2021 Volume 77 Issue 2 Pages I_79-I_84
In maritime construction, there is a high demand for wave prediction to determine whether work is possible or not. In this study, we used ConvNet (Convolutional neural network) for deep learning, and built models to predict the wave in 24 hours at the NOWPHAS observation point using the JMA air pressure data. Since there are only a few thousand data available for training, we thought there would be a shortage of data for deep learning. To solve this problem, we introduced Fine Tuning (FT) into the model, which is a method to train the model efficiently by using the learned model. And we confirmed the effectiveness of the model. A comparison of the prediction accuracy of the models shows that FT is more efficient and accurate than conventional learning methods in deep learning with a limited number of data, demonstrating the effectiveness of FT. Using deep ConvNet, it was shown that the availability of work after 24 hours can be judged with high accuracy even if the training data is only air pressure.