In the intricate world of water resource management, large-scale hydrologic models are the unsung heroes, working behind the scenes to forecast and plan our most precious resource. Yet, these models are not infallible. A recent study, led by A. Husic from the Department of Civil and Environmental Engineering at Virginia Tech, is shedding new light on the strengths and weaknesses of these models, using a novel approach that combines machine learning with traditional hydrologic science.
The study, published in the journal ‘Hydrology and Earth System Sciences’ (or ‘Hydrologie und Erdssystemwissenschaften’ in German), focuses on the National Water Model (NWM) and the National Hydrologic Model (NHM). These models are crucial for water resource planning and forecasting, but their predictive power is influenced by a complex interplay of climate, topography, soils, and land use.
Husic and his team employed a proof-of-concept approach using interpretable machine learning techniques to assess model sensitivity and process deficiency. They trained a random forest model to predict the Kling–Gupta efficiency (KGE) of streamflow predictions for 4,383 stream gauges across the conterminous United States. The team then used interpretable Shapley values to explain the local and global controls that 48 catchment attributes exert on KGE prediction.
The results were revealing. “We found that soil water content is the most impactful feature controlling successful model performance,” Husic explained. This suggests that soil water storage is particularly challenging for hydrologic models to resolve, especially in arid locations.
The study identified several nonlinear thresholds beyond which predictive performance decreases. For instance, soil water content less than 210 mm, precipitation less than 900 mm per year, road density greater than 5 km per square kilometer, and lake area percent greater than 10% all contributed to lower KGE values.
These findings have significant implications for the energy sector, which relies heavily on accurate water resource planning and forecasting. Improved hydrologic models could lead to more efficient water use in energy production, from cooling power plants to hydraulic fracturing in oil and gas extraction.
Moreover, this research demonstrates the potential of using data-driven techniques to interrogate process-based models. As Husic put it, “This study shows the utility of combining machine learning with traditional hydrologic science. It has broad applicability and potential for improving the next generation of large-scale hydrologic models.”
The study’s innovative approach could pave the way for more accurate and reliable hydrologic models, benefiting not only the water resource management sector but also the energy industry and beyond. By identifying key areas for improvement, this research is a significant step towards enhancing our ability to predict and plan for our water resources in an increasingly uncertain climate.