Standardization in building an ANN-based mooring line top tension prediction system

Abstract

ANN-based mooring line top tension prediction systems are trained on ship motion and mooring line top tension time histories from multiple wave states with a certain simulation length. In the previous studies, selection of the wave states and the simulation length differs between the studies and they are not standardized. Also, a plain neural network is mostly used. In this paper, tension prediction performances with respect to a distribution shape of the wave states, a number of the wave states, and the simulation length are first studied. Then, the prediction performances with respect to Batch Normalization (BN) and Learning Rate Decay (LRD) are studied, in which BN and LRD are very common components in modern neural network models. Lastly, a guideline for selecting the wave states and the simulation length is proposed, and BN and LRD are proven to be advisable to use to improve the prediction performance.