Use of neural network models to predict industrial bioreactor effluent quality

Area of Science:

• Environmental Engineering
• Biotechnology
• Chemical Engineering

Background:

• Engineered bioreactors are vital for treating crude oil refining wastewater.
• Predicting bioreactor efficiency is challenging due to variable waste streams.
• Deterministic modeling is unsuitable for complex bioreactor systems.

Purpose of the Study:

• To compare data reconstruction techniques for neural network modeling in industrial wastewater remediation.
• To identify the most effective methods for handling missing data in bioreactor systems.
• To assess the predictive accuracy of different water quality parameters.

Main Methods:

• Utilized a neural network model for predictive time-series analysis.
• Implemented and compared four data reconstruction techniques: interpolation, moving average, mean replacement, and median replacement.
• Evaluated model performance based on predictive capabilities for various water quality parameters.

Main Results:

• Interpolated and moving average methods yielded superior predictions compared to mean and median replacement.
• Mean and median replacement methods showed significantly poorer predictive performance.
• pH was the most accurately predicted water quality parameter, while ammonia and total phenolics were least accurately predicted.

Conclusions:

• Interpolation and moving average are recommended for data reconstruction in large-scale bioreactor remediation processes.
• Commonly used mean/median replacement methods are less suitable for this application.
• The study highlights differential predictability among water quality parameters, with pH being the most reliable.