Methods and devices for predicting air traffic flow

Abstract

This invention discloses a method and apparatus for predicting air traffic flow, relating to the field of data processing technology. It generates a multi-resolution time-series feature encoding sequence by encoding historical air traffic flow with multi-resolution time-series features. This sequence includes significant multi-resolution time-series features corresponding to the historical air traffic flow. Simultaneously, the multi-resolution time-series feature encoding sequence is concatenated with time-invariant and time-varying features to enhance the time-series characteristics of both features, making it easier for the air traffic flow prediction model to capture the temporal correlations in air traffic flow changes. Furthermore, normalized historical air traffic flow is decomposed based on Fourier transform to obtain time-invariant and time-varying sequences, enabling the air traffic flow prediction model to perceive different time-series frequencies. Finally, by integrating the aforementioned mechanisms into the air traffic flow prediction model, a performance breakthrough and improved prediction accuracy are achieved.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to a method and apparatus for predicting air traffic flow. Background Technology

[0002] With the continuous development of the aviation industry, the increase in air traffic volume has placed higher demands on air traffic management. Air Traffic Flow Forecasting (ATFP) technology plays a crucial role in this context, providing important support for improving airspace management efficiency and optimizing flight scheduling. Accurate ATFP can help air traffic controllers predict airspace traffic conditions in advance, thereby implementing more efficient traffic management measures, avoiding congestion, reducing delays, improving the utilization efficiency of airspace resources, and ensuring smooth and safe flight operations.

[0003] Existing ATFP methods primarily focus on the air traffic variation characteristics inherent in each air traffic data sequence itself. Specifically, they directly model the input air traffic data sequences to capture patterns in air traffic variation. However, simply modeling the air traffic data sequences themselves often fails to capture the complex relationships between air traffic changes, leading to information loss and decreased accuracy. Furthermore, existing ATFP methods often rely on the prediction model itself to implicitly mine the temporal features of air traffic, operating in a "black box" state. These "black box" temporal features are often not fully utilized by the prediction model, resulting in persistent performance bottlenecks, low accuracy in air traffic prediction, and insufficient interpretability.

[0004] Therefore, providing a method and apparatus for predicting air traffic flow, enabling existing prediction models to overcome performance bottlenecks, possess time-series interpretability, and improve the accuracy of air traffic flow prediction in complex airspace scenarios is an urgent problem to be solved. Summary of the Invention

[0005] This disclosure provides a method and apparatus for predicting air traffic flow. The main purpose is to provide a method and apparatus for predicting air traffic flow, enabling existing prediction models to overcome performance bottlenecks and improve the accuracy of air traffic flow prediction in complex airspace scenarios.

[0006] According to a first aspect of this disclosure, a method for predicting air traffic flow is provided, comprising:

[0007] Receive air traffic data and preprocess it to obtain historical air traffic data;

[0008] The historical air traffic flow is subjected to multi-resolution temporal feature encoding to generate a multi-resolution temporal feature encoding sequence.

[0009] The normalized historical air traffic flow is decomposed based on Fourier transform to obtain time-invariant and time-varying sequences.

[0010] Feature extraction is performed on the time-invariant sequence and the time-varying sequence respectively to obtain time-invariant features and time-varying features;

[0011] The multi-resolution temporal feature encoding sequence is concatenated with time-invariant features and time-varying features respectively to obtain temporally enhanced time-varying sequence embedding data and temporally enhanced time-invariant sequence embedding data.

[0012] Based on the Long Short-Term Memory network, time-varying sequence embedding data and time-invariant sequence embedding data are learned respectively to obtain time-invariant time series features and time-varying time series features;

[0013] The attention scores corresponding to the time-invariant and time-varying time-series features are weighted to obtain the weighted time-invariant and time-varying time-series feature sequences.

[0014] Based on the KAN prediction network, air traffic is predicted from the weighted time-invariant feature sequence and the weighted time-varying feature sequence, respectively, to obtain the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence.

[0015] The sum of the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence is inversely normalized to obtain the actual air traffic flow prediction result.

[0016] Optionally, the step of performing multi-resolution temporal feature encoding on the historical air traffic to generate a multi-resolution temporal feature encoding sequence includes:

[0017] The historical air traffic flow is encoded using time-resolution temporal feature encoding to obtain a time-series feature encoding sequence;

[0018] The historical air traffic flow is encoded using hourly time-series features to obtain an hourly time-series feature encoding sequence;

[0019] The historical air traffic flow is encoded using time-series features at week-day resolution to obtain a week-day time-series feature encoding sequence. The multi-resolution time-series feature encoding sequence includes the time-series feature encoding sequence, the hour-series feature encoding sequence, and the week-day time-series feature encoding sequence.

[0020] Optionally, the decomposition of the normalized historical air traffic based on Fourier transform to obtain time-invariant and time-varying sequences includes:

[0021] The normalized historical air traffic flow time-domain sequence is transformed into a frequency-domain complex sequence based on the discrete Fourier transform.

[0022] The frequency domain complex sequence is filtered for high-frequency components using a low-pass filter, and the filtering result is converted from the frequency domain to the time domain to obtain the time-invariant sequence.

[0023] The time-varying sequence is obtained by calculating the difference between the time-domain sequence and the time-invariant sequence.

[0024] Optionally, the step of extracting features from the time-invariant sequence and the time-varying sequence respectively to obtain time-invariant features and time-varying features includes:

[0025] The time-invariant sequence and the time-varying sequence are linearly mapped to obtain an aligned time-invariant sequence aligned to a first preset sequence length and an aligned time-varying sequence aligned to a second preset sequence length.

[0026] High-dimensional features are extracted from the aligned time-invariant sequence and the aligned time-varying sequence respectively using convolutional layers to obtain the time-invariant features and the time-varying features.

[0027] Optionally, the step of concatenating the multi-resolution temporal feature encoded sequence with time-invariant features and time-varying features respectively to obtain time-enhanced time-varying sequence embedding data and time-enhanced time-invariant sequence embedding data includes:

[0028] High-dimensional features are extracted from the time-series feature encoding sequence, the hour-series feature encoding sequence, and the week-series feature encoding sequence, and the three feature extraction results are concatenated to obtain a latent time-series representation.

[0029] The latent temporal representation is concatenated with the time-invariant feature to obtain the temporally enhanced time-invariant sequence embedding data;

[0030] The latent temporal representation is concatenated with the time-varying features to obtain the time-enhanced time-varying sequence embedding data.

[0031] Optionally, receiving and preprocessing air traffic data to obtain historical air traffic flow includes:

[0032] The air traffic data, which is acquired and updated in real time, is converted into a structured data format that can be processed subsequently.

[0033] Extract key data fields from the converted air traffic flow data and remove duplicate data.

[0034] Based on the deduplicated key data fields, a time series is constructed according to the historical time step length to generate the historical air traffic.

[0035] Optionally, before receiving and preprocessing air traffic data to obtain historical air traffic data, the method includes:

[0036] A pre-defined deep learning framework is trained to obtain an air traffic prediction model, which is used to execute the air traffic prediction method.

[0037] The air traffic prediction model includes a frequency sensing module, a sequence modeling module, and a KAN prediction module.

[0038] The frequency sensing module is used to decompose and extract features from the historical air traffic input to the air traffic prediction model to generate time-series features, which include time-invariant features and time-varying features.

[0039] The sequence modeling module used for the temporal feature enhancement is used to fuse the temporal features with the multi-resolution temporal feature encoded sequence, and to extract temporal dependencies from the fusion result based on the long short-term memory network;

[0040] The KAN prediction module based on the attention mechanism is used to map the weighted time-invariant time-series feature sequence and the weighted time-varying time-series feature sequence to the prediction dimension and output the predicted air traffic flow sequence, which includes the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence.

[0041] Optionally, training a preset deep learning framework to obtain an air traffic prediction model includes:

[0042] The air traffic prediction model is trained using the constructed training set. Backpropagation is performed using a loss function to update the network parameters of the air traffic prediction model. The learning rate is dynamically adjusted using the Adam optimizer. The formula for the loss function is as follows:

[0043] L dd =L time +αL freq

[0044] Among them, L dd Let L be the loss function. time For time-domain loss, L freq For frequency domain loss, α represents the scaling factor, used to control the impact of frequency domain loss on the total loss;

[0045] The formula for calculating the time-domain loss is as follows:

[0046]

[0047] Where, x ij Represents the true value of the time-domain sequence. The predicted value of the time-domain sequence is represented by T, the length of the predicted time-domain sequence is represented by N, the number of time-domain sequence samples in a training batch is represented by i and j, which are subscript variables, where i is the length index and j is the sample index.

[0048] The formula for calculating the frequency domain loss is as follows:

[0049] y = DFT(x),

[0050]

[0051] Where DFT(·) denotes Discrete Fourier Transform, x is a time-domain sequence, and y is a frequency-domain sequence. To predict time-domain sequences, To predict the frequency domain sequence, y ij It is the true value of the frequency domain sequence. It is the predicted value of the frequency domain sequence. The value represents the length of the frequency domain sequence after the real Fourier transform, T represents the length of the predicted time domain sequence before the real Fourier transform, n is the number of frequency domain sequence samples in a training batch, and i and j represent subscript variables, where i is the length index and j is the sample index.

[0052] According to a second aspect of this disclosure, an air traffic flow prediction device is provided, comprising:

[0053] The receiving unit is used to receive air traffic flow data and preprocess it to obtain historical air traffic flow data.

[0054] The encoding unit is used to perform multi-resolution temporal feature encoding on the historical air traffic flow to generate a multi-resolution temporal feature encoding sequence.

[0055] The processing unit is used to decompose the normalized historical air traffic flow based on Fourier transform to obtain time-invariant and time-varying sequences.

[0056] An extraction unit is used to extract features from the time-invariant sequence and the time-varying sequence respectively, to obtain time-invariant features and time-varying features;

[0057] The splicing unit is used to splice the multi-resolution temporal feature encoding sequence with time-invariant features and time-varying features respectively to obtain temporally enhanced time-varying sequence embedded data and temporally enhanced time-invariant sequence embedded data.

[0058] The learning unit is used to learn time-varying sequence embedding data and time-invariant sequence embedding data based on the long short-term memory network, respectively, to obtain time-invariant time series features and time-varying time series features;

[0059] The weighting unit is used to perform weighting processing on the attention scores corresponding to the time-invariant time-series features and the time-varying time-series features respectively, to obtain the weighted time-invariant time-series feature sequence and the weighted time-varying time-series feature;

[0060] The prediction unit is used to predict air traffic based on the weighted time-invariant time-series feature sequence and the weighted time-varying time-series feature sequence, respectively, to obtain the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence.

[0061] The inverse normalization unit is used to inverse normalize the sum of the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence to obtain the actual air traffic flow prediction result.

[0062] According to a third aspect of this disclosure, an electronic device is provided, comprising:

[0063] At least one processor; and

[0064] A memory communicatively connected to the at least one processor; wherein,

[0065] The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect above.

[0066] According to a fourth aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are configured to cause the computer to perform the method described in the first aspect above.

[0067] According to a fifth aspect of this disclosure, a computer program product is provided, comprising a computer program that, when executed by a processor, implements the method described in the first aspect above.

[0068] The air traffic prediction method and apparatus disclosed herein receive air traffic data and preprocess it to obtain historical air traffic; perform multi-resolution time-series feature encoding on the historical air traffic to generate a multi-resolution time-series feature encoding sequence; decompose the normalized historical air traffic based on Fourier transform to obtain a time-invariant sequence and a time-varying sequence; extract features from the time-invariant sequence and the time-varying sequence respectively to obtain time-invariant features and time-varying features; concatenate the multi-resolution time-series feature encoding sequence with the time-invariant features and the time-varying features respectively to obtain time-series enhanced time-varying sequence embedded data and time-series enhanced time-invariant sequence embedded data; based on length and shortness... The time-invariant memory network learns from time-varying sequence embedding data and time-invariant sequence embedding data respectively to obtain time-invariant time-series features and time-varying time-series features. Based on the attention scores corresponding to the time-invariant and time-varying time-series features respectively, weighted processing is performed to obtain weighted time-invariant time-series feature sequences and weighted time-varying time-series features. Air traffic prediction is performed on the weighted time-invariant and weighted time-varying time-series features using a KAN prediction network to obtain predicted time-invariant air traffic flow sequences and predicted time-varying air traffic flow sequences. The sum of the predicted time-invariant and predicted time-varying air traffic flow sequences is inversely normalized to obtain the actual air traffic flow prediction result. Compared with related technologies, this method generates a multi-resolution time-series feature encoding sequence by encoding the historical air traffic data using multi-resolution time-series features. This sequence includes significant multi-resolution time-series features corresponding to the historical air traffic. Furthermore, the multi-resolution time-series feature encoding sequence is concatenated with both time-invariant and time-varying features to enhance their temporal characteristics. This makes it easier for the air traffic prediction model to capture temporal correlations in air traffic changes. Additionally, the normalized historical air traffic is decomposed using Fourier transform to obtain time-invariant and time-varying sequences, enabling the air traffic prediction model to perceive different temporal frequencies. Finally, by integrating the aforementioned mechanisms into the air traffic prediction model, the model achieves a performance breakthrough and improves the accuracy of prediction results. Attached Figure Description

[0069] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0070] Figure 1 A schematic flowchart of an air traffic flow prediction method provided in an embodiment of this disclosure;

[0071] Figure 2This is a schematic diagram of the overall structure of an air traffic prediction model provided in an embodiment of the present disclosure;

[0072] Figure 3 This is a schematic diagram of the structure of a decomposition module provided in an embodiment of the present disclosure;

[0073] Figure 4 This is a schematic diagram of the structure of a KAN prediction module provided in an embodiment of the present disclosure;

[0074] Figure 5 A schematic diagram of the structure of an air traffic flow prediction device provided in an embodiment of this disclosure;

[0075] Figure 6 A schematic block diagram of an example electronic device 300 provided for embodiments of this disclosure. Detailed Implementation

[0076] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0077] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0078] Furthermore, the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such use of data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein.

[0079] The method and apparatus for predicting air traffic flow according to embodiments of the present disclosure are described below with reference to the accompanying drawings.

[0080] To provide at least one method and apparatus for predicting air traffic flow, enabling existing prediction models to overcome performance bottlenecks and improve the accuracy of air traffic flow prediction in complex airspace scenarios, this embodiment provides a method for predicting air traffic flow.

[0081] Figure 1 This is a schematic flowchart illustrating an air traffic flow prediction method provided in an embodiment of this disclosure. Figure 1 As shown, the method includes the following steps:

[0082] Step 101: Receive air traffic data and preprocess it to obtain historical air traffic data;

[0083] Step 102: Perform multi-resolution temporal feature encoding on the historical air traffic to generate a multi-resolution temporal feature encoding sequence;

[0084] Step 103: Decompose the normalized historical air traffic flow based on Fourier transform to obtain time-invariant and time-varying sequences;

[0085] Step 104: Extract features from the time-invariant sequence and the time-varying sequence respectively to obtain time-invariant features and time-varying features;

[0086] Step 105: The multi-resolution temporal feature encoding sequence is concatenated with time-invariant features and time-varying features respectively to obtain temporal enhanced time-varying sequence embedding data and temporal enhanced time-invariant sequence embedding data.

[0087] Step 106: Based on the Long Short-Term Memory Network, learn the time-varying sequence embedding data and the time-invariant sequence embedding data respectively to obtain the time-invariant time series features and the time-varying time series features;

[0088] Step 107: Based on the attention scores corresponding to the time-invariant time-series features and the time-varying time-series features respectively, perform weighted processing to obtain the weighted time-invariant time-series feature sequence and the weighted time-varying time-series features;

[0089] Step 108: Based on the KAN prediction network, air traffic is predicted on the weighted time-invariant time series feature sequence and the weighted time-varying time series feature sequence, respectively, to obtain the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence.

[0090] Step 109: Perform inverse normalization on the sum of the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence to obtain the actual air traffic flow prediction result.

[0091] The air traffic prediction method disclosed herein receives and preprocesses air traffic data to obtain historical air traffic; performs multi-resolution time-series feature encoding on the historical air traffic to generate a multi-resolution time-series feature encoding sequence; decomposes the normalized historical air traffic based on Fourier transform to obtain a time-invariant sequence and a time-varying sequence; extracts features from the time-invariant sequence and the time-varying sequence respectively to obtain time-invariant features and time-varying features; concatenates the multi-resolution time-series feature encoding sequence with the time-invariant features and the time-varying features respectively to obtain time-enhanced time-varying sequence embedding data and time-enhanced time-invariant sequence embedding data; based on long-term and short-term... The memory network learns from time-varying sequence embedding data and time-invariant sequence embedding data respectively to obtain time-invariant time-series features and time-varying time-series features. The attention scores corresponding to the time-invariant and time-varying time-series features are weighted to obtain weighted time-invariant and weighted time-varying time-series feature sequences. The KAN prediction network is used to predict air traffic flow from the weighted time-invariant and weighted time-varying time-series feature sequences to obtain predicted time-invariant and predicted time-varying air traffic flow sequences. The sum of the predicted time-invariant and predicted time-varying air traffic flow sequences is inversely normalized to obtain the actual air traffic flow prediction result. Compared with related technologies, this method generates a multi-resolution time-series feature encoding sequence by encoding the historical air traffic data using multi-resolution time-series features. This sequence includes significant multi-resolution time-series features corresponding to the historical air traffic. Furthermore, the multi-resolution time-series feature encoding sequence is concatenated with both time-invariant and time-varying features to enhance their temporal characteristics. This makes it easier for the air traffic prediction model to capture temporal correlations in air traffic changes. Additionally, the normalized historical air traffic is decomposed using Fourier transform to obtain time-invariant and time-varying sequences, enabling the air traffic prediction model to perceive different temporal frequencies. Finally, by integrating the aforementioned mechanisms into the air traffic prediction model, the model achieves a performance breakthrough and improves the accuracy of prediction results.

[0092] As a refinement of the embodiments of this disclosure, when performing step 102 to perform multi-resolution temporal feature encoding on the historical air traffic and generate a multi-resolution temporal feature encoding sequence, the following implementation methods may also be adopted, but are not limited to: performing time-resolution temporal feature encoding on the historical air traffic to obtain a time-time temporal feature encoding sequence; performing hour-resolution temporal feature encoding on the historical air traffic to obtain an hour-time temporal feature encoding sequence; performing week-resolution temporal feature encoding on the historical air traffic to obtain a week-day temporal feature encoding sequence, wherein the multi-resolution temporal feature encoding sequence includes the time-time temporal feature encoding sequence, the hour-time temporal feature encoding sequence, and the week-day temporal feature encoding sequence.

[0093] As an extension of the above embodiments, the time-series feature encoding sequence (Quarter-Of-Day, QoD) is used to capture local fluctuations in air traffic flow at different times of the day, and its calculation process is shown in formula (1):

[0094]

[0095] Here, ts represents the timestamp corresponding to the current time step, and QoD(·) represents the function that calculates the time encoding for ts. In the function expression, ts.minute and ts.hour represent the minutes and hours corresponding to the timestamp, respectively, and finally, subtracting one converts the encoding to a base-0 index.

[0096] The hourly time-series feature encoding sequence is used to assist the air traffic prediction model in capturing hourly air traffic characteristics within a day, and its calculation formula is as follows:

[0097] HoD(ts)=ts.hour-1,whereHoD(ts)∈{0,1,2,…,23}Formula (2)

[0098] Here, ts represents the timestamp corresponding to the current time step, and HoD(·) represents the function that calculates the hour code for ts. In the function expression, ts.hour represents the hour corresponding to the timestamp, and finally, subtracting one converts the code to a base-0 index.

[0099] The weekday time-series feature encoding sequence is used to help the air traffic forecasting model learn different patterns of change throughout the week (such as significant differences in traffic between weekdays and weekends), and its calculation formula is as follows:

[0100]

[0101] Here, ts represents the timestamp corresponding to the current time step, and DoW(·) represents the function that calculates the weekday code for ts. In the function expression, ts.day represents the date within the month corresponding to the timestamp. Similarly, subtracting one at the end converts the code to a base-0 index. It is important to understand that the day of the week here does not have a fixed correspondence with the traditional day of the week; it is merely an index representation used to indicate the degree of offset on a 7-day scale.

[0102] As a refinement of the above embodiments, when performing step 103, which involves decomposing the normalized historical air traffic flow based on Fourier transform to obtain a time-invariant sequence and a time-varying sequence, the following implementation methods can also be adopted, but are not limited to: transforming the time-domain sequence of the normalized historical air traffic flow into a frequency-domain complex sequence based on discrete Fourier transform; filtering the frequency-domain complex sequence for high-frequency components based on a low-pass filter, and converting the filtering result from the frequency domain to the time domain to obtain the time-invariant sequence; calculating the difference between the time-domain sequence and the time-invariant sequence to obtain the time-varying sequence.

[0103] To facilitate understanding of the above embodiments, the process of the discrete Fourier transform, combined with the formula, can be described as follows:

[0104] Y=DFT(X) formula (4)

[0105] Where DFT(·) represents Discrete Fourier Transform, X is the time-domain sequence, and Y is the frequency-domain complex sequence; furthermore, formula (4) can be specifically extended to formula (5) as follows:

[0106]

[0107] Where, for X = {x1, x2, ..., x} L Perform a discrete Fourier transform to obtain L is the length of the time-domain sequence, ω k Here, i is the normalized frequency parameter, k is the index of the frequency domain sequence (i.e., the frequency domain complex sequence), e is the natural logarithm, and n is the index of the traversal of X.

[0108] The process of filtering high-frequency components of the frequency domain complex sequence using a low-pass filter and converting the filtering result from the frequency domain to the time domain to obtain the time-invariant sequence is described in Equations (6), (7), and (8).

[0109] Filtering out high-frequency components from the frequency domain complex sequence Y and then transforming it back to the time domain yields the time-invariant sequence X. inv :

[0110] X inv=iDFT(Y⊙LPF) formula (6)

[0111] Where LPF is a pre-built low-pass filter, ⊙ represents the Hadamard product, i.e., element-wise multiplication, and iDFT(·) represents the inverse discrete Fourier transform.

[0112] The pre-constructed low-pass filter (LPF) expression is as follows:

[0113]

[0114] Among them, f c It is the cutoff frequency, when the frequency component f k When the value exceeds the threshold, the filter value is 0, and the frequency component is blocked; otherwise, the filter value is 1, and the frequency component is preserved. k is the subscript index.

[0115] The formula for the inverse discrete Fourier transform is as follows:

[0116]

[0117] in, These are the frequency components after low-pass filtering. To convert to a time-invariant sequence in the time domain, where L is the length of the time-domain sequence, ω k For the normalized frequency parameter, i is the imaginary unit, k is the index of the frequency domain sequence, i.e., the frequency domain complex sequence, e is the natural logarithm, and n is the index of the traversal of X. Please refer to the parameter interpretation of formula (3).

[0118] Time-domain sequence X and time-invariant sequence X inv The difference is the time-varying sequence X. var :

[0119] f var =XX inv Formula (9)

[0120] As a refinement of the above embodiments, when performing step 104 to extract features from the time-invariant sequence and the time-varying sequence respectively to obtain time-invariant features and time-varying features, the following implementation methods can also be adopted, for example: performing linear mapping processing on the time-invariant sequence and the time-varying sequence respectively to obtain an aligned time-invariant sequence aligned to a first preset sequence length and an aligned time-varying sequence aligned to a second preset sequence length; performing high-dimensional feature extraction on the aligned time-invariant sequence and the aligned time-varying sequence respectively based on convolutional layers to obtain the time-invariant features and the time-varying features.

[0121] To facilitate understanding of the above embodiments, this embodiment combines the above formulas (3)-(9) and explains the above embodiments in the form of formulas. The specific process is as follows:

[0122] The linear mapping is used to align the time-varying sequence and the time-invariant sequence to the predicted time series length, which includes the first preset sequence length and the second preset sequence length.

[0123] X′ var =Linear1(X var )Formula (10)

[0124] X′ inv =Linear2(X inv )Formula (11)

[0125] Among them, Linear1(·) and Linear2(·) are two independent linear layers. It is an aligned time-varying sequence. It is an aligned time-invariant sequence, where L represents the length of the history sequence and T represents the length of the prediction sequence. This represents a parameter dataset with a dimension of (L+T)×1.

[0126] The convolutional layer embeds the aligned time-varying sequence and the aligned time-invariant sequence into a high-dimensional space, i.e., a high-dimensional feature extraction representation:

[0127] X″ var =Conv1D(X′) var )Formula (12)

[0128] X″ inv =Conv1S(X′) inv )Formula (13)

[0129] Where Conv1D(·) is a one-dimensional convolutional layer with shared weights. It is a time-varying sequence embedding, also known as the time-varying feature. This is a time-invariant sequence embedding, also known as a time-invariant feature, where L represents the length of the historical sequence and T represents the length of the predicted sequence. (L+T)×D seq A parameter dataset of dimension D seq The dimension of the embedding of time-varying sequences and time-invariant sequences into the higher-dimensional space.

[0130] As a refinement of the above embodiments, when performing step 105, which involves concatenating the multi-resolution temporal feature encoding sequence with time-invariant features and time-varying features respectively to obtain time-enhanced time-varying sequence embedding data and time-enhanced time-invariant sequence embedding data, the following implementation methods can also be adopted, but are not limited to: performing high-dimensional feature extraction on the time-series feature encoding sequence, the hourly time-series feature encoding sequence, and the weekday time-series feature encoding sequence, and concatenating the three feature extraction results to obtain a latent temporal representation; concatenating the latent temporal representation with the time-invariant features to obtain the time-enhanced time-invariant sequence embedding data; and concatenating the latent temporal representation with the time-varying features to obtain the time-enhanced time-varying sequence embedding data.

[0131] To facilitate understanding of the above embodiments, this embodiment combines the above formulas (3)-(13) and explains the above embodiments in the form of formulas. The specific process is as follows:

[0132] The three multi-resolution temporal feature encoding sequences generated in step 102—time-series feature encoding sequence Q, hourly-series feature encoding sequence H, and weekday-series feature encoding sequence D—are embedded into a high-dimensional space and concatenated to obtain the latent temporal representation (LTR).

[0133] LTR=Concat(Emb1(Q),Emb2(H),Emb3(D)) formula (14)

[0134] Where Emb1(·), Emb2(·), and Emb3(·) represent three independent embedding layers, with dimensions D respectively. q D h D d Concat(·) represents a concatenation operation along the feature dimension.

[0135] Then, embed X″ into the time-varying sequence var Time-invariant timing embedding X″ inv To enhance features and better preserve temporal information, a fusion method is employed that concatenates along the feature dimensions, resulting in temporal feature-enhanced time-varying sequence embeddings, also known as time-enhanced time-varying sequence embedding data E. var And time-invariant sequence embedding with enhanced temporal features, namely the time-enhanced time-invariant sequence embedding data E inv :

[0136] E var =Concat(X″ var Formula (15) (LTR)

[0137] E inv =Concat(X″inv Formula (16) (LTR)

[0138] Where Concat(·) represents a concatenation operation along the feature dimension.

[0139] As a refinement of the above embodiments, when performing step 106, which involves learning time-varying sequence embedding data and time-invariant sequence embedding data based on a long short-term memory network to obtain time-invariant time-series features and time-varying time-series features, the following implementation methods can be adopted, but are not limited to: the long short-term memory network is used to learn the temporal correlation between the air traffic flow to obtain a high-dimensional time-varying representation after time-series feature extraction, i.e., the time-varying time-series feature H. var And the high-dimensional time-invariant representation, namely the time-invariant time series feature H inv :

[0140] H var =LSTM(E var )Formula (17)

[0141] H inv =LSTM(E inv )Formula (18)

[0142] Wherein, LSTM(·) is a Long Short-Term Memory network with hidden layer dimension D.

[0143] As a refinement of the above embodiments, when performing step 107, which involves weighting the attention scores corresponding to the time-invariant temporal features and the time-varying temporal features respectively to obtain the weighted time-invariant temporal feature sequence and the weighted time-varying temporal features, the following implementation methods can also be adopted, but are not limited to: capturing the temporal dependency association of the high-dimensional representation based on the attention mechanism to obtain the high-dimensional semantic time-varying representation after attention score weighting, which is the weighted time-varying temporal feature sequence A. var The high-dimensional semantic time-invariant representation, namely the weighted time-invariant temporal feature sequence A inv :

[0144] A var =AttBlock(H var )Formula (19)

[0145] A inv =AttBlock(H inv )Formula (20)

[0146] Where AttBlock(·) represents the attention module,

[0147] Specifically, for the attention module Attblock, the calculation process of formulas (19) and (20) is as follows:

[0148] Calculate the score S of the eigenmap for the high-dimensional representation H. f H is H inv and H var Any one of them:

[0149] S f =γ(H·W) f1 )·W f2 Formula (21)

[0150] in, It is the weight matrix used for feature mapping, and γ(·) represents the ReLU activation function;

[0151] By weighing the importance of each time step, the context-aware attention score S is obtained. t :

[0152] S t =σ(W t ·S f )Formula (22)

[0153] in, It is a weight matrix used to enhance temporal dependencies, and σ(·) represents the Sigmoid activation function;

[0154] We obtain the high-dimensional semantic representation A by weighting the attention scores:

[0155]

[0156] in, This indicates dimensional expansion according to the broadcast mechanism, where H is H inv and H var In any of them, A is the A corresponding to H. var Or A inv .

[0157] As a refinement of the above embodiments, the process of performing air traffic prediction based on the weighted time-invariant time-series feature sequence and the weighted time-varying time-series feature sequence using the KAN prediction network in steps 108 and 109, respectively, to obtain the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence; and performing inverse normalization on the sum of the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence to obtain the actual air traffic flow prediction result includes: mapping the high-dimensional semantic representation to a one-dimensional sequence based on the KAN prediction network, to obtain the predicted time-varying sequence, i.e., the predicted time-varying air traffic flow sequence P. var The time-invariant sequence, also known as the predicted time-invariant air traffic flow sequence P, inv :

[0158] P var =KANLayer(A var )Formula (24)

[0159] P inv =KANLayer(A inv )Formula (25)

[0160] Where KANLayer(·) represents the Kolmogorov-Arnold layer; unlike the existing MLPLayer, the edges of the KANLayer have a learnable, non-linear function φ(x), expressed as follows:

[0161] φ(x)=Spline(x)+SiLU(x) Formula (26)

[0162] Where Spline(·) represents the B-spline function, SiLU(·) represents the SiLU activation function as a basis, and x represents the value of the node corresponding to the edge.

[0163] Finally, the predicted time-varying sequence P var Time-invariant sequence P inv Adding them together, we get the final traffic flow prediction result P:

[0164] P = P var +P inv Formula (27)

[0165] Specifically, P is denormalized to obtain the actual air traffic flow prediction result.

[0166] To facilitate understanding of the above embodiments, this embodiment provides normalization and denormalization processes. The principles of normalization and denormalization processes involved in this disclosure are the same as those in this embodiment, including: the normalization method is z-score normalization, and the formula is:

[0167]

[0168] Where x represents the original value, μ represents the mean of the dataset, σ represents the standard deviation of the dataset, and z represents the normalized value.

[0169] The formula for inverse normalization is:

[0170] Formula (29) x=z·σ+μ

[0171] Where x represents the original value, μ represents the mean of the dataset, σ represents the standard deviation of the dataset, and z represents the normalized value.

[0172] As a refinement of the above embodiments, when receiving and preprocessing air traffic data in step 101 to obtain historical air traffic data, the following implementation methods may also be adopted, but are not limited to: converting the format of the air traffic data that is acquired and updated in real time into a structured data format that can be processed subsequently; extracting key data fields from the converted air traffic data and removing duplicate data; and constructing a time series based on the deduplicated key data fields according to the length of the historical time step to generate the historical air traffic data. To facilitate understanding of this embodiment, please refer to the following exemplary descriptions in conjunction with actual scenarios, including: receiving track data through an air traffic management system, including but not limited to integrated track, primary / secondary radar, ADS-B, etc.; preprocessing the received data (e.g., decoding raw ADS-B data) and converting it into a structured data format suitable for subsequent processing; extracting key data fields from the decoded data, including aircraft flight number, location information (longitude, latitude, flight altitude), speed, heading, etc.; counting the number of flights in the airspace according to a predetermined time step and the real-time position of the aircraft, deduplicating data according to the unique identifier of the flight (e.g., flight number), and removing invalid or incomplete data; defining the lookback window length L and the prediction time step T, constructing a time series according to the number of flights in the airspace at each time step, and generating historical air traffic flow data X. 1:L ={x1,x2,…x L The historical traffic data is formatted and standardized to ensure that all traffic indicators meet the model input requirements; the historical traffic data is stored for use by the subsequent prediction model; wherein, the trajectory data is dynamically updated in real time and the frequency of recording the trajectory data is not limited to once every 15 minutes.

[0173] As a refinement of the above embodiments, before executing step 101, the method may also adopt, but is not limited to, the following implementation methods, for example: training a preset deep learning framework to obtain an air traffic prediction model, the air traffic prediction model being used to execute the air traffic prediction method; the air traffic prediction model including a frequency sensing module, a sequence modeling module, and a KAN prediction module; the frequency sensing module being used to decompose and extract features from the historical air traffic input to the air traffic prediction model to generate time-series features, the time-series features including the time-invariant features and the time-varying features; the sequence modeling module for enhancing the time-series features being used to fuse the time-series features with the multi-resolution time-series feature encoding sequence, and extracting time-series dependencies from the fusion result based on the long short-term memory network; the KAN prediction module based on the attention mechanism being used to map the weighted time-invariant time-series feature sequence and the weighted time-varying time-series feature to the prediction dimension respectively and output a predicted air traffic sequence, the predicted air traffic sequence including the predicted time-invariant air traffic sequence and the predicted time-varying air traffic sequence.

[0174] To more clearly illustrate the architecture of the air traffic prediction model provided in the above embodiments, please refer to [link / reference needed]. Figures 2 to 4 , Figure 2 This is a schematic diagram of the overall structure of an air traffic prediction model provided in an embodiment of this disclosure. Figure 3 This is a schematic diagram of the structure of a decomposition module provided in an embodiment of the present disclosure. Figure 4 This is a schematic diagram of the structure of a KAN prediction module provided in an embodiment of the present disclosure. Figure 2 In this context, the air traffic flow sequence is the historical air traffic flow. The air traffic flow prediction model includes the temporal feature encoding module for executing step 102. The frequency sensing module includes the decomposition module, linear mapping, and convolutional layer. The frequency sensing module is used to execute the steps involved in formulas (4) to (13). The sequence modeling module includes an embedding layer, a feature fusion module, and an LSTM network. The embedding layer is used for data splicing of the three multi-resolution temporal feature encoding sequences. The feature fusion module is used to execute step 105. The LSTM network is used to execute step 106. The attention-based KAN prediction module includes an attention module and a KAN prediction module. The attention-based KAN prediction module is used to execute the processes involved in formulas (19) to (27). Figure 3 The decomposition module shown corresponds to the data processing process described in formulas (4) to (8) in its internal structure. Figure 4 The KAN prediction module with the attention mechanism shown here has an internal structure for data processing, which can be referred to in Equations (19) to (27).

[0175] In some embodiments, the historical time step (L) length of the air traffic prediction model is set to 96. Simultaneously, the prediction time step (T) is set to five different lengths: {12, 24, 48, 96, 192}. That is, observation data from the past day is used to predict traffic flow for {3 hours, 6 hours, 12 hours, 1 day, 2 days}. In the frequency sensing module, the cutoff frequency f... c Set to 10% f s The Conv1D convolutional layer has 128 channels; the embedding dimensions of Emb1(·), Emb2(·), and Emb3(·) are 128, 64, and 64, respectively. The LSTM has two layers with 128 hidden units. The prediction network consists of two KANLayers with a hidden unit count of [384, 192, 1]. The scaling factor α is set to 0.6.

[0176] As a refinement of the above embodiments, when training the preset deep learning framework to obtain the air traffic prediction model as described in the execution step, the following implementation methods can also be adopted, but are not limited to: training the air traffic prediction model using the constructed training set, using a loss function for backpropagation to update the network parameters of the air traffic prediction model, and using the Adam optimizer to dynamically adjust the learning rate. The formula for the loss function is expressed as follows:

[0177] L dd =L time +αL freq Formula (30)

[0178] Among them, L dd Let L be the loss function. time For time-domain loss, L freq For frequency domain loss, α represents the scaling factor, used to control the impact of frequency domain loss on the total loss;

[0179] The formula for calculating the time-domain loss is as follows:

[0180]

[0181] Where, x ij Represents the true value of the time-domain sequence. The predicted value of the time-domain sequence is represented by T, the length of the predicted time-domain sequence is represented by N, the number of time-domain sequence samples in a training batch is represented by i and j, which are subscript variables, where i is the length index and j is the sample index.

[0182] The formula for calculating the frequency domain loss is as follows:

[0183] y = DFT(x),

[0184]

[0185] Where DFT(·) denotes Discrete Fourier Transform, x is a time-domain sequence, and y is a frequency-domain sequence. To predict time-domain sequences, To predict the frequency domain sequence, y ij It is the true value of the frequency domain sequence. It is the predicted value of the frequency domain sequence. The length of the frequency domain sequence after the real Fourier transform is given, T represents the length of the predicted time domain sequence, n is the number of frequency domain sequence samples in a training batch, and i and j represent subscript variables, where i is the length index and j is the sample index.

[0186] In summary, the embodiments disclosed herein achieve the following effects:

[0187] 1. By performing multi-resolution temporal feature encoding on the historical air traffic, a multi-resolution temporal feature encoding sequence is generated. The multi-resolution temporal feature encoding sequence includes significant multi-resolution temporal features corresponding to the historical air traffic. Simultaneously, the multi-resolution temporal feature encoding sequence is concatenated with the time-invariant features and the time-varying features respectively to enhance the temporal features of the time-varying features and the time-invariant features. This makes it easier for the air traffic prediction model to capture the temporal correlation of air traffic changes. In addition, the normalized historical air traffic is decomposed based on Fourier transform to obtain time-invariant and time-varying sequences, enabling the air traffic prediction model to perceive different temporal frequencies. Furthermore, by integrating the aforementioned mechanism into the air traffic prediction model, the performance of the air traffic prediction model is improved and the accuracy of the prediction results is enhanced.

[0188] Corresponding to the aforementioned air traffic flow prediction method, this invention also proposes an air traffic flow prediction device. Since the device embodiments of this invention correspond to the method embodiments described above, details not disclosed in the device embodiments can be referred to the method embodiments described above, and will not be repeated here.

[0189] Figure 5 This is a schematic diagram of the structure of an air traffic flow prediction device provided in an embodiment of this disclosure, as shown below. Figure 5 As shown, it includes:

[0190] Receiving unit 21 is used to receive air traffic flow data and preprocess it to obtain historical air traffic flow data;

[0191] Encoding unit 22 is used to perform multi-resolution temporal feature encoding on the historical air traffic flow to generate a multi-resolution temporal feature encoding sequence;

[0192] Processing unit 23 is used to decompose the normalized historical air traffic flow based on Fourier transform to obtain time-invariant sequence and time-varying sequence;

[0193] Extraction unit 24 is used to extract features from the time-invariant sequence and the time-varying sequence respectively to obtain time-invariant features and time-varying features;

[0194] The splicing unit 25 is used to splice the multi-resolution temporal feature encoding sequence with time-invariant features and time-varying features respectively to obtain temporal enhanced time-varying sequence embedded data and temporal enhanced time-invariant sequence embedded data;

[0195] Learning unit 26 is used to learn time-varying sequence embedding data and time-invariant sequence embedding data based on long short-term memory network to obtain time-invariant time series features and time-varying time series features;

[0196] Weighting unit 27 is used to perform weighting processing based on the attention scores corresponding to the time-invariant time-series features and the time-varying time-series features respectively, to obtain weighted time-invariant time-series feature sequences and weighted time-varying time-series features;

[0197] Prediction unit 28 is used to predict air traffic based on the weighted time-invariant time series feature sequence and the weighted time-varying time series feature sequence, respectively, to obtain the predicted time-invariant air traffic traffic sequence and the predicted time-varying air traffic traffic sequence.

[0198] The inverse normalization unit 29 is used to perform inverse normalization processing on the sum of the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence to obtain the actual air traffic flow prediction result.

[0199] The air traffic flow apparatus provided in this disclosure receives and preprocesses air traffic flow data to obtain historical air traffic flow; performs multi-resolution time-series feature encoding on the historical air traffic flow to generate a multi-resolution time-series feature encoding sequence; decomposes the normalized historical air traffic flow based on Fourier transform to obtain a time-invariant sequence and a time-varying sequence; extracts features from the time-invariant sequence and the time-varying sequence respectively to obtain time-invariant features and time-varying features; concatenates the multi-resolution time-series feature encoding sequence with the time-invariant features and the time-varying features respectively to obtain time-series enhanced time-varying sequence embedded data and time-series enhanced time-invariant sequence embedded data; and uses long short-term time-series data... The memory network learns from time-varying sequence embedding data and time-invariant sequence embedding data respectively to obtain time-invariant time-series features and time-varying time-series features. Based on the attention scores corresponding to the time-invariant and time-varying time-series features respectively, weighted processing is performed to obtain weighted time-invariant time-series feature sequences and weighted time-varying time-series features. Air traffic prediction is performed on the weighted time-invariant and weighted time-varying time-series feature sequences using a KAN prediction network to obtain predicted time-invariant air traffic flow sequences and predicted time-varying air traffic flow sequences. The sum of the predicted time-invariant and predicted time-varying air traffic flow sequences is inversely normalized to obtain the actual air traffic flow prediction result. Compared with related technologies, this method generates a multi-resolution time-series feature encoding sequence by encoding the historical air traffic data using multi-resolution time-series features. This sequence includes significant multi-resolution time-series features corresponding to the historical air traffic. Furthermore, the multi-resolution time-series feature encoding sequence is concatenated with both time-invariant and time-varying features to enhance their temporal characteristics. This makes it easier for the air traffic prediction model to capture temporal correlations in air traffic changes. Additionally, the normalized historical air traffic is decomposed using Fourier transform to obtain time-invariant and time-varying sequences, enabling the air traffic prediction model to perceive different temporal frequencies. Finally, by integrating the aforementioned mechanisms into the air traffic prediction model, the model achieves a performance breakthrough and improves the accuracy of prediction results.

[0200] It should be noted that the foregoing explanation of the method embodiments also applies to the apparatus of this embodiment, and the principle is the same, so it is not limited in this embodiment.

[0201] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0202] Figure 6A schematic block diagram of an example electronic device 300 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0203] like Figure 6 As shown, device 300 includes a computing unit 301, which can perform various appropriate actions and processes based on a computer program stored in ROM (Read-Only Memory) 302 or a computer program loaded from storage unit 308 into RAM (Random Access Memory) 303. RAM 303 can also store various programs and data required for the operation of device 300. The computing unit 301, ROM 302, and RAM 303 are interconnected via bus 304. I / O (Input / Output) interface 305 is also connected to bus 304.

[0204] Multiple components in device 300 are connected to I / O interface 305, including: input unit 306, such as keyboard, mouse, etc.; output unit 307, such as various types of monitors, speakers, etc.; storage unit 308, such as disk, optical disk, etc.; and communication unit 309, such as network card, modem, wireless transceiver, etc. Communication unit 309 allows device 300 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0205] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, CPUs (Central Processing Units), GPUs (Graphics Processing Units), various special-purpose AI (Artificial Intelligence) computing chips, various computing units running machine learning model algorithms, DSPs (Digital Signal Processors), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above, such as methods for predicting air traffic flow. For example, in some embodiments, the air traffic flow prediction method may be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program may be loaded and / or installed on device 300 via ROM 302 and / or communication unit 309. When the computer program is loaded into RAM 303 and executed by the computing unit 301, one or more steps of the methods described above may be performed. Alternatively, in other embodiments, the computing unit 301 may be configured to perform the aforementioned air traffic prediction method by any other suitable means (e.g., by means of firmware).

[0206] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, FPGAs (Field Programmable Gate Arrays), ASICs (Application-Specific Integrated Circuits), ASSPs (Application-Specific Standard Products), SOCs (System-on-Chips), CPLDs (Complex Programmable Logic Devices), computer hardware, firmware, software, and / or combinations thereof. These various implementations may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0207] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0208] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, RAM, ROM, EPROM (Electrically Programmable Read-Only Memory) or flash memory, optical fiber, CD-ROM (Compact Disc Read-Only Memory), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0209] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (Cathode-Ray Tube) or LCD (Liquid Crystal Display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0210] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include LANs (Local Area Networks), WANs (Wide Area Networks), the Internet, and blockchain networks.

[0211] Computer systems can include clients and servers. Clients and servers are generally geographically separated and typically interact via communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. A server can be a cloud server, also known as a cloud computing server or cloud host, a hosting product within the cloud computing service ecosystem, addressing the shortcomings of traditional physical hosts and VPS (Virtual Private Server, or simply "VPS") services, such as high management difficulty and weak business scalability. Servers can also be servers for distributed systems or servers incorporating blockchain technology.

[0212] It's important to note that artificial intelligence (AI) is the study of enabling computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, and planning). It encompasses both hardware and software technologies. AI hardware technologies generally include sensors, dedicated AI chips, cloud computing, distributed storage, and big data processing. AI software technologies primarily include computer vision, speech recognition, natural language processing, machine learning / deep learning, big data processing, and knowledge graph technologies.

Claims

1. A method for predicting air traffic flow, characterized in that, include: Receive and preprocess air traffic data to obtain historical air traffic data; The historical air traffic flow is encoded using multi-resolution temporal features to generate a multi-resolution temporal feature encoding sequence. The normalized historical air traffic flow is decomposed based on Fourier transform to obtain time-invariant and time-varying sequences. Feature extraction is performed on the time-invariant sequence and the time-varying sequence respectively to obtain time-invariant features and time-varying features; The multi-resolution temporal feature encoding sequence is concatenated with time-invariant features and time-varying features respectively to obtain temporally enhanced time-varying sequence embedding data and temporally enhanced time-invariant sequence embedding data. Based on the Long Short-Term Memory network, time-varying sequence embedding data and time-invariant sequence embedding data are learned respectively to obtain time-invariant time series features and time-varying time series features; The attention scores corresponding to the time-invariant and time-varying time-series features are weighted to obtain the weighted time-invariant and time-varying time-series feature sequences. Based on the KAN prediction network, air traffic is predicted from the weighted time-invariant feature sequence and the weighted time-varying feature sequence, respectively, to obtain the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence. The sum of the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence is inversely normalized to obtain the actual air traffic flow prediction result.

2. The method according to claim 1, characterized in that, The step of performing multi-resolution temporal feature encoding on the historical air traffic flow to generate a multi-resolution temporal feature encoding sequence includes: The historical air traffic flow is encoded using time-resolution temporal feature encoding to obtain a time-series feature encoding sequence; The historical air traffic flow is encoded using hourly time-series features to obtain an hourly time-series feature encoding sequence; The historical air traffic flow is encoded using time-series features at week-day resolution to obtain a week-day time-series feature encoding sequence. The multi-resolution time-series feature encoding sequence includes the time-series feature encoding sequence, the hour-series feature encoding sequence, and the week-day time-series feature encoding sequence.

3. The method according to claim 1, characterized in that, The decomposition of the normalized historical air traffic flow based on Fourier transform to obtain time-invariant and time-varying sequences includes: The normalized historical air traffic flow time-domain sequence is transformed into a frequency-domain complex sequence based on the discrete Fourier transform. The frequency domain complex sequence is filtered for high-frequency components using a low-pass filter, and the filtering result is converted from the frequency domain to the time domain to obtain the time-invariant sequence. The time-varying sequence is obtained by calculating the difference between the time-domain sequence and the time-invariant sequence.

4. The method according to claim 2, characterized in that, The step of extracting features from the time-invariant sequence and the time-varying sequence respectively to obtain time-invariant features and time-varying features includes: The time-invariant sequence and the time-varying sequence are linearly mapped to obtain an aligned time-invariant sequence aligned to a first preset sequence length and an aligned time-varying sequence aligned to a second preset sequence length. High-dimensional features are extracted from the aligned time-invariant sequence and the aligned time-varying sequence respectively using convolutional layers to obtain the time-invariant features and the time-varying features.

5. The method according to claim 4, characterized in that, The step of concatenating the multi-resolution temporal feature encoded sequence with time-invariant features and time-varying features respectively to obtain temporally enhanced time-varying sequence embedded data and temporally enhanced time-invariant sequence embedded data includes: High-dimensional features are extracted from the time-series feature encoding sequence, the hour-series feature encoding sequence, and the week-series feature encoding sequence, and the three feature extraction results are concatenated to obtain a latent time-series representation. The latent temporal representation is concatenated with the time-invariant feature to obtain the temporally enhanced time-invariant sequence embedding data; The latent temporal representation is concatenated with the time-varying features to obtain the time-enhanced time-varying sequence embedding data.

6. The method according to claim 1, characterized in that, The process of receiving and preprocessing air traffic data to obtain historical air traffic data includes: The air traffic data, which is acquired and updated in real time, is converted into a structured data format that can be processed subsequently. Extract key data fields from the converted air traffic flow data and remove duplicate data. Based on the deduplicated key data fields, a time series is constructed according to the historical time step length to generate the historical air traffic.

7. The method according to any one of claims 1-6, characterized in that, Before receiving and preprocessing air traffic data to obtain historical air traffic data, the method includes: A pre-defined deep learning framework is trained to obtain an air traffic prediction model, which is used to execute the air traffic prediction method. The air traffic prediction model includes a frequency sensing module, a sequence modeling module, and a KAN prediction module. The frequency sensing module is used to decompose and extract features from the historical air traffic input to the air traffic prediction model to generate time-series features, which include time-invariant features and time-varying features. The sequence modeling module used for the temporal feature enhancement is used to fuse the temporal features with the multi-resolution temporal feature encoded sequence, and to extract temporal dependencies from the fusion result based on the long short-term memory network; The KAN prediction module based on the attention mechanism is used to map the weighted time-invariant time-series feature sequence and the weighted time-varying time-series feature sequence to the prediction dimension and output the predicted air traffic flow sequence, which includes the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence.

8. The method according to claim 7, characterized in that, The process of training a preset deep learning framework to obtain an air traffic prediction model includes: The air traffic prediction model is trained using the constructed training set. Backpropagation is performed using a loss function to update the network parameters of the air traffic prediction model. The learning rate is dynamically adjusted using the Adam optimizer. The formula for the loss function is as follows: L dd =L time +αL freq Among them, L dd Let L be the loss function. time For time-domain loss, L freq For frequency domain loss, α represents the scaling factor, used to control the impact of frequency domain loss on the total loss; The formula for calculating the time-domain loss is as follows: Where, x ij Represents the true value of the time-domain sequence. The predicted value of the time-domain sequence is represented by T, the length of the predicted time-domain sequence is represented by N, the number of time-domain sequence samples in a training batch is represented by i and j, which are subscript variables, where i is the length index and j is the sample index. The formula for calculating the frequency domain loss is as follows: y<DFT(x), Where DFT(·) denotes Discrete Fourier Transform, x is a time-domain sequence, and y is a frequency-domain sequence. To predict time-domain sequences, To predict the frequency domain sequence, y ij It is the true value of the frequency domain sequence. It is the predicted value of the frequency domain sequence. The value represents the length of the frequency domain sequence after the real Fourier transform, T represents the length of the predicted time domain sequence before the real Fourier transform, n is the number of frequency domain sequence samples in a training batch, and i and j represent subscript variables, where i is the length index and j is the sample index.

9. An air traffic flow prediction device, characterized in that, include: The receiving unit is used to receive air traffic flow data and preprocess it to obtain historical air traffic flow data. The encoding unit is used to perform multi-resolution temporal feature encoding on the historical air traffic flow to generate a multi-resolution temporal feature encoding sequence. The processing unit is used to decompose the normalized historical air traffic flow based on Fourier transform to obtain time-invariant and time-varying sequences. An extraction unit is used to extract features from the time-invariant sequence and the time-varying sequence respectively, to obtain time-invariant features and time-varying features; The splicing unit is used to splice the multi-resolution temporal feature encoding sequence with time-invariant features and time-varying features respectively to obtain temporally enhanced time-varying sequence embedded data and temporally enhanced time-invariant sequence embedded data. The learning unit is used to learn time-varying sequence embedding data and time-invariant sequence embedding data based on the long short-term memory network, respectively, to obtain time-invariant time series features and time-varying time series features; The weighting unit is used to perform weighting processing on the attention scores corresponding to the time-invariant time-series features and the time-varying time-series features respectively, to obtain the weighted time-invariant time-series feature sequence and the weighted time-varying time-series feature; The prediction unit is used to predict air traffic based on the weighted time-invariant time-series feature sequence and the weighted time-varying time-series feature sequence, respectively, to obtain the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence. The inverse normalization unit is used to inverse normalize the sum of the predicted time-invariant air traffic flow sequence and the predicted time-varying air traffic flow sequence to obtain the actual air traffic flow prediction result.

10. An electronic device, characterized in that, include: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-6.

Images