Scholarlink

<p>Recent progress in time series forecasting reveals a paradox: Despite sophisticated Transformers, minimalist linear models often achieve comparable or better performance, suggesting that a single uniform inductive bias may be insufficient to capture co-existing heterogeneous dynamics. We argue that progress lies not in a binary choice between linear and nonlinear approaches, but in a specialized synthesis tailored to distinct components. We introduce ADSNet, an adaptive dual-stream network following this principle. ADSNet first applies an improved Triple Exponential Moving Average (TEMA) to decompose signals into trend and seasonal components, which are processed in parallel by a lightweight Trend Stream and a Transformer-based Seasonal Stream. The Seasonal Stream is enhanced by Dual-Branch Adaptive Attention (DBAA), which adaptively fuses sparse and dense attention to handle both periodic patterns and abrupt events. A cross-attention module integrates the streams. On nine benchmarks, ADSNet achieves state-of-the-art results (best MSE 30/45; best MAE 37/45) and offers near-linear GPU scalability, &gt;12,500 samples/s throughput, and <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\approx \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>≈</mo> </math></EquationSource> </InlineEquation>2.6 ms latency for real-time smart-grid and IoT forecasting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Adsnet: an adaptive dual-stream network for multivariate time series forecasting

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • Yifan Zhang,
  • Junsan Zhang,
  • Jian Wang

摘要

Recent progress in time series forecasting reveals a paradox: Despite sophisticated Transformers, minimalist linear models often achieve comparable or better performance, suggesting that a single uniform inductive bias may be insufficient to capture co-existing heterogeneous dynamics. We argue that progress lies not in a binary choice between linear and nonlinear approaches, but in a specialized synthesis tailored to distinct components. We introduce ADSNet, an adaptive dual-stream network following this principle. ADSNet first applies an improved Triple Exponential Moving Average (TEMA) to decompose signals into trend and seasonal components, which are processed in parallel by a lightweight Trend Stream and a Transformer-based Seasonal Stream. The Seasonal Stream is enhanced by Dual-Branch Adaptive Attention (DBAA), which adaptively fuses sparse and dense attention to handle both periodic patterns and abrupt events. A cross-attention module integrates the streams. On nine benchmarks, ADSNet achieves state-of-the-art results (best MSE 30/45; best MAE 37/45) and offers near-linear GPU scalability, >12,500 samples/s throughput, and \(\approx \) ≈ 2.6 ms latency for real-time smart-grid and IoT forecasting.