A flight delay prediction method based on space-time convolution-attention graph neural network
Patent Information
- Application Number
- CN202611123936.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-28
- Publication Date
- 2026-10-09
AI Technical Summary
现有方法仅在通道维度施加Squeeze-and-Excitation注意力,忽略了不同机场、不同时段、不同天气条件下特征显著性的时空异质性——某一时刻某机场的关键延误特征可能随传播过程迅速转移至其他机场,单一的通道重标定无法适应这种动态变化
[0020](一)预测精度显著提升。 本发明通过自适应邻接矩阵与多类先验图融合的多图扩散卷积、因果TCN与多头自注意力分层混合的时序建模、以及通道-空间双重注意力特征校准三项正交设计,在真实大规模数据集上取得了领先的预测性能。在美国交通统计局(BTS)70个机场、7年数据集上,到达延误的均方根误差(RMSE)达到10.0184分钟,阈值准确率(±20min)达95.10%,出发延误的RMSE达6.6061分钟,较现有最优方法分别降低4.4%和6.6%;在中国携程50个机场、2年稀疏数据集上,到达与出发延误的±20min阈值准确率分别达93.37%和93.26%,均为全场最优。在±15min严格阈值下,美国到达延误准确率达88.50%,较次优方法领先3.44个百分点,表明本发明在严格精度要求下仍具有优异表现。
Smart Images

Figure CN122886878A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a method for predicting flight delays, and more particularly to a method based on a spatiotemporal convolutional attention graph neural network. Background Technology
[0002] Flight delay prediction is a core technical aspect of air traffic flow management, flight scheduling, and airline operation optimization. Existing prediction methods mainly fall into two technical routes: one focuses on a single airport or airport cluster, employing statistical regression models (such as ARIMA and queuing theory models) and temporal deep learning models (such as LSTM, BiLSTM-Attention, and CNN-LSTM hybrid architecture) to uncover the temporal evolution of delays. While this type of method has continuously refined its temporal modeling, it lacks sufficient characterization of the spatial propagation mechanism at the airport network level. Spatial dependencies rely on manual features or pre-clustering to simplify the topology, and it lacks unified end-to-end modeling capabilities.
[0003] Another approach, represented by Spatiotemporal Graph Neural Networks (STGNNs), abstracts the airport network into a weighted graph, captures spatial dependencies through graph convolutional networks, and models temporal dependencies by combining temporal modules (such as RNNs, TCNs, and self-attention). In recent years, representative methods such as Spatiotemporal Propagation Networks (STPNs) have introduced a space-time separable graph convolutional framework into flight delay prediction, achieving network-level delay prediction through multiple relational graphs (geographical proximity, flight path traffic) and self-attention mechanisms, achieving good performance. However, existing STGNN-like methods still have the following common limitations:
[0004] First, spatial graph structures are mostly statically constructed based on prior knowledge (such as geographical distance and fixed routes), making it difficult to capture implicit spatial dependencies driven by data, such as airline alliance hub associations, shared aircraft support, and seasonal temporary routes. Although some works introduce adaptive adjacency matrices, they are treated as independent modules or as replacements for prior graphs, failing to achieve organic integration with multiple prior graphs, and are prone to overfitting in sparse data scenarios.
[0005] Secondly, the temporal modeling methods are simplistic. Most methods employ only multi-head self-attention mechanisms, which are adept at modeling global long-range dependencies but are insufficient in responding to local transient patterns such as sudden delay spikes caused by severe weather or short-term transmissions within adjacent time periods. A few methods that use TCN or RNN lack the ability to perceive long-range patterns such as intraday cycles and intraweekly rhythms globally, and lack a hierarchical collaborative mechanism between local and global temporal patterns.
[0006] Third, the feature calibration methods are limited. Existing methods only apply Squeeze-and-Excitation attention to the channel dimension, ignoring the spatiotemporal heterogeneity of feature saliency across different airports, time periods, and weather conditions. At a certain moment, the key delay features of one airport may rapidly transfer to other airports during the propagation process, and single channel recalibration cannot adapt to such dynamic changes.
[0007] The aforementioned limitations collectively restrict the model's ability to represent complex delay propagation mechanisms and its prediction accuracy, especially in practical scenarios such as large-scale networks, multi-step prediction, and sparse data. Therefore, there is an urgent need for a flight delay prediction method that can comprehensively enhance its expressive power from three dimensions: space, time, and features. Summary of the Invention
[0008] The purpose of this invention is to provide a flight delay prediction method based on a spatiotemporal convolutional attention graph neural network.
[0009] To achieve the above objectives, the present invention is implemented according to the following technical solution:
[0010] This invention includes the following steps:
[0011] S1: Obtain a multi-relationship weighted graph of the airport network, wherein the multi-relationship weighted graph contains airport nodes and various spatial relationships between airports, wherein the various spatial relationships include at least geographical distance relationships, departure-arrival flow relationships and arrival-departure flow relationships;
[0012] S2: Obtain delay feature data and weather covariate data of each airport node within the historical time window, and then map the weather covariate data into a dense vector and concatenate it with the delay feature data to obtain the initial input feature tensor;
[0013] S3: Input the initial input feature tensor into a network consisting of L layers of spatiotemporal feature extraction blocks, where each spatiotemporal feature extraction block performs the following processing:
[0014] By applying a causal one-dimensional convolution to the input feature tensor in the time dimension, local short-time temporal features are extracted to obtain a local temporal feature tensor.
[0015] Based on learnable positional encoding, a temporal adjacency matrix is learned through a multi-head self-attention mechanism to capture global long-range temporal dependencies;
[0016] S4: The local temporal feature tensor, the temporal adjacency matrix, and the spatial adjacency matrices corresponding to the various spatial relationships are subjected to multi-graph diffusion convolution operation based on tensor pattern product to obtain a preliminary updated feature tensor. The spatial adjacency matrix includes three types of prior adjacency matrices and one type of adaptive adjacency matrix. The adaptive adjacency matrix is generated by nonlinear transformation and normalization of the inner product of two sets of learnable node embedding vectors and is used to characterize data-driven implicit spatial associations.
[0017] S5: Input the initially updated feature tensor into the channel attention module and the spatial attention module in sequence for feature calibration, wherein the channel attention module applies attention weights to the feature channels, and the spatial attention module applies attention weights to the key spatiotemporal locations on the airport-time joint plane, to obtain the calibrated feature tensor output by the current layer;
[0018] S6: Input the feature tensor output by the Lth spatiotemporal feature extraction block to the output projection layer. The output projection layer uses the position code of the future time as the query and the position code of the historical time as the key to generate a non-square matrix cross-temporal temporal adjacency matrix through a multi-head self-attention mechanism, and performs the multi-graph diffusion convolution operation again to map the temporal dimension of the feature tensor from the historical step size to the prediction step size, and outputs the arrival delay and departure delay prediction results for multiple future time steps.
[0019] The beneficial effects of this invention are:
[0020] (I) Significantly Improved Prediction Accuracy. This invention achieves leading prediction performance on real large-scale datasets through three orthogonal designs: multi-graph diffusion convolution fusion of adaptive adjacency matrix and multi-class prior graphs, temporal modeling hybrid of causal TCN and multi-head self-attention layered approach, and channel-space dual attention feature calibration. On a 7-year dataset of 70 airports from the U.S. Bureau of Transportation Statistics (BTS), the root mean square error (RMSE) of arrival delays reaches 10.0184 minutes, with a threshold accuracy (±20 min) of 95.10%, and the RMSE of departure delays reaches 6.6061 minutes, representing reductions of 4.4% and 6.6% respectively compared to the existing best methods. On a 2-year sparse dataset of 50 airports from Ctrip in China, the threshold accuracy of ±20 min for arrival and departure delays reaches 93.37% and 93.26% respectively, both being the best in the dataset. Under a strict threshold of ±15 minutes, the accuracy rate for arrival delays in the United States reached 88.50%, which is 3.44 percentage points higher than the second-best method, indicating that the present invention still performs well under strict accuracy requirements.
[0021] (II) Strong Robustness in Multi-Step Prediction. This invention utilizes a cross-temporal self-attention mechanism driven by future position encoding in the output layer to generate multi-step predictions in parallel at once, avoiding the error accumulation of autoregressive recursion. Simultaneously, the hierarchical collaboration between TCN and attention provides a stable temporal representation for long-term predictions. On the Chinese dataset, when the prediction step size increases from 2 steps to 12 steps, the accuracy at the delay threshold only decreases by 2.56 percentage points, outperforming existing methods. Under long-step prediction (≥10 steps), the threshold accuracy leads the second-best method by more than 1.1 percentage points, demonstrating good adaptability to the medium- and long-term prediction temporal domain.
[0022] (III) Strong cross-dataset generalization ability and robustness to data sparsity. This invention incorporates the adaptive adjacency matrix and prior graph into a unified Kronecker product kernel summation framework. The prior graph provides structural constraints, while the adaptive graph supplements implicit associations, effectively suppressing overfitting of pure data-driven graph mechanisms in small sample sizes. On two datasets of different sizes and varying record completeness, this invention maintains leading or equal performance in core performance metrics (threshold accuracy and RMSE), making it the only method capable of consistently achieving optimal performance in both dense and sparse scenarios.
[0023] (iv) The modules complement each other, and the ablation experiments provide sufficient verification. The ablation experiments show that removing the adaptive adjacency matrix increases the departure delay RMSE by 22.0%, removing the causal TCN branch decreases the departure delay threshold accuracy by 9.09%, and removing spatial attention increases the arrival delay RMSE by 12.3%. The three modules contribute irreplaceable gains from three orthogonal dimensions: spatial relation completion, local temporal enhancement, and spatiotemporal heterogeneity discrimination. The complete model achieves the best results in all six evaluation indicators, verifying the overall effectiveness and synergy of the technical solution of this invention.
[0024] (v) It is insensitive to input length and prediction step size parameters, and has good engineering applicability. The sensitivity experiment of input sequence length shows that the performance fluctuation of the present invention is the smallest under different historical window lengths (6-24 steps) (the arrival delay MAE range is only 0.43, which is much smaller than the 0.82 and 0.44 of the comparison method), and it remains the best under all lengths; the multi-threshold accuracy experiment shows that under the three thresholds of ±15, ±20, and ±25 minutes, the present invention consistently leads in the evaluation settings with discriminative power, indicating that its overall prediction distribution deviation is small and its tail control of large errors is strong, which is suitable for the diversified decision-making needs of actual flight operation scheduling. Attached Figure Description
[0025] Figure 1 The overall architecture of the STCA-GNN model
[0026] Figure 2The results evaluate the sensitivity of STCA-GNN to the length of the input sequence;
[0027] Figure 3 This is a graph showing the sensitivity of the prediction method to the number of prediction output steps. Detailed Implementation
[0028] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. The illustrative embodiments and descriptions herein are used to explain the present invention, but are not intended to limit the present invention.
[0029] Example:
[0030] This embodiment abstracts the airport cluster as a weighted multi-relationship graph. ,in Represents the set of airport nodes. Total number of airports; Represents the set of edges between airports; This represents the set of spatial relationship types between airports. The number of relations. Using a two-dimensional vector. Airport At the point of time Arrival and departure delays, using covariate vectors Airport At the point of time The weather type. Denote the delayed observations on graph G as a matrix. Covariates are denoted as matrices (C represents the number of weather categories). The time points... arrive Historical delay observations are denoted as three-dimensional tensors Historical covariates are denoted as .
[0031] The goal of the model is to learn the mapping function. ,Will Mapping historical observations and covariates to the future Delay prediction at each time step:
[0032] (1)
[0033] The main symbols used in this invention are summarized in Table 1.
[0034] Table 1 Explanation of Main Symbols
[0035]
[0036] The overall architecture of STCA-GNN is shown in Figure 1. Inspired by the spatiotemporal separation modeling concept of STSGCN, the core idea is to independently model temporal and spatial dependencies in each feature layer, and then achieve deep coupling between the two through tensor mode product. The specific process is as follows:
[0037] 1. Input preprocessing: First, classify the weather category. After embedding and encoding, mapping to dense vectors and then combining with delay features The initial input tensor is obtained by concatenation. .
[0038] 2. Two-branch timing modeling: Input causal TCN to extract short-term temporal local features to obtain Simultaneously, it incorporates a learnable positional encoding P and learns through a multi-head self-attention mechanism. Temporal adjacency matrix This allows us to capture long-term temporal features.
[0039] 3. The local temporal features and temporal adjacency matrix are then fed into a multi-graph diffusion convolution module based on the SoK kernel, along with four types of spatial adjacency matrices. and diffusion steps Tensor pattern product operations are performed to achieve the propagation of spatiotemporal information in multiple relationships and steps, resulting in a preliminary updated feature tensor.
[0040] 4. After residual connection and PReLU activation, the preliminary features are input into the CBAM feature calibration module, where channel attention and spatial attention are applied sequentially to enhance key features and suppress redundant features, ultimately yielding the output representation of this layer. .
[0041] The aforementioned causal TCN, multi-head self-attention, multi-graph diffusing convolution, and CBAM modules together constitute a complete ST-Block feature extraction unit. By stacking four layers of ST-Block, the model progressively completes the extraction and fusion of multi-level spatiotemporal features, ultimately outputting the future... The arrival delay and departure delay of each step are predicted in parallel.
[0042] For the layer input tensor First, a causal one-dimensional convolution is applied in the time dimension. Specifically, this is achieved by padding the left side of the input sequence. Zeros (ktcn is the kernel size) are used to ensure positional accuracy. The output at this location depends only on the position. This ensures accurate input and prevents future information leaks. Time-by-time calculation is defined as:
[0043] (2)
[0044] in For the convolution kernel at position Weight parameters on, For the input sequence time steps The feature vectors. To ensure training stability under deep stacking, LayerNorm is introduced after causal convolution with residual connections, and PReLU is used as the activation function:
[0045] (3)
[0046] LayerNorm performs normalization on the feature dimension of each sample, and residual connections ensure smooth gradient propagation. At the same time, it enables TCN to automatically degenerate into an identity mapping when it cannot provide effective local information, without impairing model performance.
[0047] Flight operations exhibit significant intraday periodicity and weekly operational rhythms, making it difficult for simple causal convolutions to capture these long-term temporal dependencies. To capture these temporal dependencies, STCA-GNN first introduces learnable sine-cosine positional encoding:
[0048] (4)
[0049] (5)
[0050] in For the index of the day's times, Determined by the time resolution of the data, For location encoding dimension, This is a learnable scale factor. For each training batch, we construct two sets of location encoding tensors, one for historical periods and one for future predicted periods:
[0051] (6)
[0052] (7)
[0053] in Used for the hidden layer self-attention module to learn the historical temporal adjacency matrix; The output layer self-attention module implements a temporal dimension mapping from history to the future. The multi-head self-attention module uses positional encoding. Learning end-to-end for input Group temporal adjacency matrix For the first The first layer The calculation process for each attention point is as follows:
[0054] (8)
[0055] (9)
[0056] (10)
[0057] in For each attention head, the query / key vector dimension, For the first Layer The learnable projection matrix of the head. The advantage of the multi-head mechanism is that it can simultaneously capture various heterogeneous temporal structures, including intraday cycles, weekly cycles, and flight propagation delays.
[0058] For the output layer (the first (layer), because it is necessary to change the time series dimension from Mapped to We employ a self-attention mechanism driven by future location encoding: switching the query source to future time location encoding. The key is still encoded by the historical position. The calculation process is as follows:
[0059] (11)
[0060] (12)
[0061] (13)
[0062] Will Take transpose Then, the tensor can be mapped from (B, N, h, C) to (B, N, p, C) in the temporal pattern product, thus natively completing the temporal dimension transformation from history to the future without the need for additional flattening, fully connected, or decoder structures.
[0063] This invention employs three types of prior adjacency matrices and one type of adaptive adjacency matrix to characterize the spatial dependency between airports: geographic distance map. Using a threshold Gaussian kernel to compute the geographical proximity between computer fields, a forward flight path map is obtained. Construct departure-arrival relationships and reverse route maps based on historical flight traffic. This is a transpose of the forward flight path diagram, depicting the backward propagation relationship between arrivals and departures. To capture data-driven implicit associations, two sets of learnable node embedding matrices are introduced. and ,in The launch characteristics of the coded airport as the starting point of delay propagation. The encoding airport is used to determine its reception characteristics as a delayed receiver. The inner product of the two sets of embeddings is filtered for negative correlation using ReLU, and then normalized by row-wise Softmax to obtain the adaptive adjacency matrix. After normalization Each row can be interpreted as the transition probability distribution starting from node i, which naturally matches the probability propagation characteristics of diffusion convolution.
[0064] In a multi-relationship graph, the spatial propagation of delays can be modeled as a random walk process on a weighted multi-relationship graph. Let... Given the delay probability distribution at each airport at a certain time t, the evolution of delays from time t to t+1 can be described as follows:
[0065] (14)
[0066] in For the first The normalized transition matrix corresponding to the relationship, Let be the random walk weight of the relation and satisfy ( The single-step propagation is generalized to a multi-step diffusing convolution with learnable weights, as follows:
[0067] (15)
[0068] in For the number of diffusion steps, Representing relations On Step random walk transition matrix, The weights are transformed into learnable features. Multi-step power series terms allow the updated representation of each node to simultaneously incorporate information from neighbors with different hop counts, expanding the receptive field without significantly increasing the number of parameters. This is achieved by combining multi-graph diffusion convolution with multiple temporal transformation matrices. Jointly applied to the local temporal feature tensor of causal TCN output Based on the tensor mode product, a compact spatiotemporal coupling form can be obtained:
[0069] (16)
[0070] in It is the feature tensor after causal TCN processing. For diffusion steps, For spatial relation numbers, Let be the number of temporal attention heads. ×_S, ×_T, and ×_F represent the tensor pattern products on the spatial pattern, temporal pattern, and feature pattern, respectively. Converting the tensor form to a matrix (flattening the node dimension N and the temporal dimension T to NT dimensions) can be equivalently represented as:
[0071] (17)
[0072] Where ⊗ represents the Kronecker product. It can be seen that... This is precisely the standard form of the Sum-of-Kronecker (SoK) kernel in Gaussian process theory. Each Each term corresponds to a specific combination of "spatial propagation mode × temporal dependence mode", and the sum of multiple terms can simultaneously carry multiple heterogeneous spatiotemporal coupling rules.
[0073] To address the issues of feature map redundancy and the varying importance of different channels and spatiotemporal locations for delay prediction, STCA-GNN introduces a Convolutional Block Attention (CBAM) module. Through a cascaded mechanism of channel attention and spatial attention, it performs dual calibration on features, enhancing key information and suppressing redundant noise.
[0074] Channel attention aims to assign learnable weights to each feature channel, highlighting the channels most relevant to the delay prediction task. The specific process involves: processing the input 4D feature tensor... By simultaneously applying global average pooling and global max pooling on the (N, T) plane, complementary channel description vectors are obtained. , The two are encoded using a two-layer bottleneck MLP with shared weights (first compressed to C / r dimension, r=16 as the reduction rate, then restored to C dimension), then added together, and finally activated by Sigmoid to obtain the channel importance weights:
[0075] (18)
[0076] in , , These are learnable weights. Then, the features are multiplied channel by channel in a broadcast manner into the original feature tensor to complete the feature calibration along the channel dimension. .
[0077] Spatial attention focuses on key locations in the (N, T) plane, enhancing the feature responses of these critical spatiotemporal locations. This is applied to the features after channel calibration. Simultaneous average pooling and max pooling are performed along the channel dimension to obtain a two-dimensional spatial description graph. , The two spatial description maps are concatenated along the channel dimension and fed into a 7×7 two-dimensional convolutional layer. The result is then activated by a sigmoid function to obtain a two-dimensional spatial attention map.
[0078] (19)
[0079] Will In the form of broadcasting and Position-by-position multiplication completes spatial dimension calibration:
[0080] (20)
[0081] Finally Rearrange the elements back into the (B, N, T, C) form to obtain the final output of the ST-Block. .
[0082] The output projection layer of STCA-GNN follows the overall framework of ST-Block, achieving temporal dimension mapping and prediction output generation through two differentiated designs. Firstly, the temporal adjacency matrix is encoded by historical positions. With future location coding The common structure is a non-square matrix First, its transpose participates in the temporal pattern product, achieving compression of the temporal dimension of the feature tensor from h to p; second, the feature pattern weights The feature channels are mapped to two dimensions, corresponding to the parallel prediction of arrival delay and departure delay, respectively.
[0083] The overall computation of the output layer can be uniformly represented as:
[0084] (twenty one)
[0085] Among them, the final output This is the parallel prediction result of arrival delay and departure delay in the next p steps.
[0086] During the training phase, the model uses the root mean square error (RMSE) with a mask as the loss function to shield the gradient from the interference of missing data.
[0087] (twenty two)
[0088] in, For a binary mask, only when Only the error corresponding to the valid observation point is included in the loss accumulation, thus avoiding the impact of invalid samples on the training process.
[0089] Experiments and Analysis:
[0090] This invention was tested on two real-world, large-scale flight delay datasets from the United States and China. The US delay dataset, sourced from the Bureau of Transportation Statistics (BTS) database, collected flight records for seven years, from January 1, 2015, to December 31, 2021, and selected 70 airports with high traffic volume from 360 airports to form an airport network. Accompanying weather data included eight weather categories: normal weather, severe cold, fog, hail, rain, snow, storm, and other precipitation. The Chinese delay dataset, sourced from the Ctrip platform, collected flight records for two years, from April 30, 2015, to May 1, 2017, with 50 airports forming an airport network. Accompanying weather data included seven categories: normal weather, rain, clouds, thunderstorms, fog, storm, and snow. Both datasets aggregated arrival and departure delays to their corresponding destination and origin airports within 30-minute time windows, retaining only flights departing between 6:00 AM and 12:00 AM daily. In contrast, the Chinese dataset has a shorter observation period and a higher missing rate, which can be used to test the robustness of the model in data sparse scenarios.
[0091] This invention employs three commonly used evaluation metrics: Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Threshold Accuracy (TA). The metrics are calculated independently for arrival delay and departure delay, with this invention focusing on the TA metric. The calculation formulas for these three evaluation metrics are defined as follows.
[0092]
[0093]
[0094]
[0095] in Represents the prediction set, Indicates the prediction result. Represents the true value. This is an indicator function; it takes the value 1 when the condition is true and 0 when the condition is false. This is the absolute error threshold.
[0096] Experimental setup:
[0097] For the US dataset, delays for the next 12 time steps (6 hours) are predicted using 12 historical time steps (6 hours). For the Chinese dataset, delays for the next 12 time steps (6 hours) are predicted using 36 historical time steps (18 hours). Inputs are Z-score normalized. The datasets are split chronologically into training (70%), validation (10%), and test (20%) sets.
[0098] STCA-GNN employs a 4-layer spatiotemporally separable multi-graph convolutional structure (3 ST-Block hidden layers + 1 output projection layer), with hidden layer feature dimensions of [128, 64, 32] respectively. Each hidden layer is followed by a CBAM module, with a channel attention reduction ratio r = 16. The number of self-attention heads is I = 4, and the order of the diffusing convolution is K = 2. Weather categories are encoded using 4-dimensional embedding, with node embedding dimension d_emb = 16 and causal convolution kernel size k_tcn = 5. The model uses the Adam optimizer with a learning rate of 0.001, and the default absolute error threshold for TA is τ = 20 min.
[0099] This invention compares STCA-GNN with the following six representative methods: (1) LSTM: Long Short-Term Memory Network based on encoder-decoder framework; (2) STSGCN: Spatiotemporal Separated Graph Convolutional Network; (3) AGCRN: Adaptive Graph Convolutional Recurrent Network; (4) STPN: Spatiotemporal Propagation Network; (5) ASTGCN: Attention Spatiotemporal Graph Convolutional Network; (6) MTGNN: Multivariate Time Series Graph Neural Network.
[0100] Main experimental results and analysis:
[0101] Based on the experimental data in Tables 2 and 3, we re-analyzed the delay prediction performance of each method on the US and Chinese datasets for the next 6 hours (12 steps). The superiority of the STCA-GNN method is reflected in the following three core dimensions.
[0102] Table 2 Comparison of prediction performance on the US delay dataset
[0103]
[0104] (1) Achieving overwhelming dominance on large-scale complete datasets: On the US dataset, STCA-GNN demonstrated absolute dominance. In the arrival delay task, STCA-GNN achieved the best values in all three metrics (MAE 7.2068, RMSE 10.0184, TA 95.10%). Compared to LSTM, which does not explicitly model spatial dependencies, its RMSE decreased significantly by 1.3291, and its accuracy (TA) increased by 2.56%; even compared to the best-performing baseline method ASTGCN, MAE and RMSE still decreased by 7.4% and 4.4%, respectively. In the departure delay task, STCA-GNN's advantage was further expanded, with the lowest MAE (4.4411) and RMSE (6.6061), significantly lower than ASTGCN by 15.5% and 6.6%, respectively. This indicates that STCA-GNN's joint spatial-temporal feature modeling mechanism can be fully activated in scenarios with sufficient data, achieving accurate capture of the spatiotemporal evolution of delays.
[0105] (2) On small and sparse datasets, STCA-GNN maintains an absolute lead in core performance metrics (TA and RMSE): On the Chinese dataset (smaller in size and with more missing data), STCA-GNN demonstrates strong robustness and practical value. In terms of threshold accuracy (TA), which is most relevant to practical operation, STCA-GNN achieves the best performance across the board in both arrival delay (93.37%) and departure delay (93.26%), and significantly outperforms ASTGCN (91.85%) in departure delay by 1.41%. At the same time, in terms of RMSE, which measures extreme errors, STCA-GNN also wins both arrival delay (11.3843) and departure delay (11.7431), outperforming ASTGCN in all aspects. Although STCA-GNN is slightly inferior to ASTGCN in MAE (less than 0.03), its dual lead in TA and RMSE fully demonstrates that it can effectively suppress large error prediction and ensure high accuracy, perfectly meeting the core requirements of actual flight operation scheduling.
[0106] (3) Overcoming the overfitting defects of adaptive graph mechanisms and achieving strong generalization ability across datasets: Experimental data reveals that adaptive graph mechanisms that rely solely on data-driven approaches (such as AGCRN and STPN) suffer from severe instability. For example, AGCRN degrades significantly on US arrival delay data (MAE as high as 11.6703, TA only 86.98%), while STPN also performs poorly on Chinese datasets. In contrast, STCA-GNN successfully avoids overfitting and instability issues under small sample sizes by fusing the adaptive adjacency matrix with the prior graph structure. It is the only method that can consistently maintain a top-three ranking or even win the championship on key indicators such as TA and RMSE on two datasets of different sizes and features. This ability to maintain high-level prediction performance regardless of data size and in both dense and sparse networks highlights the rationality of its model architecture design and its excellent generalization ability.
[0107] Table 3 Comparison of prediction performance on the China Delay Dataset
[0108]
[0109] To verify the independent contributions and complementarity of the three innovative designs, this invention designed three variants for ablation experiments on a Chinese dataset (with a fixed input step size of 36 and an output step size of 12): (1) removing the adaptive adjacency matrix; (2) removing the causal TCN branch; and (3) retaining only channel attention and removing spatial attention. The experimental results are shown in Table 4.
[0110] Table 4 Ablation experiment results (Chinese dataset, h = 36, p = 12)
[0111]
[0112] As shown in Table 4, removing any module leads to a decrease in prediction performance, verifying the effectiveness and complementarity of the three designs.
[0113] (1) Adaptive Adjacency Matrix. After removal, all metrics for arrival and departure delays significantly degraded, with departure delay RMSE increasing by 22.0% and TA decreasing by 4.59%, representing the largest decrease in departure delay RMSE among the three ablation methods. This indicates that implicit spatial relationships such as airline alliance hub associations and seasonal temporary routes are crucial for characterizing network-level delay propagation, confirming the core value of incorporating data-driven adaptive graphs into a multi-relationship diffusion convolutional framework.
[0114] (2) Causal TCN Branch. The degradation after removal exhibits significant asymmetry: arrival delay RMSE only increased by 3.1%, while departure delay MAE increased by 15.3% and TA decreased significantly by 9.09%, the largest decrease among the three ablation methods. This indicates that departure delays are more driven by local transient processes such as short-term congestion at the airport, which are difficult to capture using only self-attention, which excels at global modeling. The "local-global" hierarchical collaboration between causal TCN and self-attention is indispensable.
[0115] (3) Spatial attention. After removal, arrival delay RMSE increased by 12.3%, second only to the variant that removes the adaptive adjacency matrix, while departure delay TA decreased by 2.62%. This indicates that the significance of delays has obvious spatiotemporal heterogeneity across different airports and time periods. Feature calibration of a single channel dimension is insufficient to fully characterize the delays, and spatial attention needs to be introduced to recalibrate key spatiotemporal hotspots.
[0116] In summary, the degradation patterns of each module differ significantly: the adaptive adjacency matrix has the most systematic impact, the causal TCN branch is particularly crucial for short-term modeling of departure delays, and spatial attention mainly supplements the discrimination of spatiotemporal heterogeneity. The complete STCA-GNN is optimal in all six metrics. The three modules complement each other from three orthogonal dimensions—spatial relationships, local temporal sequences, and spatiotemporal features—to jointly support the model's ability to express complex delay propagation mechanisms.
[0117] Input sequence length sensitivity:
[0118] To evaluate the sensitivity of STCA-GNN to the length of the input sequence, this section uses h = 6, 12, 18, and 24 on the US dataset, respectively, to predict the next 12 steps. The two baselines with the best overall performance (STPN and ASTGCN) are selected for comparison, and the results are shown in Figure 2.
[0119] As shown in Figure 2, the prediction performance of STCA-GNN exhibits a trend of first increasing and then decreasing with input length. Taking arrival delay as an example, when h increases from 6 to 12, MAE decreases from 7.4091 to 7.2068, RMSE decreases from 10.1714 to 10.0184, and TA increases from 93.96% to 94.25%, indicating that a longer historical window provides the model with more sufficient periodic and propagation information. When h further increases to 18, MAE increases to 7.6414, RMSE increases to 10.4296, and TA falls back to 94.01%, indicating that excessively long inputs introduce redundant information with weak correlation to the current prediction period, negatively impacting prediction accuracy. When h = 24, the various indicators are basically the same as when h = 18, and the performance improvement tends to saturate. Departure delays exhibit a consistent pattern of change, with both MAE and RMSE reaching their optimal values at h = 12 (4.4411 and 6.6061, respectively), and then gradually degrading as h increases.
[0120] A horizontal comparison reveals that STCA-GNN significantly outperforms both baselines in both MAE and RMSE across all input lengths. For example, in arrival delay MAE, STCA-GNN's worst performance (7.6414) at h = 18 still surpasses ASTGCN's best performance (8.4693) at h = 18, a 9.8% advantage. In departure delay RMSE, STCA-GNN's lead is even more significant, being 12.6% lower than ASTGCN and 8.7% lower than STPN at h = 12. Regarding TA (Temperature Targeting), the differences in departure delay among the methods are relatively small (all remaining within the 97.9%–98.2% range), while the differences in arrival delay TA are more pronounced, with STCA-GNN leading STPN by approximately 1.6–2.3% across all input lengths.
[0121] It is worth noting that the response patterns of the two baselines to input length differ from those of STCA-GNN. STPN exhibits significant performance fluctuations in arrival delay (MAE fluctuates between 8.82 and 9.64), demonstrating a high sensitivity to input length. ASTGCN, on the other hand, achieves its optimal arrival delay MAE (8.4693) at h = 18, which is not entirely consistent with the optimal length of STCA-GNN (h = 12). In contrast, STCA-GNN not only maintains its optimal performance across all lengths but also exhibits smaller performance fluctuations (the arrival delay MAE range is only 0.43, compared to 0.82 for STPN and 0.44 for ASTGCN), indicating that the proposed model can effectively utilize long-range historical information without being significantly affected by redundant noise, demonstrating good robustness to input sequence length.
[0122] Prediction step sensitivity:
[0123] This section evaluates the performance of each method on the Chinese dataset with a fixed input step size of 36 (i.e., using historical data from the past 18 hours) for prediction steps p = 2, 4, 6, 8, 10, and 12, to analyze the decay patterns of each method in different prediction time domains. The report here is for the single-step prediction error at step p (not the average error over 12 steps as described in Section 4.4), therefore the values for the same method monotonically increase with increasing p. The changes in each metric with the prediction step size are shown in Figure 3.
[0124] As shown in Figure 3, with the increase of prediction step size, the MAE and RMSE of all methods gradually increase, while the TA gradually decreases, which is consistent with the general law of error accumulation in the time domain in multi-step prediction. However, there are significant differences in the decay rate and absolute level of different methods.
[0125] In terms of threshold accuracy (TA), STCA-GNN achieves the highest value across all step sizes for departure delays and the vast majority of step sizes for arrival delays. When p increases from 2 to 12, its arrival and departure delay TA decreases from 94.82% and 94.07% to 92.26% and 92.64%, respectively, a reduction of 2.56 and 1.43%, which is less than LSTM (3.07 / 2.27) and MTGNN (3.36 / 2.72). At longer prediction step sizes (p ≥ 10), STCA-GNN's TA advantage is even more pronounced: at p = 12, its arrival delay TA leads the second-best method STSGCN by 1.10%, and its departure delay TA leads the second-best method AGCRN by 0.69%. This long-term robustness stems primarily from two aspects: First, the hierarchical hybrid temporal module of causal TCN and multi-head self-attention simultaneously characterizes both short-term local propagation and long-term global cycles, providing a more stable temporal representation for long-term predictions; Second, the model generates future p-step predictions in parallel at once through self-attention driven by the future position encoding of the output layer, avoiding the error accumulation caused by the gradual recursion of autoregressive methods.
[0126] In terms of MAE and RMSE, STCA-GNN performs particularly well with short to medium prediction step sizes. Regarding arrival delays, STCA-GNN achieves the lowest MAE (9.2955 and 9.5352) and the lowest RMSE (11.4645 and 11.7029) among all methods at p = 6 and p = 8. Regarding departure delays, STCA-GNN achieves the lowest RMSE (11.0247, 11.5228, and 11.7228) across the p = 2 to p = 6 range. At p = 12, STCA-GNN's point error index is slightly worse than some baselines (e.g., arrival delay MAE is 9.8393, slightly higher than MTGNN's 9.5611), but the difference is small and TA remains the highest. Overall, STCA-GNN ranks among the top performers in both point error and threshold accuracy, with its lead in the TA dimension being the most stable.
[0127] It should be noted that while AGCRN's arrival delay TA curve is nearly horizontal (fluctuating between 91.05% and 92.01%), its RMSE is consistently the highest among all methods (maintaining above 13.3), and its MAE is also significantly higher than other methods. This indicates that its predictions approach a constant, and the so-called "stable" performance is actually a degenerate one, not true long-term robustness. Considering multiple indicators, STCA-GNN, while maintaining a point error level comparable to the optimal method, provides the highest threshold accuracy for each prediction step size. This characteristic is particularly important for medium- to long-term flight delay prediction, a scenario with greater operational decision-making value.
[0128] Performance under multi-threshold TA:
[0129] Table 5. Prediction performance of seven baseline methods under multiple thresholds.
[0130]
[0131] To evaluate the robustness of the model under different accuracy requirements, this invention examines the prediction accuracy of seven methods on two datasets in China and the United States at three threshold levels of ±15, ±20, and ±25 min. The results are shown in Table 5.
[0132] As the thresholds were relaxed, the performance evaluation (TA) of all methods steadily improved, but the sensitivity of different methods to threshold tightening varied significantly. In the three discriminative evaluation settings (arrivals in China, departures from China, and arrivals in the United States), STCA-GNN achieved the highest TA in most threshold ranges, with its advantage being most pronounced under the strictest ±15-minute threshold: the TA for arrival delays in China reached 83.18%, leading the second-best method by 1.36%; the TA for arrival delays in the United States reached 88.50%, leading the second-best method MTGNN (85.06%) by 3.44%, representing the largest lead across all settings. Under the ±20-minute threshold, STCA-GNN achieved 93.37% and 93.26% for arrival and departure delays in China, respectively, leading the second-best method by approximately 1%. Under the ±25-minute threshold, the differences between methods narrowed, with STCA-GNN still tied for the highest at 96.86% on the China dataset.
[0133] Notably, AGCRN and STPN exhibit significant accuracy shortcomings (TA of only 65%–67%) under a strict threshold of ±15 min, while their TA rapidly recovers to the 92%–95% range when the threshold is relaxed to ±25 min. This characteristic of "catching up when the threshold is relaxed, and exposing the shortcomings when the threshold is strict" further confirms the earlier judgment that their predictions approach a constant. In contrast, STCA-GNN maintains a robust lead under both strict and relaxed thresholds, indicating that its prediction distribution not only has a small overall bias but also stronger control over high-error samples at the tail.
[0134] It should be noted that the target accuracy (TA) for all methods in assessing US departure delays generally saturates within the 96%–98% range, with minimal differences between methods. This setting is not sensitive to the model's discriminative ability. Overall, in discriminative evaluation settings, STCA-GNN maintains a stable lead across all threshold levels, with its advantage becoming more significant under strict thresholds. This indicates that the proposed method possesses good practical value and robustness under varying accuracy requirements.
[0135] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.
Claims
1. A method for predicting flight delays based on a spatiotemporal convolutional-attention graph neural network, characterized in that, Includes the following steps: S1: Obtain a multi-relationship weighted graph of the airport network, wherein the multi-relationship weighted graph contains airport nodes and various spatial relationships between airports, wherein the various spatial relationships include at least geographical distance relationships, departure-arrival flow relationships and arrival-departure flow relationships; S2: Obtain delay feature data and weather covariate data of each airport node within the historical time window, and then map the weather covariate data into a dense vector and concatenate it with the delay feature data to obtain the initial input feature tensor; S3: Input the initial input feature tensor into a network consisting of L layers of spatiotemporal feature extraction blocks, where each spatiotemporal feature extraction block performs the following processing: By applying a causal one-dimensional convolution to the input feature tensor in the time dimension, local short-time temporal features are extracted to obtain a local temporal feature tensor. Based on learnable positional encoding, a temporal adjacency matrix is learned through a multi-head self-attention mechanism to capture global long-range temporal dependencies; S4: The local temporal feature tensor, the temporal adjacency matrix, and the spatial adjacency matrices corresponding to the various spatial relationships are subjected to multi-graph diffusion convolution operation based on tensor pattern product to obtain a preliminary updated feature tensor. The spatial adjacency matrix includes three types of prior adjacency matrices and one type of adaptive adjacency matrix. The adaptive adjacency matrix is generated by nonlinear transformation and normalization of the inner product of two sets of learnable node embedding vectors and is used to characterize data-driven implicit spatial associations. S5: Input the initially updated feature tensor into the channel attention module and the spatial attention module in sequence for feature calibration, wherein the channel attention module applies attention weights to the feature channels and the spatial attention module applies attention weights to the key spatiotemporal locations on the airport-time joint plane, to obtain the calibrated feature tensor output by the current layer; S6: Input the feature tensor output by the Lth spatiotemporal feature extraction block to the output projection layer. The output projection layer uses the position code of the future time as the query and the position code of the historical time as the key to generate a non-square matrix cross-temporal temporal adjacency matrix through a multi-head self-attention mechanism, and performs the multi-graph diffusion convolution operation again to map the temporal dimension of the feature tensor from the historical step size to the prediction step size, and outputs the arrival delay and departure delay prediction results for multiple future time steps.
2. The flight delay prediction method based on spatiotemporal convolutional-attention graph neural network according to claim 1, characterized in that, The three types of prior adjacency matrices include a geographic distance map constructed based on the geographic distance between airports, a forward route map constructed based on historical flight departure-arrival traffic, and a reverse route map constructed based on historical flight arrival-departure traffic. The geographic distance map uses a threshold Gaussian kernel function to compute the geographic proximity between airports. The forward route map constructs a transition matrix with departure airports as rows and arrival airports as columns. The reverse route map is the transpose of the forward route map.
3. The flight delay prediction method based on spatiotemporal convolutional-attention graph neural network according to claim 1, characterized in that, The adaptive adjacency matrix is generated as follows: each airport node in the airport network is mapped to a transmit embedding vector and a receive embedding vector, the transmit embedding vectors of all airports are used to form a source node embedding matrix, the receive embedding vectors of all airports are used to form a target node embedding matrix, the inner product between the transpose of the source node embedding matrix and the target node embedding matrix is calculated, negative associations are filtered by the ReLU activation function, and then Softmax normalization is performed row by row to obtain the adaptive adjacency matrix.
4. The flight delay prediction method based on spatiotemporal convolutional-attention graph neural network according to claim 1, characterized in that, The multi-graph diffusion convolution operation is represented as follows: in, For the local temporal feature tensor of the current layer, For the first spatial relationships Step-normalized random walk transition matrix, For the first The temporal adjacency matrix corresponding to each attention head Transform weights for learnable features. For diffusion steps, This represents the total number of spatial relationships. For the number of attention heads, , , These represent the tensor pattern products over the spatial pattern, temporal pattern, and feature pattern, respectively. This is the activation function.
5. The flight delay prediction method based on spatiotemporal convolutional-attention graph neural network according to claim 4, characterized in that, The multi-graph diffusion convolution operation, in matrix form, is equivalent to the sum of Kronecker kernels: in, Represents the Kronecker product, each Each item corresponds to a combination of a spatial propagation mode and a temporal dependency mode.
6. The flight delay prediction method based on spatiotemporal convolutional-attention graph neural network according to claim 1, characterized in that, The causal one-dimensional convolution is padded to the left of the input sequence. The zeros at the current position make the output at the current position depend only on the input at the current position and the positions before it. The causal one-dimensional convolution is followed by a normalized layer, a residual connection, and a PReLU activation function.
7. The flight delay prediction method based on spatiotemporal convolutional-attention graph neural network according to claim 1, characterized in that, The position encoding adopts sine-cosine encoding. Position encoding tensors are constructed for historical time periods and future prediction time periods respectively. The historical position encoding is used in the hidden layer self-attention module to learn the historical temporal adjacency matrix, and the future position encoding is used in the output projection layer to generate a cross-temporal temporal adjacency matrix from history to the future.
8. The flight delay prediction method based on spatiotemporal convolutional-attention graph neural network according to claim 1, characterized in that, The channel attention module simultaneously applies global average pooling and global max pooling to the input four-dimensional feature tensor on the airport-time joint plane. The pooling results are then encoded and summed in a two-layer bottleneck multilayer perceptron with shared weights. After Sigmoid activation, channel importance weights are obtained. These channel importance weights are then multiplied by the input feature tensor channel by channel in a broadcast manner. The spatial attention module simultaneously applies average pooling and max pooling to the channel-calibrated feature tensor along the channel dimension. The pooling results are concatenated along the channel dimension and fed into a two-dimensional convolutional layer. After Sigmoid activation, a spatial attention map is obtained. This spatial attention map is then multiplied by the channel-calibrated feature tensor position by position in a broadcast manner.
9. The flight delay prediction method based on spatiotemporal convolutional-attention graph neural network according to claim 1, characterized in that, In the output projection layer, the query of the cross-temporal temporal adjacency matrix comes from the future time position code, and the key comes from the historical time position code. The transpose of the matrix participates in the temporal pattern product, compressing the temporal dimension of the feature tensor from the historical step size to the prediction step size. The feature pattern weight maps the feature channel to two dimensions, corresponding to the parallel output of arrival delay and departure delay, respectively.
10. The flight delay prediction method based on spatiotemporal convolutional-attention graph neural network according to any one of claims 1-9, characterized in that, During the model training phase, the root mean square error with a mask is used as the loss function: in, For the number of airport nodes, To predict the step size, It is a binary mask, only when the airport At any moment The value is 1 if a valid observation exists, and 0 otherwise. This represents the actual delay value. To predict the delay value.
