[关键词]
[摘要]
提出一种基于Stacking算法集成模型的NO〖HT5”〗x排放预测方法。考虑不同算法的训练机理和观测角度,将门控循环单元(gated recurrent unit,GRU)、XGBoost(eXtreme gradient boosting)和随机森林(random forest,RF)等多个学习能力强、差异度大的模型进行融合,得到一个具有两层结构的集成模型,通过弹性网(elastic network,EN)克服DCS采集的数据集内存在的共线性和群组效应,然后构造特征变量作为集成模型的输入。以某电厂的历史运行数据进行测试,结果表明Stacking集成模型的预测均方误差为6.945 mg/m3,相比单模型降低了13.350%~52.186%,根据其准确的预测结果可以更好的调整设备运行参数,保证排放的污染物浓度控制在合适的范围内。
[Key word]
[Abstract]
An stacking ensemble model forecasting NOx emission was proposed.Considering the difference of data observation and training principles,a twolayer stacking model embedded various machine learning algorithms such as gated recurrent unit (GRU),eXtreme gradient boosting (XGBoost) and random forest (RF) was established to realize the NOx emission prediction.The elastic network (EN) was applied to extract the data from the DCS and eliminate the coupling of each feature variable,and the extracted data was used as input of stacking ensemble model.With the data of thermal power plant history as practical examples,the results show that the root mean square error of stacking ensemble model is 6.945 mg/m3,which is 13.350%~52.186% lower than that of single models.The operation parameters of the equipment can be better adjusted according to accurate prediction results of stacking ensemble model,which is of great significance to ensure the NOx concentration in the appropriate range.
[中图分类号]
TP181
[基金项目]
国家自然科学基金(61503237);上海市科委发电过程智能管控工程技术研究中心基金资助项目(14DZ2251100)