Machine Learning Soft Sensors for Energy Efficient Potable Reuse

1.Applicability Many water quality parameters important in potable reuse are expensive or slow to measure. 'Soft sensors' are one way of applying machine learning (ML) for water quality and treatment. By modeling challenging water quality variables, soft sensors can reduce monitoring costs or make treatment more adaptive. Reuse systems are engineered to remove a wide variety of chemicals and pathogens to a high degree. Reuse systems are then thoroughly monitored at site-specific critical control points for numerous chemicals and pathogens to ensure regulatory compliance and public acceptance. Most chemicals and pathogens do not have online sensors. The online sensors that are available for specific chemicals or pathogens are often expensive and not fully real-time, with sample frequencies on the order of hours not minutes. Because of the absence of real-time data for chemicals or pathogens that could inform process performance, processes within reuse systems are generally designed and operated with large safety factors. These large safety factors expend more energy than is truly necessary to purify the water. As examples: (1)N-nitrosodimethylamine (NDMA) can drive the UV dose in UV advanced oxidation processes (UV/AOP) downstream of reverse osmosis (RO). UV doses are set conservatively based on maximum historical NDMA concentrations. This consumes more energy than truly necessary to remove NDMA to regulatory targets if the NDMA concentration in the UV/AOP feed were known. (2)Ozone residual is monitored to calculate the concentration*time (CT) for ozone disinfection. However, ozone sensors commonly have greater than +/-20% error (Chen et al. 2020). This results in much random noise in the calculated CT, requiring large safety factors to maintain the desired ozone LRV at all times. This entails higher energy for unnecessarily high ozone doses. Total organic carbon (TOC) could represent a more accurate basis for ozone dosing, particularly at lower ozone doses. However, TOC instruments measure more slowly than ozone sensors, which would lead to less adaptive control. The water sector is developing innovations in real-time sensors which may one day solve these inefficiencies. In the meantime, advances in process control through machine learning could provide the next best solution. This presentation will cover two examples of soft sensors for potable reuse: (1) NDMA in RO permeate and (2) TOC in ozone influent. 2.Methodology All three soft sensors were developed in the programming language R with open-source packages using historical data already available at the utilities. Datasets were split into fully separate training and testing sets to develop and evaluate the machine learning models, respectively. Numerous types of ML models were screened and optimized with the best summarized below. 2.1NDMA Case Study A soft sensor was developed for NDMA in RO permeate. One hundred sixty-two NDMA datapoints, measured every three hours over three weeks, were provided by Orange County Water District from their Groundwater Replenishment System. Features (i.e., inputs or independent variables) included ammonia, pH, turbidity, total chlorine, and pressure. 2.2TOC Case Study A soft sensor was developed for TOC upstream of ozone in a carbon-based reuse system. Hampton Roads Sanitation District provided data from their SWIFT Research Center, a 1 MGD carbon-based reuse demonstration facility. Data was provided at 5-minute intervals over three months. Features included UV transmittance, pH, and ammonia. TOC was actually measured at approximately 2.5-hour intervals, so the true TOC measurements were extracted from the interpolated data. 3.Results 3.1NDMA Case Study A random forest ML model predicted the RO permeate NDMA with a root mean square error (RMSE) of 3.8 ng/L. Basing the UV fluence on the predicted NDMA could save 26% UV energy. Even dosing the UV fluence higher based on a 95% confidence interval around the prediction would result in 13% energy savings for a facility targeting 0.69 ng/L NDMA for surface water augmentation reuse. 3.2TOC Case Study Boosted trees (bstTree) predicted TOC before ozone with a RMSE of 0.35 mg/L (Figure 1). This was much more accurate than a linear model based on UV transmittance (RMSE = 0.709 mg/L) or assuming the last known value of TOC (RMSE = 0.528 mg/L), which would essentially be the status quo. The TOC ML models performed their best when using the last known TOC measurements as one of the inputs, which could be the case when using the model to fill in gaps between the TOC measured every 2.5 hours. Nevertheless, even without the last known TOC as an one of the inputs, ML predicted TOC more accurately than simply assume it would be equal to the last known TOC. This outcome raises the prospective of replacing the TOC instrument with ML, instead of using ML only to fill in gaps. Conclusions/Outcomes Both soft sensors succeeded in predicting the modeled contaminants more accurately than simpler approaches like linear models. Furthermore, these soft sensors outperformed the simpler methods by margins that could have meaningful consequences in terms of saving cost or energy by basing operation on the predicted values. Further research is needed to scale-up these soft sensors by applying them in real-time or transferring them to similar reuse facilities.