Multi-objective optimization can help identify efficient and appealing designs of urban drainage systems. However, their application to large-scale problems is hindered by the computational cost of urban drainage simulation. We propose a novel disaggregation approach that allows simulating a portion of a drainage network while the remaining part is represented by a surrogate model that maps changes in the region of interest to hydraulic head time-series at synthetic nodes shared with the remaining part of the network. The proposed approach is demonstrated with an application to the many-objective optimization of sustainable urban drainage systems in two urban areas. The design problem's decision variables include the types of sustainable drainage systems, their combination within a subcatchment, their surface areas and spatial distribution, whereas the objectives include the minimization of capital cost, flood volume, flood duration, and total suspended solids or average peak runoff. The results show that the proposed disaggregation-emulation approach can provide an accurate representation of the system dynamics while significantly reducing the computational time compared to a model that simulates the whole network dynamics. Two alternative surrogate models are considered based on multilayer perceptron (MLP) and generalized regression neural networks (GRNN). MLP is found to be more accurate compared to GRNN at the cost of a larger computational time for the training process.
