Comparison of Monte Carlo Algorithm, genetic algorithms and artificial neural networks for calibration of water supply networks using the epanet2toolkit

Abstract

A comparison of three water network calibration algorithms was performed using the R epanet2toolkit library. This coupling makes it possible to explore EPANET's hydraulic simulation and evaluation potentials and data analysis in R, with the main result of the work being the comparison of three calibration methods. In the calibration process by the Monte Carlo Algorithm, 100,000 roughness values ​​were randomly generated for each pipe section within the range of 0.008 to 0.09 and new pressure values ​​were generated with these roughnesses, while the calibration by the Genetic Algorithms method was used the rpy2 package that allows the use of R in Python, having 10,000 generations per simulation with 5% chance of mutation and 50% chance of crossover, admitting a deviation of ± 2 m.c.a for each pressure and the reduction of the average error. Finally, the Neural Network calibration also used the rpy2 package, with the network demand defined as the input layer and the output layer as the roughness of the pipes and for the hidden layer the input layer plus four neurons was defined. The results showed that in the smallest network the best performance was obtained by the Genetic Algorithms, followed by Monte Carlo, while the Neural Network had the worst result, and in the most complex network the Neural Network results obtained the best result, followed by the Genetic Algorithms and Monte Carlo. Thus, the potential of using Neural Networks for the calibration of more complex networks is observed, as well as its use combined with optimization techniques for the operation of water distribution networks, taking care to avoid situations of overfitting or underfitting.