Plain Language Summary
Analysis and improvement of urban water networks requires hydrodynamic models. Since these models are computationally expensive, researchers and engineers often resort to fast alternatives known as surrogate models. With the rise of artificial intelligence, machine learning methods have been increasingly used for surrogate modelling of urban water networks. In this study, we thoroughly reviewed recent papers on the field to outline the current state-of-the-art and propose future research directions. While many successful applications already exist, we found that these models have three main limiting factors: i) they need large amounts of data, ii) they are not explainable, and iii) they are too specific to each case. We argue that researchers can overcome these limitations by considering recent advancements in artificial intelligence and implement modeling techniques that better leverage the structure of the underlying data. Other promising direction include developing comprehensive benchmark databases and leveraging surrogate models for more complex applications.
1 Introduction
Urban water networks (UWNs) comprise drinking water distribution and urban drainage systems (WDS and UDS). The former are responsible for supplying drinking water to cities and the latter for evacuating wastewater and stormwater runoff. These infrastructures are a fundamental part of the city and are directly linked to its development (Brown et al., 2009). Each of these systems faces challenges to improve and maintain quality service in a dynamic urban environment under a widening range of climatic conditions; especially, in a climate-changing situation. Designing, optimising, and intervening in these systems requires approximating their hydraulic behaviour. Several models have been developed in the past years for simulating UWNs. Traditional modelling approaches are either based on accurate description of the physical processes or rely on simplified conceptual approaches; nonetheless, the former usually entail computationally expensive calculations while the latter lack fidelity. Applications such as optimisation, real-time modelling, and uncertainty analysis need an efficient model for evaluating the performance of a system multiple times or as fast as possible. Consequently, they require short execution times while maintaining a sufficient level of detail.