Machine Learning In The Dark: Disinfection With <55% UV Transmittance

Introduction: UV disinfection systems have gained popularity in wastewater disinfection applications as they can inactivate microorganisms while avoiding the injection of a chemical and any risk of a high disinfectant demand or disinfection byproducts. UV disinfection process design must account for UV Transmittance (UVT) which is a critical parameter that characterizes the ability of UV photons to pass through water and reach a microbiological target. Monitoring of UVT is achieved by online or benchtop readings, but online data can be skewed by sensor drifting or water quality impacts. Black & Veatch worked with two wastewater facilities where UVTs <55% were observed and employed machine learning to analyze correlations between measured UVT and various water quality parameters including carbonaceous biological oxygen demand (CBOD) and total organic carbon (TOC). UVT prediction allows for historical data to be verified whether the sensor was reading correctly or for optimization of UV system operation. Materials and methods: This study collected UVT and additional water quality data from each respective treatment facility secondary effluent at various sample collections intervals over the last 5 years. UVT was collected with online and benchtop instruments while other water quality data were collected via analytical methods accepted by 40 CFR criteria. Data cleaning and analysis was performed in R using various open-source packages and algorithms. Results and discussion: Review of Industry Guidance. Although there is no EPA requirement of a minimum design UVT for wastewater disinfection, guidelines have been followed in various states. The 2012 NWRI Guidelines set minimum allowable UVTs for wastewater applications based on the type of upstream treatment with a 55% minimum for applications with media filtration (REF2). In addition, 10 States standards recommends UV disinfection in wastewater be limited to effluent UVTs greater than 65% (REF3). However, UV disinfection has been implemented in states across the US with UVTs of 55% or lower, where various benchtop and full-scale data supports viable inactivation of microbes relevant to each projects operating permit. This study considered two facilities which are known to produce lower UVTs than conventional nitrifying treatment, Pure Oxygen Activated Sludge and Trickling Filter, from projects in Michigan and California, respectively. Machine Learning Techniques. Machine learning consists of algorithms and statistical models to analyze and draw inferences from patterns in data and can be utilized for various applications relevant to water treatment. This study reviews industry accepted techniques for employing machine learning for numerical predictions models based on multivariate regression and provides an in-depth review of the two algorithms employed in this work, Random Forests (RF) and Support Vector Regression (SVR). Case Study 1. A pure oxygen activated sludge treatment facility in Michigan undergoing a disinfection system evaluation was analyzed to consider UV disinfection to replace previous disinfection systems using sodium hypochlorite and ozone. Black & Veatch implemented an online UVT probe for a period of 12 months where the plant experienced various low UVT trends (<50%) that were impacted by ferric chloride addition, expected industrial source discharges, and sensor drifting. RF and SVR machine learning algorithms were employed to predict UVT based on CBOD, Solids and Temperature. Model results were successful in predicting UVT with an R mean squared value (RMSE) of 0.84 and 0.83, respectively (Figures 2 and 3). To mitigate the impact of excessive drifting of the online UVT probe, the model was further refined to predict periods where the UVT probe may have drifted out of calibration. Outcomes of these predictions could help promote instrument calibration or ensure historical data is accurate and whether or not it should be included in further analysis for process design of a UV disinfection system. Case Study 2. A regional wastewater facility in California undergoing a plant upgrade to employ UV disinfection for Title 22 non-potable reuse was also evaluated. As a part of process design and commissioning of the new Title 22 UV disinfection system, Black & Veatch analyzed 2+ years of water quality data where UVTs less than 50% were observed, which has been documented for trickling filter secondary effluent (REF4). Upstream trickling filter backwashing events were thought to have especially produced organic loads which impacted the UVT sensor and may have altered data collection after the system upset had already passed. TOC and TSS data were used to train the RF and SVR models to predict UVT across the monitoring period, with the SVR achieving a higher success in prediction with an RSME of 0.86 (Figures 5 and 6). Conclusions: 1.Applications where UVT is <55% can be managed contrary to some guidance and provide effective disinfection. 2.Machine Learning can verify analyzer trends to identify correct/incorrect historical data and ensure appropriate process design. 3.Machine Learning can be integrated with control approaches to accelerate response of UV disinfection systems based on anticipated water quality. 4.Multivariate water quality constituent regression could be paired with UV validation equations to predict performance and improve understanding of disinfection operations.