The mean of an ensemble of forecasts is traditionally defined as the
Arithmetic Mean (AM) of all ensemble members. In the past two decades,
this product gained significant use in weather forecasting. As a
statistical construct, AM offers a nonlinear filtering of unpredictable
forecast features across ensemble members that is reflected in a
Root-Mean-Square (RMS) forecast error below that of individual
forecasts. This is achieved at the cost of smearing out features that
in different members appear at different positions and with different
shapes and amplitudes.
In the proposed Developmental Testbed Center (DTC) Visitor Project, we
will develop an algorithm and software for an alternative ensemble mean
that we call Feature-based Mean (FM). In FM, all forecast features
appear at the mean of their position in the individual members,
represented with an amplitude that is the mean amplitude of features
aligned in all members. Preliminary results show that the FM retains
more small-scale features and larger amplitudes than the AM. In
addition, the FM can reduce about 10% RMS error for short to mid-term
forecasts of extreme events relative to the AM. After the development
of the algorithm, real-time FM graphics will be demonstrated to Weather
Prediction Center (WPC) forecasters and the FM software will be
contributed to the DTC software repository.