Improving Water Forecasting with Bayesian Averaging of Multiple forecasts

It is theoretically well established that averaging skillful forecasts from multiple models improves upon forecast from each individual model. In this presentation, we describe the Multi-model Merging Module (MMM) under development that performs optimal blending of operational streamflow forecast from multiple sources, and share the initial results for the Middle Atlantic River Forecast Center’s (MARFC) service area. The MMM’s blending mechanism is based on the Bayesian Model Averaging (BMA), and it supports characterization and communication of forecast uncertainty via an integrated suite of algorithms and diagnostic displays. We use BMA to combine up to 6 different streamflow forecasts from the RFC’s Community Hydrologic Prediction System, the National Water Model, the Hydrologic Ensemble Forecast Service, and the Meteorological Model Ensemble Forecast System forced by GEFS, NAEFS and SREF for lead times of 1 to 10 days. We share the initial evaluation results and findings from ten-fold cross validation, including error statistics for all amounts of discharge and for increasingly large amounts.