发电技术 ›› 2026, Vol. 47 ›› Issue (2): 336-344.DOI: 10.12096/j.2096-4528.pgt.260211

• 新型电力系统 • 上一篇    

基于频域增强Transformer的多用户短期负荷预测方法

宋立业1, 孟凡宇1, 陈郁林2, 宋信萍1   

  1. 1.辽宁工程技术大学电气与控制工程学院,辽宁省 葫芦岛市 125105
    2.浙江大学海南研究院,海南省 三亚市 572024
  • 收稿日期:2025-04-19 修回日期:2025-06-18 出版日期:2026-04-30 发布日期:2026-04-21
  • 作者简介:宋立业(1972),男,博士,副教授,研究方向为智能电网优化运行等,372492761@qq.com
    孟凡宇(2000),男,硕士研究生,研究方向为数据挖掘、电力负荷预测,15658744@qq.com
    陈郁林(1992),男,博士,副研究员,研究方向为新能源分布式控制、信息物理安全、虚拟电厂,本文通信作者,chenyl2017@zju.edu.cn
  • 基金资助:
    海南省自然科学基金项目(524RC532)

Multi-User Short-Term Load Prediction Method Based on Frequency-Domain Enhanced Transformer

Liye SONG1, Fanyu MENG1, Yulin CHEN2, Xinping SONG1   

  1. 1.Faculty of Electrical and Control Engineering, Liaoning Technical University, Huludao 125105, Liaoning Province, China
    2.Hainan Institute of Zhejiang University, Sanya 572024, Hainan Province, China
  • Received:2025-04-19 Revised:2025-06-18 Published:2026-04-30 Online:2026-04-21
  • Supported by:
    Hainan Provincial Natural Science Foundation(524RC532)

摘要:

目的 负荷预测对于电力系统规划和运行至关重要。然而,新能源的高比例接入引发的源荷不确定性增大,使得新型电力系统负荷的波动性增加,导致多用户负荷准确预测的难度增大。为提高新型电力系统短期负荷预测的精度,提出一种基于频域增强Transformer的多用户短期负荷预测方法。 方法 利用Transformer中的编码器捕捉多负荷序列的特征信息;通过离散余弦变换(discrete cosine transform,DCT)获取其中的频域信息,并用通道注意力机制进行数据增强;利用Transformer中的解码器整合特征信息输入到全连接层中,得到预测数据。为验证所提方法的优越性,采用Electricity dataset数据集进行算例分析,并将所提方法与5种常用负荷预测方法进行对比分析。 结果 该方法与传统的Transformer负荷预测模型相比,误差降低了21.8%。 结论 所提方法在准确性和稳定性方面表现优异,在实际应用中具有可行性。

关键词: 电力负荷预测, 新能源, 源荷不确定性, 深度学习, Transformer模型, 离散余弦变换, 通道注意力机制, 数据增强

Abstract:

Objectives Load prediction is crucial for power system planning and operation. However, the high proportion of new energy increases source-load uncertainty, leading to greater volatility in the load of new-type power systems, which makes it more difficult to accurately predict multi-user load. To improve the accuracy of short-term load forecasting in new power systems, a multi-user short-term load prediction method based on frequency-domain enhanced Transformer is proposed. Methods The encoder in the Transformer is used to capture the feature information of multi-load sequences. The frequency-domain information is obtained by the discrete cosine transform (DCT), and the channel attention mechanism is used for data enhancement. The decoder in the Transformer is used to integrate the feature information which is then fed into the fully connected layer to obtain the prediction data. To verify the superiority of the proposed method, the Electricity dataset is used for case analysis, and a comparative analysis is conducted between the proposed method and five commonly used load forecasting methods. Results Compared with the traditional Transformer load prediction model, the error of the proposed method is reduced by 21.8%. Conclusions The proposed method demonstrates excellent accuracy and stability, showing feasibility for practical applications.

Key words: power load forecasting, new energy, source-load uncertainty, deep learning, Transformer model, discrete cosine transform, channel attention mechanism, data enhancement

中图分类号: