1. What is non-recurrent congestion?
Non-recurrent congestion refers to traffic congestion caused by events such as road accidents, extreme weather, and large-scale events. It is not a regular occurrence and can significantly impact traffic conditions. Research shows that accidents are the principal reason for 72% of non-recurrent congestion cases. A one-minute reduction in delay triggered by an accident can cost up to 1200 euros in very congested conditions. Trafc accident post-impact (TAPI) models have been introduced to address this issue and guide traffic management centers and road users during post-accident periods. The TAPI model aims to predict accident duration, which is divided into reporting time, dispatching time, response arrival time, and road clearance time. The proposed model also considers the recovery time, which is the period between the reported accident time and when traffic conditions return to normal. The sequential framework proposed in this study utilizes readily available variables such as real-time traffic and weather factors in the first stage, and updates as new accident details become available in the second stage. The study also investigates the potential of using eXtreme Gradient Boosting (XGBoost) as an artificial intelligence model to enhance the predictive capability of TAPI models. Shapley Additive Explanations (SHAP) is employed as a tool for knowledge generation and identifying factors that significantly impact the outputs.
read more
2. What statistical models are used for TAPI prediction?
Statistical models used for TAPI prediction include log-logistic, log-normal, and Weibull distributions. Regression models are the primary choice among statistical methods. Hazard-based duration models, such as parametric accelerated failure time (AFT) models, are also used to determine significant variables in different duration time phases. These models help in understanding the relationship between accident duration and other factors, aiding in predicting the period between crash occurrence and road clearance.
read more
3. How do tree models identify important variables?
Tree models, such as decision trees, random forests, and gradient boosting, are non-linear methods that can intrinsically identify and select the most important variables. These models analyze data and determine the most significant variables based on their impact on the target variable. In the context of TAPI studies, researchers have chosen tree-based approaches like decision trees, random forests, and gradient boosting to improve prediction accuracy. For instance, Ma et al. employed a gradient boosting decision tree model and found it superior to conventional models and other machine learning methods, including random forest, support vector machine, and backpropagation neural network. Similarly, Lin and Li demonstrated that random forest outperforms support vector machines in forecasting short-period congestions. However, previous studies have not extensively explored the application of XGBoost models in TAPI framework prediction. In this study, XGBoost models were utilized to compare their prediction performance with prior results, highlighting the potential of tree-based approaches in enhancing prediction accuracy.
read more
4. What are hybrid models in traffic accident prediction?
Hybrid models in traffic accident prediction combine different models to achieve more accurate predictions. Researchers have developed various hybrid models, such as combining the MP5 tree model with a hazard-based duration model, using Bayesian optimization algorithm to optimize random forest model parameters, and employing Bayesian averaging model to handle uncertainties. These models integrate data from multiple sources, like video detectors and weather stations, to extract relevant climatic and traffic variables. By leveraging these diverse models and data sources, hybrid models enhance the accuracy of traffic accident predictions, enabling better decision-making for traffic management and safety measures.
read more