Actionable Prediction of Final Clarifier Performance Using The Data You Already Collect

[b]BACKGROUND[/b] In order to achieve more stable and proactive operation, many WRRFs are leveraging new sensors and modelling software to increase their predictive capacity. While our near-term predictive capacity for certain metrics, such as ammonia or DO, have matured significantly, secondary clarifier (SC) performance remains challenging to predict using conventional process models. While some common models [1] [2] can capture baseline behavior, they are not sensitive to daily variations in sludge settleability and hydraulics. Research has shown that the flocculent settling coefficient (r-floc) can be used to improve model accuracy, but calibrating the r-floc traditionally required specialized testing such as the threshold of flocculation (TOF) test [3] [4]. Recently, the authors reported that calibrating r-floc against historical effluent data can enhance model accuracy, but this may require additional sampling or measurement [5]. Since traditional approaches require specialized settling tests or complex sampling and calibration processes that may not be feasible for many utilities, this study proposes a practical and scalable method of estimating settling behavior using only routine operational data for daily automatic calibration. This unlocks real-time ESS prediction, offering operators a practical soft sensor for clarifier performance. [b]METHODOLOGY[/b] A daily autocalibration framework was developed around a 1-D clarifier model built using SUMO (Dynamita SARL) using the triple exponential extension to the double exponential (DE) model [7]. Flow, mixed liquor suspended solids (MLSS), sludge volume index (SVI), clarifier underflow, and clarifier configuration were the routine inputs to the model. A python orchestrator performed automatic r-floc calibration, keeping ESS within 10% of actual values while optimizing r-floc. Applied across six facilities, this method captured variations in sludge characteristics, hydraulics, and upstream biology. When settling data was unavailable, a modified Daigger [8] correlation was used to relate SVI data to estimate Vesilind parameters of hindered settling (r-hin) and maximum settling velocity (Vo); otherwise site-specifc Vo and r-hin estimates were used. The approach is demonstrated for a 9.8 MGD Facility (Facility A) which recorded high quality operational data and strict effluent total phosphorus limits. The calibrated r-floc parameter from the 1-D clarifier was an input to a calibrated whole plant model to evaluate ESS prediction based on the biokinetic model for Facility A. To address facilities with infrequent ESS sampling, a machine learning (ML) model using extreme gradient boosting (XGBoost) predicted r-floc from input features. Heavy-tailed features were log transformed, while others remained on natural scale. The chronological input dataset was then split (80% training/20% test) to prevent look-ahead bias. Hyperparameters were tuned via Bayesian methods and performance was evaluated on test set. [b]RESULTS[/b] The daily automated calibration generated varying r-floc estimates based on clarifier performance. The Facility A whole plant model predicted ESS within ±15% of actual, per Figure 1. Comparing r-floc values to observations and model outputs validated the method and unlocked rfloc's soft-sensing capabilities, per Figure 2. Poor settling periods, marked by extreme SVI, elevated selector zone readily biodegradable chemical oxygen demand (rbCOD) or operational disturbances like toxicity lowered r-floc, while site upgrades like upgraded sludge withdrawal system increased the r-floc. The ML model, trained on input features like calibrated r-floc estimates, SVI, RAS ratio, surface overflow rate (SOR), surface loading rate (SLR), along with model outputs like selector zone rbCOD, stored poly hydroxy alkanoates (PHA) and sludge iron content showed comparative performance and feature relevance is summarized in Table 1. [b]DISCUSSION[/b] The r-floc soft sensor showed slight shifts between test and control groups in full-scale pilots at Facility B (low DO) and C (hydrocyclones). Clarifier upgrades, including sludge withdrawal mechanisms and floor baffles at Facility A, resulted in improved r-floc. Also, an industrial discharge to Facility D led to deflocculation, and poor flocculation in a three sludge BNR process at Facility E lowered the r-floc. Figure 3 demonstrates the r-floc soft sensor's sensitivity to these impacts, both biological and hydraulic, across diverse plants. Thus, producing an ML-corrected r-floc metric can both improve mechanistic model accuracy, and serve as a novel performance indicator. This may be especially valuable for plants with low-resolution effluent data collection. [b]CONCLUSION[/b] This paper demonstrates how clarifier models can be automatically calibrated using only water quality data already being collected at a WRRF, in order to achieve accurate real-time predictions of effluent solids and improve operations. Daily, ML-enabled calibrations of the model's flocculant settling coefficient (r-floc) provided early diagnostic insights and opportunities for proactive operational adjustments, all without additional lab work or instrumentation. The r-floc was sensitive to a broad range of biological and hydraulic drivers, making it a robust and predictive parameter. Overall, this framework offers a practical tool for better clarifier reliability, predictability, and plant-wide resilience.