dependency relationship, and the completion time of the previous support node directly affects the progress of the subsequent support node. If the subsequent support node prediction only uses the initial static features, the dynamic running state of the previous support node will be ignored, resulting in prediction deviation. Therefore, by splicing the time prediction value of the support node and the timing context feature, the dynamic information of the previous support node is injected into the subsequent node input, so that the feature has real-time updating capability and conforms to the dynamics of the support process.

[0089] Specifically, the time prediction value of the i-th node is a 1-dimensional scalar, which is converted into a 1-dimensional vector; the timing context feature is a fixed dimension vector (such as 64 dimensions), and the dimensions can be spliced into 1+64=65 dimensions. The dynamically enhanced features after splicing are subjected to Min-Max normalization, which is mapped to the [0, 1] interval to avoid too large differences in the numerical ranges of the previous prediction value and the timing context feature, which affects the calculation accuracy of the model.

[0090] The above splices the time prediction value of the previous support node and the timing context feature, so that the subsequent support node prediction can adapt to the dynamic state of the previous node in real time, solving the problem that the traditional isolated prediction cannot cope with the previous delay, and reducing the deviation caused by delay transmission.

[0091] S5, if the time prediction values of all support nodes are obtained, a support time prediction sequence of the target flight is formed, otherwise, i=i+1 is updated and step S3 is returned.

[0092] The flight support nodes are in the fixed order of the last round of stops, opening the cabin door, boarding, closing the cabin door, removing the round of stops, and actual takeoff, which need to be predicted in sequence according to the serial number i until all N support nodes are completed. Therefore, through the iteration judgment and serial number updating mechanism, the prediction of each support node is based on the dynamic information of the previous support node, and finally the prediction sequence covering the whole process is output, which conforms to the actual needs of the airport phased scheduling.

[0093] The above completes the whole node prediction through iterative control, so that the support time prediction sequence conforms to the actual support process of the node in sequence, which is convenient for dispatch personnel to monitor the progress in stages and improves the scheduling flexibility.

[0094] In a specific embodiment, the flight support time prediction method further includes the following steps:

[0095] S6, according to the time prediction value and the true value of all support nodes, the model evaluation index is calculated.

[0096] S7, if the model evaluation index does not meet the preset condition, adjusting the parameters of the feature extraction model, the long short-term memory network model and the light gradient boosting machine model according to the time prediction value and the true value of all guarantee nodes, and re-executing step S1 using the adjusted model until the model evaluation index meets the preset condition, and obtaining the final guarantee time prediction sequence.

[0097] wherein the true value is the actual completion time of the i-th guarantee node obtained from the airport A-CDM system.

[0098] The model evaluation index is the core basis for quantifying the prediction accuracy, and needs to be calculated through the error between the prediction value and the true value, and intuitively reflects the overall performance of the model in the whole node prediction, such as mean absolute error and interval prediction accuracy. The person skilled in the art knows that the mean absolute error and the interval prediction accuracy fall within the protection scope of the present application, and will not be repeated here.

[0099] The preset condition is an evaluation index threshold set based on the airport scheduling requirements, such as mean absolute error ≤ 3.64 min, interval prediction accuracy ≥ 85%, etc., which is a qualified standard for judging whether the model is usable. The feature extraction model, the long short-term memory network model and the light gradient boosting machine model may not meet the evaluation index due to unreasonable parameter setting (such as too high LSTM learning rate) or insufficient scene adaptation (such as not covering extreme weather data). Therefore, the core parameters of each model are adjusted by error, for example, adjusting the convolution channel number of the feature extraction model (such as increasing from 32 to 64 to enhance the feature extraction capability), the hidden unit number of the LSTM model (such as increasing from 64 to 128 to improve the time series dependence capture ability), the tree number of the LightGBM model (increasing from 200 to 300 to improve the feature interaction modeling precision), and re-executing the prediction process until the index meets the preset condition, realizing the iterative improvement of the model precision.

[0100] Through the above multi-index evaluation and iterative optimization of error and parameters, each model can dynamically adapt to scene changes, solving the problem of initial model parameter solidification and insufficient precision, and finally outputting a precise prediction sequence that meets the airport requirements.

[0101] The above, by constructing a multi-dimensional initial feature set and fusing a channel attention mechanism, the feature representation is more accurate and the key information is highlighted; by capturing long-term sequence dependency relationship through the long short-term memory network model, the value of time sequence related features is fully mined, and the ability to capture the delay conduction effect of the previous node is significantly improved; through the adaptive selection of the double module based on the node attribute, the prediction demand of different types of support nodes is accurately matched, which not only takes advantage of the long short-term memory network model to capture the time sequence dynamic, but also uses the ability of the light gradient boosting machine to process structured feature interaction, greatly improving the overall accuracy of the whole process node prediction; by splicing the prediction value of the ith support node with the time sequence context feature to form the dynamic enhancement feature of the subsequent support node, the transfer and fusion of the previous prediction information to the subsequent node are realized, so that the subsequent support node prediction can adapt to the running state change of the previous support node in real time, effectively reducing the prediction deviation caused by the delay of the previous support node, and improving the dynamic adaptability and timeliness of the prediction result; through the cascaded iterative whole node prediction process, the flight support time prediction sequence is more complete and conforms to the actual process, providing more valuable decision basis for airport apron monitoring and resource scheduling, and helping to improve the efficiency of flight operation support.

[0102] Embodiment Two

[0103] The embodiment two provides a flight support time prediction device, which comprises, as shown in Figure 2

[0104] The first feature extraction module 21 is configured to perform local spatial feature extraction and weight calibration based on a channel attention mechanism on the initial feature set of the target flight by a feature extraction model, and obtain a spatial feature sequence, wherein the initial feature set is a structured numerical representation for describing the time sequence stage, static attribute and running context of each support node in the target flight, and is constructed for joint modeling of space-time dynamics and resource constraints.

[0105] The second feature extraction module 22 is configured to capture the long time sequence dependency relationship of the spatial feature sequence by a long short-term memory network model, and obtain a time sequence context feature.

[0106] The time prediction module 23 is configured to select the long short-term memory network model or the light gradient boosting machine model to process the dynamic enhancement feature of the ith support node according to the feature attribute of the ith support node, and obtain the time prediction value of the ith support node, wherein i=1, 2, …, N, N is the total number of support nodes, when i=1, the long short-term memory network model is selected to process the time sequence context feature, and the feature attribute is used to represent the time sequence mode of the completion time of the support node or the bias of the static feature interaction. ​

[0107] The feature splicing module 24 is used to splice the time prediction value and temporal context features of the i-th protection node to obtain the dynamic enhancement features of the (i+1)-th protection node.

[0108] The iterative update module 25 is used to form the guarantee time prediction sequence for the target flight if the time prediction values ​​of all guarantee nodes are obtained; otherwise, it updates i=i+1 and returns to the execution time prediction module.

[0109] In one specific embodiment, the first feature extraction module 21 includes:

[0110] The data acquisition submodule is used to acquire multi-source data of the target flight, including basic flight operation data, support node execution records, and environmental impact data.

[0111] The data preprocessing submodule is used to preprocess multi-source data to obtain reference data.

[0112] The first feature extraction submodule is used to extract time-series features, static attribute features, and runtime context features from the reference data, and then concatenate them according to the feature dimensions to form an initial feature set.

[0113] In one specific embodiment, the feature extraction model includes convolutional layer units and channel attention units connected in sequence, and the first feature extraction module 21 includes:

[0114] The second feature extraction submodule is used to input the initial feature set into the convolutional layer unit to extract local spatial features and obtain the target channel feature map.

[0115] The weight coefficient acquisition submodule is used to input the target channel feature map into the channel attention unit to obtain the weight coefficients of each channel.

[0116] The third feature extraction submodule is used to multiply the weight coefficients of each channel with the target channel feature map element by element along the channel dimension to complete the feature weight calibration and obtain the spatial feature sequence.

[0117] In one specific implementation, the feature attributes are either time-series pattern-biased or static feature interaction-biased, and the time prediction module 23 includes:

[0118] The first model selection submodule is used to select the Long Short-Term Memory network model when i≥2, if the feature attribute of the i-th guaranteed node is time-series pattern biased.

[0119] The second model selection submodule is used to select the lightweight gradient booster model if the feature attribute of the i-th guaranteed node is a static feature interaction bias.

[0120] In a specific embodiment, the time prediction module 23 comprises:

[0121] The time correlation calculation submodule is configured to calculate the time correlation according to the historical completion time sequence of the i-th guarantee node and the historical completion time sequence of the (i-1)-th guarantee node when i is greater than or equal to 2.

[0122] The first feature attribute judgment submodule is configured to determine that the feature attribute of the i-th guarantee node is time sequence mode biased if the time correlation is greater than a preset correlation threshold.

[0123] The contribution degree acquisition submodule is configured to acquire a contribution degree of the i-th guarantee node to the influence of the static feature interaction on the completion time, wherein the contribution degree is used to quantify the influence strength of the interaction of the static feature on the completion time of the current node.

[0124] The second feature attribute judgment submodule is configured to determine that the feature attribute of the i-th guarantee node is static feature interaction biased if the contribution degree is greater than a preset contribution degree threshold.

[0125] In a specific embodiment, the flight guarantee time prediction device further comprises:

[0126] The model evaluation index calculation module is configured to calculate the model evaluation index according to the time prediction values and the true values of all the guarantee nodes.

[0127] The model updating module is configured to adjust the parameters of the feature extraction model, the long short-term memory network model and the light gradient boosting machine model according to the time prediction values and the true values of all the guarantee nodes if the model evaluation index does not satisfy a preset condition, and to re-execute the first feature extraction module 21 using the adjusted model until the model evaluation index satisfies the preset condition and the final guarantee time prediction sequence is obtained.

[0128] It should be noted that the information interaction, execution process and the like between the above modules are based on the same concept as the method embodiments of the present application, and the specific functions and the technical effects brought by the same can be referred to the method embodiments part, which will not be described here.

[0129] Embodiment three

[0130] Embodiment three of the present application provides a non-transitory computer readable storage medium, the non-transitory computer readable storage medium stores at least one instruction or at least one program, at least one instruction or at least one program is loaded and executed by the processor to execute the steps in embodiment one.

[0131] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by a computer program instructing relevant hardware, and the computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, the processes of the above-mentioned embodiments of each method can be included. Any reference to memory, storage, databases, or other media in the embodiments provided by the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronization link (Synchronization Link) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0132] Those skilled in the art can clearly understand that, for the convenience and brevity of description, only the above-mentioned division of each functional unit and module is exemplified, and in actual application, the above-mentioned functions can be completed by different functional units and modules according to needs, that is, the internal structure of the device is divided into different functional units or modules to complete all or part of the above-described functions.

[0133] Embodiment four

[0134] Embodiment four of the present application provides an electronic device, which includes a processor and the non-transitory computer readable storage medium in embodiment three of the present application.

[0135] The above is only the preferred embodiment of the present application, and does not limit the present application in any form. Although the present application has been disclosed as the above preferred embodiment, it is not intended to limit the present application. Any skilled person in the art can make some changes or modifications to the above-mentioned disclosed technical content to obtain equivalent embodiments with equivalent changes, without departing from the technical solution of the present application. Any simple modification, equivalent change and modification of the above-mentioned embodiments according to the technical essence of the present application are still within the scope of the technical solution of the present application.

Claims

1. A method for predicting flight support time, characterized in that, The flight support time prediction method includes: S1, Local spatial features are extracted from the initial feature set of the target flight through a feature extraction model and weight calibration is performed based on the channel attention mechanism to obtain a spatial feature sequence. The initial feature set is a structured numerical representation used to describe the temporal stage, static attributes and operating context of each support node in the target flight, and is constructed for joint modeling of spatiotemporal dynamics and resource constraints. S2, The long-term sequence dependency of the spatial feature sequence is captured by the long short-term memory network model to obtain the temporal context features; S3, based on the feature attributes of the i-th guarantee node, select a long short-term memory network model or a lightweight gradient booster model to process the dynamic enhancement features of the i-th guarantee node, and obtain the time prediction value of the i-th guarantee node, where i=1,2,...,N, and N is the total number of guarantee nodes. When i=1, the dynamic enhancement feature is the temporal context feature, and the long short-term memory network model is selected to process the dynamic enhancement feature. When i≥2, the dynamic enhancement feature is the feature obtained by splicing the time prediction value corresponding to the (i-1)-th guarantee node and the temporal context feature. The feature attributes are used to characterize the bias of the completion time of the guarantee node towards sequential temporal patterns or towards static feature interactions. S4, concatenate the time prediction value of the i-th protection node and the temporal context feature to obtain the dynamic enhancement feature of the (i+1)-th protection node; S5. If the time prediction values ​​of all support nodes are obtained, the support time prediction sequence of the target flight is formed; otherwise, i is updated to i+1 and the process returns to step S3.

2. The flight support time prediction method according to claim 1, characterized in that, S1 includes the following steps: S11, Obtain multi-source data of the target flight, wherein the multi-source data includes basic flight operation data, support node execution records and environmental impact data; S12, preprocess the multi-source data to obtain reference data; S13, extract time-series features, static attribute features, and runtime context features from the reference data, and concatenate them according to feature dimensions to form the initial feature set.

3. The flight support time prediction method according to claim 2, characterized in that, The feature extraction model includes sequentially connected convolutional layer units and channel attention units. S1 includes the following steps: S14, the initial feature set is input into the convolutional layer unit to extract local spatial features and obtain the target channel feature map; S15, Input the target channel feature map into the channel attention unit to obtain the weight coefficients of each channel; S16, multiply the weight coefficients of each channel element-wise with the target channel feature map according to the channel dimension to complete the feature weight calibration and obtain the spatial feature sequence.

4. The flight support time prediction method according to claim 1, characterized in that, If the feature attributes are time-series pattern-biased or static feature interaction-biased, S3 includes the following steps: When i≥2, if the feature attribute of the i-th guaranteed node is time-series pattern biased, then the long short-term memory network model is selected. If the feature attribute of the i-th guaranteed node is static feature interaction biased, then the lightweight gradient booster model is selected.

5. The flight support time prediction method according to claim 4, characterized in that, S3 includes the following steps: S31, When i≥2, calculate the time correlation degree based on the historical completion time series of the i-th guarantee node and the historical completion time series of the (i-1)-th guarantee node; S32, if the time correlation degree is greater than the preset correlation degree threshold, then the feature attribute of the i-th guarantee node is determined to be time-series pattern biased. S33, obtain the contribution of the completion time of the i-th guaranteed node to the static feature interaction, wherein the contribution is used to quantify the influence of the static feature interaction on the completion time of the current node. S34, if the contribution is greater than the preset contribution threshold, then the feature attribute of the i-th guarantee node is determined to be static feature interaction bias type.

6. The flight support time prediction method according to claim 1, characterized in that, The flight support time prediction method also includes the following steps: S6. Calculate the model evaluation index based on the predicted and actual time values ​​of all support nodes. S7. If the model evaluation index does not meet the preset conditions, adjust the parameters of the feature extraction model, the long short-term memory network model and the lightweight gradient booster model according to the time prediction and actual values ​​of all guarantee nodes, and re-execute step S1 using the adjusted model until the model evaluation index meets the preset conditions and the final guarantee time prediction sequence is obtained.

7. A flight support time prediction device, characterized in that, The flight support time prediction device includes: The first feature extraction module is used to extract local spatial features from the initial feature set of the target flight through a feature extraction model and perform weight calibration based on the channel attention mechanism to obtain a spatial feature sequence. The initial feature set is a structured numerical representation used to describe the temporal stage, static attributes and operating context of each support node in the target flight, and is constructed for joint modeling of spatiotemporal dynamics and resource constraints. The second feature extraction module is used to capture the long-term sequence dependencies of the spatial feature sequence through a long short-term memory network model to obtain temporal context features; The time prediction module is used to select a Long Short-Term Memory Network model or a Lightweight Gradient Boosting Machine model to process the dynamic enhancement features of the i-th guarantee node based on the feature attributes of the i-th guarantee node, and obtain the time prediction value of the i-th guarantee node, where i=1,2,...,N, and N is the total number of guarantee nodes. When i=1, the dynamic enhancement feature is the temporal context feature, and the Long Short-Term Memory Network model is selected to process the dynamic enhancement feature. When i≥2, the dynamic enhancement feature is the feature obtained by concatenating the time prediction value corresponding to the (i-1)-th guarantee node and the temporal context feature. The feature attributes are used to characterize the bias of the completion time of the guarantee node towards sequential temporal patterns or towards static feature interactions. The feature splicing module is used to splice the time prediction value of the i-th protection node and the temporal context feature to obtain the dynamic enhancement feature of the (i+1)-th protection node; The iterative update module is used to form a support time prediction sequence for the target flight if the time prediction values ​​of all support nodes are obtained; otherwise, it updates i=i+1 and returns to execute the time prediction module.

8. The flight support time prediction device according to claim 7, characterized in that, The first feature extraction module includes: The data acquisition submodule is used to acquire multi-source data of the target flight, wherein the multi-source data includes basic flight operation data, support node execution records and environmental impact data; The data preprocessing submodule is used to preprocess the multi-source data to obtain reference data; The first feature extraction submodule is used to extract time-series features, static attribute features, and runtime context features from the reference data, and to concatenate them according to feature dimensions to form the initial feature set.

9. A non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores at least one instruction or at least one program segment, characterized in that, The at least one instruction or the at least one program segment is loaded and executed by the processor to implement the flight support time prediction method as described in any one of claims 1-6.

10. An electronic device, characterized in that, Includes a processor and the non-transitory computer-readable storage medium as described in claim 9.

Images