最佳時空分段下之高速公路事故頻次模式

Abstract

進行事故分析與預測之前，在資料整合階段必需將每一筆事故資料依據其所屬路段做空間及時間上之分割，以避免平滑效果產生導致許多重要變異資訊損失，由於空間中相鄰路段之道路幾何與交通環境變化不大，易導致路段間存在自我相關之問題，傳統路段分割方式如：固定長度分段、兩兩交流道分段等亦未考量事故因素致使同質性路段無法被歸類成同一路段，而早期時間上並無特別進行切割，傳統上皆使用固定時間長度一個月或一年為單位，此作法亦無考慮到時間上的自我相關之問題，此作法將嚴重影響模式參數推估之正確性。 基此，本研究利用基因演算法以樣本間同質性最高為目標，提出最佳時空分割方法：樣本間事故件數變異最大，進行時空分割，並與另外三種常用的事故樣本時空分割方式進行比較，再根據求得之分割樣本結果利用卜瓦松迴歸、負二項迴歸、零膨脹卜瓦松迴歸、零膨脹負二項迴歸建構事故頻次模式，並將時空最佳化分割下之事故頻次模式與傳統時空分割下之模式進行績效指標比較。結果顯示，以MAPE值而言，以本研究所提出之最佳時空切割之負二項迴歸模式MAPE為最低，擁有最佳模式預測能力。若以空間自相關值觀之，亦以最佳時空之樣本分割方式擁有最低之自相關值，最後以不同年份我國國道一號高速公路事故相關資料進行驗證，以證明事故頻次模式可適用於其它年份。

Prior to conducting traffic accident analysis and prediction, it is an essential step to spatially and temporally divide the study roadway system into several segments. Most of previous studies tend to subjectively divide the study roadway without considering the spatial and temporal correlation and similarity among segments, making the observations fail to distribute independently and leading to an estimation bias. Traditional roadway segmentation methods, such as fixed length method and interchange segmentation method are mostly subjectively and tend to result into a high correlation problem among neighboring segments. Similar problem can be found in temporal segmentation. Based on this, this study proposes a genetic algorithms-based spatiotemporal segmentation model to optimally divide roadway into segments so as to maximize variation/within variation among segmentation groups. To investigate the performance and applicability of the proposed model, the proposed segmentation is compared with another three subjective segmentation methods. Based on the results of different segmentation methods, the crash frequency are respectively modeled by Poisson regression, Negative Binomial regression, Zero-Inflated Poisson regression and Zero-Inflated Negative Binomial regression. Results show that in terms of mean absolute percentage error (MAPE) and spatial correlation coefficient, the Negative Binomial model under the proposed spatiotemporal segmentation performs best. The performance of the proposed model is further validated by the crash dataset in other years.