Short-term traffic flow prediction based on faded memory Kalman Filter fusing data from connected vehicles and Bluetooth sensors

This paper proposes a Kalman Filter (KF) technique to predict the short-term flow at urban arterials based on the information of connected and Bluetooth equipped vehicles. Online traffic flow prediction using real-time data derived from different sensors is still an open research subject. To this end, a Kalman Filter model is developed to predict traffic flow based on two sources of real-time data, i.e. Connected Vehicles (CVs) and Bluetooth data. We also apply a Faded Memory Kalman Filter (FMKF) by considering more weights for new measurements to overcome the issue of inaccuracy in the prediction model and to predict the traffic flow with more resolution. At first, based on training data from Vissim traffic simulator, parameters of the KF's equations are calibrated using a machine learning based and big data processing. Performance of the conventional and faded memory KF models are then validated and compared using some test data pertaining to different rates of connected vehicles and Bluetooth-equipped vehicles (BVs). We use a pilot study of the city of Melbourne, Australia for numerical tests. The results indicate significant superiority of the FMKF over the KF in various traffic situations, as such the prediction error in some cases has reduced up to 60%. This paper contributes to the literature in three folds: (i) It uses a computationally efficient flow prediction algorithm based on synthesizing data from CVs and BVs (ii) It proposes to use an adaptive form of KF (i.e. FMKF) to compensate for the prediction error originating from modelling error. Hence, the model can perform well for a range of traffic conditions (iii) The proposed model works well even with low penetration rates (PRs) of the CVs or BVs (say 20%).