基于宽度混合森林回归的城市固废焚烧过程二噁英排放软测量

二噁英是城市固废焚烧过程排放的痕量有机污染物. 受限于相关技术的复杂度和高成本, 二噁英排放浓度检测的大时滞已成为制约城市固废焚烧过程优化控制的关键因素之一. 虽然具有低成本、快响应、高精度等特点的数据驱动软测量模型能够有效解决上述问题, 但二噁英建模方法必须要契合数据的小样本、高维度特性. 对此, 提出了由特征映射层、潜在特征提取层、特征增强层和增量学习层组成的宽度混合森林回归软测量方法. 首先, 构建由随机森林和完全随机森林构成的混合森林组进行高维特征映射; 其次, 依据贡献率对全联接混合矩阵进行潜在特征提取, 采用信息度量准则保证潜在有价值信息的最大化传递和最小化冗余, 降低模型的复杂度和计算消耗; 然后, 基于所提取潜在信息训练特征增强层以增强特征表征能力; 最后, 通过增量式学习策略构建增量学习层后采用Moore-Penrose伪逆获得权重矩阵. 在基准数据集和城市固废焚烧过程二噁英数据集上的实验结果表明了方法的有效性和优越性.

Soft Sensing Method of Dioxin Emission in Municipal Solid Waste Incineration Process Based on Broad Hybrid Forest Regression

Dioxin is a trace organic pollutant emitted from municipal solid waste incineration process. Limited by the complexity and high cost of relative technology, the big time delay of dioxin emission concentration detection has become one of the key factors restricting the optimize control of municipal solid waste incineration process. Although the data-driven soft sensing model with the characteristics of low cost, fast response and high precision can effectively solve the above problems, the dioxin modeling method must fit the small sample and high-dimensional characteristics of the modeling data. In this paper, a broad hybrid forest regression soft sensing method is proposed, which consists of feature mapping layer, latent feature extraction layer, feature enhancement layer and incremental learning layer. Firstly, a hybrid forest group composed of random forest and completely random forest is constructed for high-dimensional feature mapping. Secondly, the latent features extraction of the fully connected mixed matrix is carried out according to the contribution rate, and the information measurement criterion is used to ensure the maximum transmission and minimize redundancy of potential valuable information. Thus, the model complexity and computational consumption are reduced. Then, the feature enhancement layer is trained based on the extracted potential information to enhance the feature representation ability. Finally, the incremental learning layer is constructed by using incremental learning strategy, and the weight matrix is obtained by using the Moore-Penrose pseudo inverse. The experimental results on high-dimensional benchmark and dioxin datasets of municipal solid waste incineration process show the effectiveness and superiority of the proposed method.