CBL - Campus del Baix Llobregat

Projecte llegit

Títol: Comparative analysis of traditional and AI-based methods for air passenger demand forecasting

Estudiants que han llegit aquest projecte:

Director/a: KULJANIN, JOVANA

Departament: FIS

Títol: Comparative analysis of traditional and AI-based methods for air passenger demand forecasting

Data inici oferta: 07-05-2024     Data finalització oferta: 07-05-2024


Estudis d'assignació del projecte:
    GR ENG SIST AEROESP
Tipus: Individual
 
Lloc de realització: ERASMUS
 
Paraules clau:
Air passenger demand, forecasting, SARIMA, econometric modeling, OLS, 2SLS, neural networks, LSTM, SHAP, time series, artificial intelligence, model interpretability.
 
Descripció del contingut i pla d'activitats:
The air travel demand is a principal step which has a strong impact on decision making process in both airline and airport planning. The accurate forecast on the number of passengers will be an important input to a variety of decision at different planning horizons. For example, air travel demand forecasts provide valuable insights that inform tactical planning decisions across various aspects of airline operations, helping airlines optimize resources, maximize revenue, and provide high-quality service to passengers. Additionally, by understanding future demand trends, airlines can make strategic decisions about fleet size, composition, and deployment to meet future passenger needs.

Air travel demand forecasting requires a combination of quantitative and qualitative methods, incorporating historical data, economic indicators, market research, and industry expertise to predict future passenger demand accurately. In addition to the mentioned techniques, the air travel demand can be accurately predicted with the application of machine learning models. In essence, machine learning algorithms can analyze large volumes of historical data and identify complex patterns and relationships that may not be apparent with traditional statistical methods.
The project will focus on applying different types of methods in forecasting air travel demand.

The specific objective of the project are as follows:
1. Perform a thorough literature review on the application of air travel demand forecasting methods;
2. Classify the papers according to different planning horzions, level of study (e.g., route level, city pair level, country level), regions of the world;
3. Identify the databases available for performing this type of analysis;
4. Collect and process the data in order to perform the selected techniques/method;
5. Compare and discuss the performance of different techniques based on the obtained results; Acknowledge the advantages and limitations for each of them;

The TFG has been carried out within an Erasmus+ mobility program and, specifically, in the Beihang University (China ).
 
Overview (resum en anglès):
This thesis conducts a comparative analysis of traditional and artificial intelligence based forecasting methods to forecast passenger demand on two important U.S. domestic routes: Los Angeles to San Francisco (LAX-SFO) and Los Angeles to New York (LAX-JFK). It compares the accuracy and the explanatory power of three major forecasting methodologies: time series (SARIMA), econometric modeling (OLS and 2SLS), and artificial intelligence (feedforward neural networks and LSTM networks).


The results indicate that the LSTM method emerges as the best model on the LAX-SFO route, highlighting that the LSTM networks can capture nonlinear relationships and temporal patterns. SARIMA, on the other hand, outshines the other methods in forecasting the LAX-JFK time series, where historical trends are more dominant than the external economic factors used as determinants, thus making AI and econometric modeling less potent.


The thesis also interprets the outputs of the AI models through SHAP, an explainable artificial intelligence method, and compares these explanations with the coefficients derived from OLS to determine the important variables. Focusing on the trade-off between interpretability and accuracy in favor of models that harness the predictive power of AI while retaining the transparency of traditional methods.

© CBLTIC Campus del Baix Llobregat - UPC