A Novel Voronoi-Based Spatio-Temporal Graph Convolutional Network for Traffic Crash Prediction Considering Geographical Spatial Distributions

Accurately predicting the probability of crashes is crucial for preventing traffic crashes and mitigating their impacts. However, the imbalance in crash data, irregular road network structures, and heterogeneity in multi-source data pose significant challenges. To address these issues, this study introduces a spatio-temporal graph convolutional network traffic crash prediction model based on Voronoi diagrams that considers geographical spatial distribution. Initially, this study introduces a spatial partitioning method based on Voronoi diagrams, grounded on the geographic spatial distribution characteristics of traffic crashes. It constructs a novel graph structure with spatial units within Voronoi diagrams as nodes and the shared length of different road types between units as edges. This graph structure integrates the spatial distribution characteristics of crashes with the graph structure, substantially contributing to addressing the zero-inflation problem inherent in spatial units constructed on a grid basis. Subsequently, the study employs a GCN (Graph Convolutional Network) and Transformer encoder to build the VSTGCN (Voronoi-Based Spatio-Temporal Graph Convolutional Network) crash prediction model, evaluating its effectiveness using real data from New York City. Comparisons with eight baseline models demonstrate that VSTGCN outperforms them in all evaluation metrics. Moreover, the paper conducts model ablation studies from different perspectives, such as feature modules and graph structure composition, revealing that the chosen spatial, temporal, and spatio-temporal features significantly influence the model's predictive performance, with spatial features having the most substantial impact. Finally, the novel graph structure based on Voronoi diagrams proposed in this study shows a clear advantage in model effectiveness compared to traditional graph structures. This research can effectively handle complex crash data structures and accurately predict crash probabilities, providing a reliable basis for developing measures to prevent crashes and alleviate their impacts.