Prediction of the Occurrence of Delays on Takeoffs from Guarulhos-Sp Airport.
Flight delay, Prediction, Meta-analysis, Machine Learning, DEA.
A common problem at airports around the world is delays on commercial flights. The growing demand for air transport makes these delays more and more recurrent, adding costs and requiring constant adjustments in flight management. Defining the most effective method to predict the occurrence of these delays is a recurring theme in air traffic operation research. In these studies, the specificity of the analyzed area (whether an airline, an airport or the entire operation of a country), the complexity of the output sought (regression or classification prediction) or the size of the database used require more detailed analysis methods. robust, with machine learning being a common alternative to the use of classical statistical methods. Among the prediction models by machine learning, Artificial Neural Networks stand out due to the modular capacity of the method, allowing adaptation to the applied purpose. However, the divergence between the scenarios studied attributes different accuracy checks between the responses obtained in the studies carried out. Thus, this study aims to define the most appropriate scenarios and methods to more accurately estimate the occurrence of delays at airports. Therefore, bibliometric review and meta-analysis methods were used to define a baseline to assess the accuracy of the methods. Then, the researched studies were classified through Data Envelopment Analysis by metafrontier, defining the ideal scenarios for a better prediction response. The analyzes indicate that the Neural Network of the MultiLayer Perceptron (MLP) type, have better effectiveness in the predictive responses for analyzes in routes or airlines, regardless of the reason for the delay. Thus, a comparative case study was applied for the defined forecasting scenario, verifying the predictive capacity of the model for the São Paulo International Airport and a national airline operating there. The results indicate that the universal analysis of delays, through a predictive classification system regarding the delay time, proved to be more effective. Therefore, the specificity of the scenarios and the adequacy of the causes for the analyzed area are more impactful in defining the results of a flight delay forecast than the mass of data obtained.