A COMPARISON OF TWO TRAINING INPUT DATA WITH AND WITHOUT HWTS IN LSTM-BASED PREDICTION MODELS

Abstract

Recently, the sea water temperature has been increased rapidly and its values was highest in 2019 compared to the last decade. These phenomena can cause the abnormally HWTs and extensive damage to the maritime economy. In this study, therefore, we proposed a water temperature prediction model based on the long short-term memory (LSTM) that is a useful deep learning model for predicting time series data, in order to prevent maritime economic damages by predicting abnormally HWTs in advance. We compared the models trained in two different cases: 1) to predict the water temperature of a given pixel, the water temperature data of that pixel are used as training input data and 2) to predict the water temperature of a given pixel, the data of different pixel with HWTs are used as training input data. In here, European Centre for Medium‐Range Weather Forecasts (ECMWF) sea surface data were used as the training input data. As a result of 1-day prediction, the models trained by input data with and without HWTs gained 0.996, 0.339, 1.619 and 0.982, 0.736, 2.974, in terms of R2, root mean square error (RMSE) and mean absolute percentage error (MAPE), respectively. It was found that the accuracy of the model trained by input data with HWTs was better.