Evaluating Medium-Range Forecast Performance of Regional-Scale Circulation Patterns

Abstract

Abstract Accurate subseasonal-to-seasonal (S2S) weather forecasts are crucial to making important decisions in many sectors. However, significant gaps exist between the needs of society and what forecasters can produce, especially at weekly and longer lead times. We hypothesize that by clustering atmospheric states into a number of predefined categories, the noise can be reduced and, consequently, medium-range forecasts can be improved. Self-organizing map (SOM)-based clustering was used on daily mean sea level pressure (MSLP) data from the North American Regional Reanalysis to categorize the synoptic-scale circulation for eastern North America from 1979 to 2016 into 28 discrete patterns. Then, using two goodness-of-fit metrics, the relative skill of four different forecasting methods over a 90-day lead time was studied: 1) a circulation pattern (CP) forecast, 2) raw forecast output from the Climate Forecast System (CFS) operated by the National Centers for Environmental Prediction (NCEP), 3) a simple climatology forecast, and 4) a simple persistence forecast. As expected, forecast skill of both the CP forecast and the raw CFS forecast generally decreased rapidly from the first day, coming to parity with the skill of climatology after 10–12 days when using correlation, and at 7–10 days when using the root-mean-square error (RMSE). Most importantly, this study found that the CP forecast was the most skillful forecast method over the 8–11-day lead time when using RMSE. On a spatial basis, the skill of the CP forecast and the raw CFS decreases latitudinally from north to south. This study thus demonstrates the potential utility of categorical or circulation pattern–based forecasting at 1–2-week lead times.