Imagining and Building the Next Generation of the US Department of Agriculture Natural Resources Conservation Service Operational Water Supply Forecast System for the American West: Machine Learning Goes Mainstream
Abstract
Water supply forecasts (WSFs) are the basis for managing the complex water and power infrastructure of the American West. The US Department of Agriculture's Natural Resources Conservation Service (NRCS) operates the largest stand-alone WSF system in the region. NRCS and its precursors were among the first to begin monitoring snowpack and issuing WSFs in the 1920s, and in the 1990s NRCS revolutionized the field by introducing principal components regression (PCR) with prediction intervals based on cross-validated standard error; the method has spread to other institutions across North America. PCR is superior to standard multiple linear regression, and relative to extended streamflow prediction using process simulation models, it provides similar predictive skill and more reliable quantitative prediction uncertainty estimates at much lower cost. However, the system's age is showing and operational experience has revealed some issues. We began with an extensive review and assessment, including documentation of the system and capabilities and protocols that have been progressively added to it; more comprehensive statistical diagnostics; review of WSF state of the art and identification of potential alternatives; and assessment of climate change implications. A new system was then designed, built, and tested, which will become the basis for NRCS operational WSFs. PCR is a proven, trusted, and appropriate framework, and the new approach can be viewed as retrofitting PCR with modern data science. This multi-model metasystem combines techniques from artificial intelligence, advanced nonparametric statistical modeling, and evolutionary computing, selected for specific characteristics, such as ability to seamlessly accommodate nonlinear relationships and complex (heteroscedastic and non-Gaussian) prediction bounds, integrate physical knowledge, and improve ease and automation of model-building and operation. Testing shows improved accuracy and robustness.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.H31E..05F
- Keywords:
-
- 1847 Modeling;
- HYDROLOGY;
- 1873 Uncertainty assessment;
- HYDROLOGY;
- 1906 Computational models;
- algorithms;
- INFORMATICS;
- 1942 Machine learning;
- INFORMATICS