Weather-Effect Disentanglement Network Boosts Traffic Forecasts in Severe Weather

Global: Weather-Effect Disentanglement Network Boosts Traffic Forecasts in Severe Weather

Model Overview

A dual‑branch Transformer model named WED‑Net has been introduced to improve urban spatio‑temporal traffic prediction when extreme weather events occur. By explicitly separating intrinsic traffic dynamics from weather‑induced variations, the system targets the rarity and complexity of conditions such as heavy rain.

Architectural Design

The network employs self‑attention within each branch to capture baseline patterns and cross‑attention to fuse weather‑related signals. Memory banks store historical representations, while an adaptive gating mechanism determines the contribution of each branch for the final forecast.

Weather Discriminator

To reinforce disentanglement, a discriminator module is trained to recognize distinct weather conditions, encouraging the primary branches to encode complementary information.

Causal Data Augmentation

Researchers also propose a causal augmentation technique that perturbs non‑causal components of the data while preserving causal relationships. This approach is intended to enhance out‑of‑distribution generalization, particularly for scenarios that are under‑represented in training sets.

Empirical Evaluation

Experiments on taxi‑flow datasets from three major cities demonstrate that WED‑Net delivers consistently lower prediction errors under severe weather compared with prior methods. Consequently, the model shows promise for applications in safer mobility, disaster preparedness, and urban resilience.

Availability

The implementation and trained models have been released publicly at https://github.com/HQ‑LV/WED‑Net, enabling further research and potential deployment.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.