环境工程学报2026，Vol.20Issue(3)：685-696,12.DOI:10.12030/j.cjee.202507054
耗氯量机器学习预测模型对污水厂余氯监测频率的适应性
Adaptability of machine learning-based predictive models for chlorine demand to residual chlorine monitoring frequency in wastewater treatment plants
摘要
Abstract
Currently,a number of wastewater treatment plants(WWTPs)in China still rely on manual residual chlorine monitoring.Due to the low monitoring frequency,precise control of disinfectant dosage remains challenging.This study initially selected WWTP A in the southwestern region,which is equipped with an online residual chlorine analyzer,as the research object.The effluent indicators(including water temperature,flow rate,NH3-N,CODCr,TP,and TN)along with the chlorine dosage were used as inputs to predict chlorine demand.These inputs were derived from high-frequency online monitoring data from WWTP A.A comparative analysis was conducted to evaluate the predictive performance of four machine learning models(i.e.,Backpropagation(BP)Neural Network,Long Short-Term Memory(LSTM),Random Forest(RF),and Support Vector Regression(SVR))in predicting chlorine demand at non-monitoring time points under different residual chlorine monitoring frequencies of once per 1,2,4,6,and 8 hours.The results indicated that the LSTM model achieved the highest prediction accuracy at a monitoring frequency of once per hour;the RF model exhibited superior performance at monitoring frequencies of once per 2 to 4 hours The BP model outperformed other models when the monitoring frequency dropped below once per 6 hours,and the SVR model consistently demonstrated the lowest accuracy across all monitoring frequencies.To ensure the robustness of the findings,additional validation was conducted using datasets from WWTP B and WWTP C,which relied on manual residual chlorine monitoring at frequencies of once per 6 hours and once per 8 hours,respectively.The results demonstrated that the BP model exhibited sustained optimal prediction performance under low-frequency residual chlorine monitoring conditions,and its prediction accuracy can be significantly improved via the optimization by the Particle Swarm Optimization(PSO)algorithm.This study provides a reference for selecting suitable machine learning models to predict chlorine demand under different residual chlorine monitoring frequencies,particularly in scenarios with low-frequency manual monitoring,thereby supporting the precise control of disinfectant dosing in WWTPs.关键词
消毒/BP神经网络/LSTM神经网络/随机森林/支持向量回归Key words
disinfection/BP neural network/LSTM neural network/random forest/support vector regression分类
建筑与水利引用本文复制引用
彭喜林,毛泽鸿,郭佳鑫,马明良,郑星宇,姚杰,唐宏,姚娟娟..耗氯量机器学习预测模型对污水厂余氯监测频率的适应性[J].环境工程学报,2026,20(3):685-696,12.基金项目
重庆市技术创新与发展应用专项资助项目(CSTB2022TIAD—GPX0035) （CSTB2022TIAD—GPX0035）
重庆市再生水利用效益评估及政策建议研究资助项目(CQSLK-2022017) （CQSLK-2022017）
