4.3
Introduction on MOS Approaches for KMA Digital Forecast

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Tuesday, 19 January 2010: 2:00 PM
B305 (GWCC)
JunTae Choi, Korean Meteorological Administration, Seoul, South Korea; and Y. Seo and M. S. Lee

Since October 2008, Korean Meteorological Administration (KMA) has initiated a new forecast suite, digital forecast which provides 12 weather elements in digital form for Korean peninsula. As the digital forecast is official forecast in Korea, the forecast is principally generated by weather forecaster. For the forecaster, high resolution objective forecast guidance and an interactive digital forecast editor are prepared.

KMA digital forecast was born from USA' digital forecast, NDFD, but there is a distinct difference. While USA digital forecast being be produced by a spatial editing procedure in most, KMA digital forecast is finalized through three steps which are point forecasting, objective analysis to convert point forecast to gridded forecast and trimming gridded forecast. Therefore, the process of the point forecast can greatly influence on the forecast accuracy.

To improve the digital forecast accuracy, several objective guidance are provided for forecaster, such as directive interpretation of NWP products, PPM and MOS. Usually MOS is regarded as the best guidance among the three guidance and MOS guidance is prepared for eight elements in twelve elements in KMA digital forecast.

In view of objective analysis, more points of forecast guidance shows better gridded forecast. But MOS can be implemented only at observation point. There are 103 synoptic observation stations in Korean peninsula and the number can not meet the resolution of KMA digital forecast, 5km. KMA adapted two approaches to overcome this problem.

One approach is enlargement of point of MOS by including AWS(Automatic Weather Station) which is a kind of mesonet and a typical unmanned operation. The good quality of observation is essential to MOS accuracy and KMA picked up 150 AWS well managed. Temperature and probability of precipitation MOS were implement at about 250 point whose spatial resolution is 20km which is tolerable in KMA digital forecast. The temperature accuracy at the AWS was a little lower than that at the synoptic station when the MOS equations were verified with independent sample.

The other approach is regional MOS equation. AWS was designed for monitoring severe weather and equipped 3 to 5 instruments. Therefore some elements such as RH and sky cover are not observed at the synoptic station whose spatial resolution is too coarse to meet the requirement of digital forecast. For these elements, a regional MOS equation was derived and applied at each grid point in the same region.

MOS in KMA is basically developed with multiple linear regressions between predictor and predictand and the best set of predictors is screened from numerous predictors by stepwise selection method. Several additional techniques were embedded to improve MOS accuracy. MOS is one of statistical model and the MOS can remove out the systemic error of NWP during sampling period. But NWP can not always simulate perfectly and it can predict deadly wrong scenario. This case data may contribute to make NWP meaningless or MOS prediction to merge to climatic value. Therefore clearly wrong cases in NWP prediction sample were excluded when MOS equations were derived. This technique worked well on POP because we can easily find out wrong prediction such as a dry sunny prediction for rainy day.

For sky cover, preprocess and post process were developed. Sky cover in digital forecast is forecasted in four categories which are clear, scatter, broken and overcast. This element is not fully continuous value. More worse is that the frequency is concentrated at the both sides categories. When this element was applied at simple linear regression, the derived equation tended to predict only median categories. This problem was removed with the modification of element definition. The cases of median categories in observation were excluded in the sample and the both sides categories were evaluated as 0 and 1 respectively. This method resolved the problem dramatically. Finally MOS prediction and corresponding observation were compared to find out optimum threshold to convert law MOS prediction to final MOS prediction. This process is called by post process.

These techniques mentioned above seem to be artificial, but the verification consistently showed the improvement of MOS. The detailed result will be present at the conference.