Abstract: Carbon content and temperature accurate prediction is the key to control endpoint of converter steelmaking.Due to the difference of raw material quality in actual production,the global single model can not accurately predict endpoint carbon content and temperature with large fluctuation of furnace number sample.For this reason,a Local Nearest Neighbour Density Peak Clustering (LNN-DPC) weighted ensemble learning soft measurement method was proposed.The improved peak density clustering algorithm was applied to classify training data after dimensionality reduction to form local sample subset,then one-to-one correspondence between subset and original data was constructed to generate Gaussian process regression sub-model,and the entropy-weighted subset “centroid” was obtained by measuring under original data subset.The model with strong correlation degree of test samples was selected as local model by gray correlation analysis,and weighted ensemble strategy of correlation degree was proposed to output carbon content and temperature prediction results.Simulation results of actual converter steelmaking production process data showed that the prediction accuracy of carbon content reached 85.2% within ±0.02 % error range,temperature reached 84.8% within ±10℃ error range.

Key words: converter steelmaking, ensemble learning, t-distributed stochastic neighbor embedding, local nearest neighbour density peak clustering, gray correlation analysis, Gaussian process regression

摘要： 转炉炼钢终点控制的关键是碳温准确预报。针对实际生产中因原料品质差异导致的炉次样本波动性较大所造成全局单一模型无法精确预测终点碳温的问题,提出一种局部最近邻密度峰值聚类算法(LNN-DPC)加权集成学习软测量方法。首先,采用改进的峰值密度聚类算法划分降维后的训练数据形成局部样本子集,构建子集与原始数据间的一一对应关系生成高斯过程回归子模型,并在原始数据子集下度量得到熵值加权的子集“质心”;其次,通过灰色关联分析选择与测试样本关联度较强的模型作为局部模型,提出关联度加权集成策略输出碳温预测结果。在实际转炉炼钢生产过程数据仿真结果下,碳含量在±0.02%的误差范围内精度达到85.2%,温度在±10℃的误差范围内精度达到84.8%。

关键词: 转炉炼钢, 集成学习, t-分布随机邻域嵌入算法, 局部最近邻密度峰值聚类算法, 灰色关联分析, 高斯过程回归