A Multi-scale Representation Learning Framework for Long-Term Time Series Forecasting

TL;DR

This work tackles the challenge of long-term time series forecasting by proposing MDMixer, an MLP-based framework that explicitly disentangles multi-scale temporal information through a dual-branch trend–seasonal design and parallel multi-granularity predictors. A Multi-granularity Parallel Predictor (MPP) and a Multi-granularity Iterative Mixer (MIM) enable coarse-to-fine, channel-aware fusion via the Adaptive Multi-granularity Weighting Gate (AMWG), while a composite loss with a multi-granularity alignment term guides intermediate representations. Empirical results on eight LTSF benchmarks show that MD Mixer achieves state-of-the-art MAE improvements (e.g., −4.64% vs TimeMixer) with substantially improved training efficiency and interpretability, outperforming Transformer-based and other ML baselines. The approach demonstrates strong generalization to linear models when extended with the dual-branch decomposition and offers insights into how different granularities contribute to channel-specific forecasts, making it practically impactful for multi-variate, real-world forecasting tasks.

Abstract

Long-term time series forecasting (LTSF) offers broad utility in practical settings like energy consumption and weather prediction. Accurately predicting long-term changes, however, is demanding due to the intricate temporal patterns and inherent multi-scale variations within time series. This work confronts key issues in LTSF, including the suboptimal use of multi-granularity information, the neglect of channel-specific attributes, and the unique nature of trend and seasonal components, by introducing a proficient MLP-based forecasting framework. Our method adeptly disentangles complex temporal dynamics using clear, concurrent predictions across various scales. These multi-scale forecasts are then skillfully integrated through a system that dynamically assigns importance to information from different granularities, sensitive to individual channel characteristics. To manage the specific features of temporal patterns, a two-pronged structure is utilized to model trend and seasonal elements independently. Experimental results on eight LTSF benchmarks demonstrate that MDMixer improves average MAE performance by 4.64% compared to the recent state-of-the-art MLP-based method (TimeMixer), while achieving an effective balance between training efficiency and model interpretability.

Figures (9)

• Figure 1: The core concept of MDMixer: (Top-Left) Parallel prediction of dynamic patterns at different temporal granularities using MPP; (Top-Right) Adaptive, channel-specific fusion of multi-granularity information, with weights determined by AMWG; (Bottom) Generation of long-term forecasts that closely track the ground truth. We observe that coarse-grained predictions capture overall seasonal and trend patterns, while fine-grained predictions extract short-term fluctuations.
• Figure 2: Overview of MDMixer architecture. Multivariate time series are decomposed into trend and seasonal components. The respective branches process them using MLP-based and Linear-based modules for prediction. The Multi-granularity Parallel Predictor (MPP) and the Multi-granularity Iterative Mixer (MIM) are responsible for multi-granularity prediction and fusion, respectively. Predictions at the same granularity level from both the trend and seasonal branches are summed and aligned with the downsampled target sequence through the computation of alignment loss. The Adaptive Multi-granularity Weighting Gate (AMWG) takes the combined patch embeddings as input and produces dynamic, channel-specific weights to aggregate the multi-granularity predictions.
• Figure 3: Ablation of MPP, MIM, AMWG, and alignment loss on ETTm1 and Weather datasets.
• Figure 4: MSE scores with varying heads number $H \in \{2,4,8,12,16\}$.
• Figure 5: MSE scores with varying alignment loss weight $\alpha \in \{0.001,0.01,0.05,0.1,0.2\}$.