Aviation internet traffic and bandwidth prediction method and system based on deep learning
By using deep learning models to predict flight traffic and bandwidth, the problem of accurately determining passengers' internet access needs for flights has been solved, enabling the rational allocation of network resources and improving passenger experience.
Patent Information
- Application Number
- CN202511137359.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-14
- Publication Date
- 2025-11-07
AI Technical Summary
Flights cannot accurately determine passengers' internet access needs, leading to wasted or insufficient bandwidth resources and affecting passengers' internet experience.
A progressive two-stage time-series prediction model for aviation internet traffic and bandwidth based on deep learning is adopted. It combines multi-dimensional flight information, historical open IP number sequences, and historical traffic and bandwidth sequences. The prediction is performed through a Transformer-LSTM hybrid neural network and a time-series block Transformer architecture. Real-time hybrid update and MC Dropout mechanisms are introduced to generate traffic and bandwidth prediction results.
It enables accurate prediction of traffic and bandwidth at future time points during flight, avoiding network congestion or resource idleness, improving network service quality and operational efficiency, ensuring communication link stability and reducing operating costs.
Smart Images
Figure CN120915682A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of aviation Internet, and in particular to an aviation Internet traffic and bandwidth prediction method and system based on deep learning. BACKGROUND
[0002] During flight, in order to avoid the interference of network signals of passengers' electronic devices on the aircraft, passengers usually need to set the electronic devices to be powered off or in a flight state, but this makes passengers unable to communicate with the outside during the flight, resulting in passengers missing important events. At the same time, due to the inability of the electronic devices to be connected to the network, the electronic devices lose the entertainment attribute, making the flight more boring.
[0003] To solve this problem, in the prior art, there is a scheme of providing network services to passengers by network devices carried by the aircraft, so that passengers can realize Internet connection by linking the network service functions provided by the aircraft.
[0004] However, the aircraft cannot accurately determine the demand of all passengers for traffic, resulting in that the bandwidth provided by the aircraft is either too large, causing resource waste, or too small, causing resource shortage, resulting in poor network use experience of passengers. SUMMARY
[0005] In view of the problem that the demand of passengers for network connection cannot be determined during flight in the prior art, the present application aims to provide an aviation Internet traffic and bandwidth prediction method and system based on deep learning, so as to at least partially solve the above problems.
[0006] To achieve the above-mentioned purpose, the technical scheme of the present application is as follows: In a first aspect, the present application provides an aviation Internet traffic and bandwidth prediction method based on deep learning, which comprises the following steps: obtaining multi-dimensional flight information, historical open network IP number sequence and historical traffic and bandwidth sequence, wherein the multi-dimensional flight information includes route information, model information and seat occupancy rate information; processing the multi-dimensional flight information, the historical open network IP number sequence and the historical traffic and bandwidth sequence by a progressive two-stage aviation Internet traffic and bandwidth time sequence prediction model to obtain traffic and bandwidth data at a future time node; The progressive two-stage aviation Internet traffic and bandwidth time sequence prediction model comprises a first-stage open network IP number prediction model and a second-stage traffic and bandwidth prediction model.
[0007] In some preferred embodiments, the open network IP number prediction model adopts a hybrid neural network architecture based on Transformer-LSTM, and after splicing the preprocessed multi-dimensional flight information and the historical open network IP number sequence, the sequence is input into a Transformer encoder to extract global context, and the LSTM network outputs the open network IP number prediction result of the future time node; The traffic and bandwidth prediction model adopts a Transformer architecture based on time series blocking, and by processing the open network IP number of the future time node obtained by prediction and the historical traffic and bandwidth sequence, the traffic and bandwidth data of the future time node are predicted.
[0008] In some preferred embodiments, the historical open network IP number sequence and the historical traffic and bandwidth sequence are screened from the flight history database based on the multi-dimensional flight information; And during the flight of the flight, a real-time hybrid prediction updating mechanism is introduced, the real open network IP number sequence and the real traffic and bandwidth sequence that have occurred are respectively input into the progressive two-stage aviation Internet traffic and bandwidth time series prediction model instead of the historical open network IP number sequence and the historical traffic and bandwidth sequence, and then the traffic and bandwidth data that have not occurred are dynamically predicted and updated.
[0009] In some preferred embodiments, the pre-processing process of the multi-dimensional flight information includes normalization and numericalization, and then a learnable mapping is used to obtain an input vector with the same dimension as the historical open network IP number sequence.
[0010] In some preferred embodiments, the route information obtains a route semantic information embedding vector through a multi-dimensional weighted embedding method, including: Based on the route information, static data corresponding to the airports at both ends of the route are obtained, and the static data includes the economic conditions of the country where the airport is located, the economic conditions of the city where the airport is located, the positioning of the city where the airport is located, and the passenger throughput of the airport; The numerical indicators in the static data are normalized, and the categorical indicators in the static data are discretely encoded to generate numerical static features corresponding to the airports at both ends of the route; The numerical static features are input into an embedding network and are respectively mapped into a departure airport embedding vector and a destination airport embedding vector with a predetermined dimension; The departure airport embedding vector and the destination airport embedding vector are input into a fusion network for weighted fusion with learnable weights to generate a route semantic information embedding vector representing the multi-dimensional attributes of the entire route.
[0011] In some preferred embodiments, the time-series block-based Transformer architecture comprises a plurality of Transformer encoder modules; the continuous historical traffic and bandwidth sequence is divided into a plurality of subsequences by sliding window block processing, position encoding is embedded for each subsequence to mark the sequence of time dimension, and then each subsequence is input into a respective Transformer encoder module, global context vectors between the predicted open network IP number of the future time node and the historical traffic and bandwidth sequence are extracted by a multi-head attention mechanism, and traffic and bandwidth data of the future time node are predicted and output by a fully connected network or a decoder.
[0012] In some preferred embodiments, the progressive two-stage aviation Internet traffic and bandwidth time series prediction model introduces an MC Dropout mechanism, which comprises embedding a dynamic Dropout unit in the fully connected layer and the recurrent layer of the progressive two-stage aviation Internet traffic and bandwidth time series prediction model. The dynamic Dropout unit adopts an adaptive deactivation probability that exponentially decays with training rounds in the model training stage, and remains in an active state in the model training stage to update network parameters in the backpropagation process and evaluate the generalization performance of the model in the validation stage. The dynamic Dropout unit remains in the same Dropout active state as in the model training stage in the model inference stage, and performs N independent forward propagation samplings on the input data of the same flight. The distribution characteristics of the prediction results obtained based on N independent forward propagation samplings are calculated to obtain a 95% confidence interval, and the confidence interval is output together.
[0013] In some preferred embodiments, the method further comprises the following steps: During the flight of the flight, when the flight enters or is about to enter a signal weak airspace, or when the predicted traffic and bandwidth data exceeds a preset traffic and bandwidth threshold, a satellite communication link switching instruction is generated, and a redundant bandwidth reservation instruction is generated according to the confidence interval.
[0014] In some preferred embodiments, the progressive two-stage aviation Internet traffic and bandwidth time series prediction model introduces an adaptive learning rate adjustment mechanism, which increases the update frequency of the model when the traffic and bandwidth prediction error exceeds a preset error threshold.
[0015] In a second aspect, the present application also provides an aviation Internet traffic and bandwidth prediction system based on deep learning, which comprises an acquisition module, a processing module and a storage module, wherein the processing module is connected with the acquisition module and the storage module. The storage module is used for storing executable computer program codes; the acquisition module is used for acquiring multi-dimensional flight information and transmitting to the processing module; the processing module is used for executing the method as described above by calling the executable computer program codes in the storage module.
[0016] According to the above technical solution, the application has the beneficial effects that the multi-dimensional flight information, the historical open network IP number sequence and the historical traffic and bandwidth sequence are processed by the progressive two-stage aviation Internet traffic and bandwidth time sequence prediction model, so that the traffic and bandwidth of the future time node can be predicted during the flight of the flight, which provides a basis for the airlines and aviation network service providers to make more reasonable planning and network resource allocation, thereby avoiding network congestion or resource idling, improving the aviation Internet service quality and overall network operation efficiency. BRIEF DESCRIPTION OF DRAWINGS
[0017] Figure 1 FIG. 1 is a flowchart of a deep learning-based aviation Internet traffic and bandwidth prediction method according to an embodiment of the application. DETAILED DESCRIPTION
[0018] The specific embodiments of the application will be further described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is used to help understand the application, but does not constitute a limitation on the application. In addition, the technical features involved in the various embodiments of the application described below can be combined with each other as long as they do not conflict with each other.
[0019] Embodiment 1 Referring to Figure 1 FIG. 1 is a flowchart of a deep learning-based aviation Internet traffic and bandwidth prediction method according to an embodiment of the application, Figure 1 The execution subject of the method shown in FIG. 1 can be a software and / or hardware device. The execution subject of the application can include but is not limited to at least one of the following: user equipment, network equipment, etc. Among them, the user equipment can include but is not limited to computers, smart phones, personal digital assistants (Personal Digital Assistant, PDA) and the above-mentioned electronic devices, etc. The network equipment can include but is not limited to a single network server, a server group composed of multiple network servers or a cloud composed of a large number of computers or network servers based on cloud computing. Cloud computing is a kind of distributed computing, which is a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this.
[0020] A deep learning-based aviation Internet traffic and bandwidth prediction method comprises the following steps: Obtain multi-dimensional flight information, historical open IP number sequence, and historical traffic and bandwidth sequence, and the multi-dimensional flight information includes route information, aircraft model information, and seat occupancy rate information. The multi-dimensional flight information, the historical open IP number sequence, and the historical traffic and bandwidth sequence are processed by the two-stage air Internet traffic and bandwidth time series prediction model to obtain traffic and bandwidth data at a future time node. The progressive two-stage air Internet traffic and bandwidth time series prediction model includes a first-stage open IP number prediction model and a second-stage traffic and bandwidth prediction model.
[0021] The historical open IP number sequence and the historical traffic and bandwidth sequence are filtered from a flight history database based on multi-dimensional flight information, which can be performed in the following manner: First, the multi-dimensional flight information is encoded to obtain a fixed-dimensional feature vector. The route feature uses a subsequent route semantic information embedding vector, the aircraft model feature uses an ICAO aircraft model code one-hot or embedding vector, and the seat occupancy rate feature is normalized to a [0, 1] scalar. These features are then concatenated to obtain the fixed-dimensional feature vector.
[0022] Then, the flight history database is searched using an approximate nearest neighbor index structure such as Faiss IVF, HNSW, ScaNN, and the like to retrieve the most similar flight, and the historical open IP number sequence and the historical traffic and bandwidth sequence are extracted from the most similar flight.
[0023] The first-stage open IP number prediction model uses a Transformer-LSTM hybrid neural network architecture. The preprocessed multi-dimensional flight information and the filtered historical open IP number sequence are concatenated and fed into a Transformer encoder to extract global context, and an LSTM network outputs the open IP number prediction result at a future time node, such as a minute-level open IP number. This embodiment uses a multi-head attention mechanism to extract the complex internal correlations and dependencies between the input features (i.e., the preprocessed multi-dimensional flight information) and the historical open IP number sequence, extracts the time series feature representation rich in context information between the historical open IP number sequence and the input features through the Transformer network, and inputs the extracted time series features into the LSTM network to predict the open IP number at the future time node using the historical open IP number sequence and the input features.
[0024] The process of preprocessing multi-dimensional flight information includes normalization and numericalization, and then obtaining an input vector with the same dimension as the historical open network IP number sequence (also referred to as the input feature mentioned above) through a learnable mapping. The input vector is fused with the historical open network IP number sequence through time dimension broadcasting or a gating mechanism, thereby meeting the input requirements of the Transform-LSTM hybrid neural network.
[0025] The route information of the embodiment obtains a route semantic information embedding vector through a multi-dimensional weighted embedding method, specifically including the following steps: Step S11. Obtain static data corresponding to the airports at both ends of the route based on the route information. The static data includes the economic conditions of the country where the airport is located, the economic conditions of the city where the airport is located, the positioning of the city where the airport is located, and the passenger throughput of the airport. Step S12. Normalize the numerical indicators in the static data, and discretize and encode the categorical indicators in the static data to generate numerical static features corresponding to the airports at both ends of the route. Step S13. Input the numerical static features into an embedding network to map them into departure airport embedding vectors and destination airport embedding vectors with a predetermined dimension. Step S14. Input the departure airport embedding vector and the destination airport embedding vector into a fusion network to perform weighted fusion with learnable weights, thereby generating a route semantic information embedding vector representing the multi-dimensional attributes of the entire route.
[0026] The multi-dimensional weighted embedding method of the embodiment comprehensively considers multiple dimensional information related to the airports at both ends of the route. Compared with the traditional one-hot encoding method which can only represent discrete information identifiers of the airport, the multi-dimensional weighted embedding method can fuse richer and more business-meaningful features, significantly improve information density and expression ability, and help the model to more deeply understand the economic and geographical characteristics, market potential and scale differences behind the route, thereby learning more complex patterns and improving the performance of the prediction task.
[0027] The traffic and bandwidth prediction model of the second stage adopts a time series block-based Transform architecture. By processing the open network IP number at the future time node obtained through prediction and the historical traffic and bandwidth sequence, the traffic and bandwidth data at the future time node, such as minute-level traffic and bandwidth data, are predicted.
[0028] The time series block-based Transformer architecture includes multiple Transformer encoder modules. The continuous historical traffic and bandwidth sequences are divided into a plurality of subsequences by sliding window block processing, for example, the continuous 60-minute historical traffic and bandwidth sequences are divided into n subsequences with a length of m by sliding window block processing. Position encoding is embedded for each subsequence to mark the sequence of the time dimension, and then input into each Transformer encoder module. The global context vector between the predicted future time node open IP number and the historical traffic and bandwidth sequence is extracted by the multi-head attention mechanism, and the traffic and bandwidth data of the future time node is predicted and output by the full connection network or the decoder.
[0029] In the embodiment, the traffic and bandwidth prediction model of the second stage is stacked with multiple Transformer encoder modules. Each Transformer encoder module can accurately capture the traffic response law when the open IP number increases or decreases sharply when processing traffic features of different time granularities. For example, when the open IP number presents a stepwise growth, the traffic and bandwidth prediction model of the second stage can quickly identify the nonlinear jump trend of traffic; when the IP number is maintained in a stable interval, the periodic characteristics in the subtle fluctuations are focused on.
[0030] Before training the progressive two-stage aviation Internet traffic and bandwidth time series prediction model, the open IP number and traffic and bandwidth data during the flight process of historical flights are collected, and the route information, aircraft information and seat occupancy rate information are integrated. After data cleaning and preprocessing, a data set is formed, including a training set, a validation set and a test set. The progressive two-stage aviation Internet traffic and bandwidth time series prediction model introduces an adaptive learning rate adjustment mechanism. When the traffic and bandwidth prediction error exceeds the preset error threshold, the update frequency of the model is improved, thereby improving the accuracy of the model prediction, and accurately matching the real-time and stability requirements of bandwidth resource scheduling in the actual network operation scene, thereby providing double protection for the dynamic regulation of aviation network bandwidth.
[0031] During the flight process, the embodiment of the application further introduces a real-time hybrid prediction update mechanism. The real open IP number sequence and the real traffic and bandwidth sequence that have occurred are input into the progressive two-stage aviation Internet traffic and bandwidth time series prediction model as the historical open IP number sequence and the historical traffic and bandwidth sequence, respectively, and then the traffic and bandwidth data that have not occurred are dynamically predicted and updated.
[0032] That is, before the flight has taken off or just taken off, it has not generated a real open network IP number sequence and a real traffic and bandwidth sequence. Through multi-dimensional flight information, the most similar historical flight is screened from the flight history database, and then the historical open network IP number sequence and the historical traffic and bandwidth sequence are extracted from the historical flight. But when the flight takes off (for example, 30 minutes after taking off) and generates a real open network IP number sequence and a real traffic and bandwidth sequence, the real sequence that has occurred is used instead of the historical sequence, and the latest generated real sequence is used for model reasoning every 30 minutes during the subsequent flight of the flight, thereby improving the timeliness and accuracy of prediction. The real-time hybrid prediction updating mechanism solves the problem that static prediction cannot adapt to sudden conditions (such as passengers concentratedly using the network) during flight and error accumulation, so that the prediction result is continuously optimized with the flight process, realizing the dual guarantee of efficiency and accuracy, and meeting the high timeliness requirement of aviation network real-time scheduling. Of course, in other embodiments, the update time step can also be other lengths, such as 10 minutes, 20 minutes, etc. The real open network IP number sequence and the real traffic and bandwidth sequence used for model reasoning do not necessarily equal the update time step described above. For example, the sequence length can be 20 minutes, 30 minutes, 40 minutes, 50 minutes, 60 minutes, etc.
[0033] In this embodiment, the progressive two-stage aviation Internet traffic and bandwidth time sequence prediction model introduces the MC Dropout mechanism, including embedding a dynamic Dropout unit in the full connection layer and the cycle layer of the progressive two-stage aviation Internet traffic and bandwidth time sequence prediction model.
[0034] The dynamic Dropout unit adopts an adaptive deactivation probability that decays exponentially with the training round in the model training stage, and remains in the active state in the model training stage to update the network parameters in the back propagation process and evaluate the model generalization performance in the validation stage. The dynamic Dropout unit remains in the same Dropout active state as the model training stage in the model inference stage, and performs N (for example, N≥50) independent forward propagation samplings on the input data of the same flight. Based on the distribution characteristics of the prediction results obtained by N independent forward propagation samplings, the 95% confidence interval is calculated, and the confidence interval is output together with the prediction result of the progressive two-stage aviation Internet traffic and bandwidth time sequence prediction model.
[0035] The embodiment breaks through the limitation of traditional point prediction by introducing the MC Dropout mechanism, not only enables the model to output the point prediction value of the minute-level traffic and bandwidth data, but also reflects the prediction uncertainty through the confidence interval width, solves the problem that the traditional prediction result lacks reliability measurement. In addition, the redundant bandwidth reservation instruction can be generated according to the confidence interval, on the one hand, the stability of the communication link can be ensured through the bandwidth redundancy, and the network service interruption caused by insufficient bandwidth resources can be avoided, on the other hand, the waste of bandwidth resources can be avoided, and the operating cost can be significantly reduced.
[0036] The method provided by the embodiment of the application further includes the following steps: During the flight of the flight, when the flight enters or is about to enter the signal weak airspace, or the predicted traffic and bandwidth data exceed the preset traffic and bandwidth threshold, a satellite communication link switching instruction is generated, so that the data transmission mode is switched to the satellite communication or the mode dominated by satellite communication in time. The signal weak airspace can be determined in advance according to experience, or can be determined according to the real-time feedback of the signal receiving device.
[0037] The method provided by the embodiment of the application can realize fine scheduling of network resources through the minute-level traffic and bandwidth prediction results. For example, during the peak period of a popular route, the bandwidth allocation can be adjusted in advance according to the predicted traffic peak value, so as to avoid network congestion caused by passengers using high-frequency flow-consuming applications such as video calls and online games, and to ensure the quality of service, and also to optimize the in-flight network package pricing strategy according to the prediction data, for example, to launch a discount package during the traffic valley period to stimulate passengers to use it and improve the utilization rate of network resources.
[0038] Based on the prediction of traffic and bandwidth at future time nodes, the traffic peak can be predicted in advance and the resources can be allocated, so that passengers can obtain a smoother and more stable experience when using the network during the flight, whether it is browsing the web, watching videos or working communication, and the network congestion can be reduced, making the network use in air travel more convenient and efficient.
[0039] In addition, the method provided by the embodiment of the application can also provide data support for long-term planning of the aviation network. By analyzing the traffic bandwidth demand law of different routes, models and time periods, the airline and network service provider can make scientific decisions on network deployment of new routes, upgrading and modification of existing network equipment, etc., realize long-term optimization of resources, and promote the continuous upgrading of aviation Internet services.
[0040] Embodiment 2 An aviation Internet traffic and bandwidth prediction system based on deep learning, the system comprises an acquisition module, a processing module and a storage module, the processing module is connected with the acquisition module and the storage module.
[0041] The storage module is configured to store executable computer program code; the acquisition module is configured to acquire multi-dimensional flight information and transmit the multi-dimensional flight information to the processing module; and the processing module is configured to execute the method provided in Embodiment 1 by calling the executable computer program code in the storage module.
[0042] It should be noted that, in the description of the present application, the terms "upper", "lower", "left", "right", "front", "back" and the like indicate the orientation or positional relationship of the structure of the present application shown in the drawings, and are only for the convenience of describing the present application simply, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.
[0043] For "first" and "second" in the technical solution, it is only a distinction for the same or similar structure, or the corresponding structure with similar functions, and is not an arrangement of the importance of the structure, nor a sorting, or a comparison of size, or other meanings.
[0044] In addition, unless otherwise explicitly specified and limited, the terms "mounting", "connection" should be understood broadly, for example, the connection can be fixed connection, or detachable connection, or integral connection; can be mechanical connection, or electrical connection; can be direct connection, or indirect connection through intermediate medium; can be the communication inside two structures. For those skilled in the art, the specific meaning of the above terms in the present application can be understood according to the general idea of the present application and the specific circumstances of the present solution.
[0045] The embodiments of the present application are described in detail above in combination with the drawings, but the present application is not limited to the described embodiments. For those skilled in the art, various changes, modifications, replacements and variations of the embodiments without departing from the principles and spirits of the present application still fall within the protection scope of the present application.
Claims
1. A method for deep learning based aerial internet traffic and bandwidth prediction, characterized in that: The method comprises the following steps: Obtaining multi-dimensional flight information, historical open IP number sequences, and historical traffic and bandwidth sequences, wherein the multi-dimensional flight information comprises route information, aircraft model information, and seat occupancy rate information; Processing the multi-dimensional flight information, the historical open IP number sequences, and the historical traffic and bandwidth sequences through a progressive two-stage aviation Internet traffic and bandwidth time series prediction model to obtain traffic and bandwidth data at a future time node; The progressive two-stage aviation Internet traffic and bandwidth time series prediction model comprises a first-stage open IP number prediction model and a second-stage traffic and bandwidth prediction model.
2. The method of claim 1, wherein: The open IP number prediction model adopts a hybrid neural network architecture based on Transformer-LSTM, concatenates the preprocessed multi-dimensional flight information and the historical open IP number sequences, inputs them into a Transformer encoder to extract global context, and outputs a prediction result of open IP number at a future time node through an LSTM network; The traffic and bandwidth prediction model adopts a Transformer architecture based on time series blocking, processes the predicted open IP number at the future time node and the historical traffic and bandwidth sequences to obtain traffic and bandwidth data at the future time node.
3. The method of claim 1, wherein: The historical open IP number sequences and the historical traffic and bandwidth sequences are filtered from a flight history database based on the multi-dimensional flight information; During flight, a real-time hybrid prediction updating mechanism is introduced, and real open IP number sequences and real traffic and bandwidth sequences that have occurred are respectively input into the progressive two-stage aviation Internet traffic and bandwidth time series prediction model to replace the historical open IP number sequences and the historical traffic and bandwidth sequences, thereby dynamically predicting and updating traffic and bandwidth data that have not occurred.
4. The method of claim 1, wherein: The pre-processing of the multi-dimensional flight information comprises normalization and numericalization, and then a learnable mapping is performed to obtain an input vector with the same dimension as the historical open IP number sequences.
5. The method of claim 4, wherein: The route information obtains a route semantic information embedding vector through a multi-dimensional weighted embedding method, comprising: Based on the route information, static data corresponding to the two end airports of the route are obtained, including the economic conditions of the countries where the airports are located, the economic conditions of the cities where the airports are located, the positioning of the cities where the airports are located, and the passenger throughput of the airports; The numerical indicators in the static data are normalized, and the categorical indicators in the static data are discretely coded to generate numerical static features corresponding to the two end airports of the route; The numerical static features are input into an embedding network to be respectively mapped into departure airport embedding vectors and destination airport embedding vectors with a predetermined dimension; The departure airport embedding vectors and the destination airport embedding vectors are input into a fusion network to be weighted and fused with learnable weights to generate a route semantic information embedding vector representing the multi-dimensional attributes of the entire route.
6. The method of claim 2, wherein: The time-series block-based Transformer architecture comprises a plurality of Transformer encoder modules; the continuous historical traffic and bandwidth sequence is divided into a plurality of subsequences through sliding window block processing, position encoding is embedded for each subsequence to mark the sequence of time dimension, and then each subsequence is input into each Transformer encoder module, a global context vector between the predicted future time node open IP number and the historical traffic and bandwidth sequence is extracted through a multi-head attention mechanism, and traffic and bandwidth data of the future time node are predicted and output through a full connection network or a decoder.
7. The method of claim 1, wherein: The progressive two-stage aviation internet traffic and bandwidth time series prediction model introduces an MC Dropout mechanism, which includes embedding a dynamic Dropout unit in the full connection layer and the recurrent layer of the progressive two-stage aviation internet traffic and bandwidth time series prediction model; The dynamic Dropout unit adopts an adaptive deactivation probability that exponentially decays with training rounds during the model training stage, and remains in an active state during the model training stage to update network parameters during the backpropagation process and evaluate the generalization performance of the model during the validation stage; The dynamic Dropout unit remains in the same Dropout active state as the model training stage during the model inference stage, and performs N times of independent forward propagation sampling on the input data of the same flight; The distribution characteristics of the prediction results obtained based on N times of independent forward propagation sampling are calculated to obtain a 95% confidence interval, and the confidence interval is output together.
8. The method of claim 7, wherein: The method further comprises the following steps: During the flight of the flight, when the flight enters or is about to enter a signal weak airspace, or the predicted traffic and bandwidth data exceeds a preset traffic and bandwidth threshold, a satellite communication link switching instruction is generated, and a redundant bandwidth reservation instruction is generated according to the confidence interval.
9. The method of claim 1, wherein: The progressive two-stage aviation internet traffic and bandwidth time series prediction model introduces an adaptive learning rate adjustment mechanism, which increases the update frequency of the model when the traffic and bandwidth prediction error exceeds a preset error threshold. 10.A deep learning based aeronautical internet traffic and bandwidth prediction system, characterized in that: The system comprises an acquisition module, a processing module and a storage module, the processing module is connected with the acquisition module and the storage module; The storage module is used for storing executable computer program codes; the acquisition module is used for acquiring multi-dimensional flight information and transmitting the multi-dimensional flight information to the processing module; the processing module is used for executing the method according to any one of claims 1-9 by calling the executable computer program codes in the storage module.