DMSC: Dynamic Multi-Scale Coordination Framework for Time Series Forecasting

TL;DR

This paper tackles TSF by addressing the limitations of static, fragmented multi-scale modeling. It introduces DMSC, a Dynamic Multi-Scale Coordination framework comprising EMPD for input-driven patch decomposition, TIB for triad dependency modeling, and ASR-MoE for temporal-aware, hierarchical fusion. Through a multi-layer progressive cascade, DMSC forms a coarse-to-fine feature pyramid, enabling adaptive integration of cross-scale information and cross-variable interactions. Extensive experiments on 13 real-world benchmarks demonstrate SOTA performance and superior efficiency, with ablations confirming the necessity of each component. The proposed approach offers a practical, scalable solution for robust TSF across diverse domains and high-dimensional data.

Abstract

Time Series Forecasting (TSF) faces persistent challenges in modeling intricate temporal dependencies across different scales. Despite recent advances leveraging different decomposition operations and novel architectures based on CNN, MLP or Transformer, existing methods still struggle with static decomposition strategies, fragmented dependency modeling, and inflexible fusion mechanisms, limiting their ability to model intricate temporal dependencies. To explicitly solve the mentioned three problems respectively, we propose a novel Dynamic Multi-Scale Coordination Framework (DMSC) with Multi-Scale Patch Decomposition block (EMPD), Triad Interaction Block (TIB) and Adaptive Scale Routing MoE block (ASR-MoE). Specifically, EMPD is designed as a built-in component to dynamically segment sequences into hierarchical patches with exponentially scaled granularities, eliminating predefined scale constraints through input-adaptive patch adjustment. TIB then jointly models intra-patch, inter-patch, and cross-variable dependencies within each layer's decomposed representations. EMPD and TIB are jointly integrated into layers forming a multi-layer progressive cascade architecture, where coarse-grained representations from earlier layers adaptively guide fine-grained feature extraction in subsequent layers via gated pathways. And ASR-MoE dynamically fuses multi-scale predictions by leveraging specialized global and local experts with temporal-aware weighting. Comprehensive experiments on thirteen real-world benchmarks demonstrate that DMSC consistently maintains state-of-the-art (SOTA) performance and superior computational efficiency for TSF tasks. Code is available at https://github.com/1327679995/DMSC.

Figures (7)

• Figure 1: The framework of DMSC. EMPD dynamically segments input series into hierarchical patches, TIB jointly models three dependencies, and ASR-MoE fuses multi-scale predictions adaptively.
• Figure 2: Visualization of different experts on ETTh1 datasets, the look-back and prediction length are set to 96.
• Figure 3: Forecasting results with varying look-back length on Electricity dataset. Look-back lengths are set to {48, 96, 192, 336, 720}, and prediction length is fixed to 96.
• Figure 4: Model efficiency analysis under 96-look-back length and 96-prediction length on Weather(21 variates) datasets. Batch size is set to 128.
• Figure 5: Model efficiency analysis under 96-look-back length and 96-prediction length on Traffic(862 variates) datasets. Batch size is set to 16.