基于时空注意力卷积神经网络的交通流量预测

为充分挖掘交通流量的复杂时空动态相关性以提高交通流量预测精度，引入空间注意力机制与膨胀因果卷积神经网络，提出一种基于时空注意力卷积神经网络的交通流量预测模型（spatio-temporal attention convolutional neural network,STACNN）.首先，由膨胀因果卷积与门控单元构建的门控时间卷积网络模块用于获取交通流量的非线性时间动态相关性，避免在训练长时间序列时发生梯度消失或梯度爆炸；其次，采用空间注意力机制为路网中的交通传感器节点自动分配注意力权重，动态关注不相邻节点之间的空间关系，并结合图卷积神经网络提取路网的局部空间动态相关性特征；然后，通过全连接层获取最终的交通流量预测结果；最后，利用高速公路交通数据集PEMSD4、PEMSD8进行了60 min的交通流量预测实验.实验结果表明：与基线模型中具有良好性能的时空图卷积网络（spatio-temporal graph convolutional network,STGCN）模型相比，提出的STACNN模型预测结果的平均绝对误差（mean absolute error,MAE）在两个数据集上分别提...

Traffic Flow Prediction Based on Spatial-Temporal Attention Convolutional Neural Network

In order to fully exploit the complex spatial-temporal dynamic correlation of traffic flow and improve the accuracy of traffic flow prediction, a spatial attention mechanism and an dilated causal convolutional neural network are introduced. A traffic flow prediction model STACNN based on spatial-temporal attention convolutional neural network is proposed. Firstly, the gated temporal convolution network block constructed by dilated causal convolution and gating unit is used to obtain the nonlinear temporal dynamic correlation of traffic flow and avoid gradient disappearance or gradient explosion when training long-term sequences. Secondly, the spatial attention mechanism is used to automatically assign attention weights to the traffic sensor nodes in the road network, which can dynamically pay attention to the spatial relationship between non-adjacent nodes, and combine the graph convolutional neural network to extract the local spatial dynamic correlation of the road network. Then, the final traffic flow prediction result is obtained through the fully connected layer. Finally, a 60-minute traffic flow prediction experiment is carried out using two highway traffic datasets PEMSD4 and PEMSD8. The experimental results show that: compared with the spatio-temporal graph convolutional network (STGCN) model with good performance in the baseline model, the MAE (mean absolute error) value of the prediction results of the proposed STACNN model on the two datasets is improved by 2.79% and 1.18%, the MAPE (mean absolute percentage error) value increased by 1.00% and 0.46%, and the RMSE (root mean square error) value increased by 3.8% and 1.25%, respectively. In addition, introducing dilated causal convolutional neural network and spatial attention mechanism have positively contributed to extraction of spatial-temporal dynamic correlation features.