基于自注意力机制的污水处理厂精确加药模型预测控制  

MODEL PREDICTIVE CONTROL FOR ACCURATE DOSING IN WASTEWATER TREATMENT PLANTS BASED ON SELF-ATTENTION MECHANISM

作　　者：裴力锋 陈伟杰 徐敬生 吕路[1,2] PEI Lifeng;CHEN Weijie;XU Jingsheng;LÜLu(State Key Laboratory of Pollution Control and Resource Reuse,School of Environment,Nanjing University,Nanjing 210046,China;Joint Institute of International Environmental Technology Innovation of Changzhou Hi-Tech District and Nanjing University,Changzhou 213125,China)

机构地区：[1]南京大学环境学院污染控制与资源化研究国家重点实验室,南京210046 [2]南京大学常高新国际环保产业技术研究院,江苏常州213125

出　　处：《环境工程》2023年第11期84-92,140,共10页Environmental Engineering

基　　金：国家水体污染控制与治理科技重大专项(2017ZX07204)。

摘　　要：污水处理过程精确控制是解决污水处理厂运行成本高和出水水质不达标问题的有效途径,但由于污水处理过程具有非线性、时变性和时滞性等特点,常规的控制方法难以满足需求。提出了一种模型预测控制方法,以实现污水处理厂药剂的精确投加。所构建的模型基于自注意力机制提取输入序列的有效信息,对参数之间复杂的非线性关系进行建模,从而提高预测准确度。使用江苏省某污水处理厂的反硝化滤池数据进行模型训练和测试,并建立与卷积神经网络(CNN)、长短期记忆网络(LSTM)的对比实验。结果表明,对于出水TN、COD和NH_(4)^(+)-N的预测,自注意力模型的平均误差率(MER)、均方误差(MSE)指标均为最好,实现了较为准确的预测。采用粒子群优化算法计算加药量,并将预测控制模型应用于该污水处理厂。2022年1,2月的模型应用结果表明,出水TN的达标率达到95%以上,优化后的加药量比污水处理厂的月平均加药量分别减少28.72%和21.78%。Precise control is an effective way to solve the problems of high operating costs and substandard effluent quality of wastewater treatment plants,but due to the non-linear,time-varying,and time-lagging nature of the sewage treatment process,conventional control methods can hardly meet the demand.This study proposed a model predictive control method to achieve accurate dosing of chemicals in sewage treatment plants.The constructed model was based on the self-attention mechanism to extract valid information from the input sequence and to model the complex non-linear relationships between features,thus improving the prediction accuracy.Data from a denitrification filter at a wastewater treatment plant in Jiangsu Province were used for training and testing.Compared with CNN and LSTM,the results showed that the MER and MSE metrics of the self-attention model is the best for the prediction of effluent TN,COD and NH_(4)^(+)-N,achieving more accurate prediction.The predictive control model using a particle swarm optimization algorithm to calculate the dosage was applied to this wastewater treatment plant.The results showed that the effluent achieved more than 95%compliance for TN and the optimized dosage was 28.72%and 21.78%lower than the monthly average dosage of the sewage treatment plant respectively.

关 键 词：深度学习 模型预测控制 可解释机器学习 自注意力机制 粒子群优化 精确加药 

分 类 号：X703[环境科学与工程—环境工程]