Automating wastewater characteristic parameter quantitation using neural architecture search in AutoML systems on...

Area of Science:

• Environmental Science
• Machine Learning
• Water Quality Analysis

Background:

• Wastewater treatment plants (WWTPs) require continuous monitoring for optimal operation.
• Predicting wastewater quality parameters is challenging due to complex biochemical processes.
• Rapid assessment of wastewater quality can significantly enhance operational efficiency.

Purpose of the Study:

• To apply Auto Machine Learning (AutoML) models for accurate wastewater quality parameter prediction.
• To demonstrate the efficiency of neural network (NN) regression models optimized via Neural Architectural Space (NAS).
• To predict concentrations of biochemical oxygen demand (BOD), chemical oxygen demand (COD), ammonia (NH3-N), total dissolved solids (TDS), total alkalinity (TA), and total hardness (TH).

Main Methods:

• Utilized spectral reflectance data in the visible-near-infrared range (400-2000 nm) as input.
• Employed various search algorithms (random search, grid search, Bayesian Optimization, Hyperband search) within NAS to identify optimal NN architectures.
• Trained and validated NN models for both single and multiple target variable predictions (2-6 parameters).

Main Results:

• NN architecture optimized via Bayesian Optimization demonstrated superior performance, achieving high R² values for all tested parameters in single-target predictions.
• Multi-target predictions also showed high accuracy, with R² values exceeding 0.988 for BOD and COD in two-target predictions.
• Hyperband search provided improved Total Hardness (TH) prediction accuracy in six-target scenarios compared to Bayesian Optimization.

Conclusions:

• AutoML, particularly NAS with Bayesian Optimization, effectively automates NN architecture selection for wastewater quality prediction.
• The developed models offer a robust and efficient method for predicting multiple wastewater parameters simultaneously.
• This approach eliminates the need for manual hyperparameter tuning, streamlining the development of predictive models for wastewater analysis.