Sensitivity of Mesoscale‐Model Forecast Skill to Some Initial‐Data Characteristics, Data Density, Data Position, Analysis Procedure and Measurement Error

Observing‐system simulation experiments (OSSE) were performed in order to determine the effect of horizontal and vertical data resolution, data location and measurement error on mesoscale‐forecast accuracy. For the data‐density experiments, data sets with various horizontal and vertical resolutions were extracted after 12 h of a 36‐h mesoscale model forecast of an East Coast cyclogenesis event. After reanalysis of the extracted data back to the model grid, the model was restarted and the experimental forecasts were compared to the original control forecast for the 12‐h to 36‐h period. In addition to these data‐density experiments, other OSSEs were performed in which the data were extracted at different locations, and in which measurement error was imposed in the simulated data sets. Initial and forecast errors were quantified using average‐error statistics as well as the amplitude and position of major meteorological features. Results show that the error, defined with respect to the control run, generally decreases with increasing forecast time for most variables in terms of both the mean error and the amplitude and position parameters. The specific error reduction (or growth) rates, however, are significantly dependent on the variable, the model level and the data resolution. Major contributors to the observed error reduction with time seem to include nonlinear effects, geostrophic adjustment and surface forcing, in addition to the use of identical lateral boundary conditions for the control simulation and OSSE's. The significant reduction in position and amplitude errors during the first 8 h of the forecast cannot be reasonably attributed to primarily lateral boundary-condition effects because it is observed immediately, near the center of the domain. This error‐reduction effect is especially noteworthy because some people have frequently expressed the hope that model dynamics might mitigate some of the effect of initial-condition error, particularly in regard to mesoscale-model initialization with synoptic‐scale data. Even though earlier modeling evidence has suggested that such an effect exists, this study quantifies it for this particular case in terms of the influence of the above‐noted mechanisms on forecast error associated with a range of specific meteorological features. When high vertical resolution (14 levels) is provided by the soundings, a significant reduction in forecast error results when the horizontal resolution is doubled from a sounding spacing of 360 km to a spacing of 180 km. When poor vertical resolution (3 levels) is employed in the soundings, however, the forecast errors are not significantly influenced by the horizontal data resolution, i.e., increasing the horizontal data resolution from synoptic scale to mesoscale, by using 36 times more soundings, does not have a positive impact. In fact, initial and forecast errors are slightly smaller with the coarse horizontal resolution, presumably because the smoothing associated with the horizontal objective analysis procedure caused a reduction in the vertical interpolation and extrapolation errors. It is found that the specific locations of the soundings in a regularly spaced synoptic-scale data network are not important in the error statistics for most variables, however, two forecasts initialized with offset data grids showed significant local differences in their error patterns. When measurement error, typical of a rawinsonde, is imposed on wind and temperature data with synoptic-scale resolution, the initial and forecast errors are increased by only 10%–20%, indicating that most of the error results from the objective analysis procedure and the data density. Also, this percentage increase in mean error on the grid is much smaller than the mean error imposed at the data points, indicating that the smoothing from the objective analysis procedure is important in reducing the effect of measurement error. Finally, it is appropriate to note the strengths and weaknesses of this particular experimental design using OSSE's, where the same model is employed to generate both the experimental forecasts and the reference atmospheric data set. The strength, of course, is that it is straightforward to relate forecast error to initial-condition error because the reference atmosphere is model generated, thus “eliminating” the error associated with imperfect model physics and numerics. This also means that the total model‐system error is not being estimated. Nevertheless, because numerical‐prediction error growth is difficult to interpret, it does seem reasonable to isolate the source of error when this is experimentally possible, as done in this study.