June-Beom Lee, "Development of Prediction Models of Ship Power and Ocean Environmental Data Based on Deep Learning", M.Sc. Thesis, Seoul National University, 2021.02.26
이준범, "딥 러닝을 이용한 해기상 및 소요 마력 예측 모델 개발", 석사학위논문, 서울대학교, 2021.02.26
첨부 '1' |
---|
Abstract | ute that minimizes the ship’s required power, it is necessary to predict the sea level and the ship’s required power accordingly. The weather forecasting companies such as the European Center for Medium-Range Weather Forecast (ECMWF) and Hybrid Coordinate Ocean Model (HYCOM), which provide ocean environmental data, typically make short-term forecasts of around six weeks. Therefore, when a long-term prediction is needed, it is necessary to predict ocean environmental data on its own. In the case of the prediction of ship’s required power, a numerical method using model test results is traditionally used. However, this method is difficult to accurately predict the ship’s required power due to the model test’s uncertainty. To solve this problem, an onboard test must be conducted, but this is expensive and time-consuming. Therefore, in this study, the ocean environmental data and ship’s required power were predicted using deep learning. In relation to the prediction of ocean environmental data, many studies have used deep learning. However, most of them made predictions for limited ocean environmental data such as wave height, wave period, and wave direction. In addition, there was a limitation in that prediction was made for a specific sea area rather than the entire sea area. Therefore, in this study, the ocean environmental data of the entire sea area was imaged for prediction. In addition, convolutional LSTM suitable for training time-series image data was utilized. In addition, AutoEncoder was used to solve delayed prediction that mainly occurs in the problem of prediction of time series data. Also, to find the optimal model, the performance of the model was evaluated by changing the size of input data (look back step) and prediction time of output data (look forward step). Finally, based on this model, the entire sea area was divided into 12 areas, and the prediction of ocean environmental data was carried out. Regression analysis was used in many studies related to the prediction of a ship’s required power. However, the regression analysis showed limitations in the accuracy of prediction in the case of complex problems such as the prediction of the ship’s required power affected by maritime and ship operating conditions. Therefore, in this study, deep feedforward neural network (DFN), a deep learning model suitable for numerical prediction, was used. In addition, various methods were used to increase prediction accuracy. First, in the case of wind and wave-related data in ocean environmental data, the effect was confirmed by performing pre-processing with relative values for the ship. Second, the effect was confirmed by changing the DFN model structure according to the characteristics of the DFN input data. Third, we analyzed the prediction accuracy according to the combination of five hyperparameters: number of hidden layers, number of hidden nodes, learning rate, dropout, and gradient optimizer for the DFN learning model. Fourth, k-means clustering was conducted to develop an independent model for predicting the ship’s required power according to a sea state and ship operational conditions. As described above, the performance of various prediction models was compared and analyzed. |
---|---|
Publication Date | 2021-02-26 |
June-Beom Lee, "Development of Prediction Models of Ship Power and Ocean Environmental Data Based on Deep Learning", M.Sc. Thesis, Seoul National University, 2021.02.26
이준범, "딥 러닝을 이용한 해기상 및 소요 마력 예측 모델 개발", 석사학위논문, 서울대학교, 2021.02.26