Abstract:

In order to solve the problem of excessive or insufficient dosing of coagulant in water plants, and to be compatible with and respond to the change of raw water quality in time, the intelligent dosage prediction model of coagulant with the integrated learning framework was constructed. In the data preprocessing stage, the turbidity removal rate was introduced to filter the original data set to reduce the proportion of unreasonable data in the training data. In characteristic engineering, the nonlinear relationship among inlet water quality,quantity and coagulant dosage was established by feedforward control theory. Meanwhile, the time sequence characteristics were obtained by autocorrelation and partial correlation coefficient analysis and used as model input. The system models were divided into long-term model, medium-term model, and short-term model. Each model used the integrated learning framework of Stacking. The system model outputted the predicted value of coagulant addition through weight distribution. The results showed that the introduction of time series features significantly improved the prediction performance of the model. With the integrated learning framework, the prediction and evaluation indexes MAPE and R ² of the mixed model dosage were 3.78% and 0.96 respectively, and the predicted dosage of coagulant was about 3.12% less than the actual value. It can provide a feasible solution for water plants to achieve accurate dosing and reduce drug consumption

Key words: coagulant, turbidity removal rate, time sequence characteristics, stacking, integrated learning

摘要：

为解决水厂混凝剂投加过量或不足的问题，同时兼顾并及时响应原水水质变化，构建集成学习框架下的混凝剂智慧投加预测模型。在数据预处理阶段，引入除浊率对原始数据集进行筛选，降低不合理投加数据在训练数据中的占比。特征工程中，以前馈控制理论建立进水水质、水量与混凝剂投加流量非线性关系，同时采用自相关、偏相关系数分析得出时序特征并作为模型输入。系统模型划分长期模型、中期模型、短期模型，各模型均采用Stacking集成学习框架，系统模型通过权重分配输出混凝剂投加预测值。结果表明，时序特征的引入使得模型预测性能得到显著提升，集成学习框架下的混合模型投加预测评估指标平均绝对百分比误差（MAPE）、R ²分别为3.78%、0.96，混凝剂预测投加流量与实际值相比节省约3.12%，可为水厂实现精准投加、降低水厂药耗提供一种可行的方案。

关键词: 混凝剂, 除浊率, 时序特征, stacking, 集成学习

CLC Number: 

• X703