Li, Feiyi

Feiyi Li is a speaker at WEFTEC 2026.

Titles from this speaker

Description: A Full-Scale Operational Digital Twin Driven by Hybrid Modeling for Carbon Dosing:...

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

Abstract

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.

This paper was presented at WEFTEC 2026 in New Orleans, Louisiana.

Presentation time

14:30:00

14:45:00

Session time

13:30:00

15:00:00

SessionSuccessful hybrid digital twin adoption in full-scale facilities

Session locationErnest N. Morial Convention Center

TopicFacility Operations and Maintenance, Municipal Wastewater Treatment Design, Research and Innovation

Author(s)

Li, Feiyi, Han, Mofei, Chen, Jun

Author(s)F. Li1, M. Han1, J. Chen1

Author affiliation(s)CSD Water Service, 1CSD Water Service Co. Ltd., 1CSD Water Service, 1

SourceProceedings of the Water Environment Federation

Document typeConference Paper

PublisherWater Environment Federation

Print publication date Sep 2026

DOI10.2175/193864718825160314

Volume / Issue

Content sourceWEFTEC

Word count19

Description: AI-Driven Wastewater Multi-Pump Station Optimization Using Soft-sensor Enabled...

AI-Driven Wastewater Multi-Pump Station Optimization Using Soft-sensor Enabled Predictive Tool

Abstract

Wastewater pumping stations account for 15-70% of electricity consumption in wastewater treatment plants, depending on different situations. Wastewater pumping stations with multiple pumps and mixed actuation (fixed-speed plus VFD units) face strongly time-varying inflow and a discrete–continuous, non-convex scheduling space. Energy-efficient operation is thus intrinsically a coupled prediction–optimization problem: without reliable inflow forecasts and tractable global search, controllers tend to be conservative or myopic, leading to avoidable energy waste. This paper proposes an AI-driven framework that combines a soft-sensor enabled inflow predictor with optimization algorithms to generate implementable, energy-minimizing pump schedules. Methodology The method is developed on the basis of three components: the physical plant, the predictive soft sensor, and the scheduling optimizer, which interact with one another in accordance with specific rules (Figure 1). The basic water balance equation and pump outflow equation formed the basis of the prediction for the soft-sensor, and machine learning models were tested to refine estimates. The evaluated architectures included: Random Forest (RF), Gradient Boosting (GB), and Long Short-Term Memory (LSTM) neural networks. Input features include wet-well water level, pump operational status, frequency settings, and temporal derivatives (level change rate) to capture dynamic behavior. The predicted inflow trajectory becomes the exogenous input to the scheduling optimizer. The scheduling problem is formulated to minimize energy consumption while maintaining operational feasibility. In this case, decisions include discrete pump commitment (on/off for four fixed-speed pumps) and continuous frequency set-points for two VFD pumps, yielding a mixed-integer, non-convex search space. Four solvers are evaluated under the same objective and constraints: (i) Genetic Algorithm (GA), (ii) Particle Swarm Optimization (PSO), (iii) a Hybrid scheme implementing GA warm-start followed by PSO local refined search, and (iv) a Greedy heuristic. All four solvers are supposed to simultaneously serve the goals of satisfying the flow rate requirement, minimizing energy consumption, and controlling the water level within range. Results and Discussion A real wastewater treatment plant dataset is used for validation. Two of the three models for the soft sensor achieves excellent inflow prediction accuracy (R² > 0.99), demonstrating that level-and-discharge-only sensing can provide high-fidelity inflow information without direct inflow metering. Specifically, Gradient Boosting (GB) achieved the highest accuracy with an R² of 0.9983, an RMSE of 8.19 m³/h, and an MAE of 1.10 m³/h. Random Forest (RF) followed closely (R² = 0.9958, RMSE = 12.73 m³/h, MAE = 1.57 m³/h), as shown in Figure 2. In contrast, the LSTM model exhibited lower performance, potentially due to the complexity of capturing long-term dependencies in the inflow profile. The ensemble model also demonstrated robust performance (R² = 0.9979, RMSE = 8.99 m³/h, MAE = 1.17 m³/h), though it did not surpass the standalone GB model. These results indicate that for this specific application, tree-based architectures (GB and RF) are more effective than LSTM in handling the system's non-linear dynamics. In the pump optimization section, a ten - day retrospective study reveals that the water level and flow rate remained unchanged before and after optimization (data lines overlapped), demonstrating the algorithms' precise flow rate management capability (Figure 1). The pump operation strategy shows that the number of running pumps decreased from 3 to 2 within the initial 136 hours, leading to energy savings. The results of different algorithms indicate that the Genetic Algorithm (GA), the Particle Swarm Optimization (PSO), and the Hybrid warm-start + local search method consistently produce the optimal schedules. In contrast, the Greedy heuristic often converges to solutions that are locally optimal but globally sub-optimal, reflecting the non-convex mixed-integer structure of the problem. By utilizing the most effective optimizer, the overall energy consumption is reduced by 25.1% compared to the baseline operation (Figure 4). Overall, the proposed framework couples a soft-sensor with predictive scheduling, turning wet-well level and outflow signals into forward-looking inflow information and thereby shifting pump-station operation from reactive responses to anticipation-based decision-making. In this closed loop, the optimization controller leverages global search to escape poor local minima and local refinement to efficiently tune both discrete pump switching and continuous VFD set-points, providing constraint-compliant and energy-efficient operation despite the mixed-integer, non-convex nature of the problem. Conclusion This work presents a prediction-informed optimization methodology for multi-pump wastewater stations in which a soft sensor transforms readily available measurements (wet-well level and discharge flow) into accurate inflow forecasts that directly drive a mixed-integer scheduling optimizer. Real-plant backtesting confirms high predictive performance and substantial energy savings, and it shows that global metaheuristics are essential for reliably solving non-convex multi-pump coordination problems.

This paper was presented at WEFTEC 2026 in New Orleans, Louisiana.

Presentation time

09:00:00

09:30:00

Session time

8:30:00

9:30:00

SessionFrom Reactive to Predictive: AI Enabled Wastewater Operations

Session locationErnest N. Morial Convention Center

TopicFacility Operations and Maintenance, Municipal Wastewater Treatment Design

Author(s)

Han, Mofei, Chen, Jun, Li, Feiyi

Author(s)M. Han1, J. Chen1, F. Li1

Author affiliation(s)CSD Water Service Co. Ltd., 1CSD Water Service, 1CSD Water Service, 1

SourceProceedings of the Water Environment Federation

Document typeConference Paper

PublisherWater Environment Federation

Print publication date Sep 2026

DOI10.2175/193864718825160307

Volume / Issue

Content sourceWEFTEC

Word count11

Description: An Essential Tool for WRRF Modeling: A Realistic and Complete Influent Generator for...

An Essential Tool for WRRF Modeling: A Realistic and Complete Influent Generator for Flow Rate and Water Quality Based On Machine Learning

Abstract

Modelling, automation and control are widely used for water resource recovery facility (WRRF) optimization. For this, a reliable dynamic influent generator (IG) is essential as key input. An IG model is required to provide the flowrate and pollutant concentration dynamics for the inlet of the WRRF, with different time resolution and scale, depending on the modelling objective. In this study, a data-driven IG model, using different machine learning methods, is proposed and tested. The IG model is able to generate daily and/or hourly flowrate as well as pollutant concentration profiles (COD, TSS and nutrients).

SpeakerLi, Feiyi

Presentation time

13:50:00

14:10:00

Session time

13:30:00

15:30:00

SessionNext Generation Data Analytics

Session number201

TopicFacility Operations and Maintenance, Intelligent Water, Municipal Wastewater Treatment Design, Research and Innovation

Author(s)

F. Li,P. Vanrolleghem,A. Amaral,