There are several approaches to using Model Output Statistics (MOS) to improve numerical weather predication (NWP) forecasts. The standard MOS approach is to use a long period of record as a training set with one set of calculated regression equations using one model for the entire period. For most cases and months the basic (non-regime) NWP - MOS system produces more accurate forecasts than the raw NWP forecasts. But there are times during transition months when the NWP-MOS approach can make the final forecast worse than the raw NWP forecast. Another problem when trying to provide forecasting services to users in remote locations is the lack of a long period of record to use as a training period.

To address the long period of record problem an approach that is typically referred to as Dynamic MOS uses only a recent history, typically a few weeks, as the training set with a continual recalculation of the MOS regression equations. This approach solves the problem of using MOS for stations without a long period of record, and it tends to help reduce the problem of MOS making the final forecasts less accurate than the raw NWP forecast. However, on average the dynamic MOS approach does not reduce the NWP model bias or MAE as much as the standard MOS.

Another approach to improve NWP forecasts, typically called regime MOS but also called Stratified MOS in some literature, uses a subset of cases to build the training set during a defined period based upon some criteria. If the regimes are selected correctly, a regime MOS can very effectively reduce NWP bias for the days standard MOS does poorly. However, the need for a long period of record becomes even more of a problem because you need a reasonable number of cases to define the regimes and create the training sets from the entire period of record.

To find a solution that addresses both the period of record problem and the tendency of standard MOS to make extreme forecasts worse, a hybrid Dynamic Regime MOS (DRMOS) has been adapted by MESO for use in operational forecasting. DRMOS uses the regime approach to define a subset of training days and continually recalculates the regression equations. Regimes are defined each day by how well that day matches days in an archived sample. The matches are based upon a set of pre-defined "matching parameters" but there are no pre-defined sets of regimes. A new training sample is created each day based upon the sample of historical cases that most closely matches the forecast days' matching parameters.

MESO has shown that the DRMOS approach was effective in improving both the transition and non-transition season solar irradiance forecasts for several Surface Radiation (SURFRAD) Network sites. The results from this research showed that the DRMOS was effective in improving both the transition and non-transition season forecasts. It would seem that the DRMOS utilizes information that both captures the current meteorological forcing and utilizes information of the past performance of the model. The improved performance is due to the fact that the model error characteristics for the time of the forecast are captured better with DRMOS than with the other two methods.

MESO has adapted DRMOS for use in improving wind power forecasts and the initial results are promising. The focus of this presentation will be the results of a comparison of the performance of DRMOS with other types of MOS when used for wind forecasting.