Accelerating optimal operation of water distribution systems with surrogating models

Pump operation in water distribution networks (WDNs) represents a major portion of energy consumption in water supply systems, rendering operational optimization a critical task. Conventional optimization approaches depend on hydraulic simulators to evaluate objective functions and constraints, yet the associated computational burden can restrict their use in real-time or near-real-time contexts. This study evaluated the feasibility and implications of substituting hydraulic simulators with artificial neural network (ANN)-based metamodels within pump operation optimization processes. Three feedforward multilayer perceptron neural networks were developed to predict pump energy consumption, tank levels, and minimum network pressure, and were integrated into the particle swarm optimization (PSO) algorithm as replacements for the hydraulic simulator. The proposed PSO–ANN framework was assessed using the Anytown benchmark network and compared with the conventional PSO algorithm coupled with the EPANET simulator. A total of 100 independent optimization runs were conducted for each approach, enabling a statistically robust comparison. The results indicated that the ANN-based approach consistently yielded hydraulically feasible solutions, with minimum pressures above required thresholds and tank levels maintained within acceptable operational ranges. However, the surrogate-based optimization exhibited a more conservative behavior, resulting in average operational costs approximately 8% higher than those achieved with the simulator-based approach. Despite this reduction in economic optimality, the significant reduction in evaluation time highlights the potential of ANN-based metamodels for real-time applications and decision-support contexts in which computational efficiency is paramount.