Big Data for Specific Places in Hydrologic Modeling

Abstract:

Unlike traditional streamflow models (e.g., conceptual and process-based models), deep learning models are most accurate when trained on large-sample, geographically diverse datasets [1]. Yet water managers and other forecast users often care about individual locations. How should we train models on big data when the objective is to achieve the highest possible accuracy at specific sites?

The approach our company has developed is to first train a “base” model using data from a large-sample dataset, then further train the models by freezing some of the layers in the deep learning architecture and reducing the learning rate to fine-tune to specific basins. These tuned models produce streamflow forecasts at both 10 day through 6 month horizons that have the highest local accuracy that we have been able to achieve and work well even when there is limited observation data available for training in a particular basin. This allows us to develop a single model structure applicable across most of the continental US and Europe that works in both gauged and ungauged basins [2], while also quickly producing high-accuracy predictions for specific locations.

In addition, individual users often require forecasts that emphasize accuracy in specific ways. For example, fish conservationists need accurate low flow prediction, and most customers require estimates of uncertainty. We describe how this approach can be tuned to match a range of forecasting metrics and requirements.

This abstract is intended to convey both innovations that increase the applicability of deep learning hydrology models to real-world applications, as well as lessons learned from our company’s experience providing operational streamflow forecasts for utility and environmental customers.

[1] Nearing, Grey, Frederik Kratzert, Alden Keefe Sampson, Craig S. Pelissier, Daniel Klotz, Jonathan Frame, Cristina Prieto, et al. 2020. “What Role Does Hydrological Science Play in the Age of Machine Learning?.” EarthArXiv. February 21. doi:10.31223/osf.io/3sx6g.

[2] Kratzert, Frederik, Daniel Klotz, Mathew Herrnegger, Alden K. Sampson, Sepp Hochreiter, and Grey S. Nearing. "Toward improved predictions in ungauged basins: Exploiting the power of machine learning." Water Resources Research 55, no. 12 (2019): 11344-11354.