Leveraging emerging sensing technology and machine learning to improve and expand hydrological forecasting to hyper-local scales with NWM-coupled adaptive sensor networks

Objective:

Complete an assessment of NWM performance in montane headwater catchments and explore the nature and distribution of NWM error

Abstract:

Much of the northeastern U.S. is dominated by montane headwater catchments, and recent
flooding events illustrate the need for accurate and timely forecasts for such systems. However,
model forecast accuracy is often reduced in mountainous regions due to sparse gaging, complex
topography, and spatially heterogeneous rainfall/runoff patterns. This project aims to improve
the performance and expand the capacity of the National Water Model (NWM) forecasting in
montane headwater catchments by achieving three main objectives: 1) assess the NWM
performance in montane headwaters, which will improve our understanding of the combinations
of geophysical and hydro-climatic forcings that govern streamflow at the subwatershed scale
during different events and across seasons, and develop a machine learning correction algorithm
that improves flow forecasts in these systems that could be operationalized, 2) deploy a
distributed water level sensing network in focal Vermont watersheds in different river corridor
environments upstream and downstream of gages that will allow us to use the NWM to forecast
water level across different reach environments, and 3) predict water levels at different locations
within each focal watershed based on the high-frequency water level data and NWM flow
forecasts using machine learning. Upon completion of this project, users and managers of streams
in montane watersheds will be able to easily access hyper-localized water level forecasts based
on short-range NWM discharge predictions that are post-processed via adaptive sensing and
machine learning models. Ultimately, we intend to develop an operational workflow that would
allow other communities across the country to improve the performance of these forecasts and
leverage relatively low-cost sensor technologies to provide distributed NWM-derived water level
forecasts across environments and infrastructures of concern. Improving (through correction
algorithms) and expanding (through distributed water level forecasting) the forecast capacity at
these sites and providing a template for others to do so will improve operational workflows and
extend water resources predictions, capabilities, and applications. Furthermore, the approaches
developed here will be particularly suitable for the modular and model-agnostic environment
envisioned for the NWM Next Generation Water Resources Modeling Framework (NextGen).