Traffic State Prediction of Perturbed and Non-perturbed Traffic Scenarios

Traffic state prediction is a regression problem of predicting traffic states such as travel speed, volume, occupancy and density. However, we observe that existing traffic state prediction models are often trained and tested on non-perturbed traffic scenarios. It lacks the robustness needed for handling perturbed traffic scenarios. This work addresses this gap in the stated problem statement using simulated data based on perturbed traffic scenarios. Through this work, we established the need to assess the performance of the regression models on the affected road segments rather than on the entire road network. In doing so, the performance of the regression models on the perturbed traffic scenarios has become apparent. In addition, we have confirmed that it helps to train the regression model using a mixture of traffic data from both types of traffic scenarios. For our experiments, we built a micro-simulation of a medium-sized traffic network based on the morning peak traffic scenario. We evaluated multiple regression models for multiple prediction horizons on the studied traffic scenarios. Our results based on the affected road segments reveal that Spatial-Temporal Graph Convolutional Network (ST-GCN) has the best mean absolute percentage error (MAPE) for all prediction horizons. Hence, it is established that ST-GCN is a more robust regression model for predicting traffic states under perturbed and non-perturbed traffic scenarios.