應用混合深度學習模式於高速公路多時段交通速率預測

APPLICATION OF HYBRID DEEP LEARNING MODELS IN FREEWAY MULTI-PERIOD TRAFFIC SPEEDS PREDICTION

過去傳統依靠經驗法則的判斷方式，常因錯誤的判斷導致不僅無法有效紓解交通問題，嚴重者恐使問題加劇，因此更突顯預測技術與工具掌握的重要性。本研究認為在進行全時段交通預測時，除考量全時段整體預測的準確性外，亦須能夠同時考量尖峰時段預測的準確性。因此提出一新穎之混合模型，結合長短期記憶網路（LSTM）、注意力機制及雙向長短期記憶網路（Bi-LSTM），分別針對6種情境，包含全日全時段、全日尖峰時段、工作日全時段、工作日尖峰時段、週末全時段、週末尖峰時段進行交通速率預測。該模式有效地透過LSTM記憶及遺忘的特性，將交通流可用的資訊保留下來後，再利用Attention機制賦予各個重要資訊權重，最後利用Bi-LSTM綜合考量兩個方向（前向、後向）的交通流序列資訊進行預測。結果顯示本研究使用之混合模型，不論於不同日態或不同時段，其預測結果皆比其他模型來得可靠。並且可明顯改善尖峰時段之交通速率預測準確性。

Traffic speed prediction plays a crucial role in traffic congestion mitigation. With the advent of deep learning, transportation researchers are empowered to forecast traffic with an unprecedented accuracy. However, in the context of short-term prediction, such accuracy fails to hold consistent at hourly level; In particular, at peak hour periods. To bridge the gap, this study proposes a novel hybrid model that integrates LSTM, Attention Mechanism, and Bi-LSTM to for traffic prediction in six scenarios:“One week-24 hours”,“One week-Peak hours”,“Weekday-24 hours”,“Weekday - Peak hours”,“Weekend-24 hours”,“Weekend-Peak hours”. By leveraging the strengths of each constituent model, the proposed model demonstrates its capability of capturing peak hour traffic patterns as compared to previous models.