Research on the Impact of Optimizer Design on Generalization Performance in Ship Trajectory Prediction Based on LSTM

Abstract

Ship trajectory prediction using the Automatic Identification System (AIS) is a core task in intelligent shipping. Focusing on straight-line navigation segments, this paper constructs a unified multi-target LSTM prediction framework that simultaneously forecasts four key variables: longitude, latitude, speed, and course. Through five-fold cross-validation and under identical network architectures, we compare the performance of three models: standard LSTM, attention-enhanced LSTM, and LSTM optimized with Lookahead-RAdam. Experimental results demonstrate that the LSTM model incorporating the Lookahead-RAdam optimization strategy significantly outperforms the other approaches across all evaluation metrics. Further analysis reveals that optimizer design has a far greater impact on model generalization than network architecture complexity. This study proposes an efficient paradigm for maritime spatiotemporal sequence modeling: “simple architecture + advanced optimization.”