How can we best use climate information and hydrologic initial conditions to improve seasonal streamflow forecasts?
Abstract:Over the last decades, a number of forecasting centers around the world have offered seasonal streamflow predictions, using methodologies that span a wide range of data requirements and complexity. In the western United States, two primary approaches have been adopted for operational purposes: (i) development of regression equations between future streamflow and in situ observations (e.g. rainfall, snow water equivalent), and (ii) ensemble hydrologic model simulations that combine initial watershed moisture states with historically observed weather sequences for the forecast period (e.g., Ensemble Streamflow Prediction, ESP). Nevertheless, none of these methodologies makes use of analyzed or forecast climate information, which might increase the skill of seasonal predictions. Further, there is a need to better understand the marginal benefits of using more complex methods (from statistical to dynamical) and different types of information.
In this work, we provide a systematic intercomparison of various seasonal streamflow forecasting techniques, including: (1) a dynamical approach based on conceptual hydrologic modeling and ESP, (2) statistical regression using climate information and/or initial hydrologic conditions, (3) an ESP trace weighting scheme based on analog climatic conditions, and (4) combination of dynamical and statistical forecasts (i.e. hybrid approach). Climate information is taken from the NCEP CFSv2 reanalysis and reforecast datasets. These methods are tested for predicting spring (e.g., May-September) runoff volumes at case study basins located in the US Pacific Northwest, and results obtained for several initialization times are evaluated in terms of accuracy, probabilistic skill and statistical consistency. Preliminary results show that for earlier initialization times (October 1 to December 1), statistical and hybrid techniques that make use of climate information outperform ESP in terms of correlation and probabilistic skill. Although ESP at times provides the best correlation between forecasts and observations after January 1, improved probabilistic skill is obtained through statistical and hybrid techniques that instead rely on initial hydrologic conditions.