最佳空間分段下之高速公路事故頻次模式

Abstract

事故頻次分析必須將事故資料依空間加分割，但由於在空間上相鄰路段的幾何設計及交通環境大致相同，易導致樣本間之自我相關性(autocorrelation)問題；其次，空間上因路段分割未能配合地理環境、沿線土地使用及道路幾何設計等因素，而忽略樣本空間異質性(heterogeneous)問題。以上兩項問題將嚴重影響模式參數推估之正確性，在實務上亦會因路段分割未能配合環境因素而難以施行。 基此，本研究利用基因演算法分別以空間自我相關最小及空間同質性最高兩項目標，分別進行路段分割，並與另外兩種常用的路段分割方式：固定路段模式（1公里路段視為一個樣本）及交流道模式（兩兩交流道間之路段視為一個樣本），進行比較。路段分割後，再以卜瓦松迴歸模式及負二項迴歸模式作為構建事故頻次模式之基礎。最後，以中山高速公路作為實例應用。結果顯示，以MAPE值而言，固定路段模式表現最佳，而本研究所提出的空間自我相關最小模式及空間同質性最高模式則次之。但若以空間自我相關係數觀之，主觀路段分割模式均呈現相當高的自我相關係數，而本研究所提出的兩種空間分割模式則能維持相當低的自我相關係數，避免可能產生的推估誤差。

To conduct an aggregate crash frequency analysis, the whole study area (e.g. a freeway or a city network) should be segmented into several smaller areas each of which is then treated as one sample so as to model building. Most of previous studies tend to subjectively segment the study area without considering the spatial correlation and similarity among segmentations, resulting into the potential estimation errors due to correlation among samples or difficult identification of accident-prone locations. These two problems will seriously affect the accuracy of model parameter estimates in practice will be divided by sections of failing to meet these environmental factors and difficult to apply. In this regard, this study develops Poisson regression and negative binomial crash frequency models under two optimal spatial segmentations (by genetic algorithms): one aims to minimize the autocorrelation of consecutive segmentations and another aims to maximize the similarity of explanatory variable values within one segment. The proposed two models are also compared with another two subjective segmentation models: fixed length model (every one kilometer segment is treated as a sample) and interchange segmentation model (the segment between two consecutive interchanges is treated as a sample). A case study on Taiwan No.1 freeway is conducted. Results show that in terms of mean absolute percentage error (MAPE), the fixed length model performs best due to its largest sample size, followed by the proposed two models. However, in terms of spatial correlation coefficient, the proposed two models exhibit rather low autocorrelation but the subjective segmentation models have much higher correlation coefficient.