Machine learning framework for modeling flocculation kinetics using non-intrusive dynamic image analysis.

Area of Science:

• Environmental Engineering
• Water Treatment Technologies
• Machine Learning Applications

Background:

• Optimizing flocculation processes is crucial for effective and sustainable water treatment.
• Current methods face challenges in accurately predicting floc evolution.
• Machine learning offers a potential solution for improving water treatment efficiency.

Purpose of the Study:

• To develop a machine learning (ML) model for predicting floc evolution during water treatment.
• To devise a framework for the adoption of ML in large-scale water treatment systems.
• To compare the performance of different ML models and a traditional time series model for flocculation prediction.

Main Methods:

• Experimental study of flocculation kinetics using non-intrusive image acquisition.
• Implementation of Multilayer Perceptron (MLP), Long-Short Term Memory (LSTM), and Auto Regressive Integrated Moving Average (ARIMA) models.
• Batch assay data with varying velocity gradients (Gf 20 and 60 s⁻¹) and flocculation times were analyzed.

Main Results:

• Flocculation kinetics showed rapid initial growth followed by a plateau.
• ARIMA model performed poorly, with negative test accuracy.
• MLP achieved high accuracy (R² 0.86-1.0 training, 0.92-1.0 testing).
• LSTM model provided the best prediction accuracy (R² 0.92-1.00) and accurately predicted floc numbers across all conditions.

Conclusions:

• ML models are sensitive to flocculation dynamics, requiring careful selection.
• The LSTM model is highly effective for predicting floc length and number in water treatment.
• The developed ML framework is replicable for water treatment modeling, promoting smart technology adoption.