Scientific progress on seasonal streamflow forecasting

Forecasting streamflow over the next month, season and multiple seasons is a challenging task, because the predictability of climate over these forecast horizons is low. On the other hand, initial conditions of soil moisture, groundwater and other water stores in a catchment can have some relatively predictable effects on streamflow months ahead. State-of-the-art forecasting methods aim to (1) quantitatively capture, as much as possible, both sources of streamflow predictability and (2) statistically represent the remaining predictive uncertainty in a reliable manner. Under the Water Information Research and Development Alliance, we have been working towards these aims by developing a number of forecasting methods for seasonal streamflow forecasting. We developed both statistical and dynamic forecasting approaches. The Bayesian joint probability (BJP) modelling method is used to statistically capture observed relationships between future streamflow and predictors that indicate the initial catchment and climate states. The BJP method may use a single model by selecting the best predictors, or use multiple models through model averaging. A dynamic approach involves forcing a hydrological model with downscaled rainfall forecasts from a global climate model to produce streamflow forecasts. The statistical and dynamic forecasts were found to be similarly reliable in ensemble spread, but varied in skill by catchment and season. Merging statistical and dynamic forecasts was able to take advantage of the better performing model and provide a consistent set of forecasts to the users. Models were initially developed for forecasting total streamflow volume in the following three months. To meet user demand, a model for generating forecast guided stochastic scenarios (FoGSS) was developed. The model generates forecasts of monthly volume of streamflow out to 12 months, in the form of ensemble time series. As forecast skill decreases with lead time, the forecasts become more like natural stochastic scenarios that follow the historical distribution of streamflow.