Maneuver-Aware Residual Multi-Scale LSTM for Short-Term Vessel Trajectory Prediction in Port-Approach Waters

Abstract

Accurate short-term vessel trajectory prediction is important for traffic monitoring and collision-risk screening in port-approach waters, where vessels frequently turn, accelerate, decelerate, and merge into traffic lanes. This study develops a maneuver-aware residual multi-scale long short-term memory (LSTM) framework for Automatic Identification System (AIS)-based 10–30 min trajectory prediction. The method predicts residual displacement relative to the last observed position, constructs maneuver-aware features from local displacement, course variation, speed variation, turning rate, and acceleration-like terms, and compares fixed and maneuver-guided multi-scale fusion strategies. Experiments are conducted on public AIS data from San Francisco Bay and adjacent approach waters using Maritime Mobile Service Identity (MMSI)-level train/validation/test splits and three random seeds. The largest observed gains come from residual prediction and maneuver-aware features. In the three-seed main evaluation, the fixed multi-scale LSTM (Fixed-MS-LSTM) provides the strongest 10 min accuracy, while the maneuver-guided multi-scale residual LSTM (MGMS-RLSTM) achieves lower average displacement error (ADE) at 20 and 30 min and learns distinct temporal-scale preferences across straight, turning, and speed-changing samples. Encounter-oriented closest point of approach (CPA) and time to closest point of approach (TCPA) evaluation further shows that the residual multi-scale models support more accurate CPA/TCPA-based high-risk screening under the evaluated benchmark. These findings indicate that maneuver-guided fusion can be characterized as a horizon-dependent scale-selection mechanism that complements the fixed multi-scale counterpart.