Using Sensor Data to Dynamically Map Large-Scale Models to Site-Scale Forecasts: A Case Study Using the National Water Model

Abstract

There has been an explosive growth in the ability to model large water systems. While these models are effective at routing water across massive scales, they do not yet forecast the street-level information desired by local decision makers. Simultaneously, the increasing affordability of sensors has made it possible for even small communities to measure the state of their watersheds. However, these real-time measurements are often not attached to a predictive model, thus making them less useful for applications like flood warnings. In this paper, we ask the question: how can highly localized forecasts be generated by fusing site-scale sensor measurements with outputs from large-scale models? Rather than altering the larger physical model, our approach uses the outputs of the unmodified model as the inputs to a dynamical system. To evaluate the approach, a case study is carried out across the U.S. state of Iowa using publicly available measurements from over 180 water level sensors and outputs from the National Water Model. The approach performs well across a third of the studied sites, as quantified by a high normalized root mean squared error. A performance classification is carried out based on Principal Component Analysis and Random Forests. We discuss how these results will enable stakeholders with local measurements to quickly benefit from large-scale models without needing to run or modify the models themselves. The results are also placed into a broader sensor-placement context to provide guidance on how investments into local measurements can be made to maximize predictive benefits.