Machine Learning Design Tools for Carbon-Based Reuse

More and more utilities are considering potable reuse to diversify water supply in response to population growth and climate change. Preliminary cost estimates-both capital and O&M-are critically important for alternatives assessment and securing funding. Meaningful cost estimates require reasonable assumptions about the effectiveness of each process for key contaminants such as per- and polyfluoroalkyl substances (PFAS). Literature averages or other benchmarks do not provide an accurate, tailored estimate since wastewater effluent quality and the resulting advanced treatment performance vary site to site. Data-driven models such as machine learning (ML) would provide a better basis for estimating treatment performance as a function of feed water quality and operational settings.

#This presentation explores a suite of ML models that inform water treatment design and operation and maintenance (O&M) decisions. Two models are presented in this ML model suite: biofiltration and granular activated carbon (GAC). The biofiltration ML model focuses on predicting total organic carbon (TOC) removal given process information and upstream water quality. The GAC ML model (Koyama et al., 2024) focuses on predicting the removal of micropollutants such as per- and polyfluoroalkyl substances (PFAS) given inputs such as chemical species and concentration, feed water quality parameters (e.g. pH, TOC) and empty bed contact time (EBCT). This GAC PFAS ML model is critical because perfluorooctanoic acid (PFOA) typically drives GAC changeout in carbon-based reuse for compliance with the 2024 EPA PFAS drinking water rule.

##The biofiltration ML model was trained on more than 200 data points collected from peer reviewed literature. The type of water matrices covered in this database includes surface water and wastewater effluent, spanning a feed TOC range of 1 to 20 mg per liter. A key modeling technique that made the ML model accurate was capturing the pretreatment unit process(es) used prior to biofiltration. For instance, ozone to TOC dosing ratio was used as an input variable and ranged from 0.3 to 3.9, with majority of data points falling between 0.5 to 1.5 (Figure 1). Ozone pretreatment can enhance biofiltration TOC removal even when the TOC remains similar. Other key input variables such as EBCT ranged from 7 to 100 minutes with majority of data points falling between 20 to 50 minutes (Figure 2) and temperature ranged from 15 to 25 degrees Celsius. The target variable to predict, TOC removal percentage, had most data distributed between 3 to 60 % (Figure 3).

#For the GAC ML model, a database consisting of over 600 points (Koyama et al., 2023) was used for training. This database consists of 60 unique compounds (pharmaceuticals, volatile organic contaminants, PFASs), 3 GAC types by base material (20 unique GAC products), and 49 water matrices (groundwater, surface water, and treated wastewater). The model for predicting bed volumes to reach 10% breakthrough (BV10) was developed based on the gradient boosting machine algorithm (Friedman, 2001). A key input is TOC, since TOC tends to foul GAC or compete with PFAS for sorption sites.

#The biofiltration TOC removal model had a testing set prediction accuracy of 7 percentage point root mean squared error (RMSE) and 5 percentage point mean absolute error (MAE), when predicting the TOC removal percentage (Figure 4). To put this into perspective, if a biofilter's removal performance is predicted to be 25% by the ML model, the actual removal range can be assumed to be between 20 to 30%, based on the MAE.

#The ML model for BV10 had an MAE of 0.1 log units on the testing set prediction (Figure 5). To put this into perspective, the ML model predicted bed volume to ten percent breakthrough with 1.6 months of error assuming an empty bed contact time of 20 minutes.

#The biofiltration TOC ML model and the GAC PFAS ML model were applied in tandem to cost estimate four reuse scenarios for a small, rural community. The four scenarios included carbon-based reuse vs reverse osmosis (RO) based reuse, and lower TOC effluent vs blending in effluent from a lagoon with higher TOC. Biofilter TOC removal was first estimated, which then provided the input TOC values for the GAC model for PFAS removal. Since one of the scenarios involved sending high-TOC effluent to carbon-based reuse, the ML model predicted the PFAS would breakthrough much faster than in average carbon-based reuse facilities. This faster breakthrough resulted in a 10-times greater O&M cost per flow for high-TOC carbon-based reuse compared to low-TOC carbon-based reuse and 3-times of that for high-TOC RO-based reuse. By adopting ML models in the initial stage of reuse planning, we avoided a likely $5M/year underestimate in in GAC change-out costs. Simultaneously, the ML models produced these results months sooner and tens of thousands of dollars less compared to bench-scale testing GAC with pre-ozonated, pre-biofiltered effluent.

#We will highlight the efficiency of ML process model tools in cost estimation by benchmarking the ML models against other methods such as conducting physical experiments or estimating based on average literature values of TOC and PFAS removal. The adoption of ML models presented in this study showcases a novel way of conducting traditional engineering services for potable reuse applications.