A Functional Data Analysis Approach to Traffic Volume Forecasting

Traffic volume forecasts are used by many transportation analysis and management systems to better characterize and react to fluctuating traffic patterns. Most current forecasting methods do not take advantage of the underlying functional characteristics of the time series to make predictions. This paper presents a methodology that uses functional principal components analysis to create high-quality online traffic volume forecasts. The methodology is validated with a data set of 1755 days of 15 min aggregated traffic volume time series. Compared with 365 randomly selected days, the functional forecasts are found to outperform traditional seasonal autoregressive integrated moving average-based methods in both count deviation and root mean squared error. In addition, through the functional data analysis approach the full exploitation of the continuous nature of the data can be achieved.