MFENet: A Spatiotemporal Traffic Flow Prediction Framework Based on Multi-dimensional Feature Embedding

Traffic flow prediction serves as a core component of Intelligent Transportation Systems (ITS) and is critical for ensuring system reliability and safety. However, the complex spatiotemporal dynamics of traffic introduce significant challenges to accurate prediction. In existing research, traditional statistical models, machine learning approaches and deep learning frameworks have been widely adopted. Among them, deep learning-based spatiotemporal graph models have emerged as the dominant paradigm. Although effective, these models typically depend on intricate computational frameworks to model temporal and spatial correlations, resulting in lower computational efficiency and increased training expenses. To address these issues, a multi-dimensional feature embedding mechanism is proposed to enable lightweight and efficient spatiotemporal dependency modeling. This mechanism incorporates periodic embeddings to integrate daily and weekly patterns, dynamically represents spatial correlations, and introduces adaptive embedding to optimize the fusion of temporal and spatial features. The proposed model MFENet is designed to reduce model complexity while maintaining high prediction accuracy. Empirical evaluations across four practical traffic flow datasets validate that our proposed approach not only outperforms existing baseline models in forecasting accuracy but also enhances computational efficiency.