关键词: 智能交通, 交通流数据修复, 图卷积网络, 城市路网交通流数据, 自注意力机制

Abstract: Focusing on the urban traffic flow imputation problem, this paper proposes a self- attention graph autoencoder (SA- GAE, Self- Attention Graph Auto- Encoder) based on the Graph Convolutional Networks and Multihead-Attention. The model includes the STGCN (Spatial-temporal Graph Convolutional Networks) network which captures the spatial- temporal correlation of traffic flow based on the topological graph structures and graph signals. In this network, the LSTM (Long Short-Term Memory) network is used to learn the temporal relationship in the data, the road self-attention and the first- order adjacent road attention coefficient are calculated through the road attention network, and the graph signal is reorganized by the graph convolution network to achieve the goal of precise imputation of missing data. The Multihead-Attention network is used to calculate the attention weight of the data and reorganize the data. The Multihead-Attention network can capture the spatial correlation in the second-order and highorder neighbor road traffic flow data and extract the relationship between the known data after the missing period and the missing data. The time relationship is added to the model in the form of residual chain as a supplement to the function of the STGCN. Experiments show that in multiple missing mode scenarios, the model can learn the topological relationship of road network, capture the temporal regularity in traffic data, understand the temporal and spatial correlations contained, and effectively repair the missing parts of the data.

Key words: intelligent transportation, traffic flow data imputation, graph convolutional network, urban road network traffic flow data, self-attention mechanism