Benchmarking feed-forward randomized neural networks for vessel trajectory prediction

Abstract

The burgeoning scale and speed of maritime vessels present escalating challenges to navigational safety. Perceiving the motions of vessels, identifying anomalies, and risk warnings are crucial. Central to addressing these challenges is the analysis of vessel trajectories, which are pivotal for anomaly detection and risk mitigation. This study introduces an innovative approach to time series vessel trajectories, focusing on the Chengshantou waters. We implement and rigorously compare seven feed-forward neural network models, including random vector functional link neural network without direct links (RVFLwoDL), deep RVFLwoDL (DRVFLwoDL), ensemble deep RVFLwoDL (edRVFLwoDL), random vector functional link neural network (RVFL), deep RVFL (DRVFL), ensemble deep RVFL (edRVFL), and broad learning system (BLS). Our evaluation, utilizing diverse error metrics and datasets from various waterways, reveals the superior performance of the RVFL-based models with direct links in trajectory prediction. The findings underscore the critical role of direct links in enhancing the representational and generalization capabilities of RVFL models, thus offering robust and reliable prediction solutions.