Abstract
Due to the rapid growth of the number of passengers over the few recent decades, air traffic forecasting has become a crucial tool for digital transportation systems, playing a fundamental role in the planning and development of traffic management and control systems. The main goal of forecasting in air transport is to predict traffic conditions in a network, on the basis of its past behavior, in order to improve safety and reduce airspace congestion. Nevertheless, air traffic time series often present an intricate behavior because of their irregular trends and strong seasonalities. In this paper, the methodology based on time series decomposition and artificial neural networks (ANNs) is thus reviewed and reconsidered within this framework of air traffic management. In this respect, a hybrid approach coupling feedforward neural networks with a nonlinear least squares-based regression curve fitting is developed for the multistep-ahead prediction. Empirical experiments are conducted in order to demonstrate the effectiveness of the proposed model on passenger traffic real datasets. The results show that, despite its simplicity, the base model is capable of generating accurate forecasts, with a performance comparable with that of powerful state-of-the-art forecasting models. In addition, there is evidence that trend pretreatment (wholly or partially) would rather degrade the forecasting accuracy.