Scholarlink

Accurate traffic forecasting is essential for intelligent transportation systems, challenged by complex spatiotemporal dynamics in urban networks. We propose METAGCRN, integrating adaptive graph convolutions, temporal transformers, and memory-augmented learning to address these challenges. The framework features: 1) Dynamic graph convolution with Chebyshev approximation for evolving spatial relationships, 2) Transformer-based temporal modeling with positional encoding, and 3) Attention-driven memory retrieval of traffic patterns. Evaluations on METR-LA and PEMS-BAY demonstrate state-of-the-art performance, showing 3.4% MAE improvement over GW-Net in 60-minute predictions. The architecture exhibits enhanced robustness to network dynamics, particularly during peak-hour anomalies. An open-source implementation supports urban mobility applications and future extensions.

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

Memory-Enhanced Transformer Adaptive Graph Convolutional Recurrent Network for Traffic Flow Forecasting

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Cheng Jiang,
  • Chun Wang

摘要

Accurate traffic forecasting is essential for intelligent transportation systems, challenged by complex spatiotemporal dynamics in urban networks. We propose METAGCRN, integrating adaptive graph convolutions, temporal transformers, and memory-augmented learning to address these challenges. The framework features: 1) Dynamic graph convolution with Chebyshev approximation for evolving spatial relationships, 2) Transformer-based temporal modeling with positional encoding, and 3) Attention-driven memory retrieval of traffic patterns. Evaluations on METR-LA and PEMS-BAY demonstrate state-of-the-art performance, showing 3.4% MAE improvement over GW-Net in 60-minute predictions. The architecture exhibits enhanced robustness to network dynamics, particularly during peak-hour anomalies. An open-source implementation supports urban mobility applications and future extensions.