Scholarlink

<p>Traffic congestion poses significant economic and environmental challenges worldwide, necessitating accurate traffic flow prediction for effective management. This paper proposes the Causal-Transformer model, a novel deep learning framework that integrates Granger causality analysis with a spatio-temporal causal attention mechanism. Traditional models often overlook causal relationships and struggle with nonlinear spatio-temporal dependencies in traffic data. The Causal-Transformer dynamically selects causal features through Granger causality testing, suppresses redundancy via semi-orthogonal projections, and employs an encoder-decoder architecture to capture global spatio-temporal interactions. Evaluated on real-world highway surveillance data, our model outperforms state-of-the-art methods such as LSTM, GRU, and hybrid architectures like CNN-LSTM, AttentionRNN, and CausalConv, achieving 17-57% lower MSE and 6-24% lower MAE. By explicitly modeling causal dependencies and optimizing video data utilization, this work advances real-time traffic prediction and offers a robust solution for intelligent transportation systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Causal-transformer: enhancing traffic flow forecasting with Granger causality-informed spatio-temporal attention

  • 期刊论文

  • 发表日期：

    2026年3月31日

查看全文
  • Haiyue Liu,
  • Zisen Mao

摘要

Traffic congestion poses significant economic and environmental challenges worldwide, necessitating accurate traffic flow prediction for effective management. This paper proposes the Causal-Transformer model, a novel deep learning framework that integrates Granger causality analysis with a spatio-temporal causal attention mechanism. Traditional models often overlook causal relationships and struggle with nonlinear spatio-temporal dependencies in traffic data. The Causal-Transformer dynamically selects causal features through Granger causality testing, suppresses redundancy via semi-orthogonal projections, and employs an encoder-decoder architecture to capture global spatio-temporal interactions. Evaluated on real-world highway surveillance data, our model outperforms state-of-the-art methods such as LSTM, GRU, and hybrid architectures like CNN-LSTM, AttentionRNN, and CausalConv, achieving 17-57% lower MSE and 6-24% lower MAE. By explicitly modeling causal dependencies and optimizing video data utilization, this work advances real-time traffic prediction and offers a robust solution for intelligent transportation systems.