【摘要】 提出一种基于异常检测、特征选择和机器学习算法的电弧炉炼钢终点温度预测的数据挖掘策略，并通过包含31-1输入-输出对的1 235炉实际生产数据验证该策略的有效性。采用逻辑回归、k近邻、决策树、极端梯度增强这4种监督学习算法开发预测模型，综合考察数据量、无监督学习异常检测算法、特征选择方法和监督学习建模算法对预测精度的影响。研究结果表明：极端梯度增强算法的预测效果优于其他3种，在±5、±10和±15℃的温度误差范围内，极端梯度增强算法的命中率分别超过40%、80%和95%；数据质量和数据量都会影响监督学习预测算法模型的预测效果，无监督学习算法自编码器能大幅度提高数据质量和预测精度，排列重要度特征选择可以简化模型结构，需要从数据、特征和算法等多个角度对预测问题进行全面分析。更多还原

【Abstract】 A data mining strategy for accurately predicting the end-point temperature of molten steel(EPT-MS) in electric arc furnace(EAF) steelmaking based on anomaly detection, feature selection and machine learning algorithm was proposed. Real EAF steelmaking process data of 1 235 heats with 31-1 input-output pair were used to verify the effectiveness of this strategy. Four supervised learning prediction algorithms including logistic regression(LR), k-nearest neighbors(kNN), decision tree(DT) and extreme gradient boosting(XGBoost) were used for prediction model development. The prediction performance was evaluated considering influencing factors including the data volume, unsupervised learning anomaly detection algorithms, feature selection methods, and supervised learning algorithms. The results show that XGBoost outperforms the other three algorithms and shows hit rates over 40%, 80% and 95% within temperature error bounds of ±5, ±10 and ±15 ℃. Both data quality and volumes affect the supervised learning prediction algorithm models. The unsupervised anomaly deletion algorithm AE enhances data quality affecting prediction accuracy greatly. The permutation importance feature selection can simplify the model structure. The EPT-MS prediction problem should be analyzed more comprehensively from multiple perspectives of data, features and algorithms.更多还原