Post-processing NWM output with spatially distributed turbidity sensing to forecast turbidity loading and source for reservoir operation management

Objective:

Develop turbidity forecasts based on NWM output in a comparative approach to existing sediment loading forecast models used by New York City Department of Environmental Protection (NYC DEP) and explore the capacity of machine learning models to improve those forecasts

Abstract:

Reservoir operations and management increasingly extend beyond considerations of
water volume and water level to include water quality, especially in situations where reservoir
water is unfiltered or treatment of a particular contaminant is onerous. A primary concern
regarding reservoir water quality is often turbidity – a measure of water clarity that is chiefly
impacted by how much sediment and other material is suspended in the water column – which
can impede reservoir operations when certain levels are exceeded. Because the National Water
Model (NWM) has been shown to have substantial utility for reservoir operations by providing
flow forecasts that inform anticipation of future water volumes, it is sensible to leverage this
forecasting capability to provide insight into future turbidity levels. Additionally, many prior
studies have suggested that turbidity is primarily influenced by water discharge and other
hydrologic parameters forecasted by the NWM (such as rainfall) as well as easily obtainable
watershed characteristics (such as geology), indicating turbidity prediction may be readily
achievable by coupling the hydrologic forecasting capability of the NWM to empirical models of
turbidity production. In this project, we employ such an approach in the Esopus Creek
catchment in the Catskills Mountains of New York State, which drains to the Ashokan Reservoir
of New York City water supply system and has been extensively monitored for the past decade.
We build off prior research to construct a machine learning-based model of turbidity as a
function of antecedent conditions, storm hydrology, and watershed characteristics. In initial
testing, this model, which leverages the distributed, high-resolution sensor network present in
the Esopus watershed, outperforms the current model used by the New York City Department
of Environmental Protection (NYC DEP) in reservoir operations. These preliminary results
support the utility of our proposed approach and suggest that machine-learning models built on
understanding of watershed processes can be fed NWM forecast products to successfully
anticipate future turbidity loading. Work in the second and third year of the project will
continue to fine-tune turbidity models with additional site-specific data from the Esopus sensor
network to further improve forecasting capabilities and enhance forward-thinking reservoir
operations. In particular, the next steps will aim to forecast not only turbidity levels, but also
anticipate where in the watershed turbidity will be produced. This type of source-specific
forecasting has substantial importance for both operational planning and management efforts
(e.g., erosion mitigation) to suppress sediment loading to Esopus streams. Overall, results of this
project will emphasize the capabilities of the NWM to extend beyond hydrologic forecasts and
provide a blueprint for others interested in leveraging such abilities to produce water quality
predictions.