Data for thesis chapter "Suitability assessment of GNN-based metamodels for BGI allocation"

Abstract

This dataset supports Chapter 5, Suitability assessment of GNN-based metamodels for BGI allocation, of the PhD thesis Metamodeling of Urban Drainage Systems with Graph Neural Networks (Delft University of Technology), by Alexander Garzón, Zoran Kapelan, Jeroen Langeveld and Riccardo Taormina. The chapter assesses whether graph neural network (GNN) metamodels are suitable surrogates of SWMM for optimising the allocation of Blue-Green Infrastructure (BGI) in an urban drainage system. Suitability is assessed along two dimensions, accuracy and ranking preservation, for the total combined sewer overflow (CSO) volume of the Loenen (NL) system under a climate-adjusted (2085) design storm with a 2-year return period. This dataset contains the SWMM model of the Loenen system (.inp), the available BGI areas per subcatchment (.xlsx), the design storm and the runoff lookup tables used by the metamodels (.csv), and the weights (.pt) and hyperparameters (.yaml) of the baseline and adapted metamodels. It also contains the 500 evaluated BGI proposals and the 100 fine-tuning proposals (.csv). The results are the CSO volumes, rankings, accuracy and ranking-preservation metrics (.csv), and the figures of the chapter (.pdf). The raw simulations (about 35 GB) are not included; they can be regenerated with the corresponding code. This work is supported by the TU Delft AI Labs & Talent Programme.