张淑芬, 李雨欣, 屈昌盛,等.改进GRU的钢铁生产能耗特征自适应提取模型[J].河南理工大学学报（自然科学版）,2025,44(4):1-10.

ZHANG S F, LI Y X, QU C S, et al. Improved GRU-based adaptive extraction model for energy consumption characteristics in steel production [J]. Journal of Henan Polytechnic University(Natural Science) , 2025, 44(4): 1-10.

改进GRU的钢铁生产能耗特征自适应提取模型

张淑芬1,2,3, 李雨欣1,2, 屈昌盛1,2, 谷铮1,2, 高瑞1,2

1.华北理工大学 理学院，河北 唐山  063210；2.河北省数据科学与应用重点实验室，河北 唐山  063210；3.唐山市数据科学重点实验室，河北 唐山 063210

摘要: 目的 针对传统模型在钢铁生产工序中能耗特征提取不充分的问题， 方法 提出一种自适应特征提取模型AGRU-Attention。该模型通过引入自适应机制，实现对输入特征权重的动态调整和对GRU（gated recurrent unit，GRU）模型的改进。首先，利用自适应门控单元（adaptive gate unit，AGU）动态调整特征权重，使模型更准确地关注到能耗预测特征因子；其次，通过门控循环单元层充分提取调整后的能耗特征；最后，利用注意力机制对提取出的特征进行加权处理，并将加权后的特征输入到全连接层进行预测输出。为验证所提模型的自适应能力，使用不同来源和数据量的2个数据集，将本文模型与线性回归、支持向量机、CNN、LSTM模型和文献［19］提出的序列到序列GRU模型进行对比。 结果 结果表明，AGRU-Attention模型在2个数据集中的预测精度明显优于其他模型。相较于AGRU模型，AGRU-Attention模型在数据集1上的MSE，RMSE，MAE分别降低了99.99%，99.71%，99.67%；在数据集2上的MSE，RMSE，MAE分别降低了98.64%，88.36%，91.27%。这表明AGRU-Attention模型在生产能耗预测中准确性更高。   结论 所提模型不仅实现了对输入特征权重的自适应调整，准确提取了不同数据集特征，而且通过注意力机制对特征进行加权，显著提升了模型预测的准确性。 

关键词:自适应特征提取;注意力机制;钢铁生产;能耗数据

doi:10.16186/j.cnki.1673-9787.2024070025

基金项目:国家自然科学基金资助项目（U20A20179）

收稿日期:2024/07/05

修回日期:2024/09/30

出版日期:2025/06/19

Improved GRU-based adaptive extraction model for energy consumption characteristics in steel production

Zhang Shufen1,2,3, Li Yuxin1,2, Qu Changsheng1,2, Gu Zheng1,2, Gao Rui1,2

1.School of Science， North China University of Science and Technology， Tangshan  063210， Hebei， China；2.Hebei Provincial Key Laboratory of Data Science and Application， Tangshan  063210， Hebei， China；3.Tangshan Key Laboratory of Data Science， Tangshan  063210， Hebei， China

Abstract: Objectives To address the insufficient energy consumption feature extraction in traditional models for different steel production processes，   Methods An Adaptive Feature Extraction Model （AGRU-Attention） was proposed. The model improved the GRU structure by introducing an adaptive mechanism that dynamically adjust the weights of input features. First， the Adaptive Gating Unit （AGU） was used to dynamically adjust the weights of the features， allowing the model to more accurately focus on the features crucial for energy consumption prediction. Secondly， the adjusted features were fully extracted through the GRU layer. Finally， the extracted features were weighted using the attention mechanism and the weighted features were input into the fully-connected layer for the predictive output. To verify the adaptive ability of the proposed model， two datasets with different sources and data volumes were compared with linear regression， support vector machines， CNN， LSTM and the sequence-to-sequence GRU model proposed in the literature ［19］.  Results The results showed that the prediction accuracy of the AGRU-Attention model was significantly better than the other models in the two different datasets. Compared with the GRU model， the AGRU-Attention model  achieves significant reductions of 99.99% in MSE， 99.71% in RMSE， and 99.67% in MAE on Dataset 1. On Dataset 2， the proposed model demonstrates reductions of 98.64% in MSE， 88.36% in RMSE， and 91.27% in MAE，respectively. It verified that the model had a higher practical application value in predicting production energy consumption.   Conclusions The proposed model not only realized adaptive adjustment of input feature weights to accurately extract features from different datasets， but also weighted the features through the attention mechanism，significantly improved the accuracy of the model’s predictions. 

Key words: adaptive feature extraction; attention mechanism; steel production; energy consumption data