Air cargo transport demand forecasting using ConvLSTM2D, an artificial neural network architecture approach

Abstract

The prediction of air traffic demand (passengers and cargo) in a regional/national air transport system is essential. Knowing the behavior of future demand helps, on the one hand, the design and execution of air transport public policies, which, for example, help to focus, guide and prioritize investment (public and private) for the expansion / modernization of airport infrastructures (or development of new airports), act on tariff policies, implement changes in regulatory policy, etc.; on the other hand, it helps airport managers to plan the airport. Therefore, in this paper, a short-term forecast (5 years) of the demand for air cargo transport was carried out, applied to a specific case study (Colombia), taking into account the most severe pandemic period (the year 2020). To perform the forecast, an approach based on Machine Learning/Deep Learning (ML/DL) method comprising a hybrid of convolutional and recurrent memory neural networks (that allow space-temporal non-linear analysis, such as multi-variable spaces and temporal multi-steps), is presented. The analysis developed here establishes the optimal length of the prediction period; on the other hand, the proposed methodology allows the identification of the most relevant socioeconomic features in the prediction of air cargo demand (domestic and international), i.e., interpreting the ML/DL results obtained through the variational analysis of different combinations of features. The results show that international air cargo demand is strongly dependent on Gross Domestic Product (GDP) and PCG (Per Capita GDP), while domestic air cargo demand is significantly dependent on PCG. Finally, the results show, for the case study country, very rapid recovery of air cargo demand at pre-pandemic rates (behavior already found in other recent studies and research).