Author(s): Victor Saenger; Lupicinio Garcia; Miguel Angel Diaz; Marta Ganzer; Sergio Montes; Susana Gonzalez

Linked Author(s):

Keywords: Virtual sensing; Water quality; Real-time monitoring; Machine learning

Abstract: Drinking water facilities deal with strict quality control regulations that oftentimes rely on data coming from expensive real-time sensors. One of the parameters that should be controlled in drinking water is the total amount of Trihalomethanes (THMs). Within this context, facilities such as Water Treatment Plants (WTP) as well as Water Distribution Networks (WDN), need to keep total THMs levels of tap water below a legal threshold of 100µg/L. This fact pinpoints the necessity to develop cheaper and reliable solutions capable of monitor real-time quality drifting events, not only in transiting water within a plant, but also at larger network-wise scales. To tackle these issues, we developed an accurate virtual THM sensor that quantifies with high accuracy, real-time THM levels at two different locations allowing us to validate the scalability of the solution. Locally, we developed this sensor at Barcelona’s Sant Joan Despí WTP, which treats 35% of all water in Barcelona’s metropolitan supply network (Petrovic et al., 2003). To address the problem from a network point of view, we applied the same reasoning at single location within the distribution network of Marbella, Spain. The algorithm powering the sensor is based on a simple neural network that predicts THM levels based on metrics from less costly sensors. This predictive power is reflected by low easy-to-interpret mean SMAPE (Symmetric Mean Absolute Percentage Error) values (Kreinovich et al., 2014), which in our k-fold performance pipeline yielded average scores lower than 10% across folds in the case of the WTP and well below 20% at the WDN sampling point. This latter fact is even more surprising considering a much lower sampling size of around 80 timepoints compared with 1000 timepoints for the former (WTP) case. This scalable and reliable real-time solution remarks the importance of using virtual sensors in operational conditions as well as proper virtual sensing tuning. References: Kreinovich, V., Nguyen, H. T., & Ouncharoen, R. (2014). How to estimate forecasting quality: A system-motivated derivation of symmetric mean absolute percentage error (SMAPE) and other similar characteristics. Petrovic, M., Diaz, A., Ventura, F., & Barceló, D. (2003). Occurrence and removal of estrogenic short-chain ethoxy nonylphenolic compounds and their halogenated derivatives during drinking water production. Environmental science & technology, 37(19), 4442-4448.

DOI: https://doi.org/10.3850/IAHR-39WC2521711920221666

Year: 2022