Embedding Neural Networks in Multiobjective Genetic Algorithms for Water Distribution System Design

Abstract

This paper investigates the use of artificial neural networks (ANN) in combination with genetic algorithms (GA) to improve computational efficiency in solving multi-objective water distribution system design problems. Offline training of ANNs have been adopted in previous studies, in which the ANNs are trained only once, before the optimisation process starts, and then are used as a substitute of the full hydraulic model during the whole optimisation process. This has two main problems: it is very computationally expensive to generate a large number of samples, which are normally required to cover the whole domain of the decision space; and the ANNs cannot be trained with more samples in the most concerned (optimum) regions of the decision space, resulting in a reduced predictive accuracy in optimal solutions. This study presents a new ANN-GA algorithm for multi-objective optimal design of water distribution systems, in which the ANN is trained periodically to directly approximate the network performance measures during the optimisation process. The methodology is tested on two case studies: the New York Tunnels network and the Anytown network, and the optimal solutions obtained are compared to the corresponding solutions from the NSGA— II method. Results show that the method proposed can achieve a significant reduction in computation time, without loss in accuracy of the optimal solutions.