| 引用本文: | 金锋,陈薇琳,赵博识,赵珺,王伟.基于知识蒸馏的钢铁高炉煤气系统建模方法[J].控制理论与应用,2024,41(3):428~435.[点击复制] |
| JIN Feng,CHEN Wei-lin,ZHAO Bo-shi,ZHAO Jun,WANG Wei.A knowledge distillation-based modeling method for blast furnace gas system in steel industry[J].Control Theory and Technology,2024,41(3):428~435.[点击复制] |
|
| 基于知识蒸馏的钢铁高炉煤气系统建模方法 |
| A knowledge distillation-based modeling method for blast furnace gas system in steel industry |
| 摘要点击 2447 全文点击 78 投稿时间：2022-09-30 修订日期：2024-01-18 |
| 查看全文 查看/发表评论 下载PDF阅读器 |
| DOI编号 10.7641/CTA.2023.20864 |
| 2024,41(3):428-435 |
| 中文关键词 知识蒸馏 时间序列 高炉煤气系统 钢铁企业 |
| 英文关键词 knowledge distillation time series blast furnace gas system steel industry |
| 基金项目 国家重点研发计划项目(2017YFA0700300), 国家自然科学基金项目(62125302, 61833003, U1908218, 62103075), 辽宁省“兴辽英才计划”项目 (XLYC2002087), 大连市科技人才创新支持计划项目(2022RG03)资助. |
|
| 中文摘要 |
| 钢铁企业高炉煤气系统具有波动性大、时变性强、不确定性高等特点, 对其未来产消趋势进行准确的建模
预测有助于企业的高效决策与节能减排. 文章提出了一种基于知识蒸馏的高炉煤气系统建模方法, 为了提高训练
过程中的拟合精度, 在教师网络中建立了基于长短期记忆网络的序列到序列模型来提取样本的中间特征. 进而, 提
出了融入教师模型中间特征的知识蒸馏策略, 建立了考虑中间特征蒸馏损失与预测均方误差的损失函数, 对知识蒸
馏过程及预测偏差进行评估. 通过国内大型钢铁企业高炉煤气系统实际运行数据的实验验证, 表明了本文所提建
模方法的有效性, 可为后续的能源系统优化调度提供支撑. |
| 英文摘要 |
| Blast furnace gas system in steel enterprises has the characteristics of high volatility, time-variability and great
uncertainty, accurately modeling of its future generation and consumption flow plays a crucial role in efficiently decisionmaking,
energy-saving and emissions reduction. In this study, a knowledge distillation-based modelling method for blast
furnace gas system is proposed. Based on a long and short-term memory network, a sequence-to-sequence model is built
in the teacher network to extract the intermediate features of the samples. And then, a knowledge distillation strategy is
constructed which incorporates the intermediate features of the teacher model. Besides, in order to evaluate the capability
of feature extraction, a new loss function is established by both considering that of the knowledge distillation process and
the regression error of the actual energy data. Validation experiments are carried out by employing real-world data from
the blast furnace gas system of a typical steel enterprise, and the results indicate the effectiveness of the proposed method
when facing with the modeling problem, so as to provide powerful support for the optimal scheduling of the energy system. |
|
|
|
|
|