1. An Algorithm Combining for Objective Prediction with Subjective Forecast Information
JunTae Choi and SooHyun Kim
NIMS, Korean Meteorological Administration, Rep. of Korea 1 3
Introduction
Experiment
As direct or post-processed output from numerical weather prediction (NWP) models has begun to show acceptable performance compared with the predictions of human forecasters, many national weather centers have become interested in automatic forecasting systems based on NWP products alone, without intervention from human forecasters. The Korea Meteorological Administration (KMA) is also starting to develop an automatic forecasting system for dry variables. The forecasts will be automatically generated with a post processing model (MOS) based on NWP predictions. However, MOS cannot always produce acceptable predictions, and sometimes its predictions are rejected by human forecasters. In such cases, a human forecaster should manually modify the prediction consistently at points surrounding their corrections, using some kind of smart tool to incorporate the forecaster’s opinion.
This study introduces an algorithm to revise MOS predictions by adding a forecaster’s subjective forecast information at neighboring points to save human forecaster workforce.
automatic forecast methods
※ (forecastneighbor – MOSneighbor) is portion by forecaster contribution and F is derived with simple linear regression
training and verification data
- MOS prediction : single point MOS based regional model(UM)
- forecast : point forecast generated by human forecaster
- period : 2013 ~ 2014
sensitivity with forecast methods
CNTL
forecasttarget = F(forecastneighbor)
FMOS
forecasttarget = MOStarget + F(forecastneighbor – MOSneighbor)
where (forecast – MOS)target = F(forecast – MOS)neighbor
OMOS
where (OBS – MOS)target = F(OBS – MOS)neighbor
MOS Bias when NWP model misses precip.
temperature relative humidity wind speed
human forecaster’s contribution
Fig. 2. Difference of mean over absolute error(ball location) between this study and MOS. Ball size represents number of case. Upper(lower) ball represent the cases outperforming(underperforming) MOS.
Projection(hour) 2 4
Approach
Result & Discussion
As KMA short term forecast is produced in digital forecast, a forecaster modify MOS prediction to produce official forecast. In this time, the forecaster tend to add similar increment to the MOS prediction when the forecast points are close to each other.
Two forecast point groups
- group A : manual forecast points, 45 synoptic stations
- group B : automatic forecast points, 47 unmanned sites
 the forecast at point in group B is automatically estimated with MOS prediction
and subjective information by human forecaster at neighbor station in group A
Verification with the existing forecasts.
The RMSE of temperature predictions produced by OMOS was slightly lower than that of the original MOS predictions, and close to the RMSE of subjective forecasts. For wind speed and relative humidity, the new algorithm outperformed human forecasters.
Unexpectedly, OMOS trained with observation and MOS showed slightly better score than that of FMOS. KMA forecaster work in four-shift schedule and four kinds of bias might influence on the regression relationship F.
The OMOS will be refined with blending guidance(Choi and Kim, 2015) which greatly outperformed any existing single MOS.
Fig. 3. RMSEs of MOS, OMOS, official forecast (human forecaster)
Fig. 4. Relative frequency according to error windows
automatic forecast
KimhwaMOS + f(dx)
manual forecast
forecaster contribution
CheorwonMOS + dx
Fig. 1. Distribution of manual forecast points(red)
and automatic forecast points(blue)