Raw data quality control is an important research content of data preprocessing in Intelligent Transportation Systems (ITS). Outlier detection is one of the core steps of data quality control to ensure the reliability of original data adopted by many ITS analysis technologies such as traffic state identification and estimation. In this paper, a time series pattern extraction and confidence interval estimation based outlier detection strategy is proposed for traffic flow outlier detection. The analyzed traffic flow data includes sectional traffic volume and average speed series with 5 minutes time interval that collected from urban arterials in Kunshan City, China. The fluctuation characteristics of traffic flow time series are analyzed to give an intuitive display of time series outliers. Then an STL model proposed by Chatfield is used for traffic flow time series decomposition and wave pattern extraction. Taking the extracted series pattern as modeling objects, an ARIMA model is introduced to estimate the dynamic confidence interval of the pattern series. Finally, the outliers of traffic flow time series can be identified and kicked off from the estimated dynamic confidence interval. Through a comparative experimental study, it was shown that the proposed strategy can accurately and reliably detect the outliers of traffic flow time series.

Traffic flow series, Outlier detection, Time series decomposition, Confidence interval estimation