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

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.

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.