Scholarlink

<p>Estimated Time of Arrival (ETA) prediction underlies route planning, dispatch, and arrival notifications, yet many systems estimate ETA from traffic history without explicit knowledge of a vehicle’s planned route. This work examines whether a vehicle-centered dynamic graph with route awareness improves ETA prediction. We present DSTRA-GNN (Dynamic Spatio-Temporal Route-Aware GNN), which integrates vehicle-level dynamics and route intent in a Mixture-of-Experts framework. Vehicles are graph nodes, interaction edges model local traffic context, and a GRU-based temporal aggregator processes sequential graph snapshots. We evaluate on SUMO simulation data and Porto-G, a graph representation of the Porto taxi trajectory dataset (ECML-PKDD 2015). Ablations over six variants show consistent, compounding gains from route awareness and temporal context. On duration-matched Porto-G, MAE drops from 93.3&#xa0;s (static graph only) to 68.1&#xa0;s (27% reduction), performing competitively with trajectory-based baselines such as DeepTTE (68.6&#xa0;s) and MetaTTE (69.5&#xa0;s). On SUMO, MAE drops from 80.1&#xa0;s to 54.8&#xa0;s (32% reduction) under the shared protocol range; on the broader p95-filtered SUMO population, the full model remains the strongest evaluated variant and outperforms applicable classical and graph-native baselines at 72.4&#xa0;s MAE. Vehicle-centered dynamic graphs with explicit route state improve ETA prediction across simulated and real-world data, positioning graph-snapshot ETA modeling as a competitive alternative to trajectory-native sequence models.</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

Accurate ETA Prediction Using Dynamic Route-Aware Graph Neural Networks for Improved Urban Mobility and Well-Being

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Guy Tordjman,
  • Nadav Voloch

摘要

Estimated Time of Arrival (ETA) prediction underlies route planning, dispatch, and arrival notifications, yet many systems estimate ETA from traffic history without explicit knowledge of a vehicle’s planned route. This work examines whether a vehicle-centered dynamic graph with route awareness improves ETA prediction. We present DSTRA-GNN (Dynamic Spatio-Temporal Route-Aware GNN), which integrates vehicle-level dynamics and route intent in a Mixture-of-Experts framework. Vehicles are graph nodes, interaction edges model local traffic context, and a GRU-based temporal aggregator processes sequential graph snapshots. We evaluate on SUMO simulation data and Porto-G, a graph representation of the Porto taxi trajectory dataset (ECML-PKDD 2015). Ablations over six variants show consistent, compounding gains from route awareness and temporal context. On duration-matched Porto-G, MAE drops from 93.3 s (static graph only) to 68.1 s (27% reduction), performing competitively with trajectory-based baselines such as DeepTTE (68.6 s) and MetaTTE (69.5 s). On SUMO, MAE drops from 80.1 s to 54.8 s (32% reduction) under the shared protocol range; on the broader p95-filtered SUMO population, the full model remains the strongest evaluated variant and outperforms applicable classical and graph-native baselines at 72.4 s MAE. Vehicle-centered dynamic graphs with explicit route state improve ETA prediction across simulated and real-world data, positioning graph-snapshot ETA modeling as a competitive alternative to trajectory-native sequence models.