Hybrid Dynamical-Statistical Seasonal Forecasts with Weather Types

Dynamical forecast models provide a foundation for seasonal forecast systems, but systematic errors may arise for various reasons, including insufficient spatial resolution, insufficient ensemble size, and errors in physical parameterizations. Despite these flaws, the ability of dynamical models to simulate the sources of prediction skill (e.g., ENSO) and their large-scale circulation responses allows us to draw from empirical predictor/large-scale circulation relationships to compensate for these shortcomings. In this study we use the framework known as weather types (WTs) to act as the mediator for a hybrid dynamical-statistical seasonal forecast system. WTs are large-scale, quasi-stationary circulation patterns that, in this application, are determined by k-means clustering of geopotential height. We generate seasonal forecasts for December – February over eastern North America by taking dynamical model forecasts of WTs and then using empirical relationships to translate these WT forecasts into probabilistic temperature and precipitation forecasts. We use hindcasts from both a lower resolution (CM2.1) and higher resolution (FLOR) dynamical forecast model from the Geophysical Fluid Dynamics Laboratory (GFDL). This application of WTs essentially serves as a pattern-dependent bias correction and downscaling approach. We evaluate the performance of the hybrid dynamical-statistical forecasts in the context of more conventional post-processing methods.