Date of Completion

11-28-2018

Embargo Period

5-27-2019

Keywords

wastewater treatment; nitrogen removal; random forest; genetic algorithm

Major Advisor

Ranjan Srivastava

Associate Advisor

Jeffrey McCutcheon

Associate Advisor

Baikun Li

Associate Advisor

Kristina Wagstrom

Associate Advisor

Emmanouil Anagnostou

Field of Study

Environmental Engineering

Degree

Doctor of Philosophy

Open Access

Open Access

Abstract

Robust and effective nutrient removal in wastewater treatment is critical to protection of human health and the environment. However, it can be very challenging under variable influent conditions such as those experienced at the University of Connecticut (UConn) Water Pollution Control Facility. The highly seasonal nature of the UConn student population results in rapid transitions between very high and very low loading rates, which can cause short-term permit violations while the system adjusts. The microbial community responsible for most nutrient removal at this and many other facilities is complex and very responsive to influent and environmental conditions. This responsiveness allows the system to adapt to new conditions but can also present problems if certain sensitive components are inhibited. This work sought to improve understanding of this system and, in turn, facilitate more effective, efficient, and consistent treatment.

Nutrient concentration profiles in the treatment basin were monitored daily over two years in order to better characterize the system. The high-resolution data yielded immediate benefits in terms of enhanced treatment when the WPCF staff modified process control techniques using the new data. The resulting dataset was crucial in creating the machine learning-based model that is the core of this work. The model developed uses two random forests in a unique, multi-layer deep learning-type structure to accommodate the combination of independent and partially-dependent variables that describe this system. The resulting model predicted 13 different effluent quality parameters under variable influent, environmental, and process control conditions.

Finally, a genetic algorithm was used in conjunction with the predictive model to determine the optimal process control settings under specific strain conditions. The framework developed was highly flexible, allowing easy modification should the permit or other needs change. The process control settings determined via the genetic algorithm were both reasonable for the system and well supported by WPCF process control practices observed outside the study period. Potentially even more valuable than the suggestion of optimal treatment conditions, this combination of tools might also aid in guiding policy and purchasing decisions.

Share

COinS