ASTPKEFormer: Adaptive Spatiotemporal Prior Knowledge Embedding-Induced Transformers for Traffic Data Forecasting

Abstract

Traffic forecasting is fundamentally challenging due to the complex and dynamic spatiotemporal dependencies inherent in road networks. Although existing prediction models are able to achieve certain results on this task, existing Transformer-based models usually rely on simple embedding strategies and do not fully utilize the prior knowledge embedded in traffic patterns and network topology. To address these limitations, we propose ASTPKEformer, a prior knowledge-guided Transformer framework for traffic prediction. Based on the pure Transformer spatiotemporal self-attention mechanism, this model first uses a multi-scale temporal channel alignment module to generate discriminative feature embeddings. Then, it incorporates temporal prior embeddings and spatial graph structure prior embeddings to provide learning guidance for the model in both temporal and spatial dimensions. To further enhance the representation capability, an embedding cross-fusion mechanism is introduced to strengthen the interaction between the previous embeddings and the adaptive spatiotemporal embedding. Extensive experiments on six real-world traffic datasets demonstrate that ASTPKEformer consistently outperforms state-of-the-art (SOTA) baselines, validating its effectiveness and strong generalization ability.