Hybrid dynamical-analog ensembles for forecasts with late medium-range lead times

The fundamental idea of analog forecasting is to find within an archive of observations or analyses a state that is similar to the current day's state, and to use the documented evolution of the chosen archived state to infer how the current day's state is apt to evolve. Analog forecasting was originally employed for deterministic forecasting purposes, and interest in the concept waned with the rapid advancement of numerical weather prediction and the recognition that because of the chaotic nature of the atmosphere even very good analogs are liable to display an evolution that is considerably different from that of the current state. However, the recent shift in the meteorological community toward probabilistic prediction has led to renewed interest in the analog forecasting concept, centered on the formation of distributions of analogs, or so-called analog ensembles. Analog ensembles have yielded improvement in short-range forecasts of ceiling and visibility at airports (Hansen 2000) and tropical cyclone tracks (Fraedrich et al. 2003). They have also proven to be of value in generating seasonal predictions of precipitation (Clark and Deque 2003).

This talk will present the idea that analog ensembles can be used to considerable advantage in formulating probabilistic forecasts with late medium-range lead times. The use of analog ensembles for this purpose is facilitated by seeking analogs to each member of a given dynamical ensemble, and by shifting the analog selection process from the initial point of a particular dynamical member's forecast trajectory to a point on the medium-range portion of that member's trajectory. On this portion of the trajectory the error variance of the dynamical member's state is sufficiently large so as to allow a sizeable number of analogs to be found even in the relatively small analysis and observation archives that exist today. Following the selection of analogs for each dynamical member, the entire collection of analogs can be combined with the dynamical ensemble to form a hybrid ensemble for prediction purposes. Since the members of the analog ensemble evolve with real atmospheric physics and dynamics, these members stand to reveal information about the forecast distribution's statistical moments that is unattainable from the members of the dynamical ensemble as a result of numerical model errors. Furthermore, the analog ensemble members provide for an increase in the sampling of the forecast distribution at very low computational cost.

As demonstration of the hybrid dynamical-analog ensemble concept, hybrid 500 hPa geopotential height ensembles are derived using the National Center for Environmental Prediction (NCEP) Global Forecast System (GFS) dynamical ensemble and a 30-year archive of NCEP reanalyses. The hybrid ensembles are constructed for the 240h, 288h, and 336h forecast lead times, and their performance is evaluated using a large set of diagnostics, including ensemble mean root-mean-square error, proportion of excessive outliers and ensemble range, measures of the spread-skill relationship, Brier score, and relative operating characteristic (ROC). The performance evaluation is carried out for a variety of hybrid ensembles that differ in terms of the similarity measure used in analog selection and in the total number of analogs selected. Results indicate that the hybrid ensembles show improvement relative to the dynamical ensembles in almost every diagnostic. Improvements in the Brier score and ROC are particularly pronounced for probabilistic forecasts of large geopotential height anomalies. This result is especially important given the burgeoning interest in the prediction of extreme events. The results encourage further investigation and investment in the concept of hybrid dynamical-analog medium-range ensembles.