Hayato Horiuchi

Application of Decision Tree Machine Learning Model with Feature Importance to Wave Height Prediction

Tokuzo Hosoyamada

Wave height forecasting up to a week in advance is necessary to determine whether or not maritime operations are possible, but the spatial resolution and accuracy of current forecasting systems are inadequate. Attempts to improve forecast accuracy using machine learning are in progress, but feature selection and model interpretation are inadequate, and machine learning performance has not been fully demonstrated. Therefore, in this study, we use machine learning to predict wave heights up to one week ahead in low wave height conditions, evaluate its performance, and propose an appropriate machine learning method. We also examine how different wind speed data affect the accuracy of wave height prediction. The models used are TabNet and XGBoost. As input data, NOWPHAS is used for wave height observations and JRA-55 for meteorological analysis. The forecasting system creates features from the input data and trains them into a training model. It then selects features that the learned model deems particularly important and re-trains them. Predictions are then made using this model and the selected features. During feature creation, wind speed data is selected and the optimal wind speed locations are adopted for each forecast condition. As a result of this study, regarding the selection of wind speed locations, wind speed locations located in the west are more important for long-term forecasts, but the importance of these locations is broad and reduced. In addition, the nature of the wind changes depending on the geographical characteristics and meteorological conditions of the forecasted location. For wave height forecasting, the shorter the forecast period, the higher the forecast accuracy, and the more accurate the forecast accuracy becomes with the selection of wind speed locations. However, it is necessary to consider the treatment of outliers and the characteristics of each wave height observation point when evaluating forecast accuracy.