SHORT-TERM TRAFFIC FLOW FORECASTING FOR URBAN ROADS USING SPACE-TIME ARIMA

With the development of technologies in telemetric, the interests and applications of Intelligent Transportation Systems (ITS) have been growing in the recent years. The applications in Advanced Traveler Information System (ATIS) and Advanced Traffic Management Aystem (ATMS) require the forecasting of traffic pattern to the near future. In contrast to the strategic models which predict over a period of month or year for long-term planning purposes, short-term models forecast traffic conditions within a day or a few hours that capture the dynamics of traffic, and are suitable for traffic management and information systems. A wide variety of short-term traffic forecasting problems have been investigated during past decades. An earlier comprehensive overview was given by Vlahogianni et al. (2004). Usage of input data would separate the forecasting techniques into univariate and multivariate approaches (Makridakis and Wheelwright, 1978). Univariate methods assumed that utilizing the historical time series data of a single variable is sufficient to detect the basic pattern for forecasting. Among the univariate models, autoregressive integrated moving average (ARIMA) model has been successfully applied in many areas and proved for its advantages over some other forecasting methods (Williams et al., 1998; Smith et al., 2002). Multivariate methods assumed there exists some relationships between two or more variables, and this pattern or trend can be extrapolated into the future (Stathopoulos and Karlaftis, 2003). This approach enables the modeling of the relationship of traffic measurements at different locations. The modelling of spatial-temporal domain of traffic data in urban area has been receiving attention in the recent years. Yang (2006) developed a spatial-temporal Kalman filter (STKF) forecasting model to compare with ARIMA and neural network (NN). An adaptive forecasting model selection strategy was proposed, which selects STKF with real-time data, but switch to use historical average method if real-time data was not available. Ghosh et al. (2009) proposed a structural time-series model methodology, which considers explicitly the trend, seasonal, cyclical, and calendar variations of traffic pattern, with the model flexibility that the immediate upstream junctions can be incorporated in the model as explanatory variables for the downstream predictions. In this paper, the space-time ARIMA (STARIMA) is used to investigate the spatial-temporal relationship and forecasting of urban traffic flow. Following the successful implements of ARIMA in the single location traffic prediction, a natural extension is to develop the multivariate version of the model with spatial-temporal domain. The space-time ARIMA model was firstly proposed by Pfeifer and Deutsch (1980), and is characterized by the autoregressive and moving average forms of several univariate time series lagged in both space and time. As compared to the ARIMA model, the calibrated model of STARIMA has the advantage with its small number of parameters, requiring fulfillment of rigorous statistical tests. A case study using the traffic data from 24 vehicle detectors in Taipei city, Taiwan are used to illustrate the performance of the model, and it is shown that STARIMA model are suitable for traffic flows forecasting in urban area.