Extensibility of a Machine Learning Model for Stormwater Basin Design and Retrofit Optimization Through a User-Friendly Web Application

Abstract

To sequester particulate matter (PM) and chemical loads, stormwater basins are a common infrastructure component of transportation land use systems in a planning/design or a retrofit phase. Basin design relies on residence time (RT), capacity–inflow ratio (CIR), and continuous stirred-tank reactor (CSTR) concepts to provide presumptive guidance for load reduction. For example, 14-day (or 21-day) RT is presumed to provide load reduction for total suspended solids (TSS) (e.g., commonly 80%), and for total phosphorus (TP) (e.g., commonly 60%). Despite such guidance, most existing basins are impaired—not meeting presumptive guidance, whether for TSS, nutrients, or chemicals (e.g., metals). RT guidance results in high initial basin design costs (land and construction). For impaired basin retrofit, RT guidance generates a significantly higher cost (area/volume increase), if such a retrofit is even feasible considering proximate infrastructure constraints. This study presents a computational fluid dynamics (CFD)-machine learning (ML) augmented web application, DeepXtorm, for cost–benefit optimization of basin design and retrofit. DeepXtorm is examined against basin costs from as-built data, RT, CIR, and CSTR (deployed in the Storm Water Management Model [SWMM]) over a range of load reduction levels of TSS (60%–90%) and TP (40%–60%). For a given load reduction requirement, DeepXtorm demonstrates up to an order of magnitude (or more) lower basin cost compared with RT, CIR, and CSTR models, and greater than an order of magnitude compared with as-built cost data. DeepXtorm represents a more robust and cost-effective tool for stormwater basin design and retrofit.