Time-Series Forecasting for Peak Hour Traffic Accidents

Globally traffic accidents cause considerable damage, injuries, and deaths, making their analysis a critical research area. Recent advances have developed various predictions with different method streams yet it is unclear what are the similarities and differences of these streams and how they suit the accident analyses in reality. This study develops time-series accident rate predictions at urban intersections to examine the performance of three streams of the models including statistical model (Negative Binomial Model), machine learning techniques (SARIMA-X) and neural network algorithms (Multi Layer Perceptron, MLP) and further analyzes the suitability of the three streams. Pearson correlation and statistical analysis are first performed to identify the relationships among the spatial-temporal variables (e.g., number of lanes). It is found that the Negative Binomial Model performs superior for the average accuracy of the accident predictions. SARIMA-X performs better for study areas with similar magnitudes of historical traffic accidents over time while MLP is more suitable for accident datasets exhibiting varied magnitudes of accident events. The results provide references and practical insights into the potential of leveraging advanced algorithms and techniques to tackle the dynamics of traffic accidents and improve road safety.