考虑交通事件影响的城市道路行程时间预测

外部环境因素对城市交通预测有较大影响，尤其在交通事件发生时，由于交通流的随机性和非线性特征，交通异常情况下的预测精度往往较低。为此，基于深度学习理论，提出一种以序列到序列模型（Sequence-to-sequence，Seq2Seq）为主体，融合外部因素特征的城市道路行程时间预测方法。利用时间序列分解算法（Seasonal and Trend Decomposition Using Loess，STL）挖掘交通历史数据的时序周期规律，结合交通事件数据深入分析交通异常产生的原因，并建立堆叠降噪自编码器模型（Stacked Denoising Autoencoder，SDAE）提取时间属性和交通事件数据的潜在特征。以北京市北四环中路和G6京藏高速路段为例，对预测模型的准确性和可行性进行验证，通过重复性交通事件和非重复性交通事件下的案例试验，对SDAE组件的有效性进行分析。研究结果表明：模型的单步和多步预测性能均优于基线模型，预测精度最高达到了87.71%；与其他输入了交通事件数据的模型相比，以SDAE作为外部组件的模型具有较好的预测性能和鲁棒性，能够适应复杂多变的交通流，在智能交通系统的短期预测上有显著的优越性，可以增强管理系统的调控能力，降低城市交通的拥堵成本。

Urban Road Travel Time Prediction Considering Impact of Traffic Event

The external environment factors have a great influence on urban traffic prediction, especially in the case of traffic event. Due to the randomness and non-linearity of traffic flow, the prediction accuracy of traffic anomalies is often low. Therefore, based on deep learning theory, a new method of urban road travel time prediction was proposed, which took the sequence-to-sequence (seq2seq) model as the main body and integrated the characteristics of external factors. This study used the seasonal and trend decomposition using loess (STL) to dig the time series cycle law of traffic history data, and deeply analyzed the causes of traffic anomaly combined with traffic event data. Finally, a stacked denosing autoencoder (SDAE) was established to extract the potential characteristics of time attribute and traffic event. Taking the segments of North Fourth Ring Middle Road and G6 Beijing-Tibet Expressway in Beijing as an example, the accuracy and feasibility of the prediction model were verified. And the effectiveness of SDAE model was analyzed through case experiments under recurring traffic event and non-recurring traffic event. The experimental results illustrate that the single step and multi-step prediction results of the model are superior to baseline models, and the highest prediction accuracy reaches 87.71%. In addition, compared with other models with traffic event data input, the model with SDAE has better prediction performance and robustness, and can adapt to the complex and changeable traffic flow. In the short-term prediction of the intelligent transportation system, the model has significant advantages, which can enhance the regulation ability of the management and reduce the congestion cost of the urban traffic.