The multi-model machine-learning metasystem (M4) is the prototype next-generation mathematical and software model for operational water supply forecasting (seasonal river flow volume prediction) built for and employed by the Snow Survey and Water Supply Forecasting (SSWSF) Program of the US Department of Agriculture (USDA) Natural Resources Conservation Service (NRCS). NRCS water supply forecasting (WSF) operations span close to 600 forecast locations across the American West and Alaska. M4 will replace the linear principal component regression (PCR) method that the SSWSF Program has used as its primary (but not sole) WSF model in this West-wide operational role since the early 1990s. M4 leverages the proven performance and established stakeholder buy-in of a broadly PCR-like architecture and philosophy but augments it with multi-model ensemble modeling for improved forecast stability and accuracy, explainable machine learning/artificial intelligence (ML/AI) for improved forecast accuracy while maintaining easily relatable geophysical ‘storylines’ around the forecast, evolutionary computing to support global feature optimization, AutoML algorithms and carefully selected defaults to make it accessible and easy for users whose expertise is in areas other than data science, more advanced statistical methods for prediction uncertainty estimation under non-Gaussian and heteroscedastic error distributions, and other advances that were directly motivated by the day-to-day operational requirements of a governmental service-delivery organization performing operational WSF at scale. It is written in the free, open-source R scientific/statistical computing language.
This initial M4 release to GitHub is a pilot-phase prototype. It has been extensively documented in the peer-reviewed literature and has experienced retrospective and live operational testing at NRCS, but it hasn’t been integrated yet into NRCS production systems or NRCS institutional protocols for issuing official WSFs. NRCS encourages applied water resource scientists and engineers and the STEM research community to explore M4 and contribute to developing future versions of it. The design characteristics of the metasystem might also make it useful for applications beyond environmental prediction.
Details and caveats are given in the User Manual available here, which also includes three publications describing the method, its rationale, and its performance. For further information or to provide feedback, contact Sean Fleming (sean.fleming@usda.gov) or Beau Uriona (beau.uriona@usda.gov).