⚡ electrical engineering

Performance Prediction and Optimal Rotational Speed Determination of a Variable-Speed Pump as Turbine Using Surrogate Modeling and SOPNN-DE

This study presents a computationally efficient, data-driven framework combining Self-Organizing Polynomial Neural Networks and Differential Evolution to optimize variable-speed Pump as Turbine control, demonstrating a 13.22% efficiency gain and a 1.23-year payback period through real-world validation in the Tehran water distribution network.

Original authors: Sama Mirmahdian, MEHDI FARTASH, MOJTABA tahani, Hadi Veisi

Published 2026-09-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sama Mirmahdian, MEHDI FARTASH, MOJTABA tahani, Hadi Veisi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, silent arteries of a modern city, water flows under pressure to reach every tap, every shower, and every fire hydrant. To keep this flow safe and steady, engineers must constantly manage the pressure, often using valves to slow the water down when it moves too fast. For decades, this slowing process has been a simple act of waste: the excess energy of the rushing water is turned into heat and sound, dissipated harmlessly into the air. However, a quiet revolution in how we manage water networks suggests that this wasted energy could instead be captured and turned into electricity. The key to this transformation lies in a device called a pump-as-turbine. It is a standard water pump, the kind found in basements and industrial plants, but run in reverse. Instead of using electricity to push water, it uses the force of the water to spin a turbine and generate power. While the concept is sound, the reality of making these machines work efficiently is tricky. Water networks are never still; the demand for water changes from hour to hour, causing the flow and pressure to fluctuate wildly. A machine designed to run at a single, fixed speed often struggles to keep up with these changes, losing efficiency and failing to capture the full potential of the energy available.

A team of researchers from universities in Iran set out to solve this problem by teaching a pump-as-turbine how to adapt in real time. They focused on a specific installation in the water distribution network of Tehran, a system that had been operating for years and provided a rich history of real-world data. Rather than relying on complex computer simulations that take hours to run or expensive new experiments, the researchers turned to a method that learns directly from the machine's own history. They gathered over a thousand records of how the pump behaved under different conditions, noting the flow rate, the pressure, the speed of the spinning motor, and the amount of electricity produced. Using a sophisticated type of computer learning known as a self-organizing polynomial neural network, they built a mathematical model that could predict exactly how the machine would perform in any given situation. Unlike many computer models that act as "black boxes," giving an answer without explaining how, this model produced clear, readable equations that described the relationship between the water flow and the machine's speed.

With this predictive model in hand, the researchers then asked a simple but powerful question: if the water flow changes, what is the exact speed the pump should spin to get the most electricity out of it? They used an optimization algorithm to search for the perfect speed for every possible flow rate, while ensuring the machine still provided enough pressure for the city's needs. The results were striking. They found that the most efficient speed was not a fixed number but a moving target that shifted smoothly as the water flow changed. When the flow was low, the pump needed to spin slower; when the flow was high, it needed to spin faster. This adaptive strategy allowed the machine to stay in its "sweet spot" of efficiency much more often than a fixed-speed machine could. In fact, under low-flow conditions, the new strategy improved the machine's efficiency by more than thirteen percent compared to running at a constant speed. The researchers also discovered that the relationship between the water flow and the ideal speed was remarkably consistent, following a nearly straight line that aligned with the fundamental laws of fluid mechanics, confirming that their data-driven approach had rediscovered a physical truth.

The study went beyond just finding the right speed; it tested how reliable this advice would be in the real world, where sensors are not perfect and measurements can be slightly off. They simulated errors in the flow measurements to see if the recommended speed would jump around wildly or stay steady. The analysis showed that the system was robust, but it also highlighted a critical dependency: the accuracy of the flow measurement. If the sensor reading the water flow was off by even a small amount, the recommended speed would shift accordingly. This finding suggests that for this technology to work perfectly, the most important investment is not in the complex software, but in ensuring the flow meters are precise. The researchers also derived a simple, direct formula from their complex optimization results. This formula acts like a quick-reference guide, allowing a control system to instantly calculate the ideal speed based on the current flow, without needing to run a slow, complex calculation every time the water level changes. This makes it possible to install the system on standard industrial controllers found in water treatment plants today.

Finally, the team looked at the economic reality of putting this idea into practice. Using the actual flow patterns of the Tehran station over a full year, they calculated how much extra electricity could be generated by switching from a fixed speed to this new adaptive speed. The numbers were compelling. The adaptive strategy would generate an additional 18,133 kilowatt-hours of electricity every year, enough to power a small number of homes continuously. At current electricity prices, this translates to an extra revenue of over 1,600 euros annually. The cost to upgrade the existing system with the necessary software and control hardware was estimated at just 2,000 euros. This means the investment would pay for itself in just over a year, a timeline that is exceptionally fast for energy projects. Over a decade, the project would generate more than 10,000 euros in net profit. The study concludes that by using data to teach old machines new tricks, water utilities can turn a necessary waste of energy into a reliable, profitable source of power, all while keeping the water flowing smoothly to the people who need it.

Technical Summary: Performance Prediction and Optimal Rotational Speed Determination of a Variable-Speed Pump as Turbine Using Surrogate Modeling and SOPNN-DE

Problem Statement
Variable-speed Pumps as Turbines (PATs) offer significant potential for energy recovery in water distribution networks (WDNs) by converting excess hydraulic head into electrical power. However, their effective operation is hindered by the lack of performance data under turbine conditions, as manufacturers typically provide curves only for pump mode. While empirical correlations and Computational Fluid Dynamics (CFD) exist, they suffer from limited generalizability, high computational costs, or an inability to provide real-time operational guidance. Furthermore, existing data-driven models often function as "black boxes," failing to provide explicit analytical relationships necessary for direct control implementation. There is also a gap in translating optimization results into simple, implementable speed recommendations and a lack of robustness analysis regarding input uncertainties, particularly flow rate measurement errors.

Methodology
The study proposes an integrated, data-driven framework combining Self-Organizing Polynomial Neural Networks (SOPNN) with the Differential Evolution (DE) algorithm. The methodology is grounded in 1,059 long-term field data records collected from an operational PAT installation in the Tehran water distribution network.

Data Preprocessing: Operational variables (flow rate, head, power, speed) were normalized against Best Efficiency Point (BEP) values to create dimensionless variables ( $Q^*$ , $H^*$ , $P^*$ , $N^*$ ), enhancing model generalizability across geometrically similar systems.
Surrogate Modeling (SOPNN): Three independent SOPNN models were developed to predict dimensionless head, power, and efficiency directly as functions of dimensionless flow rate and rotational speed. Unlike traditional approaches that calculate efficiency indirectly, this study modeled efficiency explicitly to avoid error propagation. The SOPNN algorithm automatically determined the optimal network architecture, resulting in explicit second-order polynomial equations.
Optimization (DE): A constrained Differential Evolution algorithm was employed to determine the optimal rotational speed ( $N^*_{opt}$ ) for any given flow rate. The objective was to maximize efficiency while satisfying minimum head ( $H^* \geq 0.85$ ) and power ( $P^* \geq 0.3$ ) constraints.
Analytical Speed Relationship: To facilitate real-time implementation, a secondary SOPNN model was trained on the optimization results to derive an explicit analytical equation mapping operating conditions directly to the optimal speed, eliminating the need for iterative optimization during operation.
Sensitivity and Uncertainty Analysis: A comprehensive analysis was conducted, including local gradient assessment, global parametric sensitivity, constraint perturbation, and Monte Carlo simulations to evaluate the robustness of the speed recommendations against measurement inaccuracies.

Key Contributions
The paper identifies five principal novelties:

Real-World Validation: Models were developed and validated using extensive, long-term field data from an actual WDN, capturing real hydraulic variability and noise, rather than relying on limited laboratory or synthetic data.
Direct Efficiency Modeling: Turbine efficiency was modeled directly as an explicit function of operating variables, reducing cascading prediction errors associated with intermediate head and power estimations.
Constrained Optimization: The framework integrates SOPNN surrogates with DE to identify optimal speeds under varying hydraulic conditions while strictly adhering to minimum head and power constraints.
Explicit Analytical Control Law: The study derives a low-complexity analytical relationship for recommended rotational speed as a direct function of dimensionless flow rate and efficiency, enabling rapid set-point generation suitable for PLC/VFD integration.
Comprehensive Robustness Analysis: A multi-level sensitivity and uncertainty analysis quantifies the reliability of the strategy, identifying flow rate measurement accuracy as the dominant factor influencing recommendation reliability.

Results

Performance Prediction: The SOPNN models successfully reconstructed the PAT performance hill chart, capturing the nonlinear migration of the Best Efficiency Point (BEP) ridge. The models achieved high accuracy with explicit polynomial equations.
Optimization Outcomes: The optimal rotational speed was found to vary continuously with flow rate, exhibiting a nearly constant normalized slope of 0.975, consistent with classical pump affinity laws. The variable-speed strategy achieved efficiency improvements of up to 13.22% compared to fixed-speed operation, particularly under low-flow off-design conditions.
Sensitivity Findings: Global sensitivity analysis revealed that flow rate is the dominant factor influencing efficiency variations ( $S_Q \approx 0.72$ vs. $S_N \approx 0.47$ ). Monte Carlo simulations confirmed that flow rate measurement uncertainty is the primary source of variance in the recommended speed, while the SOPNN model's intrinsic uncertainty had a lesser impact.
Economic Viability: Based on the actual annual hydraulic profile, the optimal control strategy yields an average power gain of 2.07 kW and an additional annual energy generation of 18,133 kWh. This corresponds to an annual revenue of €1,632 (at a €0.09/kWh tariff). The estimated simple payback period is 1.23 years, with a 10-year Net Present Value (NPV) exceeding €10,000.

Significance and Claims
The paper claims that the proposed SOPNN–DE framework provides a computationally efficient, interpretable, and economically attractive decision-support tool. By bridging the gap between complex optimization and practical implementation, the framework enables real-time PAT speed recommendations without the computational burden of repeated iterative optimization. The study asserts that this approach is particularly valuable for future smart water distribution systems where adaptive energy recovery is required under continuously changing hydraulic conditions. The authors emphasize that the derived analytical speed relationship allows for direct integration into industrial control systems (e.g., SCADA, PLCs), offering a practical pathway for retrofitting existing PAT installations to maximize energy recovery. The economic analysis further supports the framework's viability, demonstrating a rapid return on investment and strong cost-effectiveness for water utilities.

Technical Summary: Performance Prediction and Optimal Rotational Speed Determination of a Variable-Speed Pump as Turbine Using Surrogate Modeling and SOPNN-DE

More like this