作者投稿和查稿 主编审稿 专家审稿 编委审稿 远程编辑

计算机工程 ›› 2024, Vol. 50 ›› Issue (9): 169-178. doi: 10.19678/j.issn.1000-3428.0068641

• 人工智能与模式识别 • 上一篇    下一篇

基于MTS-BiGRU-DMHSA的工业负荷预测方法

王汝英1, 马嘉骏2, 董建强1, 刘万龙1, 张海涛1, 尹凯3, 赵博超2,*()   

  1. 1. 天津市普迅电力信息技术有限公司, 天津 300300
    2. 天津大学电气自动化与信息工程学院智能电网教育部重点实验室, 天津 300072
    3. 天津求实智源科技有限公司, 天津 300384
  • 收稿日期:2023-10-19 出版日期:2024-09-15 发布日期:2024-09-04
  • 通讯作者: 赵博超
  • 基金资助:
    国家自然科学基金青年科学基金(52307133); 国家自然科学基金联合基金(U2066207); 国网天津市电力公司科技项目(KJ22-2-04)

Industrial Load Forecasting Method Based on MTS-BiGRU-DMHSA

WANG Ruying1, MA Jiajun2, DONG Jianqiang1, LIU Wanlong1, ZHANG Haitao1, YIN Kai3, ZHAO Bochao2,*()   

  1. 1. Tianjin Richsoft Electric Power Information Technology Co., Ltd., Tianjin 300300, China
    2. Key Laboratory of Smart Grid, Ministry of Education, School of Electrical and Information Engineering, Tianjin University, Tianjin 300072, China
    3. Tianjin TransEnergy Technologies Co., Ltd., Tianjin 300384, China
  • Received:2023-10-19 Online:2024-09-15 Published:2024-09-04
  • Contact: ZHAO Bochao

摘要:

工业用电占我国全社会用电量比重大, 通过工业负荷预测了解负荷趋势和用电量信息, 有助于电网安全稳定运行, 为电力部门发电规划提供依据, 且有助于工业用户优化生产工艺和降低成本。为了兼顾工业负荷波动的不确定性以及工业用户用电行为的规律性特征, 提出一种基于多时间尺度(MTS)特征的工业负荷预测方法MTS-BiGRU-DMHSA, 利用MTS特征融合挖掘工业负荷的周期趋势特征和局部波动特征, 提升工业负荷表征的可解释性。此外, 双层多头自注意力(DMHSA)机制利用注意力权重聚焦重要特征, 在挖掘输入特征关联性的同时捕捉时序关联性, 强化重要特征变量与关键时间步的信息表达。在中国某工业企业五面受总柜实采数据上完成实验验证, 采用2种评价指标对所提方法及5种基于神经网络的预测方法进行对比分析。实验结果表明, 所提方法相较于对比方法平均误差降低逾20%, 其中4.67%得益于对MTS特征的运用。通过对比各方法计算效率证实了所提方法的综合性能优势, 可视化实验结果与对比分析也验证了其在工业负荷预测任务上的可行性。

关键词: 工业负荷预测, 神经网络, 多时间尺度特征, 注意力机制, 时间序列分析

Abstract:

Industrial electricity consumption constitutes a significant portion of total electricity usage in China. Understanding load trends and electricity consumption data through industrial load forecasting is essential for ensuring the safe and stable operation of power grids, guiding power generation planning, and aiding industrial users in optimizing production processes and reducing costs. To address the uncertainty in industrial load fluctuations while considering electricity consumption behavior of regular industrial users, a Multi-Time Scale (MTS) feature-based industrial load forecasting method, MTS Bidirectional Gated Recurrent Unit Dual-layer Multi-Head Self-Attention (MTS-BiGRU-DMHSA), is proposed. The MTS feature fusion approach mines periodic trends and local fluctuations in industrial loads, enhancing the interpretability of load characterization. Additionally, the DMHSA mechanism utilizes attention weights to highlight important features, capturing temporal correlations through input feature correlation mining, and improving the expression of significant feature variables and key time steps. This study carries out experimental validation using data from the five-sided receiving cabinet of an industrial enterprise in China. The proposed method is compared against two evaluation indicators and five neural network-based prediction methods. The comprehensive performance advantages of the proposed method are confirmed through comparisons of computational efficiency. Results demonstrate a reduction in the error index by an average of more than 20%, with a 4.67% improvement attributed to the use of MTS features. The feasibility of applying this method to industrial load forecasting tasks is validated through visualized results and comparative analysis.

Key words: industrial load forecasting, neural network, Multi-Time Scale(MTS) feature, attention mechanism, time series analysis