기계학습 기반 해양 표류체 이동 예측 모델 구축 및 개선연구

Abstract

This study applies artificial intelligence techniques, including machine learning and deep learning, to overcome the limitations of numerical model-based particle tracking prediction models. These limitations include spatial and temporal constraints, as well as fundamental limitations of numerical models. Additionally, we conducted a study on improving the accuracy of particle tracking models by applying high gradient functions and leeway. In summary, our research findings are as follows. A comparison of particle tracking models was conducted using numerical models (ROMS, MOHID) and machine learning particle tracking models using observations (linear regression, decision tree) in the vicinity of the Korea Strait. The results showed that MOHID (0.89) performed the best in the case of NCLS, while the machine learning particle tracking models performed at about 84.82% (linear regression) and about 98.68% (decision tree) compared to MOHID. The performance of five AI-based particle tracking prediction models (ET, LGBM, SVM, DNN, RBFN) was evaluated using surface float data observed in the southwestern waters of Jeju and numerical model results (ROMS, WRF). The results showed that DNN outperformed the other models with an MAE of 10.02, RMSE of 8.03, and NCLS of 0.84. Compared to the traditional numerical model (ROMS) particle tracking prediction model, the AI-based particle tracking model shows an improvement of approximately 41.1% for MAE, 44.4% for RMSE, and 12.0% for NCLS. Additionally, the AI-based model reduces the overfitting tendency of the numerical model. Furthermore, an AI-based prediction model utilizing ensemble techniques was developed to cover multiple waters. To select the base model for the ensemble, we evaluated seven AI models (RF, ET, LGBM, CB, MLR, SVM, DNN) individually for performance. We selected the top four models (ET, LGBM, MLR, SVM) that demonstrated high performance as the base model for building the ensemble model. After comparing the performance of a numerical model, seven AI models, and the ensemble particle tracking prediction model, it was found that the ensemble model generally outperformed the others. Specifically, in the case of NCLS, the accuracy increased by 26% compared to the numerical model. It is worth noting that both numerical and AI-based particle tracking models rely on first-order terms for their algorithms. We applied the theory of higher order functions (1st-4th order) and the theory of leeway, which can reflect the influence of wind, to the particle tracking prediction model to compare and analyze the movement path of surface seabirds. The first to third order terms showed similar movement tendencies, and the fourth order terms showed a decrease in the distance traveled but maintained the tendency. For NCLS, the fourth term showed an increase in results (0.5937<0.6098<0.6085<0.7791). This is a summary of three studies to be conducted in the future to improve the accuracy of particle tracking prediction. To improve the accuracy of the numerical model of seawater flow, it is suggested that the accuracy of the particle tracking prediction model can be enhanced by improving the spatial and temporal resolution in coastal areas and assimilating the numerical model data using satellite data (SST, SSHA). As numerical models and data assimilation techniques improve, the resources required are expected to increase significantly, necessitating further research. Additionally, continuous development and improvement of AI-based particle tracking prediction models are necessary to enhance accuracy. These models currently utilize variables such as ocean currents, winds, surface float paths, and derived variables. However, research is necessary to improve the accuracy of the model by learning variables that reflect the various factors in the ocean, such as wave effects caused by wind, water depth, and topography. Additionally, due to the limited availability of training data, it is necessary to supplement it with continuous observations. Improving the reliability and accuracy of the model can be achieved by enhancing the particle tracking method. This can be done by applying the high gradient function and the leeway coefficient discussed in Chapter 5 and verifying it in various seas and experimental groups. It is recommended to use derivatives of the fifth to eighth order or higher for better results, as this study only covered up to the fourth order.