Machine Learning Surrogate Modeling in R for Rapid Screening of Green Infrastructure Hydrological Performance in Urban Stormwater Management: A Proof-of-Concept Study Using Synthetic Data

Source abstract

Background: Physically-based, coupled hydrological–low-impact-development (LID) models, such as the U.S. EPA Storm Water Management Model (SWMM), estimate green infrastructure (GI) performance in detail but are computationally expensive to run across many catchment, storm, and typology combinations. Methods: This methodological proof-of-concept develops an open-source R workflow (randomForest, xgboost, caret) on a synthetic dataset of 2000 catchment–storm–typology scenarios generated from prescribed non-linear equations of imperviousness, storm return period, and the coverage of four GI typologies. None of the scenarios were generated by SWMM-LID simulation or field monitoring. Random forest, XGBoost, and a linear baseline were trained to predict synthetic peak-flow attenuation and suspended-solid (TSS) removal. Results: On the held-out synthetic test set, XGBoost and random forest reached R2 values of 0.96 and 0.92 for peak-flow attenuation (linear baseline: 0.90) and 0.94 and 0.89 for TSS reduction (linear baseline: 0.79). These values show that the models learned the synthetic response surface; they do not measure predictive skill for real GI systems. Feature importance reproduced the typology weighting embedded in the data-generating equations. Surrogate inference took about 2–4 ms (measured), compared with tens of minutes typically reported in the literature for SWMM-LID runs (not measured here); this is an indicative comparison, not a controlled benchmark. Conclusions: This study is a methodological demonstration only and is not a validated hydrological surrogate. Real application requires retraining and validation using SWMM-LID simulation ensembles or field-monitored data.