A dynamical performance-ranking method for predicting individual ensemble member's performance and its application to a smarter ensemble averaging

Ensemble forecast data becomes increasingly important in daily weather forecasts. However, given large amount of data provided by an ensemble prediction system, it's critical to extract useful information from an ensemble for forecaster to effectively use ensemble data in real time. Although each ensemble member should, by design, perform equal-likely on average, it's not so in each individual event of forecast. It is, therefore, desired to know the relative performance of each member for various reasons. Most of such existing methods to assess a member's performance are statistical based on a forecast's past performance. This study proposed a dynamical performance-ranking method, called Du-Zhou Ranking method, to predict relative performance of individual ensemble members. This approach is easier and cheaper in computation to be applied to post-processing than statistical approaches. As a demonstration, the method is then applied to a weighted ensemble averaging to see if it can improve ensemble mean forecast. Specifically, the results from this study show that the Du-Zhou Ranking method (1) works well in general especially for shorter range such as one day forecast, (2) works better in an multimodel ensemble environment than single-model ensemble one, (3) works better when model bias is small, and (4) works better when variation in performance among ensemble members is large. Comparing to simple equally-weighted ensemble averaging approach, the weighted ensemble mean forecast has smaller systematic error. This superiority of the weighted to the simple mean is especially true for small-size ensemble such as a 5-member ensemble but vanishes when ensemble size increases such as for a 21-member ensemble.