應用類神經網路於都市地區之短期交通流量預測

Abstract

近年來，隨著工商業發展，都市化現象越來越明顯，人口高度集中，使得都市地區車輛密度增加，交通壅塞程度也越來越嚴重，造成交通成本及車輛行駛時間增加，使得車輛排放更多廢氣及熱能，加劇空氣污染及都市熱島效應，使都市環境更加惡化，為了改善這些問題，如何準確預測車流量，制定良好的交通策略，並提供車輛提前進行迴避，便是一項很重要的研究課題。 現今已有許多道路交通流量預測模型被提出，並應用於各種情境，像是高速公路、圓環、以及平面道路，然而，這些研究大多使用傳統統計學方法建構模型，但已有文獻指出，此方法相對較為簡陋，無法滿足許多實際的交通網路，因為這個原因，近年來越來越多研究開始引入機器學習、深度學習等新的運算技術，並且都有較高的準確度。 因此，本研究提出一個基於長短期記憶神經網路之短期車流量預測模型，以歷史流量的時間序列作為輸入。去預測下一個時間步長的車流量。模型中使用之資料取自台北市交通工程處建構之台北交通監測系統的資料，使用其中五個汽車偵測器的資料，將其分為訓練、驗證、測試三組資料，訓練資料被用來調整網路權重及偏差值，驗證資料被用來調整網路結構，然後，測試資料被丟入模型，並得到預測之交通流量，最後，實現小波類神經演算法，並將其與所提出之長短期神經網路進行效能比較，結果顯示，使用長短期神經網路之模型，其RMSE介於6.22到10.22之間，MAPE介於7.13%到11.14%之間，相較於小波神經網路的結果，RMSE約減少了2到3，MAPE則減少2%到5％。

In the past few years as the business industries develop, the phenomenon of urbanization has become more and more popular. With the increase of human populations, the density of vehicles in the city is also increased. As people rely more on motor vehicles, the traffic flow is often backed-up, causing a great deal in transportation costs and longer travel time on the roads. With vehicles having to travel longer than before, the air quality in the city is worsening due to the exhaust gases and heat produced by the vehicles. As a result, to provide better living qualities, it is an important topic to make improvements on traffic congestion prediction, transportation management, and advancement in avoiding traffic back-ups. There are many models of traffic flow prediction being proposed for different circumstances, such as highways, roundabouts, and general in-town roads. However, most of these models are established by traditional statistic analysis, but some researches have suggested that these models are too shallow to fulfill the complication of transportation network in life. Due to this cause, in recent years, more and more studies are introducing the new technology of computation, such as machine learning and deep learning, and have a higher accuracy. As a result, this study proposed a traffic flow prediction model based on long short-term memory neural network (LSTM NN). The historical time series of traffic flow is adopted as input to predict the traffic flow in next times step. The data of this model is derived from the traffic monitoring system in Taipei City that established by Taipei City Traffic Engineering Office. The data obtained by five vehicle detectors is adopted and spited into three parts, such as training data, validation data, and testing data. The training data is adopted to adjust the weights and the bias of network, and the validation data is used to adjust the structure of network. Then, the testing data would be thrown in the model and output the predicted traffic flow. Finally, a wavelet neural network is implemented and adopted to compare the performance of proposed LSTM NN model. The results show that the RMSE of LSTM NN model with five detectors ranges from 6.20 to 10.22 and the MAPE ranges from 7.13% to 11.14%. Compared to the results of wavelet NN, the RMSE decreases by 2 to 3, and the MAPE decreases by 2% to 5%.