最佳空間分段下之高速公路事故頻次與嚴重度模式

Abstract

建構事故頻次與嚴重度模式之前，在資料整合階段必需將每一筆事故資料依據其所屬肇事路段做空間上之分割，在傳統路段分割方式如：固定長度分段法(如：每5公里為一路段)、兩兩交流道分段法(以兩兩相鄰之交流道為一路段)等方法屬於主觀分段法，且容易導致路段間存在自我相關之問題，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，將嚴重影響模式參數推估之正確性。 基此，本研究提出空間分段模式以達到最佳化路段分割目標：同時最大化F值(組間變異除以組內變異)與最小化總分割路段數等兩目標，其中組間變異定義為各路段間事故件數或事故嚴重度比例之變異；組內變異則定義為同一路段內事故件數或事故嚴重度比例之變異情形。為了有效整合兩目標值且考量各目標值介於不同之範圍，將正規化之兩目標式依不同權重設定加總，以單一目標式方式呈現，本研究以民國94年台灣國道一號高速公路事故相關資料採用基因演算法在不同權重值下進行求解，再根據求得之分割路段利用卜瓦松迴歸、負二項迴歸與多項羅吉特模式建構事故頻次與嚴重度模式，並將空間最佳化分段下之事故頻次與嚴重度模式與傳統路段分割(每1公里為路段、兩兩交流道為路段)下之模式進行績效指標比較。結果顯示，以空間最佳化分割且權重為0.5分段方法下之一般線性負二項迴歸模式擁有最佳模式預測能力。最後以民國95年國道一號高速公路事故相關資料進行驗證，結果顯示以民國94年建構之事故頻次與嚴重度模式亦可適用於其它年份。

Prior to modeing crash frequency and severity, it is an essential step to spatially divid the study roadway system into several segments. Traditional segmentation methods, such as fixed length method (e.g. segementaion at every 5km) and interchange segmentation method (a segment is formed by two adjacent interchanges). However, these segmentation methods are mostly subjectively and tend to result into high correlation problem among neighboring segments, making the observations fail to distribute independently and leading to an estimation bias. Based on this, this study proposes a spatial segementation model to optimally divide roadway into segments so as to simultaneously maximize between variation/within variation and to minimize the number of segments. The between variation is defined as the variation of the number of crashes or the probability of crash severity between segments and the within variation is the variation within the segement. To integrate two objectives and to consider rather different range values of two objectives, a single objective is set as the weighted-sum of two normialized objectives. Genetic algorithms (GAs) is employed to solve the optimal segementaion under various weight settings of variations of crash frequency and crash severity based on the accident dataset in 2005. Based on the result of segementaion, the crash frequency and severity are separately modeled by Poisson regression (or Negative Binomial regression) and Multinomial Logit model. For comparison, two traditional segementaion methods, i.e. Fixed length method (segment length=1km) and Interchange methods, are also estimated. Results show that the Negative Binomial model under the optimal segmentation with equal weight (weight=0.5) performs best. The proformance of the proposed model in terms of MAPE value is also validated by the dataset of 2006.