基于Stacking模型融合的AOD终点碳温预测

为提高AOD不锈钢冶炼终点碳温预测的准确性和可靠性,提出一种基于多个机器学习算法(RF、XGBoost、AdaBoost、KNN、SVR和岭回归)嵌入的Stacking模型融合的AOD终点碳温预测方法。基于理论基础和相关性分析确定了模型的输入特征变量,利用箱线图法对历史数据进行预处理,结合5折交叉验证和贝叶斯优化算法确定了模型的最优参数,通过对上述6种机器学习算法的逐层筛选,构建了RF+XGBoost+KNN—RF二层Stacking多模型融合的终点碳含量预测模型、RF+AdaBoost+KNN—RF+XGBoost+KNN—XGBoost三层Stacking多模型融合的终点温度预测模型。预测结果表明,RF+XGBoost+KNN—RF二层Stacking模型在终点碳质量分数误差为±0.01%的命中率为87.86%,RF+AdaBoost+KNN—RF+XGBoost+KNN—XGBoost三层Stacking模型在终点温度误差为±15℃的命中率为94.22%,相较于单一的机器学习模型,预测精度显著提高。

基于改进遗传算法的转炉炼钢过程数据特征选择

转炉炼钢生产过程数据特征选择是实现终点碳温预报的关键,针对生产过程高维数据不利于快速精确预测终点碳温的问题,提出一种改进遗传算法的转炉炼钢生产过程数据特征选择方法。首先采用皮尔逊相关系数衡量不同特征的重要贡献度,进而构造反映过程数据特征与终点碳温相关性的目标函数;然后通过目标函数定义了种群的最大、最小、平均适应度和随机个体适应度值4个变量,建立了一种自适应调节交叉变异概率机制,使得迭代寻优时种群分布更加合理的同时又提高了算法后期收敛速度,防止陷入局部最优。最后进行实际钢厂生产过程数据特征选择验证和对比实验,结果表明,特征选择平均用时为0.25 s,用于终点预报中温度误差在±5℃的精度为85.67%,碳含量预测误差在±0.01%的精度为80.67%。

Role of Black Carbon-Induced Changes in Snow Albedo in Predictions of Temperature and Precipitation during a Snowstorm

In this study the authors apply the chemistry version of the Weather Research and Forecasting model （WRF-Chem） to examine the impacts of black carbon （BC）-induced changes in snow albedo on simulated temperature and precipitation during the severe snowstorm that occurred in southern China during 0800 26 January to 0800 29 January 2008 （Note that all times are local time except when otherwise stated）. Black carbon aerosol was simulated online within the WRF-Chem. The model resuits showed that surface-albedo, averaged over 27-28 January, can be reduced by up to 10% by the deposition of BC. As a result, relative to a simulation that does not consider deposition of BC on snow/ice, the authors predicted surface air temperatures during 27-28 January can differ by -1.95 to 2.70 K, and the authors predicted accumulated precipitation over 27-28 January can differ by -2.91 to 3.10 mm over Areas A and B with large BC deposition. Different signs of changes are determined by the feedback of clouds and by the availability of water vapor in the atmosphere.