Short-Term Traffic Prediction Under Both Typical And Atypical Traffic Conditions Using A Pattern Transition Model

One of the most challenging goals of the modern Intelligent Transportation Systems comprises the accurate and real-time short-term traffic prediction. The achievement of this goal becomes even more critical when the presence of atypical traffic conditions is concerned. In this paper, we propose a novel hybrid technique for short-term traffic prediction under both typical and atypical conditions. An Automatic Incident Detection (AID) algorithm, based on Support Vector Machines (SVM), is utilized to check for the presence of an atypical event (e.g. traffic accident). If such an event occurs, the k-Nearest Neighbors (k-NN) non-parametric regression model is used for traffic prediction. Otherwise, the Autoregressive Integrated Moving Average (ARIMA) parametric model is activated for the same purpose. In order to evaluate the performance of the proposed model, we use open real world traffic data from Caltrans Performance Measurement System (PeMS). We compare the proposed model with the unitary k-NN and ARIMA models, which represent the most commonly used non-parametric and parametric traffic prediction models. Preliminary results show that the proposed model achieves larger accuracy under both typical and atypical traffic conditions.