基于广义时空图卷积网络的交通群体运动态势预测

针对当前高速公路与城市快速路交通拥堵现象愈发严重，为交通管理与控制造成巨大困难的问题，提出了一种基于广义时空图卷积网络(GSTGCN)的交通速度预测模型；基于交通数据自身具有的复杂时空特性，定义了广义交通数据图结构，同时构建了广义图的邻接关系；基于图卷积网络基础理论，采用切比雪夫近似与一阶近似简化了图卷积操作的计算成本，建立了广义图卷积算子；结合广义图卷积模块、标准卷积模块与线性全连接层，提出了用于提取复杂交通数据时间、空间特征的GSTGCN模型；利用美国威斯康星州密尔沃基市快速路网上架设的38个检测器，在21个工作日以每5 min为单位记录了车辆速度、流量和占有率数据，测试了GSTGCN模型在该数据集上的短期交通速度预测精度与训练效率。分析结果表明:相较于传统自回归求和滑动平均(ARIMA)模型、长短时记忆(LSTM)模型以及近期的STGCN模型，GSTGCN模型在交通速度的均方根误差、平均绝对误差和平均绝对百分比误差指标上分别降低了22.79%、22.97%和16.73%；此外，GSTGCN模型的训练时长比STGCN模型和LSTM模型分别降低了5.17%和75.71%。可见，GSTGCN模型能够有效处理复杂交通时空数据结构，准确预测交通速度，并为交通管控提供交通群体的运动态势信息。 

Prediction of traffic swarm movement situation based on generalized spatio-temporal graph convolution network

To address the problem that traffic congestion on highways and urban expressways is becoming more and more serious and causes great difficulties for traffic management and control, a traffic speed prediction model was proposed based on the generalized spatio-temporal graph convolution network (GSTGCN). According to the complex spatio-temporal characteristics of traffic data, the generalized traffic data graph structure was defined, and the adjacency relationships of the generalized graph were constructed. By the basic theory of graph convolution network, the Chebyshev approximation and the first-order approximation were adopted to simplify the computational cost of the graph convolution operation, and a generalized graph convolution operator was established. With the generalized graph convolution module, standard convolution module, and linear fully-connected layer, a GSTGCN model was presented to extract the spatial and temporal characteristics of complex traffic data. The vehicle speed, flow, and occupancy datum were recorded by 38 detectors at 5-minute intervals for 21 weekdays on the expressway network in Milwaukee, Wisconsin, USA. The short-term traffic speed prediction accuracy and training efficiency of the GSTGCN model were evaluated on this data set. Analysis results show that compared with the results of the traditional auto regressive integrated moving average (ARIMA) model, the long short-term memory (LSTM) model, and the recent spatio-temporal graph convolution network (STGCN) model, the root mean square error, mean absolute error, and mean absolute percentage error of the GSTGCN model in the traffic speed prediction reduces by 22.79%, 22.97%, and 16.73%, respectively. Moreover, the training time of the GSTGCN model is 5.17% and 75.71% shorter than those of the STGCN model and LSTM model, respectively. Therefore, the GSTGCN model is able to effectively deal with the complex spatio-temporal traffic data structure, accurately predict the traffic speed, and provide information on the movement situation of traffic swarm for the traffic control and management. 4 tabs, 6 figs, 31 refs.