AI-Driven Wastewater Multi-Pump Station Optimization Using Soft-sensor Enabled Predictive Tool
Wednesday, September 30, 2026 9:00 AM to 9:30 AM · 30 min. (US/Central)
New Orleans Convention Center
Traditional Technical Session
Room 354
Information
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.
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.
