Incident-Guided Spatiotemporal Traffic Forecasting

TL;DR

This work targets the challenge of forecasting traffic under external disturbances by introducing IGSTGNN, a framework that explicitly models incident-driven spatio-temporal dynamics through two core modules: ICSF for initial incident-aware spatial fusion and TIID for dynamic decay of incident impact. The approach is validated on large-scale, incident-aligned traffic datasets, achieving state-of-the-art accuracy and demonstrating the plug-and-play generality of ICSF and TIID across diverse STGNN backbones. A new benchmark dataset aligning incident records with traffic time series is released to enable reproducible evaluation. Overall, IGSTGNN offers a principled, scalable solution for incident-aware forecasting with practical implications for real-world ITS resilience and efficiency.

Abstract

Recent years have witnessed the rapid development of deep-learning-based, graph-neural-network-based forecasting methods for modern intelligent transportation systems. However, most existing work focuses exclusively on capturing spatio-temporal dependencies from historical traffic data, while overlooking the fact that suddenly occurring transportation incidents, such as traffic accidents and adverse weather, serve as external disturbances that can substantially alter temporal patterns. We argue that this issue has become a major obstacle to modeling the dynamics of traffic systems and improving prediction accuracy, but the unpredictability of incidents makes it difficult to observe patterns from historical sequences. To address these challenges, this paper proposes a novel framework named the Incident-Guided Spatiotemporal Graph Neural Network (IGSTGNN). IGSTGNN explicitly models the incident's impact through two core components: an Incident-Context Spatial Fusion (ICSF) module to capture the initial heterogeneous spatial influence, and a Temporal Incident Impact Decay (TIID) module to model the subsequent dynamic dissipation. To facilitate research on the spatio-temporal impact of incidents on traffic flow, a large-scale dataset is constructed and released, featuring incident records that are time-aligned with traffic time series. On this new benchmark, the proposed IGSTGNN framework is demonstrated to achieve state-of-the-art performance. Furthermore, the generalizability of the ICSF and TIID modules is validated by integrating them into various existing models.

Figures (5)

• Figure 1: Illustration of a conventional model's failure under incident conditions. (a) Depicts the road network under a normal scenario, along with the typical daily traffic flow at node $n_1$. (b) Depicts the same network during an incident scenario with a collision at an intersection. The chart on the right focuses on the post-incident traffic flow at node $n_1$: the ground truth (black solid line) drops sharply, while the traditional model's forecast (dark red dashed line) completely misses this change. The pink vertical dashed line indicates the incident's start time, the pink vertical block highlights the two-hour impact duration, and the shaded red area quantifies the significant prediction error.
• Figure 2: The overall architecture of our proposed IGSTGNN framework. (a) illustrates the main pipeline of the model, while (b) and (c) provide detailed views of the ICSF and TIID modules, respectively.
• Figure 3: Effectiveness of the ICSF and TIID modules. The figure compares the performance (MAE and RMSE) of baseline models in their original form (Raw) versus after integrating the modules separately (ICSF, TIID) and in combination (BOTH) on the Alameda and Contra Costa datasets.
• Figure 4: Performance comparison (in terms of average MAE and RMSE) of our ICSF module against MLP and IMP fusion methods on the Alameda and Contra Costa datasets.
• Figure 5: MAPE (%) results of the ICSF module superiority study.