LSTM及GRU模型用於預測市區交通流量之研究

Abstract

近年來，隨著各縣市智慧城市的推廣，如何改善交通混亂的議題一直受到大眾關注。若我們能有效預測交通流量，政府單位即可事先做好相關配套措施，有效舒緩交通擁塞的問題。在傳統上大多會使用ARIMA (Autoregressive Integrated Moving Average model)方法來預測交通流量，但隨著深度學習在其他領域有著突破性的發展，LSTM (Long Short-Term Memory)和GRU (Gated recurrent units)模型已被證實對於交通流量預測有良好的效益。\n對於交通流量的資料收集，本研究將撰寫程式收集「常態性交通流量資料」與「可預期之偶發性活動資料」。關於「常態性交通流量資料」，我們將使用臺北市政府資料開放平台的「車輛偵測器(VD)資料」作為資料來源。因為市區交通的狀況較為複雜，容易受到觀光盛會、演唱會、天氣等「可預期之偶發性活動」影響，本研究採用本團隊先前的研究成果[8]，透過撰寫爬蟲程式對於多個售票網站、旅遊觀光網站、中央氣象局進行資料收集。本研究將上述資料輸入至 LSTM和GRU 模型以對其進行訓練，並利用Adam Optimizer 對模型進行優化。LSTM和GRU 模型之實作，以 Google 開發之機器學習框架 TensorFlow進行。最後，我們以平均絕對誤差(Mean Absolute Error, MAE)、均方誤差(Mean Square Error, MSE)和平均絕對百分比誤差(Mean Absolute Percentage Error, MAPE)對模型之預測準確率進行評估，進而分析LSTM模型和GRU模型在市區交通流量預測之準確度及LSTM和GRU模型之效益。

In recent years, with the promotion of smart cities in each county, the issue of how to resolve the problem of traffic chaos has drawn much attention. If we can accurately predict the traffic flow, then we can alleviate the traffic congestion more effectively. ARIMA (Autoregressive Integrated Moving Average model) were used to predict traffic flow. As the deep learning method has a breakthrough in many other fields, more and more studies propose to use deep learning models to solve real-world problems, and the results approve that both LSTM (Long Short-Term Memory) and GRU (Gated Recurrent Units) models have excellent performance in traffic flow prediction.\nThe proposed method is going to use both "normal traffic flow data" and "predictable sporadic activity data." As to the "normal traffic flow data," we use the "Vehicle Detector (VD) Data" given by the Taipei City Government Information Open Platform. On the other side, the traffic is also vulnerable to predictable sporadic activities, such as citywide carnival and festival, large-scale concerts, weather, etc. At this part, we code web crawler for websites of the ticket office, tourist information, news information, Central Weather Bureau, etc. These training data is fed into the LSTM and GRU deep learning models. We then use Adam Optimizer to optimize the models. The implementations of both LSTM and GRU models are based on TensorFlow of Google. Finally, we use MAE (Mean Absolute Error) and MAPE (Mean Absolute Percentage Error) to evaluate the urban traffic flow prediction accuracy of both LSTM and GRU models.