1.3 Previous studies - Surrogate Modelling in Urban Water Networks
Previous studies have reviewed the application of metamodels in water resources. Razavi et al. (2012b) outline taxonomies, practical details, and advances of these SMs in water resources along with recommendations for future research. Among the multiple insights of this work, they highlight the non-trivial effort to choose the right metamodel approach to the problem at hand and advocate for further research on these methods, especially in their assessment and validation. Furthermore, in the same year, Razavi et al. (2012a) numerically assessed metamodeling strategies in computationally intensive optimization, showing that metamodeling is not always a reliable approach, especially for complex response surfaces. The authors also warned about the inappropriateness of neural network models when having a limited computational budget. Later, Broad et al., (2015) presented a formalized qualitative process to determine the most suitable scope for a metamodel based on the evaluation of a fitness function to maximize fidelity. Hadjimichael et al. (2016) reviewed the application of AI methods to UWS management and their integration with decision support systems. While valuable, these published reviews give low emphasis to SMs for UWNs, and do not account for the recent growth in machine learning-based surrogate models (MLSMs) driven by the rapid advancements in AI.
This study aims to fill this gap by assessing the current state of MLSMs for UWNs in order to propose future directions based on identified outstanding issues and recent developments in ML. To achieve this purpose, we applied the review methodology described in Section 2 to review 31 published applications of metamodels for water networks. The results of the review are reported and discussed in Section 3, while major current gaps are detailed in Section 4. We propose future research directions in Section 5 and provide conclusions in Section 6.
2 Materials and Methods
We conducted a semi-systematic (Snyder, 2019) review of MLSM applications for UWNs to synthesize the state-of-the-art of the field. The review integrates the multiple applications of metamodels across water network applications, and explores them in a transversal manner. First, we searched journal papers in which MLSMs were applied to UWNs. Second, we determined a set of criteria to assess the relevant characteristics when applying these metamodels to UWNs’ problems.