WRF 5121 Final Report Overview: Long-Term Evaluation of Hybrid Digital Twins at Three Full-Scale WRRFs

INTRODUCTION The Water Research Foundation (WRF) project 5121, Development of Innovative Predictive Control Strategies for Nutrient Removal, aims at developing full-scale hybrid nutrient management controllers (named here as Hybrid Optimizer, HO) by employing a hybrid approach that combines both machine learning and mechanistic models. Three full-scale pilots are deployed and evaluated for this project to demonstrate long-term controller abilities. The concept, design and preliminary results have been shared previously [1-7], which triggered wide interest and heated discussion. Over the past year, refinements were made to improve performance and evaluation approaches were standardized. Currently, all three pilots are fully deployed and operational. This paper provides an overview of three case studies. The control for three pilots is advisory only with no direct control functions due to cybersecurity concerns. Instead, operators receive daily email notifications outlining recommended operations and/or access a web-based user interface (UI), guiding their actions. The full-scale pilots are: 1. Clean Water Services Durham WRRF (CWS): Primary clarifier alum addition setpoints once a day, aiming to stabilize orthophosphate load to the downstream biological treatment system. 2. Agua Nueva WRF (AN): Dissolved oxygen setpoints twice daily for the bioreactor systems, aiming to minimize overall aeration energy consumption. 3.AlexRenew WRF (AR): HO recommends three equalization setpoints once a day to equalize the influent loadings to the bioreactor to minimize energy, chemicals, and blower staging. METHODOLOGY As shown in Figure 1, Hybrid Optimizer consists of seven major components. Descriptions of each component can be found in [1-3]. In this paper, the performance of Soft Sensor, Influent Forecaster and emulator/optimizer is mainly presented and discussed. Four different types of standard graphs are used to evaluate the performance of the HO. (1)Control Chart. Directly conveys the validity of the model, based on which warnings and alarms are derived. It integrates the measured and the modelled values into a dimensionless metric (t_score) and allows its comparison with predefined bounds. (2)Process Plot. The process plot presents the measured and the modelled values. (3)Unscaled Error Plot. The Error plot presents the residuals between the modelled and the measured. (4)XY Plot. The XY Plot presents the Modelled v. s. the Measured and the prediction intervals. (not shown here for brevity) RESULTS AND DISCUSSION The components and timelines of the three pilots are listed in Table 1. The Graphic User Interface (GUI) for two pilots are in progress. Evaluation results are available for all three, however this abstract presents only partial results. The soft sensor can be considered a full digital twin of the modelled components. Thus, it provides dynamic 15-minute frequency data on all wastewater components in the model scope. For CWS the primary influent and effluent ortho-phosphate data was used to support primary alum addition recommendations, without an actual sensor. For AN the focus was ammonia. The soft sensor demonstrates alignment with measured data (Figure 2).

*In the Control Chart, the t score falls within the warning and alarm bounds (set at 95% and 99% confidence intervals), which indicates soft sensor performance. The t-score is the normalized metric of a given set of measurements.

*The Process and Residual Plots are more intuitive for process engineers to evaluate. The modelled and measured values align closely, and their parallel trends indicate good soft sensor performance. Figure 3 shows the AN performance for primary influent and effluent ammonia. The top graph shows a good match between the composite samples. However, the primary effluent in the middle and bottom graphs shows a poor match. It was found that high influent sulfide was biasing the measurements high. The bottom graph showing the dynamic concentration variation, other than the bias, shows a good variability match. Flow forecasting is critical to predictive control. The Forecaster uses the available weather forecast and other variables to estimate plant influent flow. Figure 4 compares influent flow (forecasted vs actual) as well as the forecasted versus actual rainfall. A close examination shows that much of the influent forecast error can be attributed to the error in the precipitation forecast. The AR optimizer (Figure 5) aims to equalize ammonia load, which helps stabilize the aeration demand and optimization of aeration and chemicals. The flow and loading prediction from the HO allows the optimizer to recommend three equalization setpoints for the next 24 hours. In the top graph, the green line is the target daily average loading, red and blue show the load with and without the recommendations. CONCLUSIONS This abstract presents abbreviated results of the long-term performance of HO during three full-scale pilot studies. The HO appears to be a promising opportunity to improve plant operation using digital twin technology by giving both optimized operational recommendations and increased insight into the facility's status. The ongoing assessment of the HO performance is anticipated to yield further valuable practical experiences and insights, including the benefits and costs of such applications.