Abstract: Aiming at the problem of online information of COD index which can’t be measured directly in the wastewater treatment process, an integrated soft-sensing model of COD index based on distributed improved BP neural network and grey prediction is proposed. In order to reflect the different status in wastewater treatment process, the satisfactory clustering algorithm is utilized to cluster the data samples and form several subsets. The improved BP neural network method is used to establish prediction model for each subset. The Euclidean distance between the current input data and each clustering center is calculated and the COD index prediction value of distributed neural network is obtained through combining the output of some prediction model whose Euclidean distance is small. According to the historical data of COD index, the improved grey prediction method is used to predict the current value of COD index, which can reflect the time dependence of COD index. The dynamic weighted method is adopted to integrate the two obtained COD index prediction value. Simulation results show that the proposed integrated soft-sensing model has good prediction performance and can satisfy the precision requirements of COD index measuring in wastewater treatment process.

Key words: wastewater treatment process, Chemical Oxygen Demand（COD） index, distributed improved BP neural network, grey prediction, integrated soft-sensing model, dynamic weighted method

摘要： 针对污水处理过程中水质参数COD指标难以在线检测的问题，提出一种基于分布式改进BP神经网络和灰色预测的COD指标集成软测量模型。为反映污水处理过程的不同工况，采用满意聚类算法对数据样本进行聚类处理，将数据样本划分为若干个子样本集，利用改进BP神经网络方法分别为每个子样本集建立预测模型，计算当前输入数据与各个聚类中心的欧式距离，将欧式距离较小的部分预测模型的输出进行综合，得到分布式神经网络的COD指标预估值；为反映COD指标的时间相关性，基于COD指标历史数据采用改进灰色预测建模方法计算得到当前时刻COD指标的预估值；采用动态加权方法将获得两个COD指标预估值进行加权集成。仿真实验表明，集成软测量模型具有较好的预测性能，可以满足污水处理过程COD指标实时检测的精度要求。

关键词: 污水处理过程, 化学需氧量（COD）指标, 分布式改进BP神经网络, 灰色预测, 集成软测量模型, 动态加权方法