Manifoldly Constrained Monte Carlo Optimization and Uncertainty Estimation for an Operational Hydrologic Forecast Model

River forecasts have two broad uncertainty classes: errors associated with meteorological forecasts, and those associated with the hydrologic model. We developed a technology (dubbed Absynthe) to address the latter error class in a practical and defensible way. The technique merges the proven, Monte Carlo-based Generalized Likelihood Uncertainty Estimation (GLUE) concept for model parameter identification with: (i) multiple performance goals defined by operational and physical considerations, including matching daily, seasonal, and annual flows as well as snowpack, as expressed via individual behavioural criteria and a net likelihood function; (ii) several moving (rank-based) constraints to assure non-pathological parameter sets, containing values that are physically plausible not only for each parameter individually but also collectively; and (iii) a hard constraint on snow-free elevation bands to force the surface meteorological component of the watershed model toward correct solutions. The result is an ensemble of parameter sets reflecting model uncertainty as captured in a loosely Bayesian framework. BC Hydro will combine these with ensemble NWP weather forecasts to generate uncertainty estimates for operational hydroelectric reservoir inflow forecasts.