Comparison between TAF and MOS Focused on Wind Speed during Spring Season in 2017–2018

1. Introduction

As the aviation industry becomes global, numerous air traffic growths have been doubled every fifteen years since the mid-1970’s (ICAO, 2016b). Additionally, the frequency and intensity of high impact weather events at local airports (e.g. Incheon International Airport) are increasing due to climate change and abnormal weather conditions. Forbes magazine pointed out that the top priority of Big Data analytics target is the aviation industry (Columbus, 2014). Aeronautical meteorological data serve as catalysts to increase the value of the aviation industry.
In this study, I would like to examine the accuracy of TAF produced by Incheon International Airport, which was ranked second at 2018 the Skytrax World Airport Awards. In addition, by comparing the results of MOS computation performed by the KMA super computer to the results of MOS (Model Output Statistics), we can diagnose the possibility of automation of TAF, build the LAMP (Localized Aviation MOS Program) in 2020, and refine the data – from what is currently performed in three-hour interval to one-hour interval. The LAMP system using the logistic regression method will be constructed and the next-generation TAF automation system will be constructed automatically through take-off and landing forecasting.

2. Data and Methodology

The TAF produced by the weather forecast of the Aviation Meteorological Office (AMO) in Korea Meteorological Administration (KMA) serves at three-hour intervals and projects up to 36 hours ahead for four times (05, 11, 17, 23 UTC) a day. MOS serves at three-hour intervals as well but with up to 75 hours of projection time twice a day (00, 12 UTC). In order to check the accuracy of 00 UTC TAF and 05 UTC MOS produced at Incheon International Airport, the differences between the two forecasts were taken. Starting from the fixed projection time of 06 UTC to the following 36 hours, the numerical value of the two projections were obtained; the differences were calculated from METAR (reference observation); and the results were compared. As MOS in KMA became officially in December of 2016, the wind data of spring (March-April-May) in 2017 and 2018 were compared in this study. This paper solely presents the errors of wind speed data.
To verify the accuracy of the two forecasts, quantitative methods such as Bias, RMSE, MAE, Correlation, and visualized boxplots and scatterplots were employed. The former assists to understand numerically with accuracy, and the latter to distinguish the difference from the normal flow through its form.

3. Analysis
3-1. Numerical descriptive views
From March to May of 2018, the TAF and MOS errors were analyzed. For eleven forecast times, monthly Bias, RMSE, MAE, and Correlation were calculated at intervals of three hours: 6-hour after (06 UTC of the day), 9-hour after (09 UTC of the day), ..., 36-hour after (15 UTC of the next day). Table 1 shows four statistical parameters by three-hour interval of the projection time for each month during a course of two years. Here, ‘T’ means that the error of TAF is lesser than MOS, and ‘M’ means that the error of MOS is smaller than TAF. The ICAO (2016a) error criterion is given only when the wind speed is less than 2.5 knots. If this criterion is satisfied, RMSE and MAE are expressed as TT and MM. Correlation identified the data as TT and MM when it was 0.8 or more.
The changes in Bias and RMSE in two years (2017 and 2018) were larger in March than in other months, and April was relatively small. The reason seems to be the timing of the seasonal transition in March. The correlation coefficient of MOS in all two years was much higher than that of TAF, and all except for three cases –18-hours after, 21-hour after, and 27-hour later—were 0.8 or more. The reason is that the wind speed (e.g., 6 knots, 8 knots, or 10 knots) used by the ICAO regulation is determined for the TAF. Because of calculating intact MOS value, MOS has a correlation coefficient higher than TAF.
In Table 1, when the correlation between MOS and METAR is high, the utilization of MOS as TAF guidance and the possibility of utilization of automated TAF can be considered.

3-2. Graphical descriptive views
As seen in Figure 1, the 2-D scatterplots (TAF over X-axis, MOS over Y-axis) are easy to understand. In the correlations between TAF and MOS, frequency was concentrated at less than 10 knots and less frequent at more than 14 knots. While TAFs tend to concentrate on forecast values such as 6 knots, 8 knots, and 10 knots, MOS values tend to be expressed in various values. The boxplots has been analyzed through the difference of TAF values in MOS values, but it shows a generally balanced shape. It seems that the MOS prediction value is being produced stably. Figure 1 shows the correlations and boxplots only for 6 hours after March-August 2018 due to the ground relationship.

Tracing the dates of both large errors and small errors and examining their weather charts, the latter was when surface pressure gradient was small associated with located high pressure over Korean Peninsula, or when there was no significant variation of pressure. On the other hand, the former was when surface pressure gradient is large associated with low pressure/ frontal system approaches the Korean peninsula, described east-west contrasts/ north-south contrasts.

Acknowledgements
This study was under the auspice of the Fund of Short-term Team Training Abroad by Ministry of the Interior and Safety-ROK in 2018.

References
Columbus, L, 2014: 84% Of Enterprises See Big Data Analytics Changing Their Industries' Competitive Landscapes In The Next Year, Forbes.
International Civil Aviation Organization (ICAO), 2016a: Annex 3 to the Convention on International Civil Aviation: Meteorological Service for International Air Navigation.
_____, 2016b: 2016-230 Global Air Navigation Plan. Doc 9750-AN/963, 137pp.
Wilks, D. S., 2011: Statistical Methods in the Atmospheric Sciences. Academic Press. 676pp