Analysis and prediction of the queue length for non-recurring road incidents

The queue length of an incident refers to the number of upstream links (i.e., links in the opposite direction of traffic flow) experiencing congestion due to the incident. In our work, we fuse incident records with traffic speed data from the expressways of Singapore for computing the queue length. Moreover, we propose a hybrid classification-regression model to predict the queue length of the incidents in real-time. At first, the model acts as a binary classifier. If the queue length of an incident is predicted to be higher than a predetermined threshold value, then the congestion is assumed to be impactful. Therefore, in the second step, the model performs regression analysis to predict the queue length of these incidents for fine-tuning. We also analyze the performance of different classification and regression methods in our work. In the classification step, all the methods perform almost equally well, and we can achieve 80%–96% classification accuracy for different threshold values. However, in the regression step, Neural Network outperforms other methods. For the threshold value α = 250 m, the mean absolute percentage error is 65.78%, whereas for α = 1000 m the error value is 18.76%. Furthermore, we cluster the incidents to understand the underlying pattern of the external features and build separate regression models for each cluster of incidents.