Machine Learning Based Method for Modeling Traffic Events

Traffic events and incidents occurring on daily bases in the traffic network have been suspected of causing potentially severe delays. Also, such incidents may constitute a major disruption and may require a prompt response of traffic control authorities. Traffic incidents can be expected and predicted in the network and for a specific period of time (shortor long-term forecasting). This study focuses on how such predicted incidents can be simulated in order to predict their impacts on the transport network performance. A method has been suggested to convert real traffic events into simulation incidents. It consists of two main components; machine learning based model for event classification and event-to-incident mapping. Many machine learning algorithms have been evaluated as tools for classifying real traffic events and numerous real events of various features have been modeled. The proposed method provides the most appropriate simulation incident which represents the input real traffic event. 45% of the event classification model were identical to the observation data. Generally, the proposed method provided good prediction results when compared with the real data collected. Additionally, the average Mean Absolute Percentage Error (MAPE) of tested events was 17.8% and therefore this method can be considered as a good event-to-incident predictor.