面向绿色低碳的高炉布料矩阵智能决策方法

于泽隽; 邢宇昆; 张森; 肖文栋; 李政国

doi:10.16383/j.aas.c260347

面向绿色低碳的高炉布料矩阵智能决策方法

doi: 10.16383/j.aas.c260347 cstr: 32138.14.j.aas.c260347

于泽隽^{1, 2, 3,},
邢宇昆^4,,
张森^{1, 2, 3,},
肖文栋^{1, 2, 3,},
李政国^5,

1.
北京科技大学工业过程知识自动化教育部重点实验室北京 100083 中国
2.
北京科技大学自动化学院北京 100083 中国
3.
北京科技大学顺德创新学院顺德 528399 中国
4.
机械工业仪器仪表综合技术经济研究所北京 100055 中国
5.
新加坡科技研究局新加坡 138632 新加坡

基金项目: 国家自然科学基金(62173032,61903028,62003038), 北京市自然科学基金(J210005), 广东省基础与应用基础研究基金(2022A1515140109)资助

详细信息

作者简介:
于泽隽：北京科技大学自动化学院博士研究生.主要研究方向为复杂系统建模与控制和现代大型高炉布料系统. E-mail: d202310366@xs.ustb.edu.cn

邢宇昆：机械工业仪器仪表综合技术经济研究所助理工程师.主要研究方向为基于工业数据的质量预测及控制、污染物溯源、智能制造评测. E-mail: 15601323202@163.com

张森：北京科技大学自动化学院教授.主要研究方向为复杂系统建模与控制和现代大型高炉布料系统.本文通信作者. E-mail: zhangsen@ustb.edu.cn

肖文栋：北京科技大学自动化学院教授.主要研究方向为无线定位与导航、能量采集及网络资源管理、可穿戴计算与智慧养老、数据挖掘与智能大数据处理、物联网与信息化平台建设等. E-mail: wdxiao@ustb.edu.cn

李政国：新加坡科技研究局高级科学家.主要研究方向为视频处理与传输、计算摄影、开关与脉冲控制、传感器融合以及物理驱动深度学习. E-mail: ezgli@i2r.a-star.edu.sg

计量
- 文章访问数: 16
- HTML全文浏览量: 14
- 被引次数: 0
出版历程
- 网络出版日期: 2026-10-16

Intelligent Decision Method of Blast Furnace Charging Matrix for Green and Low-carbon

YU Ze-Jun^{1, 2, 3
,},
XING Yu-Kun^4
,,
ZHANG Sen^{1, 2, 3
,},
XIAO Wen-Dong^{1, 2, 3
,},
LI Zheng-Guo^5
,

1.
Key Laboratory of Knowledge Automation for Industrial Processes of the Ministry of Education, School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing, 100083, China
2.
School of Automation and Electrical Engineering, University of Science and Technology Beijing, Beijing, 100083, China
3.
Shunde Innovation School, University of Science and Technology, Foshan, 528399, China
4.
Instrumentation Technology and Economy Institute, Beijing, 100055, China
5.
Institute for Infocomm Research, Agency for Science, Technology and Research, Singapore, 138632, Singapore

Funds: Supported by National Natural Science Foundation of China (62173032, 61903028, 62003038), Natural Science Foundation of Beijing Municipality, China (J210005), and GuangDong Basic and Applied Basic Research Foundation(2022A1515140109)

More Information

Author Bio:
YU Ze-Jun　Ph.D. candidate at the School of Automation and Electrical Engineering, University of Science and Technology Beijing. His research interest covers complex system modeling and control, as well as modern large-scale blast furnace burden distribution systems

XING Yu-Kun　Assistant Engineer at the Instrumentation Technology and Economy Institute. His research interests include quality prediction and control based on industrial data, pollutant tracing, and intelligent manufacturing evaluation

ZHANG Sen　Professor at the School of Automation and Electrical Engineering, University of Science and Technology Beijing. Her research interests include complex system modeling and control, as well as modern large-scale blast furnace burden distribution systems. Corresponding author of this paper

XIAO Wen-Dong　Professor at the School of Automation and Electrical Engineering, University of Science and Technology Beijing. His research interests include wireless positioning and navigation, energy harvesting and network resource management, wearable computing and smart eldercare, data mining and intelligent big data processing, as well as Internet of Things and information platform development

LI Zheng-Guo　Senior Scientist at the Agency for Science, Technology and Research, Singapore. His research interests include video processing and delivery, computational photography, switched and impulsive control, sensor fusion and physics-driven deep learning

摘要

摘要: 高炉炼铁工序贡献了钢铁生产流程七成以上的碳排放, 而布料操作直接影响煤气利用率与碳排放水平, 是绿色低碳制造的重要切入点. 然而, 面对高炉内部"黑箱"环境, 现有方法难以预测布料调整对炉况的影响, 只能被动响应. 为此, 提出一种基于炉况感知与深度强化学习的高炉布料矩阵决策方法. 首先, 针对炉况粗精度分类不足以支撑布料决策的问题, 提出炉况健康度9级评分体系, 构建基于时间卷积网络(Temporal Convolutional Network, TCN)的炉况时序分类模型. 其次, 针对布料调整对炉况的影响难以预测的问题, 设计基于改进交叉注意力机制(Cross Attention, CA)的炉况多步预测模型, 捕捉布料矩阵与炉况的时序依赖关系. 最后, 将上述分类与预测模型作为状态空间与奖励函数嵌入深度强化学习框架, 以炉况长期稳定顺行为目标, 构建基于双延迟深度确定性策略梯度算法(Twin Delayed Deep Deterministic policy gradient algorithm, TD3)的布料矩阵决策模型. 在高炉实际生产数据集上的验证结果表明, 该方法能够自动调整布料矩阵, 维持炉况稳定顺行, 为高炉的智能化、低碳化运行提供了可落地的方案.
- 布料矩阵决策 /
- 炉况感知 /
- 深度强化学习 /
- 时间卷积网络 /
- 交叉注意力机制 /
- 绿色低碳
Abstract: The blast furnace (BF) ironmaking process contributes more than 70% of the carbon emissions of the entire steel production process, while the charging operation directly affects gas utilization rate and carbon emission level, making it a critical entry point for green and low-carbon manufacturing. However, due to the "black-box" environment of the BF, existing methods fail to predict the influence of charging adjustments on furnace conditions and can only react passively after deterioration. To tackle this problem, a BF charging matrix decision method based on furnace condition perception and deep reinforcement learning is proposed. First, to overcome the limitation of coarse-grained classification that cannot support charging decisions, a 9-level health scoring system for furnace conditions is established, and a temporal convolutional network (TCN) based time-series classification model for furnace conditions is constructed. Second, to address the difficulty of predicting the effect of charging adjustments on furnace conditions, a multi-step prediction model based on an improved cross-attention (CA) mechanism is designed to capture the temporal dependencies between the charging matrix and furnace conditions. Finally, the above classification and prediction models are embedded as the state space and reward function into a deep reinforcement learning framework. A charging matrix decision model based on the twin delayed deep deterministic policy gradient (TD3) algorithm is then built, where the reward objective is to ensure long-term stable and smooth operation. Validation on an actual BF production dataset demonstrates that the proposed method can automatically adjust the charging matrix, maintain stable and smooth furnace conditions, and provide a practical solution for intelligent and low-carbon BF operation.
- charging matrix decision /
- furnace condition perception /
- deep reinforcement learning /
- temporal convolutional network /
- cross attention /
- green and low-carbon

HTML全文

图 1 无钟布料原理图

Fig. 1 Schematic diagram of bell-less burden charging

下载: 全尺寸图片幻灯片

图 2 方法总体架构图

Fig. 2 Overall architecture of the proposed method

下载: 全尺寸图片幻灯片

图 3 TCN结构图

Fig. 3 Structure diagram of TCN

下载: 全尺寸图片幻灯片

图 4 炉况预测模型结构图

Fig. 4 Furnace condition-aware causal cross-attention mechanism

下载: 全尺寸图片幻灯片

图 5 $ \text{TCN}_\mathrm{matrix} $结构图

Fig. 5 Structure of $ \text{TCN}_\mathrm{matrix} $

下载: 全尺寸图片幻灯片

图 6 偏移因果掩码

Fig. 6 Shifted causal mask

下载: 全尺寸图片幻灯片

图 7 布料矩阵决策深度强化学习框架图

Fig. 7 Deep reinforcement learning framework diagram for charging matrix decision

下载: 全尺寸图片幻灯片

图 8 TCN炉况分类模型训练损失曲线

Fig. 8 Training loss curve of TCN condition classification model

下载: 全尺寸图片幻灯片

图 9 TCN炉况分类模型混淆矩阵

Fig. 9 Confusion matrix of TCN condition classification model

下载: 全尺寸图片幻灯片

图 10 TCN-CA炉况预测模型训练结果

Fig. 10 TCN-CA condition prediction model training result

下载: 全尺寸图片幻灯片

图 11 顶压预测结果

Fig. 11 Top pressure prediction

下载: 全尺寸图片幻灯片

图 12 顶温预测对比

Fig. 12 Comparison of top temperature prediction

下载: 全尺寸图片幻灯片

图 13 TD3模型训练结果

Fig. 13 TD3 model training results

下载: 全尺寸图片幻灯片

图 14 TD3与DDPG模型每轮平均回报对比

Fig. 14 Average reward per episode: TD3 vs. DDPG

下载: 全尺寸图片幻灯片

图 15 TD3布料决策模型测试结果

Fig. 15 TD3 charging decision model testing results

下载: 全尺寸图片幻灯片

图 16 不同决策方法下的煤气利用率变化对比

Fig. 16 Comparison of gas utilization rate changes under different decision methods

下载: 全尺寸图片幻灯片

表 1 布料矩阵的调整范围

Table 1 Adjustment range of charging matrix

挡位		1	2	3	4	5	6	7	8	9	10	11
上限	$ O_a $	41	40	39	38	37	36	33.5	31	29	26	23
	$ O_n $	2	3	4	1	1	1	1	1	1	1	1
	$ C_a $	42	41	40	39	38	37	34.5	32	30	27	24
	$ C_n $	2	3	4	1	1	1	1	1	1	1	1
下限	$ O_a $	36	35	34	33	32	31	28.5	26	24	21	18
	$ O_n $	1	2	3	0	0	1	1	1	0	0	0
	$ C_a $	36	35	34	33	32	31	28.5	26	24	21	18
	$ C_n $	1	2	3	0	0	1	1	1	1	0	0

下载: 导出CSV

表 2 炉况状态变量及与布料的关系

Table 2 Furnace condition state variables and their relationship with charging operation

变量	定义与单位	受布料调整的影响	对炉况的指示作用
$ P_{\text{top}} $	炉顶煤气压力($ \mathrm{kPa} $)	布料改变料面形状与煤气通道, 影响$ P_{\text{top}} $分布.	影响煤气流速与料柱稳定性, 稳定顶压利于高产.
$ \Delta P $	炉顶与炉底压力差($ \mathrm{kPa} $)	布料矩阵决定矿焦比径向分布, 改变料柱透气阻力.	反映料柱透气性, 异常压差预示悬料、崩料.
$ K $	风量与压差的比值	布料影响煤气流动路径, 进而改变$ K $.	反映煤气流与料柱的匹配程度.
$ \eta_{\text{CO}} $	煤气中${\mathrm{CO}} _2 $ 占CO与${\mathrm{CO}} _2 $总量的比值	布料影响矿焦接触效率, 改变还原反应程度.	能量效率的核心指标.
$ T_{\text{top}} $	炉顶气体平均温度($ ^\circ C $)	布料改变煤气流分布, 导致$ T_{\text{top}} $变化.	反映炉内热状态, 过高或过低均为异常.

下载: 导出CSV

表 3 炉况健康度9级评分体系

Table 3 9-level health scoring system for furnace conditions

等级		炉况描述
正常	9(优)	炉况极其稳定, 所有关键指标均处于理想范围且保持最优水平.
	8(良)	炉况稳定, 个别指标(如水温差)存在微小波动, 整体控制良好.
	7(中)	炉况较稳定, 煤气流小幅度波动, 水温差波动频繁, 煤气利用率偏低.
	6(差)	炉况稳定性一般, 压差明显偏高、温差波动较大, 控制水平较差.
休风	5	因计划检修、设备维护、炉渣清理或指标突变, 高炉停止冶炼.
异常	4(轻度异常)	指标小范围偏离正常, 需减风、停氧并密切观察.
	3(中度异常)	指标偏离较大, 可能出现炉皮过热、发红, 需立即减风停氧.
	2(高度异常)	指标严重偏离, 温差过高、炉皮破裂, 可能出现崩料.
故障	1	炉况极端异常, 料线停滞、滑料, 须立即调节, 否则引发安全事故.

下载: 导出CSV

表 4 关键超参数的设定及依据

Table 4 Key hyperparameter settings and their selection basis

模型	参数	取值	调参依据
分类模型	隐藏层层数	8	如式(5), 感受野计算公式
	模型时间步	240	对应4小时工作周期
	卷积核大小	2	经验值, 捕捉局部依赖
	空洞因子	2	经验值, 指数扩张, 扩大感受野
	学习率	0.001	验证集调优
	分类数	9	炉况健康度9级评分体系
预测模型	注意力维度 $d_k$	64	经验值^[30]
	FFN输出维度	128	经验值, 平衡表达力与计算量
	学习率	0.000 4	验证集调优
决策模型	最大训练轮次	10 000	保证充分收敛
	每轮步数	10	匹配工艺周期
	学习率	0.001	验证集调优
	折扣因子$ \gamma $	0.99	验证集调优
	软更新系数$ \tau $	0.001	验证集调优
	策略噪声$ \epsilon $	0.2	经验值^[31]

下载: 导出CSV

表 5 超参数敏感性分析结果

Table 5 Hyperparameter sensitivity analysis results

分类/预测模型			决策模型
参数	取值	性能指标	参数	取值	性能指标
学习率	0.0002	0.852	$ \gamma $	0.95	47.21%
	0.0005	0.873		0.98	48.05%
	0.001	0.885		0.99	48.36%
	0.002	0.872		0.995	48.28%
$ d_k $	32	1.68	$ \tau $	0.0005	48.22%
	64	1.42		0.001	48.36%
	128	1.44		0.005	48.29%
	256	1.52		0.01	47.98%

下载: 导出CSV

表 6 炉况分类性能对比

Table 6 Comparison of furnace condition classification performance

模型	准确率	宏平均F1分数	故障召回率	故障精确率
CNN	0.818$ \pm $0.020	0.788$ \pm $0.021	0.895$ \pm $0.018	0.882$ \pm $0.022
LSTM	0.842$ \pm $0.015	0.821$ \pm $0.017	0.948$ \pm $0.012	0.937$ \pm $0.016
ELM	0.861$ \pm $0.010	0.844$ \pm $0.011	0.962$ \pm $0.008	0.954$ \pm $0.012
TCN	0.885$ \pm $0.008	0.868$ \pm $0.009	1$ \pm $0	0.982$ \pm $0.006

下载: 导出CSV

表 7 模型训练与推理时间对比

Table 7 Comparison of training and inference time

模型	训练时间(min)	单样本推理时间(ms)
ELM	2.3 $ \pm $ 0.5	7.0 $ \pm $ 0.4
CNN	26.1 $ \pm $ 3.8	9.6 $ \pm $ 0.6
TCN	63.4 $ \pm $ 5.2	8.7 $ \pm $ 0.5
LSTM	172.5 $ \pm $ 15.3	21.5 $ \pm $ 1.2

下载: 导出CSV

表 8 炉况预测误差($ 10^{-2} $)对比

Table 8 Comparison of prediction errors($ 10^{-2} $) for furnace condition

参数	模型	MAPE	RMSE
$ P_{\text{top}} $	TCN-CA	1.419$ \pm $0.085	2.757$ \pm $0.165
	TCN	1.865$ \pm $0.130	3.343$ \pm $0.234
	TCN(无布料矩阵)	2.018$ \pm $0.151	3.758$ \pm $0.282
	LSTM	2.195$ \pm $0.176	4.008$ \pm $0.320
	ELM	2.568$ \pm $0.205	4.731$ \pm $0.378
$ \Delta P $	TCN-CA	1.880$ \pm $0.113	3.069$ \pm $0.184
	TCN	2.403$ \pm $0.168	3.797$ \pm $0.266
	TCN(无布料矩阵)	2.621$ \pm $0.183	4.156$ \pm $0.291
	LSTM	2.844$ \pm $0.199	4.526$ \pm $0.317
	ELM	3.318$ \pm $0.232	5.374$ \pm $0.376
$ K $	TCN-CA	1.670$ \pm $0.100	2.908$ \pm $0.175
	TCN	2.142$ \pm $0.150	3.545$ \pm $0.248
	TCN(无布料矩阵)	2.312$ \pm $0.162	3.894$ \pm $0.273
	LSTM	2.524$ \pm $0.177	4.265$ \pm $0.299
	ELM	2.950$ \pm $0.207	5.058$ \pm $0.354
$ \eta_{\text{CO}} $	TCN-CA	1.534$ \pm $0.092	3.994$ \pm $0.240
	TCN	2.126$ \pm $0.149	4.958$ \pm $0.347
	TCN(无布料矩阵)	2.310$ \pm $0.162	5.436$ \pm $0.381
	LSTM	2.492$ \pm $0.174	5.915$ \pm $0.414
	ELM	2.922$ \pm $0.204	7.012$ \pm $0.491
$ T_{\text{top}} $	TCN-CA	7.134$ \pm $0.428	9.174$ \pm $0.550
	TCN	9.451$ \pm $0.567	11.512$ \pm $0.691
	TCN(无布料矩阵)	10.284$ \pm $0.617	12.593$ \pm $0.756
	LSTM	11.117$ \pm $0.667	13.718$ \pm $0.823
	ELM	13.010$ \pm $0.781	16.270$ \pm $0.976

下载: 导出CSV

表 9 模型训练与预测耗时对比(240步)

Table 9 Comparison of training time and prediction time (240 steps)

模型	训练时间(min)	预测耗时(ms)
ELM	3.8 $ \pm $ 0.6	7.0 $ \pm $ 0.4
TCN(无布料矩阵)	51.2 $ \pm $ 4.5	12.0 $ \pm $ 0.6
TCN	74.6 $ \pm $ 6.1	14.2 $ \pm $ 0.7
LSTM	215.3 $ \pm $ 18.7	28.6 $ \pm $ 1.4
TCN-CA	182.6 $ \pm $ 15.8	22.5 $ \pm $ 1.1

下载: 导出CSV

表 10 动作网络输出的布料矩阵时序表

Table 10 Sequence table of the charging matrix output by Actor

时间	布料矩阵
时间	挡位	1	2	3	4	5	6	7	8	9	10	11
0:00	$ O_a $	41	40	39	38	37	36	33.5	31	29	21	18
	$ O_n $	2	2	3	0	0	1	1	1	0	0	0
	$ C_a $	42	41	40	39	38	37	34.5	32	30	25	24
	$ C_n $	1	2	3	0	0	1	1	1	1	0	0
$\dots$
4:00	$ O_a $	41	40	39	38	37	36	33.5	31	24	21	18
	$ O_n $	2	2	3	0	0	1	1	1	1	0	0
	$ C_a $	42	41	40	39	38	37	34.5	32	30	25	24
	$ C_n $	1	2	3	0	0	1	1	1	1	0	0
$\dots$
8:00	$ O_a $	41	40	39	38	37	36	33.5	31	29	21	18
	$ O_n $	2	2	3	0	0	1	1	1	0	0	0
	$ C_a $	42	41	40	39	38	37	34.5	32	30	25	24
	$ C_n $	1	2	3	0	0	1	1	1	1	0	0
$\dots$

下载: 导出CSV

表 11 测试集平均煤气利用率对比

Table 11 Comparison of average gas utilization rate on the test set

决策方法	平均煤气利用率(%)	相对提升
DDPG	46.21$ \pm $0.28	—
TD3	48.36$ \pm $0.31	+4.65

下载: 导出CSV

参考文献(35)

[1]	李壮年, 柳政根, 储满生, 陈树文. 大型高炉布料模型的研究与应用. 东北大学学报(自然科学版), 2016, 37(5): 658−662 doi: 10.3969/j.issn.1005-3026.2016.05.011 Li Zhuang-Nian, Liu Zheng-Gen, Chu Man-Sheng, Chen Shu-Wen. Research and application of burden charging model for large blast furnace. Journal of Northeastern University (Natural Science), 2016, 37(5): 658−662 doi: 10.3969/j.issn.1005-3026.2016.05.011
[2]	Chen J S, Zuo H B, Xue Q G, Wang J S. A review of burden distribution models of blast furnace. Powder Technology, 2022Article No. 117055
[3]	Yan S, Shen Y S. Model study of central coke charging on ironmaking blast furnace performance. Powder Technology, 2020, 361: 124−135 doi: 10.1016/j.powtec.2019.10.012
[4]	Zong Y C, Hu S W, Qin D B, Wang Z X, Zhang C J, Chu J H. Iron-tapping state recognition of blast furnace based on Bi-GRU composite model and post-processing classifier. IEEE Sensors Journal, 2023, 23(18): 22006−22018 doi: 10.1109/JSEN.2023.3300123
[5]	Zhang Z, Tang J, Shi Q, Chu M S, Wang M Y, Zhang Z F. Comprehensive status evaluation and prediction of blast furnace based on cascade system and combined model. International Journal of Minerals, Metallurgy and Materials, 2025, 32(12): 2942−2957 doi: 10.1007/s12613-025-3179-6
[6]	Zhu X Z, Yang C J, Lou S W, Yang Y L. Knowledge-Based Temporal GCN: A spatial–temporal fault diagnosis method for blast furnace ironmaking process with imbalanced data. IEEE Transactions on Instrumentation and Measurement, 2025, 74: 1−12 doi: 10.1109/tim.2024.3522663
[7]	Wu P, Wang Y X, Gao J F, Zhang X J, Lou S W, Yang C J. Multichannel dynamic graph convolutional network-based fault diagnosis and its application in blast furnace ironmaking process. IEEE Sensors Journal, 2023, 23(23): 29293−29302 doi: 10.1109/JSEN.2023.3325353
[8]	Xu S Y, Jiang Z H, Chen Z W, Pan D, Yu H Y, Li L L. Blast furnace condition recognizing in the ironmaking process based on prior knowledge and Platt scaling probability. In: Proceedings of the 2024 IEEE International Conference on Industrial Technology, Bristol, United Kingdom: IEEE, 2024. 1-6.
[9]	Guo K, Zhang Y X, Zhang S, Xiao W D. Classification model for blast furnace status based on multi-source information. Engineering Applications of Artificial Intelligence, 2025Article No. 109728
[10]	Liu X J, Li T S, Li H W, Zhang Y J, Li X, Yi F Y. Predictions of blast furnace permeability index based on SVM-RFE+WOA-BiLSTM model. In: Proceedings of the 2025 40th Youth Academic Annual Conference of Chinese Association of Automation, Zhengzhou, China: IEEE, 2025. 1023-1028.
[11]	Chen Q F, Li Z, Li W J, Guo Y P, An J Q, She J H. Multi-step prediction model of blast furnace permeability index based on multi-time scale. ISA Transactions, DOI: 10.1016/j.isatra.2025.11.036.
[12]	刘仕鑫, 尹怡欣, 张森. 基于核超限学习机的高炉透气性指数预测模型. 控制理论与应用, 2023, 40(1): 65−73 Liu Shi-Xin, Yin Yi-Xin, Zhang Sen. Prediction model of kernel extreme learning machine for permeability index of blast furnace. Control Theory & Applications, 2023, 40(1): 65−73
[13]	蒋朝辉, 许川, 桂卫华, 蒋珂. 基于最优工况迁移的高炉铁水硅含量预测方法. 自动化学报, 2022, 48(1): 194−206 doi: 10.16383/j.aas.c200980 Jiang Zhao-Hui, Xu Chuan, Gui Wei-Hua, Jiang Ke. Prediction method of hot metal silicon content in blast furnace based on optimal smelting condition migration. Acta Automatica Sinica, 2022, 48(1): 194−206 doi: 10.16383/j.aas.c200980
[14]	蒋珂, 蒋朝辉, 谢永芳, 潘冬, 桂卫华. 基于动态注意力深度迁移网络的高炉铁水硅含量在线预测方法. 自动化学报, 2023, 49(5): 949−963 doi: 10.16383/j.aas.c210524 Jiang Ke, Jiang Zhao-Hui, Xie Yong-Fang, Pan Dong, Gui Wei-Hua. Online prediction method of blast furnace hot metal silicon content based on dynamic attention deep transfer network. Acta Automatica Sinica, 2023, 49(5): 949−963 doi: 10.16383/j.aas.c210524
[15]	Li Z, An J Q, Guo Y P, Zhao G Y, Wu M. Prediction of gas utilization ratio based on singular spectrum analysis, improved PSO and KELM. In: Proceedings of the 2024 36th Chinese Control and Decision Conference. Xi'an, China: IEEE, 2024. 345-350.
[16]	Zhao G Y, An J Q, Guo Y P, Yang C L, He Y. A multi-time-scale prediction method for blast furnace gas utilization rate based on causality analysis. In: Proceedings of the 2023 China Automation Congress. Chongqing, China: IEEE, 2023. 8475-8480.
[17]	Zhang T K, Liu C B, Liu Z J, Tan J, Ahmat M. Temporal double graph convolutional network for CO and CO2 prediction in blast furnace gas. IEEE Transactions on Instrumentation and Measurement, 2025, 73: 1−13 doi: 10.1109/tim.2023.3341110
[18]	李艳姣, 张森, 尹怡欣, 张杰. 基于数据驱动的高炉料面优化决策模型研究. 控制理论与应用, 2018, 35(3): 324−334 doi: 10.7641/CTA.2017.70302 Li Yan-Jiao, Zhang Sen, Yin Yi-Xin, Zhang Jie. Research on optimization model of blast furnace burden surface based on data driven. Control Theory & Applications, 2018, 35(3): 324−334 doi: 10.7641/CTA.2017.70302
[19]	Li Y J, Zhang S, Zhang J, Yin Y X, Xiao W D, Zhang Z Q. Data-driven multiobjective optimization for burden surface in blast furnace with feedback compensation. IEEE Transactions on Industrial Informatics, 2020, 16(4): 2233−2244 doi: 10.1109/TII.2019.2908989
[20]	Huang X K, Yang C J, Zhang H W, Lou S W, Gao D L, Kong L Y. Data and knowledge collaborative-driven fault identification and self-healing control action inference framework for blast furnace. Expert Systems with Applications, 2024, 245: 123040 doi: 10.1016/j.eswa.2023.123040
[21]	安剑奇, 赵国宇, 何勇, 李炜俊, 郭云鹏, 吴敏. 数据驱动的多时间尺度高炉煤气利用率模型预测控制. 控制理论与应用, 2025, 42(1): 189−201 doi: 10.7641/CTA.2024.30318 An Jian-Qi, Zhao Guo-Yu, He Yong, Li Wei-Jun, Guo Yun-Peng, Wu Min. Data-driven multi-time scale model predictive control for blast furnace gas utilization rate. Control Theory & Applications, 2025, 42(1): 189−201 doi: 10.7641/CTA.2024.30318
[22]	Zhu J C, Jiang Z H, Pan D, Yu H Y, Zhou K, Gui W H. Burden surface shape modeling and charging matrix optimization for the blast furnace charging process. IEEE Transactions on Industrial Informatics, 2024, 20(11): 12705−12716
[23]	Zhu J C, Jiang Z H, Pan D, Yu H Y, Xu C, Zhou K, et al. An intelligent optimization strategy for blast furnace charging operation considering three-dimensional burden surface shape. IEEE/CAA Journal of Automatica Sinica, 2025, 12(7): 1445−1463 doi: 10.1109/JAS.2025.125192
[24]	Jiang K, Jiang Z H, Jiang X D, Xie Y F, Gui W H. Reinforcement learning for blast furnace ironmaking operation with safety and partial observation considerations. IEEE Transactions on Neural Networks and Learning Systems, 2024, 35(3): 3077−3090 doi: 10.1109/TNNLS.2023.3340741
[25]	Amorim L L, Mutz F, Paixão T M, Rampinelli V, De Souza A F, Badue C, et al. Blast furnace control by hierarchical similarity search in historical data. Engineering Applications of Artificial Intelligence, DOI: 10.1016/j.engappai.2024.109779.
[26]	Zhang Y X, Guo K, Yu Z J, Zhang S, Yang Y L, Xiao W D, et al. Soft sensor-driven spatiotemporal-periodic synergistic predictive control for blast furnace gas flow. ISA Transactions, DOI: 10.1016/j.isatra.2025.12.053.
[27]	Huang Y F, Lai X Z, Zhang K X, An J Q, Chen L F, Wu M. Two-stage decision-making method for burden distribution based on recognition of conditions in blast furnace. IEEE Transactions on Industrial Electronics, 2021, 68(5): 4199−4208 doi: 10.1109/TIE.2020.2982121
[28]	Shi L, Zhao G S, Li M X, Ma X. A model for burden distribution and gas flow distribution of bell-less top blast furnace with parallel hoppers. Applied Mathematical Modelling, 2016, 40(23-24): 10254−10273 doi: 10.1016/j.apm.2016.07.024
[29]	Bai S, Kolter J Z, Koltun V. An empirical evaluation of generic convolutional and recurrent networks for sequence modeling. arXiv: 1803.01271, 2018.
[30]	Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez A N, et al. Attention is all you need. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, USA: Curran Associates, 2017. 5998-6008.
[31]	Fujimoto S, Hoof H, Meger D. Addressing function approximation error in actor-critic methods. In: Proceedings of the 35th International Conference on Machine Learning. Stockholm, Sweden: PMLR, 2018. 1587-1596.
[32]	Lecun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 1998, 86(11): 2278−2324 doi: 10.1109/5.726791
[33]	Hochreiter S, Schmidhuber J. Long short-term memory. Neural Computation, 1997, 9(8): 1735−1780 doi: 10.1162/neco.1997.9.8.1735
[34]	Huang G B, Zhu Q Y, Siew C K. Extreme learning machine: theory and applications. Neurocomputing, 2006, 70(1−3): 489−501 doi: 10.1016/j.neucom.2005.12.126
[35]	Lillicrap T P, Hunt J J, Pritzel A, Heess N, Erez T, Tassa Y, et al. Continuous control with deep reinforcement learning. arXiv: 1509.02971, 2015.