Verification of Forecast Grids in the Central Region of the National Weather Service

The Central Region (CR) of the National Weather Service (NWS) is standardizing the population of the day 4 to 7 forecast grids in the Graphical Forecast Editor (GFE). The goals of the project are to provide consistent first-guess grids across the CR, and these grids should very often be of sufficient quality that they do not require much editing, if any. Verification of the grids for each of the thirty-eight weather forecast offices (WFO's) across CR was performed. CR encompasses a wide range of topography from flatter terrain of the Midwest to the mountainous terrain of Colorado and Wyoming, and this wide range of topography allows for a comparison of which models and blends perform better in complex terrain versus flatter terrain.

The grids verified fall into three categories. One category is deterministic forecast models (such as the GFS and ECMWF). The second category is statistical guidance, and grids of point statistical guidance were created by adjusting a forecast model field using the values in guidance. The third category is consensus models which are combinations of various models. A “bias-corrected” grid was calculated for each grid. The bias correction consists of adjusting each 2.5km by 2.5km point based on the grid biases over the past 30 days.

Using mean absolute error (MAE), a consensus of all the grids (from all three categories) verifies the best overall (smallest MAE) throughout the region. However, there are regional and temporal differences. The bias-corrected version of this grid performs the best over complex terrain. This bias-corrected grid performs the best for the shorter forecast periods, and bias correction becomes less useful for long time ranges. Bias correction is most helpful for temperature and it is not very helpful for dew point. It is hypothesized that bias correction is most helpful in complex terrain because it corrects the model fields for local terrain influences and because the coarser model grids cannot adequately represent the finer resolution terrain in the GFE grids.

While the consensus models have the lowest MAE, they do not adequately represent the variability of the weather especially in the longer time periods since they are an average of a potentially wide range of possible weather. Decision support is needed for the extended period to provide a range of possible forecasts and probability of occurrence, which is not provided by a single deterministic grid.