基于动态时间调整的时空图卷积路网交通流量预测

为深入挖掘交通流数据的复杂时空特征并建立其依赖关系，提高交通流参数的预测精度， 本文提出一种新的交通流量预测模型——基于注意力机制和残差网络的时空关系图卷积网络 (TSARGCN)。TSARGCN对输入数据进行切片，实现多分支建模，挖掘数据的时间周期性特征； 引入残差网络保证网络中信息传递的完整性；利用DTW (Dynamic Time Warping)算法计算路网 中节点之间交通流量序列在时间维度的相似程度大小，提出时间图的概念，结合路网结构中各节 点的邻近关系，提出时空关系图的概念；基于时空关系图，在每个分支结合注意力机制分别进行图卷积和时间维度卷积，捕获交通流的时空特征及其依赖关系，实现对路网交通流量数据时空关系的建模。经过在公开数据集PEMSD4上进行实验，结果表明：TSARGCN在交通流量预测中的平均绝对误差 (MAE) 达 到 19.24，均方根误差 (RMSE) 达到 27.09，比 ARIMA(Autoregressive Integrated Moving Average model)，Conv-LSTM(Convolution Long short-term memory)及 ASTGCN (Attention based Spatial-temporal Graph Convolutional Network)等知名交通流量预测算法具有更高的预测精度。

Network-wide Traffic Flow Prediction Research Based on DTW Algorithm Spatial-temporal Graph Convolution

To deeply explore the complex temporal and spatial characteristics of traffic flow data and establish their dependence relationship, a new traffic flow prediction model, TSARGCN, based on attention mechanism and residual network, is proposed to improve the prediction accuracy of traffic flow parameters. The TSARGCN slices the input data to realize the time periodicity of data mining by multi-branch modeling. A residual network is introduced to ensure the integrity of information transmission in the network. The DTW algorithm was used to calculate the similarity degree of traffic flow sequence between nodes in the road network in the time dimension, and the concept of a time graph was put forward. Based on the spatial-temporal diagram, graph convolution and time convolution were carried out in each branch combined with an attention mechanism, respectively, to capture the spatial-temporal characteristics of the traffic flow and its dependence relationship, and then the spatial-temporal relationship of the traffic flow data was modeled. Experiments on the open data set PEMSD4 show that, MAE and RMSE of the TSARGCN are 19.24 and 27.09, respectively, which are better than those of ARIMA, CONV-LSTM, and ASTGCN.