Generalising human heuristics in augmented evolutionary water distribution network design optimisation

Abstract

The use of evolutionary algorithms (EAs) for finding near optimal water distribution network (WDN) designs is well-established in the literature. Even though these methods have the ability to generate mathematically promising solutions based on defined objective function(s), the resulting solutions are not necessarily suitable for real-world application. This is because of the size, complex and non-linear nature of WDNs, which make it difficult to define important factors that a water engineer or an expert needs to consider during the design process in an objective function. Incorporating an expert in the optimization process has been used to deal with this problem and to guide an EA’s search toward obtaining more practical solutions. Accordingly, this study proposes a methodology for capturing and generalizing engineering expertise in optimizing small/medium WDNs through machine learning techniques, and integrating the resultant heuristic into an EA through its mutation operator to find the optimum design for larger WDNs. The combined interaction from different users on four small /medium benchmark WDNs from the literature were collected and used to train a decision tree model. Seven input features including current pipe diameter, velocity, upstream and downstream head deficient, pipe influence, flow and length are used to train the decision tree for predicting new diameter for a selected pipe. The resultant decision tree model is then applied to a larger network namely Modena to assess the ability of the HDH method. The results demonstrate better performance in comparison with a standard EA approach for finding minimum network cost.