Developing data-driven learning models to predict urban stormwater runoff volume

ABSTRACT

The Storm Water Management Model (SWMM) is a hydrological model for simulating and predicting runoff. Although powerful, SWMM can be computationally demanding. Therefore, we develop machine learning (ML) models to approximate the behavior of SWMM and expedite the task of predicting runoff. We perform a case study for the First Creek watershed in Knoxville, Tennessee, USA. We train ML models using rainfall data and subcatchment characteristics and apply feature engineering and clustering to objectively compare the outputs from SWMM and ML models. The results show that random forests can predict runoff volume accurately, with a Mean Absolute Error (MAE) of 0.006 (0.001) ${10}^6$ gallons, where predictions are made almost instantaneously. Hence, our proposed ML-based approach can accurately predict runoff while greatly reducing computational requirements, filling a critical need in the field.

KEYWORDS:

• Machine learning
• SWMM
• runoff volume prediction
• clustering
• SHAP values

Acknowledgements

We thank Mollie Turner for being instrumental in writing and running code for this study. We also acknowledge the City of Knoxville Stormwater Engineering Division for providing us with the SWMM model of First Creek.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

Data are available through this Github repository for this research (https://github.com/rwood25/Developing-Data-Driven-Learning-Models-to-Predict-Urban-Stormwater-Runoff).

Additional information

Funding

This material is based upon work supported by the National Science Foundation under Grant No. [CMMI-1634975].