基于有向图卷积神经网络的交通预测与拥堵管控

为解决城市快速路正面临的日益严重的交通拥堵问题，提出了一种针对城市快速路的基于有向图卷积神经网络的交通预测与拥堵管控方法，该方法能够有效利用海量交通数据进行交通预测，实现拥堵的主动管控。首先，基于交通路网的空间有向性和交通流的时空特性，定义了有向的距离影响矩阵、修正欧式距离矩阵和自由流可达矩阵，构建出有向的图卷积算子，并将其应用于长短时记忆神经网络模型中，提出了能学习交通路网时空双重特性的有向图卷积-长短时记忆神经网络（Directed Graph Convolution-LSTM，DGC-LSTM）模型；其次，基于DGC-LSTM的交通预测结果识别出拥堵产生点并将其作为拥堵管控的对象；再次，采用控制进口匝道车辆输入快速路主线的手段，针对管控对象的时空特征，设计了全圈层分时段阶梯式拥堵管控策略；最后，基于上海市快速路网上布设的2 712个检测器在122个工作日每间隔5 min记录的速度、流量和占有率信息，开展实例分析，测试了DGC-LSTM模型的预测精度以及全圈层分时段阶梯式拥堵管控策略的有效性。结果表明：与传统的循环神经网络、长短时记忆神经网络相比，DGC-LSTM模型具有更高的预测精度，能将速度预测的平均绝对误差和误差标准差分别降低38%和20%以上；基于预测结果采用的全圈层分时段阶梯式拥堵管控策略能令拥堵产生点的速度提升14 km·h^-1以上，并能使拥堵的持续时长缩短40%，可阻止拥堵从产生点开始发生大范围的蔓延，降低整个路网的拥塞程度。

Traffic Prediction and Congestion Control Based on Directed Graph Convolution Neural Network

In order to solve the increasingly serious traffic congestion problem that urban expressways are facing, a traffic prediction and congestion control method based on a directed graph convolution neural network for urban expressways was proposed. It can effectively use massive traffic data to predict traffic and realize active congestion control. Firstly, based on the spatial directionality of the traffic road network and the spatial-temporal characteristics of traffic flow, a directed distance influence matrix, a modified Euclidean distance matrix and a free flow reachable matrix were defined. A directed graph convolution operator was constructed and applied in the Long Short-term Memory Networks (LSTM). After those settings, the Directed Graph Convolution-LSTM (DGC-LSTM) which was used to predict traffic flow status was established. Next, the congestion bottleneck was identified as the object of congestion control based on the traffic prediction results. Then, the approach of controlling the on-ramp vehicles to enter the mainstream of the expressway was selected as the basic measure, and the specific stepwise congestion management and control strategy for the whole circles by time period was designed according to the temporal and spatial characteristics of the bottleneck. Finally, based on the speed, flow and occupancy recorded by the 2 712 detectors deployed on the Shanghai expressway network at intervals of 5 minutes in 122 working days, a case study was carried out to test the prediction accuracy of the DGC-LSTM model and the effectiveness of the control strategy. Results show that the DGC-LSTM model has higher prediction accuracy and can reduce the Mean Absolute Error (MAE) and error Standard Deviation (SD) of speed prediction by about 38% and 20%, respectively, compared with the traditional Recurrent Neural Network (RNN) and LSTM models; the stepwise congestion management and control strategy can raise the speed at the bottleneck by more than 14 km·h^-1, and shorten the duration of congestion by 40%, which means that it can mitigate congestion spreading on a large scale from the congestion bottleneck, and reduce the congestion degree of the entire road network effectively.