Deep Learning-Based Forecasting of Containerized Maritime Trade: A Decomposition and Attention Approach
Yasin Ozkan, Mahmut Mollaoglu
article (Uluslararası, Hakemli) | Yazar(lar): Yasin Ozkan, Mahmut Mollaoglu | Yıl: 2026 | Dergi/Kurum: IEEE ACCESS | İndeks: WOS.SCI | DOI: 10.1109/access.2026.3712037 | Anahtar Kelimeler: Forecasting, Modeling, International trade, Long short term memory, Timing, Containers, Learning (artificial intelligence), Archite
This study aims to reveal the current container trade and future forecasting estimations for T & uuml;rkiye. To address this aim, study develops a deep learning based long term forecasting framework for containerized trade by integrating time series decomposition with and attention mechanism. This study proposes a decomposition based deep learning architecture for long term forecasting of containerized maritime trade, with a primary methodological contribution centered on improving predictive stability across short- medium- and long-term horizons. The approach begins with a time-series decomposition procedure that separates container throughput data into trend, seasonal, and residual components, thereby reducing noise and isolating distinct temporal dynamics. Each decomposed component is modeled using a hybrid neural network structure, integrating Long-Short Term Memory (LSTM) layers to capture sequential dependencies and an attention mechanism to dynamically assign weights to historically relevant observations. Empirical findings demonstrate that the decomposition attention model consistently outperforms baseline approaches, particularly in medium- and long-term forecasts where traditional time-series models exhibit declining accuracy. The results confirm that integrating structural signal decomposition with attention-enhanced deep learning significantly improves robustness, reduces forecast error propagation, and enhances temporal feature extraction in container trade prediction.