考虑空间自相关的贝叶斯事故预测模型研究

Bayesian Crash Prediction Model Considering Spatial Autocorrelation

摘要:

依托美国佛罗里达州Hillsborough县的数据建立区域安全预测模型。将该县重新划分为200、500、700个交通安全分析小区；提取小区层面路网特征数据、出行数据和其他影响因素，考虑空间自相关特性，建立贝叶斯空间模型；分析小区层面影响因素的安全效应，评价不同分区规模对安全因素效应的影响。对比传统的泊松模型和泊松－对数正态模型，贝叶斯空间模型具有更高的数据拟合度；分区数目越多，空间因素在随机因素中的比重越高；同一种分区下，路网特征变量的安全效应具有鲁棒性；限速大于56公里的路段总长度是预测安全水平的主要指标。

Abstract:

This study proposed a regional safety prediction model based on the data from Hillsborough County, Florida, USA. By regionalizing the county into 200, 500 and 700 traffic safety analysis zones, we developed a Bayesian spatial model with consideration of spatial autocorrelation to relate crash rate to zonal factors including road network, trip generation etc. By the model results, we investigated the relationships between traffic safety and zone-level factors, as well as the effects of varied zoning schemes on the estimation of factor effects. Results show that compared with the traditional Poisson model and Poisson-lognormal model, the Bayesian spatial model has a better model-fitting; the greater the total zone number, the higher the spatial effects; the factor estimates are robust given a specific zoning scheme; the most significant factor affecting zonal safety is the total road length with speed limit over 35 mph. The present research contributes to the regional safety modeling in context of traffic safety planning.