Probabilistic Downscaling Methods for Developing Categorical Streamflow Forecasts using Climate Forecasts

Abstract:

Statistical information from climate forecast ensembles can be utilized in developing probabilistic streamflow forecasts for providing the uncertainty in streamflow forecast potential. This study examines the use of Multinomial Logistic Regression (MLR) in downscaling the probabilistic information from the large-scale climate forecast ensembles into a point-scale categorical streamflow forecasts. Performance of MLR in developing one-month lead categorical forecasts is evaluated for various river basins over the US Sunbelt. Comparison of MLR with the estimated categorical forecasts from Principle Component Regression (PCR) method under both cross-validation and split-sampling validation reveals that in general the forecasts from MLR has better performance and lower Rank Probability Score (RPS) compared to the PCR forecasts. In addition, MLR performs better than PCR method particularly in arid basins that exhibit strong skewness in seasonal flows with records of distinct dry years. A theoretical underpinning for this improved performance of MLR is also provided.