Authors: Mirce Morales-Velazquez, Beverley Wemple, Donna M. Rizzo, Kristen Underwood and Andrew W. Schroth – University of Vermont
Title: A Machine Learning Approach for Enhancing National Water Model Streamflow Forecasts in Montane Headwater Catchments
Abstract: Headwater streams play a crucial role in maintaining water quality, quantity, flood generation, and nutrient mobility. However, they are often overlooked and present challenges for streamflow prediction due to complex topography, highly variable rainfall-runoff patterns, and limited gaging infrastructure to develop and inform models. The National Water Model (NWM) aims to provide streamflow forecasts across the United States, yet its performance in these sites, which cover 70-80% of the river network, remains challenging. Inaccurate forecasts in these low-order streams can contribute to extensive damage to local communities as illustrated by recent catastrophic floods in Vermont in July of 2023. This study evaluates the NWM retrospective dataset’s performance in headwater catchments in the northeastern US and, considering the discrepancies found, a machine learning correction model is developed to improve NWM performance in predicting hourly streamflow in these low-order streams. The post-processing correction model is developed using a physics-informed Light Gradient Boosting Machine (LightGBM) to relate archived NWM predictions and other ancillary data to measured hydrographs. Preliminary results show a dramatic improvement in the model’s predictive performance relative to observations based on commonly used metrics. Including water level measurements, even from relatively distant sites, greatly improved model performance. This demonstrates the potential to improve forecasts by deploying low-cost water level sensors in ungaged basins of interest, without the need for timely and costly rating curve development. We take the lessons learned from the retrospective case and apply them to develop a general framework that can be applied to enhance model performance in medium-range streamflow forecasts. This approach and the general framework presented for its forecast application offer advantages such as interpretability and ease of use, enabling widespread adoption. Furthermore, its design is model agnostic and also quite malleable, making it suitable for integration into the NextGen framework of the NWM to improve performance as model development continues for the foreseeable future.