Yohei KUGUE

Machine Learning for Short and Mid term Wave Prediction System

Tokuzo Hosoyamada

 Wave forecasting is required to determine whether or not to construct offshore wind power generation facilities and operate vessels, etc., for wave height thresholds (1.0~1.2 m) in stormy weather for 3 to 7 days ahead. However, it has been pointed out that current forecasting systems may be insufficient in terms of spatial resolution and forecasting accuracy. However, most of the studies are incomplete in that the feature engineering required for machine learning and the interpretation of the reasons for the predictions and behavior of the models have not been conducted. Therefore, it cannot be said that the performance of machine learning has been truly demonstrated.
 Therefore, in this study, we use machine learning, which is considered to be a common method, to predict waves in low-wave conditions, and once again evaluate its performance and propose an appropriate machine learning method. For the model and library, TabNet and XGBoost are used, considering the handling of table data and the interpretability of the model. The input data are NOWPHAS: Significant waves and wave heights (m) from wave observations, and JRA-55: Surface 10 (m) u- and v-directional wind speeds (m/s) from long-term meteorological reanalysis data. Treat the forecast as a regression analysis or time series analysis, in which the forecast is based on changes or combinations of past data from the time of the forecast.
The first step of the proposed system is to create features from observed and analyzed values up to a certain prediction point. Next, the created features are input into the model and trained. Only the features that the learned model particularly values are selected, input into the model again, and trained. Predictions were made based on the model and selected features.
 As a result, the accuracy was improved by 30% to 40% compared to the previous case. In addition, the machine learning model focused on the standard deviation of the northwest wind as a feature, indicating that the model acquired behavior and characteristics similar to a physical model through learning. In the future, this system can be used as a basis for further development to improve the accuracy of predictions and to make spatial predictions at arbitrary points.