Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Predicting the Skies: A Novel Model for Flight-Level Passenger Traffic Forecasting

Sina Ehsani, Elina Sergeeva, Wendy Murdy, Benjamin Fox

TL;DR

The paper addresses flight-level passenger traffic forecasting by introducing a multimodal deep learning framework that integrates historical traffic, fare closure dynamics, and seasonality. It employs a 3D tensor data representation and three encoders (temporal, spatial, and seasonal) coupled with a CNN decoder, along with masking and data augmentation strategies to handle incomplete data and diversify training. Empirically, the model achieves substantial improvements in predictive accuracy, outperforming ARIMA, SARIMA, CNN, and ConvLSTM baselines by roughly one-third on MSE, with the DeepShallow and shared-weights variants delivering the strongest gains. The approach offers significant potential for airline revenue management and pricing optimization, enabling better demand forecasting, scenario analysis for fare strategies, and robust adaptation to changing market conditions.

Abstract

Accurate prediction of flight-level passenger traffic is of paramount importance in airline operations, influencing key decisions from pricing to route optimization. This study introduces a novel, multimodal deep learning approach to the challenge of predicting flight-level passenger traffic, yielding substantial accuracy improvements compared to traditional models. Leveraging an extensive dataset from American Airlines, our model ingests historical traffic data, fare closure information, and seasonality attributes specific to each flight. Our proposed neural network integrates the strengths of Recurrent Neural Networks (RNN) and Convolutional Neural Networks (CNN), exploiting the temporal patterns and spatial relationships within the data to enhance prediction performance. Crucial to the success of our model is a comprehensive data processing strategy. We construct 3D tensors to represent data, apply careful masking strategies to mirror real-world dynamics, and employ data augmentation techniques to enrich the diversity of our training set. The efficacy of our approach is borne out in the results: our model demonstrates an approximate 33\% improvement in Mean Squared Error (MSE) compared to traditional benchmarks. This study, therefore, highlights the significant potential of deep learning techniques and meticulous data processing in advancing the field of flight traffic prediction.

Predicting the Skies: A Novel Model for Flight-Level Passenger Traffic Forecasting

TL;DR

The paper addresses flight-level passenger traffic forecasting by introducing a multimodal deep learning framework that integrates historical traffic, fare closure dynamics, and seasonality. It employs a 3D tensor data representation and three encoders (temporal, spatial, and seasonal) coupled with a CNN decoder, along with masking and data augmentation strategies to handle incomplete data and diversify training. Empirically, the model achieves substantial improvements in predictive accuracy, outperforming ARIMA, SARIMA, CNN, and ConvLSTM baselines by roughly one-third on MSE, with the DeepShallow and shared-weights variants delivering the strongest gains. The approach offers significant potential for airline revenue management and pricing optimization, enabling better demand forecasting, scenario analysis for fare strategies, and robust adaptation to changing market conditions.

Abstract

Accurate prediction of flight-level passenger traffic is of paramount importance in airline operations, influencing key decisions from pricing to route optimization. This study introduces a novel, multimodal deep learning approach to the challenge of predicting flight-level passenger traffic, yielding substantial accuracy improvements compared to traditional models. Leveraging an extensive dataset from American Airlines, our model ingests historical traffic data, fare closure information, and seasonality attributes specific to each flight. Our proposed neural network integrates the strengths of Recurrent Neural Networks (RNN) and Convolutional Neural Networks (CNN), exploiting the temporal patterns and spatial relationships within the data to enhance prediction performance. Crucial to the success of our model is a comprehensive data processing strategy. We construct 3D tensors to represent data, apply careful masking strategies to mirror real-world dynamics, and employ data augmentation techniques to enrich the diversity of our training set. The efficacy of our approach is borne out in the results: our model demonstrates an approximate 33\% improvement in Mean Squared Error (MSE) compared to traditional benchmarks. This study, therefore, highlights the significant potential of deep learning techniques and meticulous data processing in advancing the field of flight traffic prediction.
Paper Structure (26 sections, 5 equations, 6 figures, 1 table)

This paper contains 26 sections, 5 equations, 6 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dataset
  4. Seasonality Features
  5. Historical Traffic
  6. Fare Closure
  7. Baselines
  8. Approach
  9. Data Processing
  10. Creating 3D Tensors Capturing Spatial Relations
  11. Leveraging Historical Traffic Patterns for Enhanced Traffic Forecasting
  12. Masking Protocols for Dataset Segmentation
  13. Data Augmentation Techniques
  14. Feature Normalization in Data Preprocessing
  15. Architecture of the Model
  16. ...and 11 more sections

Figures (6)

  • Figure 1: Figure X: 3D Tensor Visualization of Fare Closure and Traffic Data for a Specific Flight on a Given Date. This matrix represents a three-dimensional dataset where the x-axis delineates different fare classes, the y-axis corresponds to time to departure (measured in intervals), and the z-axis differentiates between traffic types—local versus flow. The fare closure data is normalized between 0 and 1 for any market, indicating the availability of fare classes over time, with 1 being full closure. Traffic data is scaled by the number of passengers booked within each price range, with the minimum set at 0. The color gradient represents these quantities, with purple signifying the minimum and yellow indicating the maximum value. This visual encoding is designed to reveal the dynamic relationship between fare availability, passenger booking behavior, and traffic type as the departure time approaches.
  • Figure 2: Data Completeness Across Different Timeframes to Departure. This figure highlights the progression of data availability as it correlates with the time remaining until a flight's departure. The visual is segmented into four panels, each representing a different stage relative to the departure time: 'Departed', 'Close to Departure', 'Mid-way to Departure', and 'Far from Departure'. Within each panel, the x-axis categorizes fare classes, while the y-axis measures time to departure in intervals. The color coding is indicative of booking data, where purple signifies no bookings (value 0), yellow indicates a high volume of bookings, and white represents missing data due to the information not being available or the flight being in the future. The contrast between panels clearly shows that data for flights 'Far from Departure' have the most white spaces, reflecting a high degree of incompleteness, which progressively diminishes as flights near the departure date, culminating in the 'Departed' panel, where the dataset is fully detailed with booking patterns.
  • Figure 3: Schematic of the Model Architecture for Flight-Level Passenger Traffic Prediction. This figure delineates the configuration of the model's three encoders and the subsequent decoder layer. The Temporal Encoder processes time series traffic data, transforming it through a sequence of operations to capture temporal dynamics. The 3D Encoder manages the spatial aspects of fare closure data, employing volumetric filters to comprehend price-related variations. The 1D Encoder is specialized for discerning seasonality patterns, ensuring cyclical trends are captured. These processed inputs are then synthesized in the Decoder layer through a series of CNN layers that integrate the temporal, spatial, and seasonal features. The model's architecture is designed to preserve and utilize spatial relationships, culminating in a passenger traffic tensor that reflects a spatially informed forecast. The visualization provides insight into the input and output data shapes, illustrating the comprehensive flow from raw data to traffic prediction.
  • Figure 4: Validation Loss Variation with Window Size for the DeepShallow Network Model. This graph illustrates how different historical window sizes influence the model's validation loss, with the trend indicating the optimal range for balancing model complexity and predictive performance. The data points represent aggregated results across a spectrum of market conditions.
  • Figure 5: Performance Comparison Between ConvLSTM+Spatial and DeepShallow Models Over a 100-Day Period. This graph depicts the absolute differences in predictions from the actual observed traffic data, highlighting the prediction accuracy for each model. The data, averaging results from 12 diverse markets, illustrates the temporal prediction trend from April 1, 2023, to July 10, 2023. The DeepShallow network’s curve demonstrates its relative prediction performance against the ConvLSTM+Spatial model across the observed period.
  • ...and 1 more figures