【摘要】 为克服时空图卷积网络在交通流预测中输入序列较短、无法捕捉全局时空关系和长期历史周期特征等问题，提出了一种预训练增强的时空Transformer模型。为充分挖掘历史序列中非线性时间依赖关系和动态空间依赖关系，利用长期历史交通流数据预训练一个编码器，采用嵌入时间周期信息的时间感知注意力提取长序列特征；构建数据驱动的图学习层生成时空图，提取输入序列的时空特征；采用自适应策略融合长序列特征和时空特征，得到预测结果。在公开数据集PeMS04和PeMS08上进行实验，结果表明，对比最近提出的几类基线模型，新模型具有更高的预测精度。更多还原

【Abstract】 To address the limitations of spatiotemporal graph convolutional networks(ST-GCNs) in traffic flow prediction, including short input sequences, failure to capture global spatiotemporal correlations, and inability to model long-term historical periodic patterns, a pre-training enhanced spatial-temporal Transformer model is proposed. To comprehensively exploit the nonlinear temporal dependencies and dynamic spatial correlations in historical sequences, we pre-train an encoder using long-term historical traffic flow data, employing a temporal-aware attention mechanism embedded with periodic temporal information to extract long-sequence features. A data-driven graph learning layer is constructed to generate a spatial-temporal graph to further mine the spatial-temporal features of the input sequence. An adaptive strategy is applied to fuse long sequence features and spatialtemporal features to obtain the prediction results. Experiments are conducted on publicly available datasets PeMS04 and PeMS08, and the results show that the new model has higher prediction accuracy compared to several recently proposed baseline models.更多还原