A Full-Scale Operational Digital Twin Driven by Hybrid Modeling for Carbon Dosing: From Reactive Operation to Predictive Control

Summary: A full-scale digital twin framework is developed to optimize external carbon dosing in a Water Resource Recovery Facility (WRRF). It is based on a plant-wide hybrid model that combines mechanistic process knowledge with data-driven learning to represent complex dynamic behavior. The architecture integrates physical assets (treatment units, sensors, IoT) with a digital layer including real-time ETL and AI-based prediction and decision support. A real-time control strategy is implemented to estimate denitrification carbon demand and compute optimal dosing under varying conditions. Full-scale application demonstrates more stable nitrogen removal, reduced carbon consumption, smoother dosing, and lower operational cost and workload, showing that hybrid digital twins enable predictive and intelligent control of wastewater treatment processes. Introduction: External carbon addition is a key strategy for biological nutrient removal in WRRFs. In practice, dosing is commonly governed by empirical rules or simple feedback control, such as fixed ratios, time-based schedules, or single-point water quality measurements. These approaches rely heavily on operator experience, are difficult to standardize, and often fail under dynamic conditions due to inherent time delays between carbon addition and nitrogen response, which limit the effectiveness of feedback control, particularly in pre-denitrification processes. As a result, carbon dosing may lag behind actual demand or become excessive, leading to unstable nitrogen removal and inefficient use of external carbon. To address these limitations, digital twin technologies have emerged as a promising framework for intelligent operation in wastewater treatment systems (Daneshgar et al., 2024; Ghorbani Bam et al., 2025). By integrating mechanistic models with data-driven learning, plant-wide hybrid modeling provides a digital representation of WRRFs that assimilates operational data and captures complex dynamics (Schneider et al., 2022; Serrao et al., 2024). This enables a predictive, model-based control framework that shifts carbon dosing from reactive feedback to proactive, data-driven optimization. In this study, a real-time external carbon dosing strategy based on a full-scale hybrid digital twin is developed to estimate denitrification demand and adjust dosing accordingly dynamically. Methodology: The proposed digital twin and control framework is implemented at the CiXi-East WRRF, with treatment capacity of 100,000 m³/d. Figure 1 illustrates the overall framework of the proposed digital twin system. The architecture consists of a physical system layer including sensors, WRRF facilities, and SCADA; a data acquisition and ETL layer responsible for real-time data collection and preprocessing; a digital twin layer integrating data-driven models with a mechanistic model powered by SUMO; and an application layer supporting real-time monitoring, forecasting, and decision support. A closed-loop interaction between the physical system and the digital twin is achieved through real-time control and decision feedback. Within this framework, a predictive external carbon dosing module is embedded as a closed-loop optimization and control component. Influent conditions are forecasted using data-driven models, while real-time operational states-such as return flows, aeration, and nitrogen concentrations in the anoxic and aerobic zone-are continuously simulated with the online mechanistic model. Based on the real-time process states simulated by the mechanistic model, the digital twin evaluates multiple dosing scenarios and determines the optimal external carbon dosing strategy. The carbon dosing strategy is determined by solving a constrained optimization problem, in which the objective is to minimize external carbon addition subject to an upper bound on the predicted effluent NOx (nitrate) concentration, typically set to 8 mg/L and configurable via the user interface. The optimal dosing setpoint is transmitted to the physical system via OPC UA and implemented through the carbon dosing pumps, thereby enabling real-time adaptive control and closing the digital twin feedback loop. In addition to the automatic predictive control mode, a manual dosing mode is retained to allow operator intervention when required. Results and future developments Figure 2 indicates that the MPC-recommended carbon dosing responds more rapidly to dynamic operating conditions than conventional manual operation, with an average reduction of 11.3% in external carbon consumption. Currently, the carbon dosing strategy relies on a mechanistic model, which requires considerable computational effort. Future work will develop a reinforcement learning-based emulator to enable faster online optimization and adaptive, data-driven predictive carbon dosing. Conclusion: This work demonstrates the full-scale application of a hybrid digital twin for predictive model-based optimization of external carbon dosing in a WRRF. The proposed framework enables a transition from conventional reactive operation to proactive, data-driven control, supported by a user-friendly graphical interface for real-time monitoring and decision support. The results highlight the potential of digital twins to achieve more stable nitrogen removal with reduced external carbon consumption. Future work will focus on developing a reinforcement learning–based emulator to further enhance control efficiency and robustness.