Predicting Ports Congestion by Utilizing State Space Models

Maritime port congestion poses significant risks to global supply chains, as ports handle over 80% of world trade by volume. Current congestion prediction methods struggle with the complex temporal dependencies inherent in port operations, limiting their effectiveness for proactive supply chain management. This study introduces state-space models (SSMs) to maritime logistics forecasting, proposing a novel architecture that combines input-adaptive dynamics with multidimensional temporal processing.We develop and evaluate three SSM variants using Automatic Identification System (AIS) vessel tracking data from 20 East Coast U.S. terminals: (1) Liquid-S4 with polynomial bilinear kernels for input-adaptive dynamics, (2) S4ND for multidimensional signal processing, and (3) Liquid-S4ND-PB—a novel combination extending liquid kernels to multidimensional state-space models for the first time. Principal component analysis reduces the feature space from over 9,800 dimensions to 9 components while retaining 99% of variance. We systematically evaluate historical window lengths (7, 14, 28, and 56 days) across terminals with varying throughput characteristics.Comprehensive evaluation against established baselines (XG-Boost, LSTM, spatial-temporal graph neural networks) reveals that SSMs achieve competitive performance, particularly for medium and low-throughput terminals. Key findings demonstrate that (1) extended historical windows consistently improve SSM prediction accuracy, (2) S4ND outperforms standard liquid-enhanced S4 models on longer temporal sequences, and (3) the proposed Liquid-S4ND-PB architecture effectively captures both short and long-term dependencies. While XGBoost maintains superiority for high-throughput terminals, SSM approaches show significant promise for diverse port environments, with S4ND achieving MAE improvements from 0.51 to 0.47 over previous best-performing methods at medium-throughput facilities.This work establishes SSMs as a viable paradigm for maritime congestion forecasting and provides the theoretical foundation for applying adaptive state-space modeling to complex spatio-temporal logistics problems. The demonstrated scalability and strong performance on extended temporal horizons position SSMs as critical tools for next-generation supply chain optimization.