Optimizing Accuracy of National Blend of Models Forecasts Using an Analog Process

Forecasts for many standard weather elements (temperature, wind, precipitation, etc.) created by the National Weather Service (NWS) are becoming increasingly reliant on the National Blend of Models (NBM) ensemble system, which includes a full suite of probabilistic forecast information. Producing accurate forecasts is important for the NWS in providing high-quality decision support services to core partners.

Using archived model data, an opportunity exists for identifying improvements to the forecasts by examining verification of similar past forecasts. This so-called “Analog Process” has been applied with great success for several years at NWS Boise, ID, for temperature, dew point, and wind forecasting.

The process consists of pattern-matching a current forecast with a set of past forecasts at each individual gridded location, determining the errors made by comparing with accompanying observations (via the Unrestricted Mesoscale Analysis (URMA) database), and subtracting those errors in advance from the current forecast. The underlying assumption is that the current forecast will make the same errors as the similar past forecasts. The past forecasts supply topographical detail to the current forecast through these errors, information not otherwise available from the models.

Since this process doesn’t perfectly match previous forecasts to the current forecast, determining the matching closeness is instead calculated as a similarity score. The similarity scores of the past forecasts are sorted to put the closest matches at the top of the list. Then the errors, weight-averaged by the similarity scores, are subtracted from the current forecast.

Several questions arise: First, is it even necessary for the past forecasts to be similar to the current forecasts? If yes, then how many past cases should be examined, and after sorting, how many should be used in the forecast adjustment? Next, what is the best way to quantify similarity -- should the current and past cases be compared directly or should the means be subtracted out first and only the residuals compared? If only one past case is used, how much of that case’s error should be subtracted from the current grid?

Answering these questions consists of testing and verifying various options and comparing results. This presentation will focus on the process for applying the analog forecast approach, with results showing improvements to specific forecasts from the NBM to assist NWS forecasters in providing high-accuracy forecasts.