Scholarlink

<p>Accurate day-ahead load forecasting is fundamental to ensuring the economic and stable operation of power systems and optimizing energy storage system scheduling. However, electricity load series exhibit multi-timescale coupling characteristics, making it difficult for a single model to simultaneously capture the dynamic variations of different frequency components. To address this challenge, this paper proposes a Multi-Scale Decomposition Gated Residual Network (MSDGRN). This model adopts an encoder-decoder framework and performs differentiated feature extraction through dual pathways: a local pathway focuses on the fine-grained features of the recent window of the historical series; a global pathway combines the decomposition of the entire historical series with multi-scale convolution to extract its trend and periodic components. In the temporal modeling stage, the encoder deeply encodes the local features; the decoder takes the global features as input and is initialized by the final state of the encoder, enabling the generation of future loads to grasp the macro trend while being aware of recent dynamic levels. In the feature fusion stage, a gated residual block is introduced to adaptively integrate the outputs of the encoder and decoder, dynamically balancing linear and nonlinear information through a gating mechanism. Experimental results on the ISO-NE public dataset show that MSDGRN achieves MAE, RMSE, and MAPE of 420.43&#xa0;MW, 642.94&#xa0;MW, and 2.77%, respectively, significantly outperforming six baseline models including LSTM, TCN, and Transformer. Peak prediction analysis and ablation experiments further verify the model’s robustness in critical decision windows and the effectiveness of each core module.</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

Day-ahead electricity load forecasting based on multi-scale decomposition gated residual network

  • 期刊论文

  • 发表日期：

    2026年5月9日

查看全文
  • Ziming Ouyang,
  • Xiangdong Chen,
  • Haoxin Li,
  • Weigui Wang,
  • Zhaohui Li

摘要

Accurate day-ahead load forecasting is fundamental to ensuring the economic and stable operation of power systems and optimizing energy storage system scheduling. However, electricity load series exhibit multi-timescale coupling characteristics, making it difficult for a single model to simultaneously capture the dynamic variations of different frequency components. To address this challenge, this paper proposes a Multi-Scale Decomposition Gated Residual Network (MSDGRN). This model adopts an encoder-decoder framework and performs differentiated feature extraction through dual pathways: a local pathway focuses on the fine-grained features of the recent window of the historical series; a global pathway combines the decomposition of the entire historical series with multi-scale convolution to extract its trend and periodic components. In the temporal modeling stage, the encoder deeply encodes the local features; the decoder takes the global features as input and is initialized by the final state of the encoder, enabling the generation of future loads to grasp the macro trend while being aware of recent dynamic levels. In the feature fusion stage, a gated residual block is introduced to adaptively integrate the outputs of the encoder and decoder, dynamically balancing linear and nonlinear information through a gating mechanism. Experimental results on the ISO-NE public dataset show that MSDGRN achieves MAE, RMSE, and MAPE of 420.43 MW, 642.94 MW, and 2.77%, respectively, significantly outperforming six baseline models including LSTM, TCN, and Transformer. Peak prediction analysis and ablation experiments further verify the model’s robustness in critical decision windows and the effectiveness of each core module.