影響高速公路事故發生之車流特性暨即時交通事故預測模式之研究

The Effects of Traffic Characteristics on Crash Occurrence and Real-time Traffic Crash Prediction Model on Taiwan Freeways

本研究應用高速公路沿線車輛偵器資料，以配對(有事故-無事故)控制實驗設計，應用條件羅吉特迴歸，分析車流特性對交通事故發生之影響，並以隨機森林建立即時交通事故預測模式，重要成果包含研擬及確認最佳車流觀測範圍及指標，鑑別交通事故發生之關鍵車流特性，建立並驗證即時交通事故預測模式及其可行性，提出實務應用作業流程。研究結果發現當車流平均速率下降、速率標準差增加、右側車道速率較左側速率高、或內側車道車流占比較中線車道高時，交通事故發生機率增加，建議交通管理機關修改高速公路車道使用規範，縮減各車道內的行車速率差異，確保行車安全及順暢，並建立交通事故預測相對應之交通事故防制機制和對策，防範事故於未然。

This study applied the matched control (crash and non-crash) experimental design and conditional logistic regression models to analyze the effect of traffic flow characteristics on traffic crashes on freeways. Machine-learning random forest models were also developed to predict the probabilities of traffic crashes by using vehicle detection systems along the freeways. The main contributions of this study include determining the optimal deployment and index in traffic flow monitoring systems, identifying the critical factors affecting traffic crashes, confirming the feasibility of the developed real-time traffic crash prediction models, and proposing the practical application procedure for traffic crash predictions. The results indicated that the likelihood of collisions increases as the mean speed decreases and the variance of speed increases. Higher speed of the right-side lanes than the left-side lanes and higher occupancy rate of the inner traveling lane than the middle traveling lanes also tend to result in higher possibility of crashes. These findings confirm that the traveling lane flow distribution and speed variance are critical factors affecting the occurrence of traffic crashes on freeways. Revising traffic regulations is needed to regulate the use of traveling lanes to decrease the speed variances within individual traffic lanes, ensuring traffic safety and efficiency. Traffic management agencies should refer to the real-time prediction model to establish traffic crash prevention countermeasures to prevent traffic crashes in advance. Keywords: Conditional logistic regression, Random Forest