High-frequency parameter updates in digital twins: Report on a 2-week challenge test for fully autonomous adaptive operation of an effluent water recovery system.

Abstract

Digital twins (DTs) that can predict the transient behaviour of a water treatment plant combined with buffer capacity, enable reductions in specific energy consumption (SEC) in response to time-of-use (TOU) electricity costs. The accuracy of a DT depends on the richness of its training data (Conejos Fuertes et al. 2020). However, process changes cannot easily be tested in production-scale facilities due to water-quality compliance and output requirements. To address this, UCI developed a heavily instrumented closed-circuit reverse osmosis (CCRO) physical twin of the OCWD plant. Experiments on the physical twin over a wide range of process parameters generate the rich dataset needed to train the DT. The trained DT can then use sparse plant-scale data to predict how process changes affect water quality, energy use, membrane fouling, etc. In our research, we developed a method to train the DT using a rich experimental data set collected from the physical twin. Then we use the DT to find a time series of process parameters that minimize the SEC over a seven-day interval based on a TOU electricity cost schedule. To simplify the exposition in this work, factors such as water quality, membrane lifetime, and brine disposal were excluded from the optimization but can be incorporated in future cost-function constraints. Even without TOU pricing, optimal CCRO setpoints are not self-evident; this framework identifies them and is broadly applicable to other treatment systems.