LMSNet: Variable-aware multi-scale framework for multivariate time series forecasting

Capturing the rich, hierarchical information embedded in complex multivariate time series (MTS) necessitates modeling at multiple temporal scales. However, prevalent multi-scale forecasting methods suffer from critical limitations: a uniform scaling approach that overlooks individual variable characteristics, inconsistent modeling of temporal patterns and inter-variable dependency leading to modeling spurious correlations, and a neglect of global context. To overcome these challenges, we introduce LMSNet, a Lightweight Multi-Scale graph learning framework. Specifically, LMSNet introduces a novel variable-specific scaling mechanism, which leverages frequency-domain analysis to assign a tailored analytical window to each variable based on its intrinsic dynamics, establishing a robust foundation for accurately modeling multi-scale inter-variable dependencies. Furthermore, LMSNet synergizes global and local information and performs concurrent temporal and inter-variable dependency modeling at every network layer, aided by a momentum update mechanism that refines deep-layer graph node representations to mitigate the risk of modeling spurious correlations. LMSNet establishes a new state-of-the-art on eight benchmark datasets. Crucially, it achieves a 2.94% reduction in MAE with reducing GPU memory usage by 83% and accelerating training by 60% compared to strong baseline-TimeMixer. Our source code is provided at https://github.com/05Pikachu24/TSF-LMSNet.