Scholarlink

Accurate power load forecasting is essential for grid stability and energy management. To address the limitations of existing models in capturing hierarchical periodic features and fusing multi-source meteorological data, this paper proposes a Transformer-based Dual-Scale Attention (DSA) framework. The model incorporates: (1) Cyclical positional encoding that explicitly represents daily/weekly/seasonal temporal structures using sine-cosine transforms; (2) Parallel intra-day and inter-day attention branches decoupling short-term dynamics (15-min resolution) and long-term trends (daily profiles) through transpose-concatenation fusion; (3) Path tokenization compressing long sequences into fixed-length segments, reducing memory overhead. Validated on 2021–2024 provincial industrial data (15-min granularity), DSA achieves 95% average accuracy for 72-h forecasts in MAPE. The model demonstrates robustness to meteorological anomalies.

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

Short-Term Power Load Forecasting Method Based on Dual-Scale Attention

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Peng Tang,
  • Yusi Wei

摘要

Accurate power load forecasting is essential for grid stability and energy management. To address the limitations of existing models in capturing hierarchical periodic features and fusing multi-source meteorological data, this paper proposes a Transformer-based Dual-Scale Attention (DSA) framework. The model incorporates: (1) Cyclical positional encoding that explicitly represents daily/weekly/seasonal temporal structures using sine-cosine transforms; (2) Parallel intra-day and inter-day attention branches decoupling short-term dynamics (15-min resolution) and long-term trends (daily profiles) through transpose-concatenation fusion; (3) Path tokenization compressing long sequences into fixed-length segments, reducing memory overhead. Validated on 2021–2024 provincial industrial data (15-min granularity), DSA achieves 95% average accuracy for 72-h forecasts in MAPE. The model demonstrates robustness to meteorological anomalies.