Full-Scale Optimization of Mechanistic Feedforward-Feedback Ammonium-Based Aeration Control at Blue Plains

1. Introduction Ammonia-Based Aeration Control (ABAC) has become an increasingly important strategy for optimizing aeration energy use and stabilizing nitrification performance in biological nutrient removal processes (Medinilla et al., 2020; Uprety et al., 2015). While conventional ABAC systems typically utilize standard feedback (FB) loops, relying solely on ammonium sensor readings to trim the DO setpoint has shown limited control ability (Rieger et al., 2014). Because sensors are often placed downstream in zones with long hydraulic retention times, the system exhibits slow and delayed responses to rapid changes in daily flow. Therefore, at Blue Plains, black-box feedforward-feedback (FF-FB) ABAC systems were implemented since 2015 to provide a mechanism to better deal with diurnal flows and small rain events. However, these rely heavily on empirical correlations between influent flow, target ammonia, and dissolved oxygen (DO), making them still sensitive to diurnal load variation and limited under severe wet-weather events. These systems commonly use nonlinear feedback loops based on ammonium sensor readings to trim the DO setpoint; however, they do not explicitly represent nitrification kinetics or account for the non-linear relationship between DO and biological growth rates (Newhart et al., 2020; Regmi et al., 2015). Consequently, they struggle to maintain stability across broad operating conditions These limitations were observed at the study facility (Lee et al., 2025). To address these limitations, a mechanistic feedforward-feedback ABAC strategy was developed and implemented based on the work of (Gagnon et al., 2023). Unlike the black-box approach, the mechanistic controller explicitly incorporates Monod DO kinetics by linearizing the nitrification factor (N.F = DO / (K_DO + DO)), enabling a more proportional relationship between controller output and expected nitrification activity (figure 1B). The mechanistic FF-FB ABAC aims to improve robustness, enhance DO-airflow control interactions, and maintain stable ammonium setpoints under variable influent loads. This study evaluates the fundamental differences between the two ABAC approaches and documents the full-scale operational adjustments, tuning, and performance outcomes. 2. Material and methods This study evaluated two ABAC implementations at a full-scale nitrification system between 2024-04-01 and 2025-11-01. The operational period from 2024-04-01 to 2025-06-01 utilized the black-box ABAC, while 2025-06-01 to 2025-11-01 implemented the mechanistic ABAC. Performance metrics included Average Absolute Error (AAE), Average Percent Absolute Error (APAE), and Percentage of Time Within Allowed Threshold (Data in Control) for DO, airflow, and ammonium concentrations (Lee, et al, 2025). 2.1 Black Box feedforward-feedback ABAC: data-driven system that modulates the DO setpoint to achieve the target ammonia concentration in a feedforward action. The feedforward component utilizes a model to predict the required DO setpoint based on influent flow rates and empirical DO, flow - NH4 relationships. A nonlinear feedback loop adjusts the DO setpoint based on the error between the online ammonium concentration and the desired setpoint (Fig, 1A). 2.2 Mechanistic feedforward-feedback ABAC: model-based control strategy designed to linearize the relationship between the output and the nitrification factor (N.F), which represents the DO Monod terms within nitrification kinetics: DO/ (KDO + DO) (Gagnon et al., 2023). The feedforward model calculates the required N.F using influent ammonia loads, nitrification kinetics, and an estimated mass of active nitrifiers. The feedback action corrects this N.F. fraction to meet NH4 setpoint and then the N.F, fraction is converted to DO setpoint to further control airflow rates and valve openings. (Fig, 1B). The linearization of the first PID loop is hypothesized to help with more robust performance in a wider range of operating conditions. 3. Results and discussion: 3.1 Process Optimization of the New ABAC Air-Valve Response Optimization: Prior to optimization, the air-valve response exhibited significant latency, characterized by a 15-20 minute delay between an airflow setpoint change and the corresponding valve actuation (Figure 2A). This mechanical lag decoupled the controller from the process during periods of rapid load variation, preventing the system from achieving DO setpoints in a timely manner. Consequently, this delay induced ammonium deviations and unstable, oscillatory control behavior. To mitigate this, deadbands and signal delays within the airflow control loops were removed. This intervention reduced the response time to approximately 3-5 minutes (Figure 2B). The reduction in actuator latency facilitated near-instantaneous tracking of the airflow setpoint (Figure 2C), improving the physical system's ability to respond to control logic commands. Optimization of Nitrification Factor (N.F) Bounds: The initial mechanistic implementation constrained the Nitrification Factor (N.F) within a relatively narrow range (0.55 < N.F < 0.88), which corresponded to a calculated DO setpoint range of 0.48-2.93 mg O2/L. While sufficient for average loading, this constraint proved restrictive under low-load conditions. Specifically, the lower bound of 0.55 prevented the controller from reducing aeration sufficiently, resulting in over-aeration. Figure 3 illustrates the performance improvements achieved after extending the lower N.F bound to 0.30. This expansion increased the controller's turndown ratio, allowing it to allocate appropriate airflow without over-aerating during periods of low ammonia loading. Furthermore, this adjustment reduced the frequency at which the N.F term saturated at its upper limit. The expanded N.F range enabled smoother transitions between low and high aeration demands. Integration of SRT Adjustments: During the early phases of mechanistic control operation, the nitrifier mass estimates utilized by the feedforward model were inconsistent with the actual biological conditions of the system. Operational data suggested that the model was overestimating the nitrifier inventory; based on kinetic principles, an overestimated biomass results in a lower calculated oxygen demand per unit of biomass, leading to systematic under-aeration. In response, the Solids Retention Time (SRT) parameter in the control model was reduced from 15 to 11 days (Figure 3). This calibration improved the alignment between the modeled nitrification capacity and the actual system performance. By correcting the biomass inventory estimate, the feedforward layer provided more accurate aeration predictions, reducing deviations in ammonium control. PID Feedback Tuning to Improve Responsiveness: Prior to tuning, the mechanistic PID feedback loop exhibited a conservative response to deviations in NH4-N concentration. This sluggishness caused delayed corrective actions, which limited the mechanistic feedforward layer (Figure 4A). To address this, the lambda parameter, which dictates the closed-loop time constant, decreased from 3.0 to 2.5. In addition, the feedback gain was increased (from 0.79 to 1.06), and the integral term was reduced (from 2314 to 1543) to minimize windup and settling time. As shown in Figure 4B, the tuned controller responded more rapidly to ammonium setpoint deviations. These adjustments reduced undershooting and prevented the N.F term from collapsing to zero during sudden load increases, resulting in faster error correction and tighter overall closed-loop control. 3.2 Performance Comparisons Between Black-Box and Mechanistic ABAC The overall performance differences between the black-box and mechanistic ABAC systems are summarized in Table 1. Focusing on ammonium control loop, the mechanistic ABAC demonstrated lower AAE and Percentage of time within 0.6 mg N/L threshold, indicating improved setpoint tracking and reduced variability. The mechanistic controller kept the system within the acceptable control band for a greater percentage of time (84 ± 5.5%), demonstrating more stable operation under the same operational conditions. The improvements can be attributed to the more robust feedforward predictions under the mechanistic framework. Unlike the black-box controller, which based DO setpoint calculations solely on influent flow, the mechanistic controller integrates ammonia load and estimated nitrifier mass, thereby better anticipating the actual oxygen demand. Conclusion: Transitioning from black-box to mechanistic FF-FB ABAC improved ammonia control stability by explicitly incorporating nitrification kinetics through a linearized nitrification factor. While performance increased, further refinement is needed to optimize feedforward predictions and support integration of nitrate-based aeration control.