Tracking time-evolving data streams and an application to short-term urban traffic flow forecasting
Francesco Masulli · 2016
Data streams have arisen as a relevant topic during the past decade. In this work we consider non-stationary data stream clustering using a possibilistic approach. The Graded Possibilistic Clustering model offers a way to evaluate “outlierness” through a natural measure, which is computed directly from the model. Both online and batch training scheme are considered, to provide two different trade-offs between stability and speed of response to change. The proposed approach is evaluated on a synthetic data set, for which the ground truth is available. Moreover, a real-time short-term urban traffic flow forecasting application is proposed, taking into consideration both spatial (road links) and temporal (lag or past traffic flow values) information. To this aim, we introduce a Layered Ensemble Model (LEM) which combines Artificial Neural Networks and Graded Possibilistic Clustering models obtaining an accurate forecast of the traffic flow rates with outlier detection. Experimentation has been carried out on two different data sets. The former was obtained from real UK motorway and the later was obtained from simulated traffic flow on a street network in Genoa (Italy). The proposed LEM model for short-term traffic forecasting provides promising results and given the its characteristics of outlier detection, accuracy, and robustness, it can be fruitful integrated in traffic flow management systems.