This is the second of two brief reports on the activities and results of the Joint Numerical Weather Prediction Unit since May 1955, and is concerned primarily with the accuracy and characteristic errors of the numerical forecasts described in the previous report. The quality of the barotropic and 3-level forecasts has been measured by several statistical indices of error, and compared with that of the subjective forecasts issued by the National Weather Analysis Center. A breakdown of these statistics shows the dependence of forecasting accuracy on length of forecast period, level, data coverage, and proximity of lateral boundaries. Various sources of systematic error are discussed with reference to the JNWP Unit's efforts to isolate and remedy them.
After almost a year of experimentation and operational numerical weather forecasting, it is concluded that the quality of the numerical 500 millibar forecasts is not significantly different from that of the best subjective forecasts prepared by methods in current use. Recent results indicate that a significant improvement can be expected in the near future. The numerical 1000 mb forecasts are worse, but recent changes of model show promise of matching the performance of subjective methods. Finally, the most glaring systematic errors of the present numerical forecasts have adequate explanation in existing theory, and can be (or have already been) corrected by generalization of the models.
Machine learning emulation of gravity wave drag in numerical weather forecasting
