Machine Learning For Stochastic Flood Model Hydrograph Typing

Stochastic rainfall/runoff models are at the forefront of hydrologic modeling state-of-practice. These models have increasingly been used by the Reclamation Technical Service Center (TSC) to estimate flood magnitudes and associated return periods, along with uncertainty, for detailed flood hazard studies such as issue evaluations (IEs) and corrective action studies (CASs). Stochastic rainfall/runoff models simulate many thousands of potential flood realizations across frequency space to estimate probabilistic floods and support risk analyses. The resulting products of these studies are typically flood frequency curves representing peaks, volumes, and water surface elevations produced from a large number of modeled hydrographs. One challenge that arises with these large datasets comes when hydrographs must be used for additional analyses beyond determination of existing hydrologic loads, such as design, modification, or operational changes. In these scenarios, working with a smaller number of hydrographs becomes necessary. The process of selecting a subset of hydrographs is currently manual, time consuming, and dependent upon the judgement of the person tasked with selection.

The current project developed an automated hydrograph classification workflow using a two-stage classification procedure, first a self-organizing map (SOM) machine learning (ML) method followed by mean shift clustering. The SOM method groups hydrographs by evaluating their similarity in the shape and magnitude. The SOM groups are further refined with the mean shift clustering operation to yield a small number of hydrograph clusters that are representative of the range of behavior at a site. The developed ML workflow is an automated process with minimal user input that runs rapidly and scales to the number of hydrographs produced by stochastic rainfall/runoff models. The ML hydrograph classification workflow was tested across multiple gage and model instances. In each of these cases, the ML workflow was robust and produced a hydrograph classification that is representative of the site.