Abstract:

Aiming at the problem that existing traffic flow prediction models fail to make full use of the spatial and temporal characteristics of traffic flow data to achieve accurate prediction, a convolutional gated recurrent unit prediction model combined with the attention mechanism（ACGRU） is proposed. The model uses the Convolution Neural Network（CNN） and the Gated Recurrent Unit（GRU） to extract the spatiotemporal characteristics of traffic flowand then uses the attention mechanism to generate the traffic flow feature representation with the probability distribution of attention. At the same time, periodic characteristics of traffic flow are extracted by weekly similarity. All the features are integrated for regression prediction. Experiments on real traffic flow data sets show that the ACGRU model proposed in this paper has higher prediction accuracy and the prediction error is reduced by 9% on average compared with other prediction models.

Key words: intelligent transportation, short-term traffic flow forecast, Convolutional Neural Network（CNN）, Gated Recurrent Unit（GRU）, spatial-temporal feature

摘要：

针对现有交通流预测模型未能充分利用交通流数据的时空特征以实现准确预测的问题，提出一种结合注意力机制的卷积门控循环单元预测模型（ACGRU）。该模型利用卷积神经网络（CNN）和门控循环单元（GRU）提取交通流的时空特征，然后使用注意力机制生成含有注意力概率分布的交通流特征表示，同时利用交通流的周相似性提取周期特征，将所有特征相互融合进行回归预测。在真实交通流数据集上的实验表明，提出的ACGRU模型具有更高的预测精度，预测误差相比其他预测模型平均降低了9%。

关键词: 智能交通, 短时交通流预测, 卷积神经网络, 门控循环单元, 时空特征