Scholarlink

<p>Accurate short-term load forecasting is essential for power system reliability and the integration of renewable energy. However, existing models often fail to simultaneously capture long-term dependencies and local fluctuations in load data. This study proposes CA-BiGRU+Autoformer, a hybrid model that combines Convolutional Neural Networks (CNN), Multi-Head Attention, Bidirectional Gated Recurrent Units (BiGRU), and the Autoformer framework. The model is designed to enhance multi-scale feature extraction by integrating local detail perception, bidirectional temporal modeling, and trend-seasonal decomposition. Experiments were conducted on two public datasets—GEFCom2014-E and the Australian Electricity Load and Price Forecasting dataset—to validate the model’s performance across different time resolutions and regional patterns. Compared with state-of-the-art baselines such as Autoformer, Informer, and LSTM, the proposed model achieves superior accuracy. On the GEFCom2014-E dataset, it reduces MSE by 11.4% in 96-step forecasting, while on the Australian dataset, it lowers MSE by 13.6% in 24-step tasks. Ablation studies confirm the effectiveness of each component. The results demonstrate that CA-BiGRU+Autoformer effectively captures complex temporal patterns and offers a robust solution for real-world short-term load 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

Research on the application of an improved Autoformer model integrating CNN-attention-BiGRU in short-term power load forecasting

  • 期刊论文

  • 发表日期：

    2025年8月4日

查看全文
  • Ruijun Tie,
  • Ming Li,
  • Cong Zhou,
  • Nanwei Ding

摘要

Accurate short-term load forecasting is essential for power system reliability and the integration of renewable energy. However, existing models often fail to simultaneously capture long-term dependencies and local fluctuations in load data. This study proposes CA-BiGRU+Autoformer, a hybrid model that combines Convolutional Neural Networks (CNN), Multi-Head Attention, Bidirectional Gated Recurrent Units (BiGRU), and the Autoformer framework. The model is designed to enhance multi-scale feature extraction by integrating local detail perception, bidirectional temporal modeling, and trend-seasonal decomposition. Experiments were conducted on two public datasets—GEFCom2014-E and the Australian Electricity Load and Price Forecasting dataset—to validate the model’s performance across different time resolutions and regional patterns. Compared with state-of-the-art baselines such as Autoformer, Informer, and LSTM, the proposed model achieves superior accuracy. On the GEFCom2014-E dataset, it reduces MSE by 11.4% in 96-step forecasting, while on the Australian dataset, it lowers MSE by 13.6% in 24-step tasks. Ablation studies confirm the effectiveness of each component. The results demonstrate that CA-BiGRU+Autoformer effectively captures complex temporal patterns and offers a robust solution for real-world short-term load forecasting.