Forecasting and Optimizing Dual Media Filter Performance via Machine Learning.

Area of Science:

• Environmental Engineering
• Water Treatment Technologies
• Machine Learning Applications

Background:

• Multi-media filters are crucial for water purification.
• Predicting filter performance is essential for operational efficiency and water quality.
• Raw water quality and plant operating variables significantly impact filter performance.

Purpose of the Study:

• To evaluate machine learning algorithms for predicting multi-media filter performance.
• To identify the optimal algorithm and time lag for accurate performance prediction.
• To assess the reliability and forecasting capabilities of the best-performing model.

Main Methods:

• Applied five machine learning algorithms: Decision Tree, Random Forest, Multivariable Linear Regression, Support Vector Regressions, and Gaussian Process Regressions.
• Trained models using seven years of data on water quality (true colour, turbidity) and operating variables (plant flow, chemical doses).
• Utilized a 1-day time lag between input variables and unit filter run volume (UFRV) for optimal prediction.

Main Results:

• Random Forest (RF) with grid search demonstrated the highest reliability in predicting UFRV (RMSE: 31.58, R²: 0.98).
• RF achieved the shortest training time and highest prediction accuracy.
• RF showed strong forecasting capabilities for extreme wet weather events (AUC > 0.8).

Conclusions:

• Random Forest with grid search and a 1-day time lag is a robust algorithm for predicting multi-media filter performance.
• This predictive capability can assist water treatment operators in real-time decision-making.
• Provides early warnings of potential turbidity breakthrough, enhancing treatment process control.