SCPaT Framework Enhances Multivariate Time Series Forecasting with Semantic Partitioning

SCPaT employs a Transformer-based architecture that uses adaptive semantic unit generation to decompose sequences. It constructs a dynamic semantic graph to model dependencies between heterogeneous temporal patterns, addressing limitations in fixed or multi-scale partitioning.