Robust dynamic spatio-temporal graph neural network for traffic forecasting

Accurate traffic flow prediction is crucial for traffic planning and the development of intelligent transportation systems. However, existing approaches often overlook the impact of abnormal signals, suffer from inadequate modeling of multi-scale temporal dependencies, and fail to fully capture spatial correlations. To address these challenges, we propose a Robust Dynamic Spatio-Temporal Graph Neural Network (RDSTGNN) aimed at enhancing prediction performance and robustness in complex traffic scenarios. The proposed model consists of two key modules: a periodic module designed for normal periodic signals, and a Gradient Spatio-Temporal Local Graph Convolutional Network (GSTLGCN) for capturing spatio-temporal dependencies. In the periodic module, we introduce an abnormality filtering gate to eliminate noise, leverage Recurrent Neural Networks (RNNs) to model short-term dependencies, and incorporate a global attention mechanism to capture long-term dependencies, thereby jointly enhancing the modeling of periodic signals. In the GSTLGCN module, mean gradients are employed to suppress abnormal disturbances. We integrate static graphs constructed from prior knowledge, adaptive graphs, and dynamic graphs to jointly model complex spatio-temporal relationships, and formulate this process as a diffusion mechanism to improve information propagation. We evaluate our model on six real-world datasets, and the experimental results demonstrate that RDSTGNN significantly outperforms existing baselines across multiple evaluation metrics, validating the effectiveness and robustness of the proposed approach.