Monday, 12 January 2009: 4:15 PM
Replacing Missing Data for Ensemble Systems
Room 125A (Phoenix Convention Center)
Numerical Weather Prediction (NWP) ensemble forecasting systems were developed to assess the uncertainty in weather forecasts while being capable of producing a deterministic forecast. Several AI systems have been developed for post-processing to produce a deterministic forecast by an optimal combination of ensemble members. A major issue with all of these methods is the difficulty in attaining complete training datasets. Without a complete dataset, it can become difficult, if not impossible, to train and verify statistical post-processing techniques. To ameliorate this problem, an analysis of the treatment of missing data in ensemble model temperature forecasts is performed with the goal of determining which method of replacing the missing data produces the lowest Mean Absolute Error (MAE) of forecasts while preserving the ensemble calibration. This study explores several methods of replacing missing data, including applying persistence, an iterative imputation technique, a fifth degree polynomial fit, a Fourier fit, ensemble member mean substitution, and substituting the three day mean deviation from the ensemble mean. The methods are evaluated according to their effect on the forecasting performance of two ensemble post-processing forecasting methods, traditional (K-means) regime clustering and a 10-day performance weighted window. The methods are also assessed by verification rank histograms to determine the calibration of the ensembles. The analysis is performed on 48-hour temperature forecasts for four locations in the Pacific Northwest. For both post-processing techniques, the three day mean deviation method of replacing the missing ensemble forecasts outperformed all other methods by producing the lowest MAEs and appropriately calibrated ensembles.