This paper reviews two alternatives for reducing structural uncertainty in medium-term hydro-climatic forecasting. The first is a static ensemble average, illustrated here using the Multiple Reservoir Inflow Forecasting System, a nonparametric probabilistic forecasting model that relates streamflow to climate predictors, and generates monthly sequences of multi-site flow from the present for the coming 12 months. Instead of forming a single predictive relationship, multiple constituent models, each having their own unique predictor variable sets, are formed. A weighted probabilistic combination of these constituent models completes the static ensemble average. The second alternative is a dynamic ensemble average that allows constituent models to change importance with time, model weights evolving as a function of these weights at preceding time steps. Dynamic model combination is demonstrated here for first combining multiple sea surface temperature anomaly forecasts to produce a global sea surface temperature anomaly field, and then using the dynamically combined sea surface temperature anomaly (SSTA) field to concurrently ascertain inflows at multiple locations in a semi-arid Australian catchment. The paper concludes by identifying scenarios under which one would expect to see improvements as a result of static or dynamic model combination, and provides suggestions for further research in this area.
Keywords: dynamic combination, model uncertainty, probabilistic forecasting, rainfall, seasonal forecasting, streamflow