Using Artificial Neural Network surrogate models to reduce the calculation cost of leak detection in water distribution networks

The leak detection parameters in the inverse transient analysis (ITA) are obtained in an inverse approach by solving a nonlinear programming problem using metaheuristic algorithms such as the genetic algorithm (GA). Despite its high capability in deriving the leak detection parameters, the ITA method is computationally complex and costly as it uses the (GA). The reason can be attributed to the random behavior and gradual evolution of this algorithm inspired by the nature. This study aimed to increase the computational efficiency by employing surrogate models in the optimization process of the ITA method. The surrogate model is in fact a simulated sample of the main model capable of approximately calculating the objective function in a fraction of a second. The way these models are integrated into the optimization model highly affects their success or failure. To this end, two algorithms incorporating population-based surrogate models, namely (Pre-selection Strategy) PS and (Best Strategy) BS, were presented. To evaluate and compare the results, a distribution network was used to identify the leak detection parameters. The results indicated an increase in the computational efficiency compared to the ITA method integrated with the (GA). The PS algorithm demonstrated the highest performance by reducing the objective function and time complexity by 58% and 78%, respectively.