DySCo: Dynamic Semantic Compression for Effective Long-term Time Series Forecasting

Xiang Ao; Yinyu Tan; Mengru Chen

cs.LGApr 1, 2026

DySCo: Dynamic Semantic Compression for Effective Long-term Time Series Forecasting

Authors: Xiang Ao, Yinyu Tan, Mengru Chen

Organizations: School of Software Engineering, Beijing Jiaotong University, Beijing, China

Abstract

Time series forecasting (TSF) is critical across domains such as finance, meteorology, and energy. While extending the lookback window theoretically provides richer historical context, in practice, it often introduces irrelevant noise and computational redundancy, preventing models from effectively capturing complex long-term dependencies. To address these challenges, we propose a Dynamic Semantic Compression (DySCo) framework. Unlike traditional methods that rely on fixed heuristics, DySCo introduces an Entropy-Guided Dynamic Sampling (EGDS) mechanism to autonomously identify and retain high-entropy segments while compressing redundant trends. Furthermore, we incorporate a Hierarchical Frequency-Enhanced Decomposition (HFED) strategy to separate high-frequency anomalies from low-frequency patterns, ensuring that critical details are preserved during sparse sampling. Finally, a Cross-Scale Interaction Mixer(CSIM) is designed to dynamically fuse global contexts with local representations, replacing simple linear aggregation. Experimental results demonstrate that DySCo serves as a universal plug-and-play module, significantly enhancing the ability of mainstream models to capture long-term correlations with reduced computational cost.

Figures & tables

Fig. 1 : Predictive performance (measured by MSE) under different lookback windows. The ”Hoped” denotes the results conforming to common sense.

Panel of Figure 2 — Fig. 2 : Competition of different lookback windows. The short lookback window is derived from the long lookback window spanning steps 600 to 700. Note how the long window captures the trend but obscures the local periodicity visible in the short window.

Fig. 3 : Framework of DySCo. The left panel illustrates the overall basic framework of DySCo, which comprises the HFED module and prediction-related components; the right panel details the workflow of the EGDS module.

Dataset	Dim	Dataset Size	Information
ETTm1 [ 21 ]	7	(34465, 11521, 11521)	Temperature
ETTm2 [ 21 ]	7	(34465, 11521, 11521)	Temperature
ETTh1 [ 21 ]	7	(8545, 2881, 2881)	Temperature
ETTh2 [ 21 ]	7	(8545, 2881, 2881)	Temperature
Electricity [ 22 ]	321	(18317, 2633, 5261)	Electricity [ 22 ]
Traffic	862	(12185, 1757, 3509)	Transportation

TABLE I : Dataset detailed descriptions. The dataset size is organized as (Train, Validation, Test).

Dataset	TimeMixer						PatchTST
	DySCo		Basic(2440)		Basic(best)		DySCo		Basic(2440)		Basic(best)
	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE	MSE	MAE
ETT(avg)	0.339	0.369	0.445	0.450	0.357	0.389	0.321	0.362	0.446	0.482	0.342	0.377
Weather	0.219	0.253	0.260	0.314	0.231	0.267	0.221	0.253	0.231	0.332	0.231	0.266
Traffic	0.388	0.258	0.442	0.321	0.405	0.285	0.400	0.271	0.623	0.441	0.412	0.280
Electricity	0.141	0.221	0.201	0.332	0.166	0.257	0.152	0.231	0.189	0.272	0.162	0.255

TABLE II : Results of Comparative Experiments. “Basic(2440)” denotes the baseline model with input length 2440. “Basic(best)” denotes the baseline with its optimal input length among

\{96,192,336,720,1440,2440\}

. Lower MSE/MAE indicates better accuracy. The best results are highlighted in bold .

Sampling Strategy	ETT (Avg)		Weather		Electricity
Sampling Strategy	MSE	MAE	MSE	MAE	MSE	MAE
Uniform Sampling	0.426	0.458	0.298	0.364	0.222	0.346
Random Sampling	0.589	0.612	0.347	0.401	0.258	0.397
Top-k Frequency	0.412	0.498	0.283	0.347	0.221	0.342
EGDS	0.371	0.413	0.241	0.291	0.185	0.289

TABLE III : Comparison of different sampling strategies across ETT(Avg), Weather, and ECL datasets (Base model: Linear). The best results are highlighted in bold .

Variant	ETT (Avg)		Weather		Electricity
Variant	MSE	MAE	MSE	MAE	MSE	MAE
w/o HFED	0.394	0.418	0.252	0.304	0.192	0.299
w/o EGDS	0.443	0.522	0.289	0.321	0.237	0.393
w/o CSIM	0.381	0.447	0.244	0.299	0.191	0.303
DySCo (Full)	0.371	0.413	0.241	0.291	0.185	0.289

TABLE IV : Ablation study of DySCo components on ETT(Avg), Weather, and ECL datasets(Base model: Linear). The best results are highlighted in bold .

Fig. 4 : Efficiency comparison of basic models versus DySCo-integrated models across three representative datasets. The evaluation includes model parameters (left), GPU memory consumption (center), and training time per epoch (right). DySCo consistently demonstrates significant reductions across all computational metrics despite utilizing a multi-branch architecture.

Fig. 5 : Visualization on ETTh1: DySCo effectively captures the long-term trends underlying the noisy historical data, whereas the baseline model struggles with information redundancy.

Fig. 6 : Visualization on Electricity: DySCo maintains high prediction fidelity consistent with the baseline in short-term periodic scenarios, demonstrating its robustness.

Fig. 7 : Performance Analysis with Varying Lookback Windows.

Long-Term Time Series Forecasting

Long-term time series forecasting predicts distant future values from historical sequences in domains such as energy, weather, traffic, and health.

See all 115 papers

DMSC: Dynamic Multi-Scale Coordination Framework for Time Series Forecasting
Aug 3, 2025Haonan Yang, Jianchao Tang, Zhuo Li +1Time Series Forecasting Multi-Scale Feature Fusion
National University of Defense Technology, Changsha, Hunan, China
TA-SparseMG: Trend-Aware Sparse Forecasting via Multi-Scale Gating for Long-Term Time Series
Jun 26, 2026Wenchao Liu, Hongbing Wang, Youji Zhu +2Long-Term Time Series Forecasting Non-Stationary Time Series Forecasting
School of Mathematical Sciences, Guizhou Normal University, Guiyang 550025, China. · School of Big Data and Computer Science, Guizhou Normal University, Guiyang 550025, China.
Recovering Lost Details: Multi-Scale Frequency Compensation for Long-Term Time Series Forecasting
Sep 21, 2026Runmin Zou, Siyi Xie, Yaohui Huang +1Long-Term Time Series Forecasting Time Series Forecasting
Central South University Changsha, Hunan, China

Aug 3, 2025J/K move · Enter open · S save

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

Haonan Yang, Jianchao Tang, Zhuo Li, Long Lan

National University of Defense Technology, Changsha, Hunan, China

Time Series Forecasting Multi-Scale Feature Fusion Time Series Decomposition Multiscale Time-Series Modeling

Open paper AskPDF arXiv

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.

Abstract

Long-Term Time Series Forecasting

Explore similar work

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

TA-SparseMG: Trend-Aware Sparse Forecasting via Multi-Scale Gating for Long-Term Time Series

Recovering Lost Details: Multi-Scale Frequency Compensation for Long-Term Time Series Forecasting