由於車輛急遽成長，使中山高速公路經常發生行車擁塞及交通事故，造成的時間延滯與社會成本損失已相當嚴重，傳統上肇事分析集中於事後資料之分析，且多著重於肇事頻率與道路幾何設計、車流特性等等的因果關係等，較少針對肇事作即時性的預測。緣此，本研究於考量事故發生處理程序之時間特性與合理性，建立二階段處理時間預測模式，依序為影像觀測與警員到場回報，以提供用路人事故資訊並幫助交控單位於已知處理時間下，進行交控策略之實施，妥善而有效率的提高高速公路運轉績效。類神經網路對於非線性問題、因子間共線性之預測上具有良好預測能力之優點，因此本研究採用類神經網路中倒傳遞演算法(Back propagation)進行處理時間預測模式的構建。 為捕捉肇事因子間特性並提昇模式的預判能力，本研究先就歷史資料探討高速公路交通事故發生之影響因子種類對於處理時間影響之特性，接著回顧各處理時間預測模式，比較各模式的特性及其預測時所發生的各種問題與限制，作為本研究模式構建之參考。本研究以倒傳遞演算法之三層(Three-Layer)、完全連結(Fully-Connected)及前向(Feed-Forward)等架構型態進行預測。肇事處理中影響的因素很多，本研究於訓練與測試範例中採用貢獻圖的概念，評估輸入層處理單元之影響性，選出較具影響之變數，而後分別動態調整隱藏層層數、學習速率及慣性因子等相關參數，以找出最佳之預測模式。最後，本研究以國內實際架設影像偵測所獲得之肇事資料進行一驗證測試與分析，在依分析之結果提出結論與建議。 With the rapidly increase of motor vehicle has incline to freeway capacity. Once the traffic accident occurred on freeway, it would cause time delay and huge social cost expenditure. Traditionally, most previous studies concentrated on the relationship between cause and effect of the accident frequencies and the following factors including road design, traffic flow, environment to discuss the severity of accident based on the aggregate data. Quite few studies has concentrated on the immediately response of traffic accident. Therefore, we take into account the procedure dealing with traffic accident and construct a two-phase accident duration forecasting model, which will provide drivers the accident information and traffic control center to take measures to promote the freeway performance. Artificial neural network (ANN) is approved with its good performance on non-linear phenomenon and the ability to make use of correlated data. We used “back propagation algorithm” of artificial neural network in constructing the two-phase accident duration model. In order to know the characteristics between accident factors and promote the forecast performance of the accident duration forecasting model, we first discussed the effect of accident factors to the accident duration from historical data. Then we reviewed several predicting models and focus on the comparison of the attributes and restrictions between each of them. We used “Three-Layer”, “Fully-Connected” and “Feed-Forward” in back propagation algorithm to forecast the accident duration. The performance of duration forecasting model is affected by many accident factors. We used the contribution graph to select the variables we need, then search for the best forecast model including "training and testing set", "input variable", "hidden layer", "transfer function", "learning rule", and "momentum factor". To evaluate the model, we performed a case study using real image detection accident data on freeway from the police records.
