The application of machine learning pretreatment models for O3-BAC process in drinking water treatment plant

Abstract

Machine learning methods have been increasingly integrated into intelligent control systems for water treatment plants. However, their practical implementation faces challenges such as insufficient online data accuracy, difficulties in real-time monitoring of critical parameters, historical issues of overdosing in chemical dosing systems, and the need for model optimization under fluctuating water quality conditions. To address these challenges, this study developed a comprehensive technical framework integrating full-spectrum online analyzers and pretreatment systems, including ceramic membrane filtration and ultrasonic cleaning. Leveraging eight months of large-scale online orthogonal experimental data, the framework encompasses three core components: robust data acquisition, construction of a predictive model library, and optimized chemical dosing strategies. The results demonstrated that the Long Short-Term Memory (LSTM) algorithm effectively captured temporal dynamics in time-series data, exhibiting superior stability during practical operation. Post-retraining, the updated LSTM achieved remarkable error reductions of 67 % in Root Mean Square Error (RMSE) and 63 % in Mean Absolute Percentage Error (MAPE) compared to the pre-update baseline. Following model optimization, the proposed framework achieved a 15 % reduction in pretreatment chemical dosing costs at the drinking water treatment plant. The main contributions of this study include the development of a rapid and interference-resistant online pretreatment system, an adaptive dosing model based on orthogonal experimental design, and a dynamic control strategy that combines classification modeling with optimization algorithms to enhance dosing accuracy and reduce operational costs.