Co-simulation-based Pre-training of a Ship Trajectory Predictor

Abstract

A ship trajectory predictor plays a key role in the predictive decision making of intelligent marine transportation. For better prediction performance, the biggest technical challenge is how we incorporate prior knowledge, acquired during the design-stage experiments, into a data-driven predictor if the number of available real-world data is limited. This study proposes a new framework under co-simulation platform Vico for the development of a neural-network-based trajectory predictor with a pre-training phase. Vico enables a simplified vessel model to be constructed by merging a hull model, thruster models, and a controller using a co-simulation standard. Furthermore, it allows virtual scenarios, which describe what will happen during the simulation, to be generated in a flexible way. The fully-connected feedforward neural network is pre-trained with the generated virtual scenarios; then, its weights and biases are finetuned using a limited number of real-world datasets obtained from a target operation. In the case study, we aim to make a 30 s trajectory prediction of real-world zig-zag maneuvers of a 33.9m-length research vessel. Diverse virtual scenarios of zig-zag maneuvers are generated in Vico and used for the pre-training. The pre-trained neural network is further finetuned using a limited number of real-world data of zig-zag maneuvers. The present framework reduced the mean prediction error in the test dataset of the real-world zig-zag maneuvers by 60.8% compared to the neural network without the pre-training phase. This result indicates the validity of virtual scenario generation on the co-simulation platform for the purpose of the pre-training of trajectory predictors.