Improving Forecast Scores by Filtering Short Waves

The enhancement of model resolution is the most popular modern approach to improving weather forecasting. As we know, high resolution models allow for the meso-scale phenomena to be seen and analyzed in more detail. However, abnormal short waves may appear with time and contribute misleading short wave energy values to the long wave terms. Therefore, finding ways to utilize the advantages of high-resolution models while maintaining the predictability of low-resolution models is a crucial issue that meteorologists working diligently to resolve.

NCEP's operational global forecast system (GFS) has two segments, each spanning one week, for weather prediction: Week one is high resolution of T574 for weather forecast and week two is coarse resolution of T190, which is used for ensemble forecast in climate prediction center. The reasoning behind reducing the resolution for the week two forecast is not only to decrease the computing resource required, but to increase accuracy of the predictions as well. Based on our previous experiments, we found that high resolution does not improve the anomaly correlation scores for week two forecast. This led us to the conclusion that the lack of increased accuracy despite the increased resolution is caused by the abnormal short waves created by the high-resolution model, which have bad feedback to long wave terms. Thus the abnormal short waves must be filtered out according to time to prolong the predictability of the high resolution model.

In this study, we will present numerical experiments with changing resolutions to filter out abnormal short waves in NCEP GFS. As a result, the sensitivity of the model's resolution for weather prediction scores can be presented. The purpose of filtering out short waves is that since we know long wave terms are the key factors that affect the scores, keeping the healthy long waves is intended to improve the scores. The primary goal of this research is to seek a simple way to enhance and prolong model predictability. The details and end results will be demonstrated in the presentation.