기계학습 기법을 고려한 표층 뜰개의 대한해협 이동 예측 연구

Abstract

This study identified the characteristics of surface ocean currents in the Korea Strait and the East Sea by analyzing the trajectories of 15 drifting buoys deployed in the coastal region where the Tsushima Warm Current (TWC) is dominant. A comparison and an analysis were made between the numerical current model-based particle tracking models and the machine learning-based particle tracking models using measured data in order to improve the prediction accuracy of particle tracking. Three main results are summarized as follows: 1) Among the 15 drifting buoys deployed into the southern part of the East Sea on 21 July 2020, the buoys (T-4∼T-9) deployed at the far locations from the coast moved northeastward during the first 4～5 days with a half turn clockwise motion from the major axis of current. After such a half turn clockwise motion, the trajectories measured by drifting buoys during 140 days were similar to those of mean surface currents shown in both schematic map of surface currents and chart of seasonal mean surface currents provided by the Korea Hydrographic and Oceanographic Agency (KHOA). Both meso-scale and small-scale eddies and inertial oscillations were also observed. Even three deployed buoys at each T-2, T-5, and T-8 showed different movements at the separation points of each tributary or at the unification points where buoys meet the eddies. 2) Two numerical models (ROMS, MOHID) and two machine learning models (Linear Regression, Decision Tree) were applied to predict the trajectory of the T-5 buoy alone due to its good data quality and quantity. Three skill assessment methods (CC, RMSE, NCLS) were used to evaluate the performance of four different models. Data used for machine learning were tide height, wind (wind direction and velocity) and the locations of T-5 buoy trajectory. Hourly tide height and wind data were obtained from the KHOA Geojedo and Gadeokdo Tide Stations and the Kyoboncho Ocean Station. Decision Tree showed the best skill scores in both CC and RMSE, and MOHID showed the best skill score in NCLS. 3) In the cumulative deviation between the observations and the predictions, the four inflection points in U-component velocity were found to have a period of about 17 to 25 hours, which are inferred owing to the tide influence. More observations and studies are required sure of it.