文献类型：期刊文献

英文题名：STPSformer: Spatial-Temporal ProbSparse Transformer for Long-Term Traffic Flow Forecasting

作者：Lu, Jun[1]; Wang, Dan[1]; Wang, Zhanquan[1]

机构：[1] East China University of Science and Technology, School of Information Science and Engineering, Shanghai, China

年份：2024

外文期刊名：Proceedings of the International Joint Conference on Neural Networks

收录：EI(收录号：20244017122515)

语种：英文

外文关键词：Traffic control - Weather forecasting

摘要：Long-term traffic forecasting has promising applications in intelligent transportation systems (ITS). However, as the prediction horizon expands, existing traffic forecasting models face a significant hurdle in long-term traffic forecasting. Therefore, a Spatial-Temporal ProbSparse Transformer (STPS-former) is proposed to address the problem. Firstly, a spatial-temporal adaptive embedding module is designed to capture the inherent spatial-temporal relationships within traffic data. Then, a spatial-temporal ProbSparse self-attention module is proposed to globally model the dynamics of traffic sequences by selecting strongly correlated nodes with minimized time and space complexity. Moreover, a spatial-temporal gated fusion module is introduced to intelligently integrate the extracted feature information from the spatial-temporal ProbSparse self-attention, to obtain accurate long-term traffic forecasting. Experimental results on two public traffic datasets show that our model can outperform mainstream traffic forecasting baselines and exhibit competitive computational efficiency, indicating that STPSformer is effective and efficient in long-term traffic forecasting. ? 2024 IEEE.