上海交通大学学报 ›› 2023, Vol. 57 ›› Issue (3): 316-325.doi: 10.16183/j.cnki.jsjtu.2021.346

所属专题: 《上海交通大学学报》2023年“机械与动力工程”专题

• 机械与动力工程 • 上一篇    下一篇

基于注意力机制与神经网络的热电联产锅炉负荷预测

万安平1, 杨洁1,2, 缪徐1, 陈挺1(), 左强1, 李客2,3   

  1. 1.浙大城市学院 机电系,杭州 310015
    2.浙江大学 机械工程学院,杭州 310027
    3.中信重工机械股份有限公司,河南 洛阳 471039
  • 收稿日期:2021-09-08 接受日期:2021-11-10 出版日期:2023-03-28 发布日期:2023-03-30
  • 通讯作者: 陈 挺(1989-),博士研究生,博士后;E-mail:chenting@zucc.edu.cn.
  • 作者简介:万安平(1983-),博士后,副教授,主要从事热电联产机组负荷预测与运行优化研究.
  • 基金资助:
    国家自然科学基金项目(51705455);浙江省教育厅教师专业发展项目(FX2021111);广东省海洋经济发展(海洋六大产业)专项资金项目(GDNRC[2022]28);杭州市农业与社会发展科研计划重点项目(20212013B04);浙江省基础公益研究计划项目(LGG20E050007)

Boiler Load Forecasting of CHP Plant Based on Attention Mechanism and Deep Neural Network

WAN Anping1, YANG Jie1,2, MIAO Xu1, CHEN Ting1(), ZUO Qiang1, LI Ke2,3   

  1. 1. Department of Mechanical Engineering, Zhejiang University City College, Hangzhou 310015, China
    2. School of Mechanical Engineering, Zhejiang University, Hangzhou 310027, China
    3. CITIC Heavy Industries Co., Ltd., Luoyang 471039, Henan, China
  • Received:2021-09-08 Accepted:2021-11-10 Online:2023-03-28 Published:2023-03-30

摘要:

热电联产机组的锅炉负荷准确预测对电厂生产管理及调度有直接作用.基于注意力机制和深度卷积-长短期记忆网络原理,提出一种新的热电联产长期负荷预测模型,该模型以锅炉出口蒸汽流量(负荷)历史数据和多维负荷影响因素为输入,对负荷进行长期预测.利用Pearson相关系数判定对原始数据进行筛选;将处理后的数据经卷积层进行特征提取和进一步降维,通过长短期记忆层进行拟合,并采取注意力机制对权值进行优化,实现对负荷的精准预测.以浙江桐乡电厂实测数据为例进行验证,结果表明所提方法的平均绝对百分比误差小于1%,能够实现锅炉负荷的精准预测,智能算法在热电联产领域的应用具有一定的借鉴意义.

关键词: 热电联产, 注意力机制, 卷积神经网络, 长短期记忆, 负荷预测

Abstract:

Accurate boiler load forecasting of cogeneration units plays a direct role in production management and dispatching of power plants. A long-term load forecasting model of combined heat and power (CHP) based on attention mechanism and the deep convolution long-short-term memory network (CNN-LSTM-AM) is proposed, which takes the historical data of boiler outlet steam flow (load) and multi-dimensional load influence factors as input to make long-term load forecasting. First, the original data is screened by Pearson correlation coefficient judgment. Then the processed data is processed by convolution layer for feature extraction and further dimensionality reduction, fitted through long-term and short-term memory layer, and optimized the weight by adopting attention mechanism, so as to achieve accurate load forecasting. The proposed model is verified by the measured data of Tongxiang Power Plant in Zhejiang Province. The results show that the MAPE of the proposed method is less than 1%. It can realize the accurate prediction of boiler load, which has a certain reference significance for the application of intelligent algorithm in the field of combined heat and power.

Key words: combined heat and power (CHP), attention mechanism (AM), convolution neural network (CNN), long-short term memory (LSTM), load forecasting

中图分类号: