運用深度學習方法預測公車旅行時間之初探

Abstract

中文摘要 近年來隨著科技蓬勃發展，運用各種深度學習演算法投入於各項研究產業中都有些重大突破，而在交通運輸產業上，也結合先進科技技術、資訊來掌握資料的蒐集、應用、傳遞，目的是希望能提供給使用者更好的服務品質，其中的核心價值在於能安全情況下即時提供民眾真正所需的需求，藉此需要透過過去的資料進行預測，分析未來可能所產生的變化，及時提供資訊服務大眾，因此本研究利用長短期記憶LSTM演算法，建立多變項模式預測6702路線各班次總旅行時間及站點間之旅行時間，在旅行時間資料中使用過去總旅行時間，預測未來7天路線之總旅行時間及區間旅行時間，觀察加入節慶、星期、雨量變數之預測變化，並將預測模式之結果透過平均絕對百誤差MAPE比較進行參數調整，得出最佳之預測模式預測未來7天旅行時間，經本研究發現最佳之預測模式平均絕對百誤差MAPE為3.67%，表示模式預測結果為極佳。

Abstract In recent years, with the vigorous development of science and technology, the use of various deep learning algorithms to invest in various research industries has some major breakthroughs. In the transportation industry, advanced technology and information are also used to master the collection, application and transmission of data. The purpose is to provide users with better service quality. The core value is to provide users with real needs in real time under safe conditions. This requires forecasting through past data and analyzing possible changes in the future. Provide information to the public in time. Therefore, this study uses LSTM algorithm to establish a multi-variable model to predict the total travel time of each shift of the 6702 route and the travel time between stations. The past total travel time is used in the travel time data. Predict the total travel time and interval travel time of the route in the next 7 days, observe the forecast changes by adding the festival, week, and rainfall variables, and adjust the parameters of the forecast model results through the average absolute error MAPE comparison to obtain the best forecast model To predict the travel time in the next 7 days, this study found that the best prediction model has an average absolute hundred error MAPE of 3.67%, indicating that the model prediction results are excellent.