QT-SimAM: Hybrid Queue-Theory Flight Delay Model

• The paper introduces a hybrid model that embeds queue-theoretic backlog surrogates into attention and LSTM modules for enhanced flight-delay prediction.
• The methodology employs an M/M/1 queue proxy to transform flight leg data into workload indicators, which modulate both a modified SimAM attention module and QMogrifier LSTM updates.
• Experimental results demonstrate state-of-the-art performance with up to 92.7% accuracy in US settings and effective transferability to EU networks without weather data.

Searching arXiv for the specified paper to ground the article in the source text. Queue-Theory SimAM (QT-SimAM) is a flight-delay prediction framework introduced in “Queue up for takeoff: a transferable deep learning framework for flight delay prediction” (Aghanya et al., 12 Jul 2025). It combines queue-theoretic backlog surrogates with a simple attention model and an LSTM-based temporal head in an end-to-end architecture intended to be both precise within a network and transferable across networks. The central premise is that flight-delay prediction should capture not only high-dimensional statistical structure in operational data but also the residual workload imposed by one flight leg on subsequent legs. In the reported formulation, this is achieved by approximating each leg with an $M/M/1$ queue, injecting the resulting workload proxy into a modified SimAM attention energy, and modulating recurrent updates through a “QMogrifier” mechanism (Aghanya et al., 12 Jul 2025).

1. Motivation and conceptual basis

The motivating problem is the operational and financial disruption caused by flight delays, together with the need for prediction models that are accurate and generalizable across different aviation networks. The framework is explicitly positioned against two incomplete alternatives. Pure deep-learning models, including CNN-LSTM systems with SimAM, are described as strong at pattern recognition but as neglecting the physical causality of cascading delays. Classical queueing theory, by contrast, is presented as formalizing how service demand and capacity interact to produce waiting times and backlogs, but as lacking the ability to learn rich, non-linear interactions from high-dimensional feature spaces (Aghanya et al., 12 Jul 2025).

QT-SimAM is designed to reconcile these two viewpoints. Its queue-theory component produces a lightweight proxy for the aircraft’s residual workload on each leg, and this proxy is used twice: first to bias attention toward legs under heavy congestion, and second to alter the LSTM gating process. The intended effect is to mitigate two stated failure modes: omission of operational backlog in purely data-driven networks, and insufficient representational capacity in naïve queueing models.

A useful way to characterize the framework is as a hybrid causal-statistical model for flight chains. This suggests that its novelty is not merely architectural composition, but the decision to make queue-derived state variables first-class inputs to both feature weighting and temporal recurrence.

2. Queue-theoretic residual workload formulation

The queue-theory component approximates each flight leg as a “customer” in an $M/M/1$ single-server queue. The formulation uses the following quantities: effective arrival rate $\lambda$, service rate $\mu$, traffic intensity $\rho=\lambda/\mu$, expected service time $E_S$, expected waiting time in queue $W_q$, and expected queue length $L_q$ (Aghanya et al., 12 Jul 2025).

For each leg $t$, two raw features are extracted from the input tensor $\mathbf{X}\in\mathbb{R}^{B\times S\times p}$: great-circle distance $M/M/1$0 and airborne time $M/M/1$1. These are converted into surrogates through

$M/M/1$2

Utilization is then capped as

$M/M/1$3

Under the $M/M/1$4 steady-state assumption, with small $M/M/1$5 for numerical stability, the expected waiting time and queue length are

$M/M/1$6

Because the model operates on three-leg chains, these quantities are normalized to $M/M/1$7 within each chain by min-max normalization across $M/M/1$8: $M/M/1$9

These normalized quantities, $\lambda$0 and $\lambda$1, are the operational backlog proxies used throughout the remainder of the system. The formulation is explicitly described as a surrogate rather than a full physical queueing model. This suggests that the queue-theoretic part is intended to provide structured inductive bias at low computational and data-integration cost, rather than a fully calibrated air-traffic simulator (Aghanya et al., 12 Jul 2025).

3. SimAM modification and recurrent modulation

The attention mechanism begins from SimAM, the Simple, parameter-free Attention Module. SimAM assigns a saliency score to each neuron in a convolutional feature map by modeling lateral inhibition among neurons within the same channel. For a feature map $\lambda$2, with channel means $\lambda$3 and variances $\lambda$4, the original energy for neuron $\lambda$5 is

$\lambda$6

with attention weight

$\lambda$7

where $\lambda$8 is the logistic sigmoid. The refined activation is $\lambda$9 (Aghanya et al., 12 Jul 2025).

QT-SimAM modifies this energy by adding chain-level congestion terms. Specifically, the attention energy is biased by the chain-average normalized waiting time $\mu$0 and queue length $\mu$1: $\mu$2 where $\mu$3 is the original variance term in SimAM and $\mu$4. The resulting attention weight becomes

$\mu$5

The stated purpose is to raise attention for channels corresponding to legs under heavy residual workload.

The same backlog proxies are also injected into the recurrent stage. After convolution and attention, the output $\mu$6 is reshaped back to a sequence $\mu$7 and fed to a QMogrifier LSTM. At each time step $\mu$8, the model computes a mixing mask

$\mu$9

uses it to modulate the input,

$\rho=\lambda/\mu$0

and then applies a standard LSTM update,

$\rho=\lambda/\mu$1

This dual insertion of queue-derived variables is the distinguishing technical feature of QT-SimAM. It is not only an attention variant; it is also a recurrent modulation scheme in which the hidden-state transition is directly conditioned on evolving workload surrogates (Aghanya et al., 12 Jul 2025).

4. End-to-end architecture and optimization

The full QT-SimAM pipeline comprises five principal modules: input embedding from raw features, including categorical embeddings and normalized numericals; a ResidualDelayLayer that computes $\rho=\lambda/\mu$2 via the $\rho=\lambda/\mu$3 surrogate; a convolutional stem consisting of three $\rho=\lambda/\mu$4D-convolution plus ReLU blocks; the QT-SimAM layer applied after each convolutional block; and a QMogrifier LSTM head, described as a two-layer LSTM that is optionally bidirectional. The output layer is a fully connected classifier over five classes (Aghanya et al., 12 Jul 2025).

The dataflow is stated as

$\rho=\lambda/\mu$5

From the final hidden state $\rho=\lambda/\mu$6, class logits are computed as

$\rho=\lambda/\mu$7

Training and inference follow standard mini-batch back-propagation with Adam. One training epoch is described procedurally as computing $\rho=\lambda/\mu$8 through the ResidualDelayLayer, obtaining convolutional features $\rho=\lambda/\mu$9, applying QT-SimAM with $E_S$0, reshaping to a sequence representation, processing the sequence with the QMogrifier LSTM, computing class logits, evaluating CrossEntropy loss, and backpropagating gradients. The reported optimizer settings are Adam with learning rate $E_S$1, batch size $E_S$2, and weight decay $E_S$3. Inference omits the loss computation and uses $E_S$4 across logits (Aghanya et al., 12 Jul 2025).

Because the system is described as end-to-end, the queue-derived variables are not a post hoc explanation layer; they are integrated into the forward path and therefore influence optimization dynamics directly.

5. Data construction, evaluation protocol, and reported performance

The experimental setup uses data from the US Bureau of Transportation Statistics for March, June, September, and December 2022, together with EUROCONTROL data from the same months. The two sources were harmonized on $E_S$5 shared features. Flight chains of length $E_S$6 were constructed per tail number per day, with filtering by minimum $E_S$7 minute turnaround and maximum $E_S$8 hour ground time. This yielded approximately $E_S$9 K chains for the US data and approximately $W_q$0 K chains for the EU data. Targets were five ordinal delay bins $W_q$1 defined at $W_q$2, $W_q$3, $W_q$4, and $W_q$5 minute thresholds (Aghanya et al., 12 Jul 2025).

The reported metrics are accuracy, precision, recall, and F1. Accuracy is given as $W_q$6. Per-class precision, recall, and F1 are defined in the standard way: $W_q$7

$W_q$8

The paper reports both in-region and cross-region transfer performance:

Setting	Variant	Reported metrics
In-Region (US, with weather feature)	QT-SimAM (uni-LSTM)	Accuracy $W_q$9, Precision $L_q$0, Recall $L_q$1, F1 $L_q$2
In-Region (US, with weather feature)	QT-SimAM (Bidirectional)	Accuracy $L_q$3, Precision $L_q$4, Recall $L_q$5, F1 $L_q$6
Transfer (US-trained on harmonized features $L_q$7 EU)	QT-SimAM	Accuracy $L_q$8, Precision $L_q$9, Recall $t$0, F1 $t$1
Transfer (US-trained on harmonized features $t$2 EU)	QT-SimAM (Bidirectional)	Accuracy $t$3, Precision $t$4, Recall $t$5, F1 $t$6

These results are reported as surpassing competing attention-based baselines, specifically CBAM-CNN and SimAM-CNN-LSTM, as well as traditional machine-learning methods such as SVM and decision trees. The abstract separately highlights accuracy $t$7 and F1 $t$8 for the bidirectional model on US BTS data, and accuracy $t$9 and F1 $\mathbf{X}\in\mathbb{R}^{B\times S\times p}$0 on EUROCONTROL for transferability assessment (Aghanya et al., 12 Jul 2025).

6. Interpretation, transferability, and stated limitations

The paper’s discussion emphasizes three main findings. First, embedding queue-theoretic backlog proxies into attention yields more focused feature weighting and improves delay prediction. Second, the QMogrifier LSTM’s direct sensitivity to evolving workload enhances temporal modeling of cascading delays. Third, the framework is described as achieving state-of-the-art accuracy, with $\mathbf{X}\in\mathbb{R}^{B\times S\times p}$1 in the US setting and $\mathbf{X}\in\mathbb{R}^{B\times S\times p}$2 in transfer to the EU, while requiring minimal domain-specific tuning (Aghanya et al., 12 Jul 2025).

Three advantages over baselines are stated explicitly. The model integrates physical causality, in the form of residual workload, with learned representations. Its attention augmentation is parameter-free, avoiding expensive gating networks. It also exhibits transferability between US and EU networks without weather data. Taken together, these points locate QT-SimAM within a broader class of hybrid predictive systems that use theory-derived state variables to regularize representation learning. This suggests that the model’s transfer behavior is tied not only to architectural choice but also to the portability of its queue surrogates across operational contexts.

The limitations are also explicit. The current queue proxies rely on distance and time surrogates, and the integration of real-time traffic and ATC data is proposed as a way to refine them. Class imbalance in severe-delay bins is identified as motivation for cost-sensitive losses or over- and under-sampling. Explainable AI techniques are proposed as a means to further dissect how queue surrogates and attention interact to produce predictions (Aghanya et al., 12 Jul 2025).

A recurring interpretive point is that QT-SimAM is neither a purely queue-theoretic model nor a conventional attention-augmented sequence model. Its defining claim is that classical queueing theory and modern attention mechanisms can co-exist within a single end-to-end framework for flight-delay prediction, with gains in both performance and interpretability relative to the baselines considered.