Monthly composite of 84-hr diurnal forecast
In an effort to evaluate and understand the diurnal nature of model forecasts, we examine the 84-hr diurnal model forecasts from the NAM (previously mesoscale Eta) model only, and the NAM, GFS (previously AVN), and NGM models verified against observations averaged by region and composited on a monthly basis to allow trends to emerge. Observations and forecasts are available for each 3-hour period. Across a particular region, for a given 3-hour observation/forecast period, the number of observations and model forecasts are consistent in number and location, although from 3-hour to 3-hour period this number may change. For this present treatment, given the large number of observations/forecasts available across the region and throughout the month, the plots traces generally appear to smoothly vary through the two diurnal cycles of the forecast period. (To require that all traces at all times be consistent could be overrestrictive with potentially many observations/forecasts being eliminated.)
X-AXIS = observation and forecast hour (0-48)
Y-AXIS = 2-meter air temperature (C), 2-meter relative humidity (%), or 10-meter wind speed squared (m^2/s^2). (Because of the statistics available from our forecast verification system, only the 10-meter wind speed squared is avaliable; however, taking the square root of the Y-AXIS values gives the wind speed itself, e.g. 4=2m/s, 9=3m/s, etc.; this is then a nonlinear scale.)
O_MEAN = observed mean (taken from NAM, although NAM, GFS, and NGM obs are all equal)
F_MEAN = forecast mean (NAM, GFS, or NGM)
observations (solid line), NAM (short dashed line), GFS (long dashed line), NGM (short-long dashed line)
Annual time series of 48-hour forecast bias
In an effort to evaluate and understand the long-term trends in model forecasts, we examine the annual time series of the bias in the 48-hour forecast (end point, or last forecast hour) from the NAM model verified against observations averaged by region for the given year. Observations and forecasts at the end of each 48-hour forecast are used to determine the bias. Across a particular region, for a given 48-hour bias, the number of observations and model forecasts are consistent in number and location, although for any given 48-hour forecast this number may change. Because of transient weather systems, and because the same points (number and location) may not be available at the end of each 48-hour forecast cycle, the day-to-day bias may appear noisy, though this variation is less in these regionalized plots than it would be had it been tracked for an individual station. Regional trends emerge when monthly to seasonal to annual periods are examined.
X-AXIS = month and day of year
Y-AXIS = forecast bias of 2-meter air temperature (C)(SCALAR_BIAS), 2-meter relative humidity (%) (SCALAR_BIAS), or 10-meter wind speed (m/s) (VECTOR_BIAS).
Zero bias (solid horizontal line), NAM model (solid line).
Model grids used in verification
|NAM/218 (12-km)||GFS/212 (40-km)||NAM/242 (11.25-km)||GFS/216 (45-km)||NGM/104 (90-km)|
Forecast Verification System (FVS) acknowledgements
Geoff DiMego - original developer of codes to compute the forecast-minus-observation differences and their partial-sums.
Mark Iredell - led effort to define FVS data base format for the partial-sums.
Keith Brill - developer of the FVS system that accesses the data base, performs the final-sums and displays the statisitcs under user control -- also corrected aspects and enhanced capability of DiMego's original computation codes.
Eric Rogers - implemented scripts to routinely run the codes to compute and store the partial-sums for most of NCEP's operational and parallel modelling systems.
Michael Ek - generation of FVS plots and development/maintenance of this web page.