A Self-Organized Neuro-Fuzzy System for Air Cargo and Airline Passenger Dynamics Modeling and Forecasting

Introduction

Knowing air cargo and airline passenger demands are important in planning and operation of airline industry. The arrangement of an air route, the purchasing of planes, the assignment of flight crew, the dispatch of flights, and the operation of revenue management all require reliable information to support right decisions. Pitfield et al. (2010) presented the airline strategies for aircraft size and airline frequency with changing demand and competition by a simultaneous-equations approach. Pai (2010) assessed the determinants of aircraft size and frequency of flights on airline routes by market demographics, airport characteristics, airline characteristics and route characteristics. Wei and Hansen (2006) proposed an aggregate demand model for air passenger traffic by airline service variables such as service frequency, aircraft size, ticket price, flight distance, and number of spokes in the network.

Many recent works have recognized the necessity of accurate forecasts for air cargo or airline passenger demand. State forecasting of a dynamical system requires the current and previous condition to forecast the system’s future states. Forecasting entails gathering historical data, analyzing patterns, and acquiring results that facilitate the prediction of future events. It is believed that past information can be modeled into a system that explains the current behavior. Classical approaches are mainly based on stochastic models, using time series analysis techniques such as autoregressive moving average (ARMA) and multiple regression models; however, the accuracy of state forecasting may be vulnerable to qualitative factors from macro-economical and political effects.

Many have focused on applying recurrent or feedforward neural network (NN) to state forecasting. NN is effective in realizing the input-output mapping and is useful as a state forecaster, but its performance is constrained by large amount of training required and by the “black box” nature opaque to the dataset. Rob and Norman (1999) developed a neural network model to forecast Japanese demand traveling to Hong Kong. Lim and McAleer (2002) proposed an integration analysis of annual tourism demand by Malaysia for Australia. Vincent (2003) also compared three different approaches: exponential smoothing, autoregressive integrated moving average and artificial neural networks to tourist arrival forecasting. Recently, there have been significant developments in neuro-fuzzy, fuzzy-neural and neural-fuzzy model integrating the advantages of neural network and fuzzy system (Chen, 2008, 2009, 2011; Chen & Lin, 2008; Hsiao et al., 2008; Chen, Wang, & Tsai, 2009; Chen, Wang, & Wu, 2009; Chen, Wu, & Wang, 2009; Chen & Wang, 2010). The neural network has the learning ability of mimicking biological neural network to deal with complex problems, while the fuzzy logic is to tune the membership functions of knowledge-based system operating in linguistic, rule-based structure. It is known that a neuro-fuzzy model is effective to system modeling and control in engineering. Yang et al. (2007) and Chen et al. (2009a, 2009b) proposed a self-organized neuro-fuzzy model for system identification, in which the five-layer network adaptively adjusts the membership functions and dynamically optimizes the fuzzy rules. Recent interests have also been in applying neuro-fuzzy model to system dynamics modeling and forecasting. Paz and Peeta (2009) proposed a fuzzy control to estimate driver response behavior in transportation system. Ang and Quek (2006) suggested a neuro-fuzzy approach in stock trading. Keles et al. (2008) presented a neuro-fuzzy model for predicting domestic debt. Lin et al. (2008) also proposed an early warning system for currency crisis by neuro-fuzzy model. Atsalakis and Valavanis (2009) applied a neuro-fuzzy inference system to forecast stock market short-term trends. All of the above works were limited to training and validation of their models, yet there is no forecasting at all. Tobias et al. (2007) proposed a gravity model including general economic activity and geographical characteristics for airline passenger volume estimation, but it is limited to city-pair model. In this paper, a self-organized neuro-fuzzy model is applied to system modeling and forecasting of air cargo and airline passenger in Taoyuan and Kaohsiung International Airport.