Optimization of the Analog Ensemble Method

An analog ensemble is constructed by matching up the current forecast from a numerical weather prediction model with similar past forecasts, then using the past verifying observation (or gridded analysis) from each match as an ensemble member. Initial results indicate that when this approach is fully localized, an analog ensemble may provide the best efficacy for production of skilled probabilistic forecasts. The analog ensemble has several very attractive features including use of a higher resolution model (since only 1 real-time forecast is needed), no need of initial condition and model perturbation strategies, and a natural ability to produce reliable forecasts (i.e., no postprocessing required). This follow-on study investigates various aspects of analog ensemble design optimization, including: sensitivity to the length of the training data set, ensemble size, individual member skills, selection of predictor variables and their relative weight.

An additional question is whether the analog ensemble can capture flow-dependent uncertainty as well as an NWP ensemble. Could the best approach be a hybrid ensemble where m analogs are found for each member of a small n-member NWP ensemble, to produce a total of m×n members? Initial results indicate that the hybrid ensemble is superior to the NWP ensemble, but that the pure analog ensemble (using a higher resolution model) still provides the most skill at the lowest cost. The power of analog ensemble may be derived from its ability to resolve smaller scale phenomena and accurately depict those phenomena's uncertainty.