【作者】 汤健； 夏恒； 乔俊飞； 郭子豪；

【Author】 TANG Jian;XIA Heng;QIAO Jun-Fei;GUO Zi-Hao;Faculty of Information Technology, Beijing University of Technology;Beijing Key Laboratory of Computational Intelligence and Intelligent System;

【机构】 北京工业大学信息学部； 计算智能与智能系统北京市重点实验室；

【摘要】 二噁英(DXN)是城市固废焚烧(MSWI)过程排放的剧毒污染物。目前实际工业过程主要通过先现场采集排放烟气样品再以实验室化验分析的方式对DXN排放浓度进行检测,存在周期长、费用高等问题。本文利用过程控制系统实时采集的过程数据,建立基于随机森林(RF)和梯度提升树(GBDT)混合集成的DXN排放浓度预测模型。首先,针对具有小样本高维特性的DXN建模数据进行样本和特征的随机采样,生成训练子集;接着,采用训练子集建立多个基于RF的DXN子模型;然后,对每个基于RF的DXN子模型进行迭代,构建多个基于GBDT的DXN子模型;最后,对基于RF和GBDT的DXN子模型的预测输出,采用简单平均加权方式获得最终输出。本方法通过集成RF和GBDT算法降低了DXN模型的预测方差和偏差。采用UCI平台水泥抗压强度和焚烧过程DXN数据验证了所提方法的有效性。更多还原

【Abstract】 Dioxins(DXN) are highly toxic pollutants emitted by Municipal Solid Waste Incineration(MSWI). At present, the actual industrial process mainly collects exhaust gas samples on site and then tests the DXN emission concentration by means of laboratory analysis. There are problems such as long cycles and high costs. This paper uses the process data collected by the process control system in real time to establish a DXN emission concentration prediction model based on a mixture of random forest(RF) and gradient boosted tree(GBDT). First, random sampling of samples and features is performed on DXN modeling data with small sample high-dimensional characteristics to generate a training subset; then, using the training subset to establish multiple RF-based DXN sub-models; then, for each RF-based DXN sub-models are iterated to construct multiple GBDT-based DXN sub-models. Finally, the predicted outputs of RF and GBDT-based DXN sub-models are combined using a simple average weighting method to obtain the final output. This method reduces the prediction variance and bias of the DXN model by integrating RF and GBDT algorithms. UCI platform cement compressive strength and incineration process DXN data verify the effectiveness of the proposed method.更多还原