Investigating the Operational Benefits of Intelligent Water Systems at Water Resource Recovery Facilities: A Plant-wide Simulation Modeling and Machine Learning Approach
Advisor: Dr. Roland D. Cusick
Abstract
Dynamic influent biodegradable Carbon (C) variations significantly impact the overall wastewater treatment and nutrient removal efficacy of water resource recovery facilities (WRRFs). As the effluent discharge quality standards of WRRFs are becoming stricter, developing real-time insights into the influent wastewater quality is necessary to replace time-consuming lab measurements. A bio-electrochemical sensor (BES) offers this capability as a novel biomonitoring tool that provides real-time amperometric electric signals in response to the biodegradable C content of wastewater. As these electric signals are obtained in uncontrolled environmental conditions and in the presence of competing factors, advanced statistical analysis, signal processing, and machine learning (ML) techniques could facilitate the integration of BESs as feedforward control components in WRRFs’ control systems.
This dissertation demonstrates the plant-wide implications of C availability on wastewater treatment and biological nutrient removal (BNR) processes using GPS-X as the main mechanistic modeling platform and discusses how an integrated ML-BES approach can be used to develop intelligent water systems. While principal component analysis (PCA) demonstrated the main environmental factors dominating the signal variance of BESs, the temporal and structural components embedded in the dynamic response of BES signals proved the ability of BESs to detect WRRF-influent organic shock loading events and dynamic plant-influent characteristics using the Facebook (FB) prophet univariate time-series forecasting model. Furthermore, a game-theory approach using SHapley Additive exPlanations (SHAP) analysis of an extreme gradient boosting (XGBoost) regression model developed using a BES signal and other online sensor measurements determined the BES to be the strongest predictor of the denitrification performance and biodegradable C metabolism in a 5-stage modified Bardenpho process. Finally, an integrated hybrid modeling approach demonstrated how a long short-term memory (LSTM) recurrent neural network (RNN) as a deep learning time-series forecasting model and input feature engineering could be used in a data-limited system to dynamically predict the oxygen injection rates in a high-purity oxygen (HPO) process and offer operational cost and electricity savings compared with a manual oxygen injection control process. Overall, this dissertation discusses the operational insights offered by intelligent water systems in the absence of conventional mechanistic insights and demonstrates how ML modeling coupled with novel biomonitoring tools can increase the efficacy of BNR and wastewater treatment processes.
