Graph-Enhanced Temporal Modeling for Long-Sequence Forecasting: Dynamic Dependency Learning and Multi-Scale Feature Fusion

Abstract

This paper proposes a Graph-Enhanced Temporal Modeling (GETM) approach to address the challenges of long sequence forecasting involving complex multidimensional dependencies, dynamically evolving temporal structures, and heterogeneous feature scales. The method first constructs a temporal feature encoding module that extracts multi-granularity temporal representations through convolutional and transform-based architectures, enabling simultaneous capture of short-term fluctuations and long-term trends. A dynamic graph learning mechanism is then introduced, which dynamically updates inter-node dependency weights via similarity projection and adaptive sparsification strategies to model the evolving relationships over time. On this basis, a cross-scale feature fusion layer is designed to integrate features across different temporal resolutions through weighted aggregation, achieving a balanced representation of global consistency and local sensitivity. Meanwhile, a temporal consistency constraint is incorporated to ensure state smoothness and dynamic continuity across time steps. The proposed model achieves high-accuracy prediction and stable generalization on multidimensional system metric sequences, verifying the effectiveness of the graph enhancement mechanism in complex time-varying scenarios. Experimental results show that GETM maintains low error and high robustness in long sequence forecasting and demonstrates superior dynamic adaptability under non-stationary conditions, providing a practical and efficient solution for structured modeling and predictive analysis of multivariate temporal signals.