Dynamic spatiotemporal graph convolutional network collaborative pre-training learning for traffic flow prediction

Accurate traffic flow prediction constitutes a fundamental pillar for intelligent transportation systems (ITS) optimization. However, the intricate spatiotemporal correlations within traffic networks pose significant challenges to precise prediction. Existing methods often rely on static relational assumptions, inadequately capturing global temporal correlations and struggling to model the complex trends and periodic patterns present in long-term traffic data. To surmount these limitations, we present a novel dynamic spatiotemporal graph convolutional network collaborative pre-training learning (DGCN-PTL). Our methodology incorporates a dual stage architecture: initially, a pre-training stage employs masked autoencoder mechanisms coupled with Transformer architectures to effectively extract temporal representations from extensive historical time series data. Subsequently, the prediction stage executes downstream forecasting through several pivotal components. We develop a dynamic graph learning module that adaptively captures evolving spatial interdependencies among network nodes across temporal intervals. Additionally, we integrate gating mechanisms with self-attention operations to augment the model's capability in characterizing both local and global temporal correlations. A dedicated feature transformation module facilitates channel adaptation and representation refinement. Comprehensive experiments across four real-world datasets substantiate DGCN-PTL's superior performance against 23 state-of-the-art baselines, achieving remarkable improvements of 5.49 % over the most competitive existing method.