NDMA soft-sensors for potable reuse: A model development study

N-nitrosodimethylamine (NDMA) is a carcinogen of significant concern in potable water treatment but real-time monitoring of NDMA is not yet feasible with current analytical techniques or mechanistic models. Measuring NDMA and its precursors is time- and labor-intensive which often results in conservative, energy-intensive NDMA treatment approaches, such as operating UV at the maximum dose, to remove NDMA under all possible conditions. To reduce the energy required for NDMA treatment, data-driven modeling was used to simulate an NDMA soft sensor for real-time UV dose control. Models were fit on process data from a full-scale reverse osmosis (RO) indirect potable reuse facility in California, USA. Statistical and machine learning (ML) methods were explored to determine the best data preparation and model structures given a large volume of instrumentation and autocorrelated laboratory samples. Energy savings were simulated based on various safety factors and regulatory limits. This work demonstrates that principal component analysis (PCA) greatly improves statistical and ML model predictions by reducing noise from highly sensorized environments including full-scale RO. Support vector machines with PCA for dimension reduction were found to achieve 13-31% energy reduction over conventional approaches when used to predict NDMA for downstream UV control, depending on the facility's NDMA regulatory limit and desired safety factor. This work illustrates the potential for utilizing instrumentation and laboratory data to precisely control critical treatment units in energy-intensive, potable reuse scenarios and proposes solutions to common challenges for full-scale RO facilities.