Aerial flight flow prediction method based on lightweight adaptive network

Abstract

The invention belongs to the technical field of air traffic management, and relates to an air flight flow prediction method based on a lightweight adaptive network, which comprises the following steps of: firstly, acquiring data and fusing the data to generate a flow time sequence matrix indexed by airport identification and timestamp; secondly, constructing a self-adaptive dynamic graph; performing mixed graph convolution on the output dynamic graph and node features to form a spatial feature tensor, and performing frequency domain enhancement on the spatial feature tensor to obtain an event enhancement sequence; then processing the event strengthening sequence through an hour-level large convolution kernel and a minute-level expansion small convolution kernel, amplifying a congestion peak based on trend-fluctuation gating after time alignment, and outputting a fusion feature sequence; and finally, carrying out recursive decoding and prediction to obtain the predicted air flight flow. According to the method provided by the invention, under the conditions of dynamically changing airspace topology and strong noise and multi-scale coupled time sequence data, a set of lightweight spatial-temporal model is constructed and updated online in a self-adaptive manner, so that high-precision multi-step prediction of the multi-element flight flow can be realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of air traffic management, and relates to an air flight flow prediction method based on a lightweight adaptive network. BACKGROUND

[0002] With the rapid development of the global economy, the demand for air transportation is increasing. The prediction of air flight flow has gradually become the focus of the research field. More accurate air flow prediction can not only effectively help airlines improve flight transportation efficiency, but also provide strong support for optimizing air traffic management strategies.

[0003] With the increasing depth of research on multivariate time series prediction, deep learning models have attracted widespread attention and recognition due to their excellent nonlinear modeling capabilities and strong complex feature extraction capabilities. In recent years, deep learning methods based on Transformer have made significant progress in time series prediction, enabling more accurate prediction of air flight flow.

[0004] In the prior art, airport-route flow prediction usually regards each airport as a node and the two-way routes in the flight schedule as static directed edges. First, a multidimensional input is formed based on ADS-B and A-SMGCS, and then a deep recurrent network or a temporal convolution network is used to capture the time evolution. The mainstream solution uses a fixed topology graph convolution or graph attention as a spatial encoder to map node features to a hidden space and then input them into LSTM, GRU or Temporal-CNN to complete multi-step extrapolation. The output includes short-term prediction results such as takeoff and landing times, and flight segment throughput.

[0005] However, in multivariate time series prediction, the interaction between different variables (dimensions) may change over time, leading to the following core challenges in modeling dynamic cross-dimensional associations: non-stationary correlation and noise interference. On the one hand, the correlation between sequences presents non-stationary time-varying characteristics, and the strength and direction of the association change dynamically over time. For air flight flow prediction, the flow dependency relationship between different airports and routes may suddenly change due to adverse weather, temporary airspace control, and other events. The static spatial assumption of traditional time series models cannot capture such dynamic patterns, resulting in limited prediction accuracy. On the other hand, the multivariate time series data collected by sensors for monitoring air flight flow has noise interference, mainly manifested in: (1) Non-uniform distribution of noise energy in the frequency domain; (2) Coupling diffusion effect between spatial dimensions; (3) Non-stationary changes in noise over time. These noises can seriously affect the accuracy of feature extraction and reduce the robustness of the prediction model.

[0006] For example, patent CN118573526A discloses a mixed coherent interference suppression beamforming method in the background of impulsive noise, which adopts nonlinear compression preprocessing and differential smoothing matrix to recover the coherent interference subspace in the MVDR framework to construct deep nulls, thereby suppressing impulsive noise. However, this scheme focuses on the spatial decorrelation of array signals, and the random non-stationary noise only depends on the fixed scale parameter mu and single difference compression, which cannot describe the noise coupling and diffusion among multiple nodes, and lacks a time-rolling update frequency energy threshold, resulting in over-suppression or residual noise in the mixed scene of peak, peak, and far-end crosstalk of airport traffic sequences.

[0007] Further, in patent CN116630704A, a ground feature classification network model based on attention enhancement and dense multi-scale is proposed, which expands the receptive field based on DenseASPP-SE hollow pyramid convolution, superimposes channels and spatial double attention at the encoding end, and performs multi-scale fusion by fixing attention weights by upsampling and splicing multi-level features at the decoding end. However, this method has static weights and lacks scenario-oriented gating adjustment, cannot switch fusion strategies in real time according to congestion intensity or prediction time interval, and only targets single images, making it difficult to handle time-coupled features in air traffic management.

[0008] Overall, the above-mentioned prior art routes can meet the conventional traffic fluctuation prediction under the assumption of static network and periodic retraining, but when encountering flight batch adjustment, continuous flow control, or severe weather refresh, the description granularity and update delay of spatial dependence and time sequence features are still limited, resulting in a significant decrease in prediction accuracy.

[0009] Therefore, it is of great significance to study an air flight traffic prediction method based on a lightweight adaptive network to solve the problems in the prior art. SUMMARY

[0010] The purpose of the present application is to solve the problems in the prior art and provide an air flight traffic prediction method based on a lightweight adaptive network.

[0011] To achieve the above-mentioned purposes, the technical solutions adopted by the present application are as follows:

[0012] The air flight traffic prediction method based on the lightweight adaptive network comprises the following steps:

[0013] S1: Collect airport takeoff and landing counts, taxi queue lengths, congestion levels, schedule changes, release instructions, and weather radar data, and fuse them to generate a traffic time sequence matrix indexed by airport identifier and timestamp;

[0014] S2: Establish a topology with airports as nodes and direct routes as directed edges, calculate edge weights according to the nearest window node traffic synchronization degree, normalized flight quantity and runway occupancy ratio, and dynamically adjust in combination with airspace capacity announcements and relevance thresholds to obtain an adaptive dynamic graph;

[0015] S3: Perform hybrid graph convolution on the dynamic graph output in step S2 and node features: one-hop propagation introduces time decay to obtain short-range embedding, multi-hop propagation superimposes distance decay to obtain long-range embedding, and congestion-time interval weights are fused based on running level gating to form a spatial feature tensor;

[0016] S4: Perform trend-residual decomposition and fast Fourier transform on the spatial feature tensor to suppress random high frequencies with an energy threshold, and reserve the main frequency using a flight peak template (which is an event frequency template mapped based on the daily flight plan peak period of the airport) and then inverse transform to obtain a denoised trend sequence and a denoised fluctuation sequence, and concatenate the two to obtain an event enhanced sequence;

[0017] Among them, the fast Fourier transform: using fast Fourier transform to convert the input time series from time domain to frequency domain, which is convenient for further frequency domain analysis and processing. Given the input time series, perform fast Fourier transform along the time dimension to obtain the corresponding frequency domain representation;

[0018] S5: Process the event enhanced sequence with a large convolution kernel at the hour level and an expanded small convolution kernel at the minute level respectively, time align, and then amplify the congestion peak based on the trend-fluctuation gating to output the fusion feature sequence;

[0019] S6: Perform recursive decoding on each node based on the fusion feature sequence, take the fusion feature sequence as the initial input of the decoder, input the previous prediction value and real-time weather vector in the loop, and output the future window ground traffic, and then obtain the predicted air flight traffic through adjacent averaging.

[0020] As a preferred technical solution:

[0021] The air flight traffic prediction method based on a lightweight adaptive network as described above, step S1 includes the following sub-steps:

[0022] S11: The take-off and landing counts output by the airport surface monitoring equipment (ADS-B, A-SMGCS, ground radar) in real time taxi queue length and ground congestion level wherein one minute is taken as the sampling interval; write the traffic index set according to the airport number i-time stamp t to obtain the airport i basic traffic time series;

[0023] S12: Write the scheduling change stream and the tower release instruction stream into the time series database in the order of arrival; extract the flight type, expected take-off and landing period, and runway-taxiway occupation time from each tower release to generate a dispatch feature vector Then, according to the airport number i, map to to depict the immediate impact of the dispatch side on the flow;

[0024] S13: Map the weather radar echo intensity high-altitude wind field vector and the cloud top height to the running level label by level mapping g(·) associate with the same time sampling record to inject the external meteorological influence dimension into the flow sequence; the mapping relationship is as follows:

[0025]

[0026] where i is the airport number, t is the timestamp, g(·) is constructed based on industry running thresholds, to describe the running level of airport i at timestamp t;

[0027] The level mapping can also be directly mapped to the airport management order;

[0028] S14: Concatenate the basic flow sequence dispatch feature vector and running level label in chronological order to form a multi-dimensional time series matrix that is, a flow time series matrix indexed by airport identifier and timestamp.

[0029] The air flight flow prediction method based on a lightweight adaptive network as described above, step S2 includes the following sub-steps:

[0030] S21: First, map the airport list to the node set Then generate the edges in the origin-destination in the route database, and establish directed edges for each direct route and write them into the edge set where e ij represents a directed physical edge from node (origin airport) to (destination airport); finally, the physical topology constrained by the airport-route accessibility relationship is completed, and the topology structure can be represented as the basic physical graph:

[0031] S22: Get node traffic according to flight plan and two-way flights The two are coupled and estimated to calculate the flow coupling intensity The flow coupling strength Assign physical edge e ij As the weight, define the first layer mixed graph as: Among them, the weight function Represents edge e ij The weight of is the flow coupling strength;

[0032] S23: Monitor the airspace control announcement stream and parse each announcement into a flight segment capacity factor like Then the corresponding edge weight If the proportion is reduced, If the notification requires a 30% reduction in traffic for a certain segment, the weight of the corresponding directed edge is scaled by 70%. If the notification requires a temporary closure of a segment, the edge weight is set to zero.

[0033] The above capacity mapping enables the topology to shrink instantly in emergency control scenarios, ensuring that the graph structure remains synchronized with the available airspace.

[0034] S24: Node traffic sequence within the most recent prediction window Δt Perform sliding correlation calculation; weight the edge (i, j) whose correlation coefficient is lower than the confidence threshold η Decrease until deleted, and simultaneously increase the weight of high-correlation edges with correlation coefficients greater than or equal to the confidence threshold η to generate an adaptively updated dynamic graph

[0035] S25: Together with the node sequence Common packaging

[0036] In the above-mentioned method for predicting air traffic flow based on lightweight adaptive network, the traffic coupling strength is calculated in step S22. The specific process is as follows: First, Perform maximum-minimum normalization within the most recent Δt to obtain a comparable scale with consistent direction. Secondly, extract the synchronous fluctuations of the traffic series of the two nodes within Δt and calculate the sliding co-movement degree to obtain the correlation component reflecting consistency. Thirdly, use the runway-taxiway occupancy ratio and flight segment distance of the two airports to attenuate the correlation component to reduce coupling misjudgment caused by excessive distance or large resource differences. Finally, superimpose the attenuated correlation component with the normalized flight volume according to the weight to obtain

[0037] The air flight traffic prediction method based on the lightweight adaptive network as described above, step S3 includes the following sub-steps:

[0038] S31: Constructing a node feature matrix is expressed as Wherein is the feature vector of the i-th airport; the dynamic graph output by step S2 is loaded synchronously together with the node feature matrix ; the weight matrix w t is normalized and a unit diagonal is explicitly added to form a self-loop normalized adjacency A set of propagation depths W = {W S , W L} is defined, wherein W S is a shallow propagation (a shallow element in the set of propagation depths), responsible for perceiving the nearest neighbor topology of the airport (such as the directly connected route network), and the depth value of the shallow propagation is d s ; W L is a deep propagation (a deep element in the set of propagation depths), responsible for perceiving the cross-regional topology (such as the indirectly connected long-distance route association), and the depth value of the deep propagation is d l ;

[0039] S32: Convolution propagation is performed on the one-hop neighborhood of depth d s ; the latest traffic vector of the direct neighbor (or "one-hop neighbor") of each airport node is aggregated by using , in which a decreasing coefficient based on the time difference δ is introduced to ensure that a one-minute level sudden increase (such as a sudden drop) is fully retained; the aggregation result is added to the local feature residual and projected to the hidden space by W S to form a short-range embedding The embedding highlights immediate resource competition features such as runway occupancy and taxiway queuing, and filters far-end weakly related noise;

[0040] S33: Convolution propagation is performed on the multi-hop neighborhood of depth d l ; the K-th power of is calculated to obtain a K-hop reachable matrix, and an exponential decay λ K is introduced for different hop distances to suppress overlong diffusion, wherein 0<λ<1 and K≤6, generally not more than 3; the far-end airport traffic features are weighted and converged using the decay matrix, and the feature mapping is completed by W L to obtain a long-range embedding

[0041] S34: For the scene of severe weather or capacity contraction, the operation level is combined with the meteorological warning intensity (real-time alert information from aviation meteorological flow, such as the level quantization value of thunderstorm, heavy precipitation, etc. alert, and the meteorological radar data collected in step S1 belong to aviation meteorological input, used to represent the severity of weather alert of airport i at time stamp t) input gate mapper generates coefficients To With Perform element-by-element scaling, re-write the gated memory pool and use the exponential smoothing factor μ for temporal smoothing to avoid sudden changes in embedding caused by alert jitter; by mapping the running level to a scaling coefficient between 0 and 1, the information flow of the restricted airport is dynamically adjusted to achieve fine-grained suppression of strong noise and airspace bursts, while the unrestricted node maintains the original weight output and does not damage the learned weights;

[0042] S35: Based on the differentiated needs of airport operation scenarios, the short-range embedding and long-range embedding after gating are segmented and adaptively fused to obtain a cross-scale vector The normalized congestion index is calculated from the real-time taxi queue length and runway occupancy rate The length of the current rolling prediction window τ is read at the same time p ; select the weight group according to the congestion-time distance joint condition, the rules are as follows:

[0043]

[0044] Where θ c is the congestion threshold; τ0 is the prediction time distance dividing point; α S and β S represent the weight coefficients of short-range (S) embedding, used for low congestion and short prediction window scenarios, α L and β L represent the weight coefficients of long-range (L) embedding, used for high congestion or long prediction window scenarios (other cases), and satisfy α S > β S , α L < β L ; α S , α L emphasize short-range influence to highlight immediate resource competition; β S , β L strengthen long-range coupling to improve long-term trend discrimination;

[0045] The segmented weight function ensures that the model prioritizes local information in light congestion or short-term prediction scenarios, and automatically increases the weight of long-range signals in heavy congestion or long-term prediction scenarios, dynamically adapting to prediction needs;

[0046] S36: Collect the stack along the time dimension to form a spatial feature tensor

[0047] In the above-mentioned lightweight adaptive network-based air traffic flow prediction method, step S4 includes the following sub-steps:

[0048] S41: The spatial feature tensor generated in step S3 Split by node index i∈{1,…,N} to obtain independent time series At the same time, record the upstream standard one-minute sampling interval and the current sliding window length L w ,Will As the starting input for frequency domain enhancement;

[0049] S42: For each Perform additive decomposition: Based on the scheduling cycle P (i) Adaptive selection window w (i) , separating the long-term trend component by sliding mean and short-term fluctuation component Isolate the masking effect of scheduling rules on instantaneous traffic;

[0050] S43: respectively and Perform a Fast Fourier Transform (FFT) to obtain:

[0051] Trend spectrum: long-term trend component Frequency domain representation, including amplitude spectrum and phase spectrum It mainly reflects the low-frequency characteristics dominated by regular factors such as scheduling plans (such as the frequency corresponding to the daily shift cycle);

[0052] Fluctuation spectrum: short-term fluctuation component Frequency domain representation, including amplitude spectrum and phase spectrum It mainly reflects the high-frequency characteristics dominated by sudden factors such as tower bursts and instantaneous resource competition (such as the frequency corresponding to minute-level surges);

[0053] S44: Amplitude spectrum of trend spectrum and the amplitude spectrum of the fluctuation spectrum Calculate the energy density E (i) (f)=|A (i) (f)| 2 , and then generate the energy threshold according to the following formula:

[0054]

[0055] Among them, κ∈(0,1) is the learnable proportional coefficient;

[0056] like Then the corresponding frequency coefficient is set to zero to weaken the random disturbance; after processing, the high-energy frequency component is retained to obtain the filtered spectrum, and the absolute value of the spectrum is calculated to obtain the amplitude spectrum;

[0057] S45: Mapping Airport Peak Flight Schedules to Event Frequency Templates Y (i) , the amplitude spectrum peak value of the trend spectrum and the fluctuation spectrum after step S44 (in the amplitude spectrum A (i) (f) calculates the local maximum value, which is the amplitude spectrum peak) to match; when the frequency f∈Υ (i) and the amplitude is greater than the event threshold When the original value is retained, otherwise continue to suppress according to the rules of step S44;

[0058] S46: Perform inverse frequency domain conversion on the trend spectrum and fluctuation spectrum processed in steps S44 and S45 to reconstruct the noise reduction trend sequence and noise-reduced fluctuation series

[0059] S47: and Add point by point to generate a reinforcement sequence containing key event information Bundle Splice into feature stack;

[0060] S48: The energy threshold saved in the previous window Compare with the current window frequency distribution and update based on the exponential smoothing coefficient ξ

[0061] In the above-mentioned method for predicting air traffic volume based on a lightweight adaptive network, step S5 includes the following sub-steps:

[0062] S51: Strengthen the event sequence generated in step S4 Arrange in ascending order by airport number and concatenate in the channel dimension to obtain a two-dimensional feature stack Where N is the number of airports, L w is the current sliding window length; then a double-branch convolution is performed for U t Bind the hyperparameter templates (convolution kernel size, stride, and padding method) of the two convolution branches, and set the sampling interval Δ and the scheduling period P sched and sliding window length L w written to the module header as structured metadata;

[0063] S52: Adopting length One-dimensional large kernel U t Convolution is performed along the time axis, and the stride of the convolution layer is set to Overlap sampling, symmetrical mirroring of filling mode, output matrix Capture cross-hourly scheduling synchronization fluctuations and slowly changing congestion trends;

[0064] S53: Parallel use of length k S The one-dimensional small kernel U t Perform point-wise convolution with stride s S =1, so that the output M t Maintain high temporal resolution with the input, where k S ≤5. To improve the response to sparse spikes, dilated convolution is added to the small kernel with a dilation coefficient of d = 2. By inserting "spaces" (zero padding) between the elements of the convolution kernel, the receptive field is expanded without increasing the number of kernel parameters. The so-called parallelism means that step S53 and step S54 process the same input data independently to generate two types of features (high-resolution details / coarse-grained trends), which are then aligned and fused.

[0065] S54: Due to G t With M t Different dimensions, first G t Fill the end of the time with zero padding to the current sliding window length L w , then to M t Apply a linear interpolation function By factor Perform isometric scaling to align the two tensors in the time dimension and output them after alignment

[0066] S55: After completing the time dimension alignment, first align the trend tensor Normalize the nodes and calculate the relative increase of each airport Where (j,·) represents the jth row (full column), i.e., the trend vector of the jth airport; then the learnable interaction coefficient υ∈(0,1) is introduced to construct the trend gating term, and is combined with the fluctuation tensor Do element-wise multiplication to get the fused matrix:

[0067]

[0068] Among them, 1 is a matrix of all 1s, which is used to retain the original fluctuation amplitude baseline; ⊙ represents the element-by-element multiplication of the airport index and the time index; R t Amplify local fluctuations at peak congestion and compress random jitter in normal periods; υ controls the amplification strength and automatically tunes it during the training phase through backpropagation.

[0069] Based on the above interaction mechanism, the spike signal can be enhanced in the peak propagation path and the noise can be suppressed in the stable section, ensuring that the subsequent decoding can grasp the spatiotemporal diffusion trajectory of the traffic peak;

[0070] S56: P tThe channel is normalized and connected to the nonlinear activation function σ(·), and the output normalized tensor F t = σ(Norm(P t )) is obtained, which suppresses isolated noise frames and enhances continuous high-flow areas, and F t is packaged as a fusion feature sequence package <F t ,t>.

[0071] The air flight flow prediction method based on the lightweight adaptive network as described above, in step S52, to prevent overfitting, batch normalization and Dropout are added after large core convolution, and the ratio is fixed at 0.1.

[0072] The air flight flow prediction method based on the lightweight adaptive network as described above, step S6 includes the following sub-steps:

[0073] S61: recursively decoding the fusion feature sequence generated in step S5 node by node to obtain the predicted sequence of each airport node;

[0074] S62: reading the wind direction, precipitation, and thunderstorm indicators from the aviation meteorological flow and compressing them into a vector According to the airspace level where the node is located, the weight β (i) The hidden state is corrected to obtain the predicted value sequence from t+1 to t+H (H time steps in the future) If the wind direction suddenly changes or heavy rain is observed, the gating weight automatically increases the proportion of meteorological components, which can reduce the short-term prediction deviation in real time and retain the original time sequence memory;

[0075] S63: obtaining the corrected is obtained by recursively decoding the hidden state ground , and then the ground flow sequence G air is obtained by averaging the flow of adjacent nodes based on the topological edge table. The air flight flow sequence G air is obtained. Both of them are written into a unified message package together with the meteorological correction residual, and pushed to the online feedback update module to realize the real-time sharing of monitoring terminals, flow management units and airline dispatching seats, and provide closed-loop data support for subsequent topology update and model reuse.

[0076] The air flight flow prediction method based on the lightweight adaptive network as described above, the specific process of step S61 is: for each airport node, an independent hidden vector is allocated, and the initial value is obtained by linear mapping of the latest feature, and the cycle number is configured according to the predicted time length H, and the future window is sequentially unfolded; the strategy used in the unfolding process is: at the kth step, the future flow prediction value generated by the model at the k-1 step in the recursive decoding stage is re-mapped by the embedding layer and decoded with the next frame feature in parallel, and the generated At the same time, the time-increasing sampling probability scheduling is used to reduce or eliminate the dependence on the real label in the training stage.

[0077] The airport network changes frequently in actual operation with NOTAM, AUP and other announcements. If the static topology is still used, it will cause the release decision to be inconsistent with the available airspace. The present application constructs an initial weighted graph based on the flight volume-flow coordination degree, and immediately reduces or deletes the corresponding edge when receiving capacity changes. At the same time, the weakly coupled edge is adjusted by using the flow correlation, realizing the self-shrinking / expansion of the topology. In this way, the model structure seen by the tower and the ATFM is always consistent with the current flyable segment, avoiding incorrect release or excessive flow control due to lagging graph.

[0078] At the same time, the time-increasing sampling probability scheduling is used to reduce or eliminate the dependence on the real label in the training stage.

[0079] Minute-level data is mixed with scheduling period, thunderstorm disturbance and random error code. Directly feeding into the time series network is easy to misjudge the exception as a trend. The present application introduces frequency domain enhancement after spatial coding: an improved strategy is proposed, that is, first divide the flow sequence into long-term trend and short-term fluctuation, and then use energy threshold and peak template double screening to retain the main frequency of scheduling and real peak, and filter out random high frequency. Then, based on the hour-level large kernel convolution, the slow-changing congestion is described, and the minute-level expanded small kernel is used to capture the tower burst, and the trend-fluctuation gate is used to dynamically amplify the peak channel.

[0080] The main frequency of scheduling added in step S4 is the main frequency corresponding to the flight scheduling period, such as the 1 / 24 hour frequency corresponding to the daily peak, which is a high-energy main frequency component in the frequency domain.

[0081] The "real peak" refers to the sudden peak value in the actual flow due to sudden situations (such as tower burst, short-term flow surge), which is different from random noise. In the frequency domain processing of step S4, the real peak may appear as a high-amplitude value at a specific frequency in the frequency domain, which needs to be distinguished from random high-frequency noise and retained by energy threshold and peak template.

[0082] In the face of the business rhythm of continuous rolling prediction, the traditional scheme needs to rely on offline retraining to adapt to new weather or adjust the strategy, and there is a window period. The lightweight recursive decoder is an efficient decoding module designed by the application for the continuous rolling prediction scene, which has the characteristics of small parameter size (more than 60% of the parameters are reduced compared with the traditional LSTM / Transformer decoder) and strong real-time (single-step decoding time ≤5ms), which can adapt to the business rhythm of "minute-level update, hour-level rolling" in air traffic flow prediction. The core function is to dynamically generate multi-step prediction results based on the fusion feature sequence combined with real-time weather feedback, without offline retraining to quickly adapt to new weather, airspace control and other sudden scenes. The application proposes to inject real-time weather vectors at the inference end based on the lightweight recursive decoder, and correct the hidden state in real time through node-level gating; and the latest prediction results and residual error are fed back to the graph update module to form a closed loop, so that the prediction is adaptive without stopping the service. The tower and the airline can obtain the flow curve synchronized with the latest wind direction and precipitation changes to solve the decision-making blind area caused by prediction freezing in a high dynamic environment.

[0083] The lightweight recursive decoder is an efficient decoding module designed by the application for the continuous rolling prediction scene, which has the characteristics of small parameter size (more than 60% of the parameters are reduced compared with the traditional LSTM / Transformer decoder) and strong real-time (single-step decoding time ≤5ms), which can adapt to the business rhythm of "minute-level update, hour-level rolling" in air traffic flow prediction. The core function is to dynamically generate multi-step prediction results based on the fusion feature sequence combined with real-time weather feedback, without offline retraining to quickly adapt to new weather, airspace control and other sudden scenes.

[0084] Beneficial effects:

[0085] (1) The air flight flow prediction method based on the lightweight adaptive network can construct and online adaptively update a lightweight space-time model under the conditions of dynamically changing airspace topology and strong noise and multi-scale coupled time series data, so as to realize high-precision multi-element flight flow multi-step prediction.

[0086] (2) The air flight flow prediction method based on the lightweight adaptive network dynamically perceives noise distribution and strengthens key events through the FFT double-threshold denoising, event template frequency preservation and trend-fluctuation interactive amplification mechanism, solves the problem of noise interference; at the same time, through the congestion and time interval adaptive gating and trend-fluctuation interactive amplification mechanism, the multi-scale information flow is dynamically adjusted, which can dynamically adjust the gating, solves the problem that the single receptive field model cannot meet the multi-level demand, and greatly improves the air flight flow prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0087] Figure 1The schematic diagram of the overall flow of the air flight flow prediction method based on a lightweight adaptive network of the present application is shown. DETAILED DESCRIPTION

[0088] The present application will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present application and not used to limit the scope of the present application. Furthermore, it should be understood that after reading the content taught by the present application, those skilled in the art can make various modifications or changes to the present application, and these equivalent forms also fall within the scope defined by the appended claims of the present application.

[0089] The air flight flow prediction method based on a lightweight adaptive network comprises the following steps as shown in the figure: Figure 1

[0090] S1: Collect airport take-off and landing counts, taxi queue lengths, congestion levels, schedule changes, release instructions and weather radar data, and perform fusion to generate a flow time series matrix indexed by airport identification and time stamp;

[0091] S11: Take-off and landing counts output in real time by airport surface monitoring equipment (ADS-B, A-SMGCS, ground radar) taxi queue lengths and ground congestion levels wherein one minute is sampled as an interval; write the flow index set according to the airport number i-time stamp t, to obtain the airport i basic flow time series

[0092] S12: Write the schedule change stream pushed by the flight plan and the tower release instruction stream according to the arrival order into the time series database; extract the flight type, expected take-off and landing period and runway-taxiway occupation time length for each tower release to generate a dispatch feature vector then map to to depict the immediate impact of the dispatch side on the flow;

[0093] S13: Perform level mapping g(·) on weather radar echo intensity upper air wind field vector and cloud top height to obtain a running level label associate to the same time sampling record to inject an external weather influence dimension into the flow sequence; the mapping relationship is as follows:

[0094]

[0095] wherein i is the airport number, t is the minute-level time index, g(·) is constructed based on industry running thresholds, ​Used to describe the operability level of airport i at time t;

[0096] S14: Basic traffic sequence Scheduling feature vector with runlevel labels Splice in chronological order to form a multivariate time series matrix That is, the traffic time series matrix indexed by airport ID and timestamp.

[0097] S2: A topology is established with airports as nodes and direct routes as directed edges. Edge weights are calculated based on the nearest window node traffic synchronization, normalized flight volume, and runway occupancy ratio. This is combined with airspace capacity notifications and dynamic adjustment of correlation thresholds to obtain an adaptive dynamic graph.

[0098] S21: First, map the airport list to nodes, assign a node to each airport and write it into the node set Then generate edges between the starting point and the end point in the route database, create directed edges for each directly accessible route and write them into the edge set Among them, e ij Represents a slave node (Original Airport) to (destination airport); finally, the physical topology with the airport-route reachability relationship as the constraint After initialization is completed, the topology can be represented as a basic physical graph:

[0099] S22: Get node traffic according to flight plan and two-way flights The two are coupled and estimated to calculate the flow coupling intensity The flow coupling strength Assign physical edge e ij As the weight, define the first layer mixed graph as: Among them, the weight function Represents edge e ij The weight of is the flow coupling strength;

[0100] Calculating flow coupling strength The specific process is as follows: First, Perform maximum-minimum normalization within the most recent Δt to obtain a comparable scale with consistent direction. Secondly, extract the synchronous fluctuations of the traffic series of the two nodes within Δt and calculate the sliding co-movement degree to obtain the correlation component reflecting consistency. Thirdly, use the runway-taxiway occupancy ratio and flight segment distance of the two airports to attenuate the correlation component to reduce coupling misjudgment caused by excessive distance or large resource differences. Finally, superimpose the attenuated correlation component with the normalized flight volume according to the weight to obtain

[0101] S23: Listen to airspace regulation announcement stream, parse each announcement as a leg capacity coefficient If , then update the corresponding edge weight by scaling down by a factor of λ (i.e. edge weight is updated to ), if , then temporarily remove the edge;

[0102] S24: Perform sliding correlation calculation on node traffic sequence within the latest prediction window Δt ; for edges (i,j) with correlation coefficient below confidence threshold η, decrease their weight until removal, for high correlation edges with correlation coefficient greater than or equal to confidence threshold η, simultaneously increase their weight, generate an adaptively updated dynamic graph

[0103] S25: Encapsulate along with node sequence together to obtain

[0104] S3: Perform mixed graph convolution on the dynamic graph and node features outputted in step S2: one-hop propagation introduces time decay to get short-range embedding, multi-hop propagation superimposes distance decay to get long-range embedding, based on running level gating to fuse congestion-time distance weight, form spatial feature tensor;

[0105] S31: Construct node feature matrix denoted as where is the feature vector of the i-th airport; load the dynamic graph outputted in step S2 along with node feature matrix synchronously; normalize weight matrix w t and explicitly add unit diagonal to form self-loop normalized adjacency Define propagation depth set W = {W S , W L}, where W S is shallow propagation, responsible for perceiving the nearest neighbor topology of the airport, the depth value of shallow propagation is d s ; W L is deep propagation, responsible for perceiving the cross-regional topology, the depth value of deep propagation is d l ;

[0106] S32: Perform convolution propagation on one-hop neighborhood of depth d s : utilize ​The latest traffic vector of each airport node is aggregated with direct neighbors, and a time difference δ-based decreasing coefficient is introduced in this process to ensure that a minute-level surge (e.g., a sudden release) is fully retained; the aggregation result is added to the local feature residual and passed through W S Projection to the hidden space forms a short-range embedding The embedding highlights immediate resource competition features such as runway occupancy and taxi queue, and filters far-end weakly related noise.

[0107] S33: For a depth d l Multi-hop neighborhood performs convolution propagation: Calculate K power to obtain K-hop reachable matrix, and introduce exponential decay λ K for different hop distances to suppress excessive spread, where 0 < λ < 1, K ≤ 6; use the decay matrix to weight the aggregation of far-end airport traffic features, and pass them through W L Complete feature mapping to obtain long-range embedding

[0108] S34: For severe weather or capacity tightening scenarios, run the level with the intensity of the weather warning Input gate mapper to generate coefficients For and Perform element-wise scaling, write to the gated memory pool, and use an exponential smoothing factor μ for temporal smoothing.

[0109] S35: Based on the differentiated needs of airport operation scenarios, segment the short-range embedding and long-range embedding after gating, and obtain cross-scale vectors Calculate the normalized congestion index from the real-time taxi queue length and runway occupancy rate At the same time, read the length of the current rolling prediction window τ p ; select the weight group according to the congestion-time distance joint condition, the rules are as follows:

[0110]

[0111] Where θ c is the congestion threshold; τ0 is the prediction time distance dividing point; α S and β S represent the weight coefficients of the short-range (S) embedding, used for low congestion and short prediction window scenarios, α L and β L represent the weight coefficients of the long-range (L) embedding, used for high congestion or long prediction window scenarios (other cases), and satisfy α S > β S , α L < β L; α S ,α L Emphasize short-range impact to highlight immediate resource competition; S ,β L Strengthen long-range coupling to improve distal trend discrimination;

[0112] S36: Collect data in the most recent window along the time dimension Stacking to form a spatial feature tensor

[0113] S4: The spatial feature tensor is first subjected to trend-residual decomposition and fast Fourier transform (FTT). The random high frequencies are suppressed using an energy threshold. The main frequencies are retained using a flight peak template and then inversely transformed to obtain a denoised trend sequence and a denoised fluctuation sequence. The two are then concatenated into an event enhancement sequence.

[0114] S41: The spatial feature tensor generated in step S3 Split by node index i∈{1,…,N} to obtain independent time series At the same time, record the upstream standard one-minute sampling interval and the current sliding window length L w ,Will As the starting input for frequency domain enhancement;

[0115] S42: For each Perform additive decomposition: Based on the scheduling cycle P (i) Adaptive selection window w (i) , separating the long-term trend component by sliding mean and short-term fluctuation component Isolate the masking effect of scheduling rules on instantaneous traffic;

[0116] S43: respectively and Perform a fast Fourier transform and obtain:

[0117] Trend spectrum: long-term trend component Frequency domain representation, including amplitude spectrum and phase spectrum It mainly reflects the low-frequency characteristics dominated by regular factors such as scheduling plans;

[0118] Fluctuation spectrum: short-term fluctuation component Frequency domain representation, including amplitude spectrum and phase spectrum It mainly reflects the high-frequency characteristics dominated by sudden factors such as tower bursts and instantaneous resource competition;

[0119] S44: Amplitude spectrum of trend spectrum and the amplitude spectrum of the wave spectrum Calculate the energy density E(i) (f)=|A (i) (f)| 2 , and then generate the energy threshold according to the following formula:

[0120]

[0121] Among them, κ∈(0,1) is the learnable proportional coefficient;

[0122] like Then the corresponding frequency coefficient is set to zero to weaken the random disturbance; after processing, the high-energy frequency component is retained to obtain the filtered spectrum, and the absolute value of the spectrum is calculated to obtain the amplitude spectrum;

[0123] S45: Mapping Airport Peak Flight Schedules to Event Frequency Templates Y (i) , the amplitude spectrum peak value of the trend spectrum and the fluctuation spectrum after step S44 (in the amplitude spectrum A (i) (f) calculates the local maximum value, which is the amplitude spectrum peak) to match; when the frequency f∈Υ (i) and the amplitude is greater than the event threshold When the original value is retained, otherwise continue to suppress according to the rules of step S44;

[0124] S46: Perform inverse frequency domain conversion on the trend spectrum and fluctuation spectrum processed in steps S44 and S45 to reconstruct the noise reduction trend sequence and noise-reduced fluctuation series

[0125] S47: and Add point by point to generate a reinforcement sequence containing key event information Bundle Splice into feature stack;

[0126] S48: The energy threshold saved in the previous window (saved when step S4 was last executed) Compare with the current window frequency distribution and update based on the exponential smoothing coefficient ξ

[0127] S5: The event reinforcement sequence is processed by a large hourly convolution kernel and a small minute-level dilated convolution kernel respectively. After time alignment, the congestion spike is amplified based on trend-fluctuation gating and the fused feature sequence is output.

[0128] S51: Strengthen the event sequence generated in step S4 Arrange in ascending order by airport number and concatenate in the channel dimension to obtain a two-dimensional feature stack Where N is the number of airports, L w is the current sliding window length; then a double-branch convolution is performed for U tBind the hyperparameter templates of the two convolution branches (convolution kernel size, stride, padding mode), and set the sampling interval Δ, the scheduling period P sched and the sliding window length L w Write the module header as structured metadata;

[0129] S52: Use a one-dimensional large kernel of length to convolve U t Perform convolution along the time axis, and set the convolution layer stride to Use overlapping sampling, and use symmetric mirroring as the padding mode. After large kernel convolution, add batch normalization and Dropout with a fixed ratio of 0.1. The output matrix is Capture cross-hour scheduling synchronous fluctuations and slow-changing congestion trends;

[0130] S53: Use a one-dimensional small kernel of length k S to convolve U t Perform point-by-point convolution, and set the stride to s S = 1 to keep the output M t high in temporal resolution, where k S ≤ 5. Add dilated convolution to the small kernel with a dilation factor d = 2. So-called parallel means that steps S53 and S54 are independently processed on the same input data to generate two types of features (high-resolution details / coarse-grained trends), which are then aligned and fused;

[0131] S54: Since G t and M t have different dimensions, first pad G t at the tail of the time dimension to the sliding window length L w , and then apply a linear interpolation function to the M t to perform equidistant scaling by a factor to align the two tensors in the time dimension. The output after alignment is

[0132] S55: After completing the time dimension alignment, first normalize the nodes of the trend tensor to calculate the relative increase of each airport where (j, ·) represents the jth row (all columns), i.e., the trend vector of the jth airport. Then introduce a learnable interaction coefficient υ ∈ (0, 1) to construct a trend gating term, and perform element-wise multiplication with the fluctuation tensor to obtain the fusion matrix:

[0133]

[0134] where 1 is an all-1 matrix used to preserve the original fluctuation amplitude baseline; ⊙ represents element-wise multiplication by airport index and time index; Rt Amplify local fluctuations in peak congestion, compress random jitter in normal segments; υ controls the amplification strength, and is automatically tuned in the training phase through backpropagation;

[0135] S56: P t Channel normalization is adopted and a nonlinear activation function σ(·) is connected, outputting normalized tensor F t = σ(Norm(P t )) to suppress isolated noise frames and strengthen continuous high-flow areas, and F t is packaged as a fusion feature sequence package <F t ,t>.

[0136] S6: Perform recursive decoding on each node based on the fusion feature sequence, take the fusion feature sequence as the initial input of the decoder, input the last predicted value and real-time weather vector in the loop, output the future window ground flow, and get the predicted air flight flow through adjacent averaging;

[0137] S61: Recursively decode the fusion feature sequence generated in step S5 node by node to get the predicted sequence of each airport node; the specific process is as follows:

[0138] Assign an independent hidden vector to each airport node, whose initial value is obtained by linear mapping of the latest feature, and the number of cycles is configured according to the prediction length H, and the future window is unfolded in sequence; the strategy used in the unfolding process is: at the kth step, the future flow prediction value generated in the k-1th step is re-mapped by the embedding layer and decoded with the next frame feature in parallel, and the prediction value is generated in turn At the same time, the sampling probability is scheduled in time increasing, so that the model can reduce or get rid of the dependence on real labels in the training stage.

[0139] S62: Read wind direction, precipitation, thunderstorm indicators from aviation weather flow and compress them into a vector According to the airspace level where the node is located, assign a weight β (i) to the hidden state Perform weighted correction to get the prediction value sequence from t+1 to t+H (future H time steps) If wind direction mutation or heavy precipitation is observed, the gating weight will automatically increase the proportion of weather components, real-time weaken the short-term prediction deviation and retain the original time sequence memory;

[0140] S63: The corrected is respectively summarized as ground flow sequence G ground , and based on the topological edge table, the adjacent node flow is averaged to get the air flight flow sequence G airThe two are written into a unified message package together with meteorological correction residual, and pushed to an online feedback updating module, so as to realize instant sharing of the monitoring terminal, the flow management unit and the airline dispatching seat.

[0141] The following describes a flight flow prediction method based on a lightweight adaptive network in the application by using specific examples, taking Beijing Capital Airport (airport number i=1), Tianjin Binhai Airport (i=2) and Shijiazhuang Zhengding Airport (i=3) as research objects, and predicting flight flow in the next one hour (H=60 minutes). The specific steps are as follows:

[0142] I. Step S1: data acquisition and fusion

[0143] 1. Data source and sampling

[0144] The acquisition time range is t=08:00-08:30 (Beijing time), the sampling interval is 1 minute, and the data includes:

[0145] (1) Airport operation data (from ADS-B, A-SMGCS system):

[0146] Takeoff and landing count: Capital Airport (30 timestamps), Tianjin Airport Shijiazhuang Airport

[0147] Taxi queue length: Capital Airport Tianjin Airport Shijiazhuang Airport q 3,t =[2,2,1,...,2];

[0148] Ground congestion level: Capital Airport Tianjin Airport d 2,t =[1,1,2,...,2], Shijiazhuang Airport d 3,t =[1,1,1,...,1] (1-4 levels, 3 levels are moderate congestion).

[0149] (2) Dispatching data (from flight planning system and tower instruction):

[0150] Schedule change: 2 flights are delayed at 08:10 in Capital Airport (marked as 1 indicates a wide-body aircraft, 08:40 is the expected takeoff time, and 10 minutes is the taxiing time);

[0151] Release instruction: 3 flights are released at 08:20 in Tianjin Airport, and (0 indicates a narrow-body aircraft) is generated.

[0152] (3) Meteorological data (from weather radar system):

[0153] Capital Airport: Weather Radar Echo Intensity (Weak thunderstorm), Upper air wind field vector (Southeast wind), Cloud top height Run level label through level mapping g(·) (Low impact);

[0154] Tianjin Airport: (No precipitation), (Normal operation).

[0155] 2. Flow time series matrix generation;

[0156] Splice data by airport identifier and timestamp to form a multi-element time series matrix (Dimension: 3 airports x 30 minutes x 6 feature categories), example segment as follows:

[0157]

[0158] II. Step S2: Adaptive dynamic graph construction;

[0159] 1. Topology initialization;

[0160] (1) Node set: V0 = {V1, V2, V3} (corresponding to the three airports respectively);

[0161] (2) Directed edge set: ε0 = {e 12 ,e 13 ,e 23}(Direct routes: Capital to Tianjin, Capital to Shijiazhuang, Tianjin to Shijiazhuang).

[0162] 2. Edge weight calculation (flow coupling strength );

[0163] Take e 12 (Capital to Tianjin) as an example:

[0164] (1) Normalized flight volume: 20 flights in both directions in the last 30 minutes, normalized by max-min to get 0.6;

[0165] (2) Flow synchronization degree: The sliding correlation degree of the flow sequence of the two airports is 0.8;

[0166] (3) Attenuation factor: runway occupancy ratio (0.7 for Capital, 0.6 for Tianjin) x flight distance attenuation (120 km, attenuation coefficient 0.9), 0.7 x 0.6 x 0.9 = 0.378;

[0167] (4) Coupling strength: (weight distribution: flight volume 30%, synchronization degree 70%).

[0168] 3. Dynamic adjustment;

[0169] (1) Spatial control: Capital→Shijiazhuang route (e 13 ) is temporarily flow-controlled, with a segment capacity coefficient κ 13 = 0.7, and the edge weight updated to 0.5 x 0.7 = 0.35 (initial φ 13 = 0.5);

[0170] (2) Correlation screening: Tianjin→Shijiazhuang (e 23 ) flow correlation coefficient 0.4 < confidence threshold η = 0.5, edge weight decreased to 0 and the edge temporarily deleted.

[0171] 4. Dynamic graph results;

[0172] Final dynamic graph where the effective edges and weights are: e 12 = 0.39, e 13 = 0.35.

[0173] Three, step S3: mixed graph convolution;

[0174] 1. Node feature matrix and adjacency matrix;

[0175] (1) Node feature matrix (x i is a 30-dimensional feature vector);

[0176] (2) Normalized adjacency matrix (including self-loop, diagonal line is 1).

[0177] 2. Short-range embedding and long-range embedding;

[0178] (1) Short-range embedding (1-hop propagation): Capital Airport aggregates Tianjin Airport features, introducing a time decay coefficient (δ = 5 minutes), resulting in (highlighting the sudden increase in takeoff and landing from 08:10 to 08:15);

[0179] (2) Long-range embedding (2-hop propagation): Capital Airport indirectly aggregates Shijiazhuang features through Tianjin Airport, with a distance decay λ 2 = 0.8 2 = 0.64, resulting in

[0180] 3. Cross-scale fusion;

[0181] (1) Capital Airport congestion index ρ 1,t = 0.8 > θ c = 0.5 (high congestion), with β weight:

[0182] (2) Stack spatial feature tensors along time axis (3 x 30 x 128).

[0183] Four, Step S4: Frequency domain enhancement;

[0184] 1. Trend-fluctuation decomposition;

[0185] (1) Long-term trend component (Capital Airport): [12, 13, 14,..., 18] (reflecting the increasing trend of the morning peak);

[0186] (2) Short-term fluctuation component [0, 2, -1,..., 0] (contains a sudden noise at 08:10).

[0187] 2. Frequency domain filtering;

[0188] (1) Energy threshold Filter out random high frequencies with energy E(f) < 1.5;

[0189] (2) Flight peak template Y (1) = {1 / 24h -1 , 1 / 1h -1} (daily and hourly frequency), keep the main frequency f = 1 / 1h -1 (corresponding to the morning peak).

[0190] 3. Event enhancement sequence;

[0191] After inverse transformation and splicing Highlight the traffic peak feature of 08:30-09:00.

[0192] Five, Step S5: Dual-branch convolution and fusion;

[0193] 1. Dual-branch processing;

[0194] (1) Hourly large kernel convolution (k L = 60): Capture cross-hour scheduling fluctuations, output G t ;

[0195] (2) Minute-level dilated small kernel convolution (k S = 3, d = 2): Capture 08:10 tower sudden release, output M t .

[0196] 2. Time alignment and gated fusion;

[0197] (1) Aligned tensor:

[0198] (2) Trend gating: Relative rise R t of the Capital Airport= 0.8, fusion matrix Amplify congestion spikes.

[0199] 3. Fusion feature sequence; activated function output F t (Dimension: 3x60), packed as prediction input.

[0200] Six, step S6: recursive decoding and prediction;

[0201] 1. Lightweight recursive decoding

[0202] (1) Initial hidden vector Generated by F t mapping;

[0203] (2) Loop input: predicted value from last step and real-time weather vectors (such as Beijing Airport t+10 thunderstorm enhancement, w t+10 = 20 m / s);

[0204] (3) Output ground flow: (Beijing Airport), (Tianjin Airport).

[0205] 2. Air traffic flow calculation;

[0206] Based on dynamic graph adjacency relationship, the average of adjacent node flow is taken:

[0207] Beijing-Tianjin air segment air traffic flow = and Average = [16, 17,..., 20.5].

[0208]

[0209] The case data is based on the actual flight operation scenario simulation of North China in July 2025, and the index verification is taken by averaging 100 times of Monte Carlo simulation. In 100 times of Monte Carlo simulation, the random seed range is 1-100, the simulation process adopts random initial weight conforming to industry standard, and the stability of the result is verified to ensure the reliability of the result.

Claims

1. A lightweight adaptive network-based air traffic flow prediction method, characterized by The steps include: S1: Collects airport takeoff and landing counts, taxi queue lengths, congestion levels, schedule changes, release instructions, and weather radar data, and fuses them to generate a traffic time series matrix indexed by airport ID and timestamp; S2: A topology is established with airports as nodes and direct routes as directed edges. Edge weights are calculated based on the nearest window node traffic synchronization, normalized flight volume, and runway occupancy ratio. This is combined with airspace capacity notifications and dynamic adjustment of correlation thresholds to obtain an adaptive dynamic graph. S3: Perform hybrid graph convolution on the dynamic graph and node features output from step S2: one-hop propagation introduces time attenuation to obtain short-range embedding, multi-hop propagation superimposes distance attenuation to obtain long-range embedding, and based on the operation level gating and congestion-time weight fusion, a spatial feature tensor is formed; S4: The spatial feature tensor is first subjected to trend-residual decomposition and fast Fourier transform. The random high frequencies are suppressed using an energy threshold. The main frequencies are retained using a flight peak template and then inversely transformed to obtain a denoised trend sequence and a denoised fluctuation sequence. The two are then concatenated into an event enhancement sequence. S5: The event reinforcement sequence is processed by a large hourly convolution kernel and a small minute-level dilated convolution kernel respectively. After time alignment, the congestion spike is amplified based on trend-fluctuation gating and the fused feature sequence is output. S6: Perform recursive decoding on each node based on the fused feature sequence, use the fused feature sequence as the initial input of the decoder, input the previous prediction value and the real-time meteorological vector in the loop, output the ground traffic in the future window, and obtain the predicted air flight traffic through adjacent averaging.

2. The method for predicting air traffic volume based on a lightweight adaptive network according to claim 1, characterized in that: Step S1 includes the following sub-steps: S11: The take-off and landing counts output by the airport surface monitoring equipment in real time Taxi queue length and ground congestion levels The sampling interval is one minute; the traffic indicator set is written according to the airport number i-timestamp t Get the basic traffic time series of airport i; S12: Write the flight schedule change flow and tower release instruction flow pushed by the flight plan into the time series database in the order of arrival; extract the flight type, expected take-off and landing time, and runway-taxiway occupancy time for each tower release to generate a scheduling feature vector Then, according to the airport number i Map to Describe the immediate impact of the dispatching side on traffic; S13: Weather radar echo intensity High-altitude wind field vector Height to cloud top Perform level mapping g(·) to obtain the running level label Will Associated with the sampling records at the same time, the external meteorological impact dimension is injected into the traffic sequence; the mapping relationship is as follows: Where i is the airport number, t is the timestamp, and g(·) is constructed based on the industry operation threshold. Used to describe the operability level of airport i at time stamp t; S14: Basic traffic sequence Scheduling feature vector with runlevel labels Splice in chronological order to form a multivariate time series matrix That is, the traffic time series matrix indexed by airport ID and timestamp.

3. The method for predicting air traffic volume based on a lightweight adaptive network according to claim 2, characterized in that: Step S2 includes the following sub-steps: S21: First, map the airport list to nodes, assign a node to each airport and write it into the node set Then generate edges between the starting point and the end point in the route database, create directed edges for each directly accessible route and write them into the edge set Among them, e ij Represents a slave node (Original Airport) to (destination airport); finally, the physical topology with the airport-route reachability relationship as the constraint After initialization is completed, the topology can be represented as a basic physical graph: S22: Get node traffic according to flight plan and two-way flights The two are coupled and estimated to calculate the flow coupling intensity The flow coupling strength Assign physical edge e ij As the weight, define the first layer mixed graph as: Among them, the weight function Represents edge e ij The weight of is the flow coupling strength; S23: Monitor the airspace control announcement stream and parse each announcement into a flight segment capacity factor like Then the corresponding edge weight If the proportion is reduced, Then temporarily remove the edge; S24: Node traffic sequence within the most recent prediction window Δt Perform sliding correlation calculation; weight the edge (i, j) whose correlation coefficient is lower than the confidence threshold η Decrease until deleted, and simultaneously increase the weight of high-correlation edges with correlation coefficients greater than or equal to the confidence threshold η to generate an adaptively updated dynamic graph S25: Together with the node sequence Common packaging 4. The method for predicting air traffic volume based on a lightweight adaptive network according to claim 3, characterized in that: In step S22, the flow coupling strength is calculated The specific process is as follows: First, Perform maximum-minimum normalization within the most recent Δt to obtain comparable scales with consistent directions. Secondly, extract the synchronous fluctuations of the traffic series of the two nodes within Δt and calculate the sliding co-movement to obtain the relevant components reflecting consistency. Thirdly, the runway-taxiway occupancy ratio and flight distance of the two airports are used to attenuate the relevant components to reduce the coupling misjudgment caused by excessive distance or large resource differences. Finally, the attenuated relevant components are superimposed with the normalized flight volume according to the weights to obtain 5. The method for predicting air traffic volume based on a lightweight adaptive network according to claim 4, characterized in that: Step S3 includes the following sub-steps: S31: Construct node feature matrix Expressed as in is the characteristic vector of the i-th airport; the dynamic graph output in step S2 Together with the node feature matrix Synchronous loading; weight matrix w t Normalize and explicitly add the unit diagonal to form a normalized adjacency with self-loops Define the propagation depth set W = {W S ,W L }, where W S It is a shallow propagation layer, responsible for sensing the nearest neighbor topology of the airport. The depth of the shallow propagation layer is d s ;W L It is deep propagation, responsible for sensing cross-region topology, and the depth of deep propagation is d l ; S32: for depth d s Perform convolutional propagation in one-hop neighborhood of: using Aggregate the latest traffic vectors of direct neighbors for each airport node; add the aggregation result to the local feature residual and pass W S Projected into the latent space to form a short-range embedding S33: for depth d l Perform convolutional propagation over multi-hop neighborhoods: Compute K-th power to obtain the K-hop reachability matrix, and introduce exponential decay λ for different hop distances K To suppress excessive diffusion, where 0<λ<1, K≤6; use the attenuation matrix to weight the flow characteristics of remote airports and pass W L Complete feature mapping and obtain long-range embedding S34: For severe weather or capacity shortage scenarios, the operating level and weather warning intensity Input gate mapper generates coefficients right and Perform element-by-element scaling, write to the gated memory pool, and use exponential smoothing factor μ for time series smoothing; S35: Based on the differentiated needs of airport operation scenarios, the short-range after gate control is embedded and long-range embedding Perform segmented adaptive fusion to obtain cross-scale vectors Calculate the normalized congestion index based on real-time taxi queue length and runway occupancy At the same time, read the length of the rolling prediction window τ p ; According to the congestion-time distance joint condition, the selection weights are reorganized, and the rules are as follows: Among them, θ c is the congestion threshold; τ0 is the forecast time interval demarcation point; α S and β S represents the weight coefficient of short-range embedding, which is used for low congestion and short prediction window scenarios, α L and β L Represents the weight coefficient of long-range embedding, which is used in high congestion or long prediction window scenarios and satisfies α S >β S , α L <β L ; S36: Collect data in the most recent window along the time dimension Stacking to form a spatial feature tensor 6. The method for predicting air traffic volume based on a lightweight adaptive network according to claim 5, characterized in that: Step S4 includes the following sub-steps: S41: The spatial feature tensor generated in step S3 Split by node index i∈{1,…,N} to obtain independent time series At the same time, record the upstream standard one-minute sampling interval and the current sliding window length L w ,Will As the starting input for frequency domain enhancement; S42: For each Perform additive decomposition: Based on the scheduling cycle P (i) Adaptive selection window w (i) , separating the long-term trend component by sliding mean and short-term fluctuation component S43: respectively and Perform a fast Fourier transform and obtain: Trend spectrum: long-term trend component Frequency domain representation, including amplitude spectrum and phase spectrum Fluctuation spectrum: short-term fluctuation component Frequency domain representation, including amplitude spectrum and phase spectrum S44: Amplitude spectrum of trend spectrum and the amplitude spectrum of the fluctuation spectrum Calculate the energy density E (i) (f)=|A (i) (f)| 2 , and then generate the energy threshold according to the following formula: Among them, κ∈(0,1) is the learnable proportional coefficient; like Then the corresponding frequency coefficient is set to zero, and the high-energy frequency component is retained after processing to obtain the filtered spectrum, and the absolute value of the spectrum is calculated to obtain the amplitude spectrum; S45: Mapping Airport Peak Flight Schedules to Event Frequency Templates Y (i) , match the peak values ​​of the amplitude spectrum of the trend spectrum and the fluctuation spectrum after processing in step S44; when the frequency f∈Y (i) and the amplitude is greater than the event threshold When the original value is retained, otherwise continue to suppress according to the rules of step S44; S46: Perform inverse frequency domain conversion on the trend spectrum and fluctuation spectrum processed in steps S44 and S45 to reconstruct the noise reduction trend sequence and denoised fluctuation series S47: and Add point by point to generate a reinforcement sequence containing key event information Bundle Splice into feature stack; S48: Save the energy threshold in the previous window Compare with the current window frequency distribution and update based on the exponential smoothing coefficient ξ 7. The method for predicting air traffic volume based on a lightweight adaptive network according to claim 6, characterized in that: Step S5 includes the following sub-steps: S51: Strengthen the event sequence generated in step S4 Arrange in ascending order by airport number and concatenate in the channel dimension to obtain a two-dimensional feature stack Where N is the number of airports, L w is the current sliding window length; then a double-branch convolution is performed for U t Bind the hyperparameter templates of the two convolution branches and set the sampling interval Δ and the scheduling period P sched and sliding window length L w written to the module header as structured metadata; S52: Adopting length One-dimensional large kernel U t Convolution is performed along the time axis, and the stride of the convolution layer is set to Overlap sampling, symmetrical mirroring of filling mode, output matrix Capture cross-hourly scheduling synchronization fluctuations and slowly changing congestion trends; S53: Parallel use of length k S The one-dimensional small kernel U t Perform point-wise convolution with stride s S =1, so that the output M t Maintain high temporal resolution with the input, where k S ≤5, add dilated convolution to the small kernel with dilation factor d = 2; S54: First to G t Fill the end of time with zero padding to the current sliding window length L w , then to M t Apply a linear interpolation function By factor Perform isometric scaling to align the two tensors in the time dimension and output them after alignment S55: After completing the time dimension alignment, first align the trend tensor Normalize the nodes and calculate the relative increase of each airport where (j,·) represents the trend vector of the jth airport; then the learnable interaction coefficient υ∈(0,1) is introduced to construct the trend gating term and is combined with the fluctuation tensor Do element-wise multiplication to get the fused matrix: Among them, 1 is a matrix of all 1s, which is used to retain the original fluctuation amplitude baseline; ⊙ represents the element-by-element multiplication of the airport index and the time index; R t Amplify local fluctuations at peak congestion and compress random jitter in normal periods; υ controls the amplification strength and automatically tunes it during the training phase through backpropagation. S56: P t Channel normalization is used and nonlinear activation function σ(·) is connected to output the normalized tensor F t =σ(Norm(P t )), and F t Encapsulated as fusion feature sequence package <F t ,t>.

8. The method for predicting air traffic volume based on a lightweight adaptive network according to claim 7, characterized in that: In step S52, batch normalization and dropout are added after the large kernel convolution, and the ratio is fixed to 0.

1.

9. The method for predicting air traffic volume based on a lightweight adaptive network according to claim 8, characterized in that: Step S6 includes the following sub-steps: S61: recursively decode the fused feature sequence generated in step S5 node by node to obtain a prediction sequence for each airport node; S62: Read wind direction, precipitation, and thunderstorm indicators from aviation weather streams and compress them into vectors Assign weight β according to the airspace level of the node (i) Hidden state Perform weighted correction to obtain a sequence of predicted values ​​from t+1 to t+H S63: The corrected Summarize the ground traffic sequence G according to the airport nodes ground , then make an average mapping of the adjacent node traffic based on the topological edge table to obtain the air flight traffic sequence G air ; Write the two together with the meteorological correction residuals into a unified message package and push it to the online feedback update module to achieve instant sharing with the monitoring terminal, traffic management unit and airline dispatch desk.

10. The method for predicting air traffic volume based on a lightweight adaptive network according to claim 9, characterized in that: The specific process of step S61 is: assign an independent hidden vector to each airport node Its initial value is obtained by linear mapping from the features of the most recent frame. The number of loop steps is configured according to the predicted duration H, and the future windows are sequentially expanded. The strategy adopted in the expansion process is: in the kth step, the future traffic prediction value generated in the k-1th step is converted to After being remapped by the embedding layer, it is decoded and input in parallel with the next frame features to generate progressively At the same time, the sampling probability is scheduled incrementally over time, so that the model can reduce or get rid of its dependence on real labels during the training phase.

Images