Channel-Aware Residual BiLSTM for Ship Trajectory Prediction and Collision Risk Assessment in Restricted Waters

Restricted waterways combine curved fairways, limited maneuvering space, and dense encounters, complicating short-term ship trajectory prediction and collision risk assessment. This study proposes a channel-aware residual Bidirectional Long Short-Term Memory (BiLSTM) framework. The Nonlinear Randomly Reuse-based Mutated Whale Optimization Algorithm (NRRMWOA) configures its hyperparameters; a constant-velocity (CV) residual prior, channel look-ahead geometry, global–local fusion, and an Artificial Potential Field–Potential Collision Risk (APF-PCR) risk head are incorporated. Evaluation uses Automatic Identification System (AIS) data from two restricted-water regions. The fused system yields average and final displacement errors (ADE and FDE) of 103.70/217.16 m in Study Area 1 and 116.15/243.67 m in Study Area 2. Compared with the strongest metric-specific model, the ADE/FDE reductions are 5.47%/6.15% and 8.27%/8.07%, respectively; all differences are significant. The alert+ area under the receiver operating characteristic curve (AUC) values are 0.988 and 0.993. Under the tested conditions, the CV residual, channel constraints, and expert fusion reduce trajectory error and supply forecasts for stepwise collision risk assessment.