基于有向图卷积与门控循环单元的短时交通流预测方法

为了充分挖掘快速路交通流时空特性，解决当前城市快速路交通流预测存在交通流时空特性挖掘不充分等问题，进一步提高城市快速路短时交通流的预测精度与效率，研究了基于有向图卷积神经网络和门控循环单元的组合模型（directed graph convolution network-gate recurrent unit，DGC-GRU）的城市快速路短时交通流预测方法。该方法提出空间相关性矩阵并将其引入图卷积神经网络中，构建有向图卷积神经网络用于表征交通流的有向性和流动性。将交通流参数输入有向图卷积神经网络后得到有向图卷积算子，并将有向图卷积算子引入门控循环单元，通过有向图卷积神经网络捕捉交通流的空间特性，通过门控循环单元捕捉交通流的时间特性，输出快速路交通流预测结果。选取西雅图环形快速路感应器检测数据进行实例分析，对比模型预测效果。结果表明：在数据集与参数设置均相同的情况下，DGC-GRU交通流预测模型的训练收敛速度更快，且平均绝对误差（mean absolute error，MAE）和平均绝对百分比误差（mean absolute percentage error，MAPE）均优于对比模型，与传统的GRU、GCN、DGC-LSTM模型相比，DGC-GRU模型能够将MAE和MAPE指数分别降低33.01%、5.76%、1.32%和27.75%、1.15%、7.76%，表明DGC-GRU交通流预测模型能够有效挖掘城市快速路网中的交通流时空分布特征，具有良好的预测精度与效率。

A Method of Predicting Short-term Traffic Flows Based on a DGC-GRU Model

In order to study the spatiotemporal characteristics of traffic flows on urban expressways which have not been fully explored in previous studies, a shortterm prediction method for traffic flow on urban expressways is proposed, in order to improve the prediction accuracy and efficiency based on a combined model of directed graph convolutional neural networks and gated recurrent units (DGC-GRU). The proposed method uses a spatial correlation matrix, which is also combined with a graph convolutional neural network. A directed graph convolutional neural network (DG-CNN) is developed to characterize the directionality and variability of traffic flows. Traffic flow parameters are input into the DG-CNN to obtain the directed graph convolution operator, and the directed graph convolution operator is introduced into the gated loop unit. The DG-CNN and the gated loop unit are used to capture spatial and temporal features of traffic flow, respectively, and are combined to predict traffic flow on expressways. Traffic flow data collected from a Ring Expressway of the City of Seattle is used for experiment analysis, in order to compare the performance of the proposed prediction models. Study results show that the convergence speed of the proposed DGC-GRU model is faster than other baseline models, and the mean absolute error (MAE) and mean absolute percentage error (MAPE) of the DGC-GRU model are smaller than those of the baseline models, given the same dataset and parameter settings. Compared with traditional GRU, GCN, and DGC-LSTM models, the DGC-GRU model reduces the MAE by 33.01%, 5.76%, 1.32%, and MAPE by 27.75%, 11.15%, 7.76%, respectively, which indicate that the DGC-GRU model can effectively study the spatiotemporal characteristics of traffic flows of the urban expressway network and has a better performance on prediction accuracy and efficiency than the compared models.