TCFO - Maritime and Port Analytics

The spot freight rates of Liquefied Natural Gas (LNG) carriers have become increasingly volatile due to rising demand for cargo transportation and dynamic market conditions. This study aims to enhance the accuracy of freight rate volatility predictions using artificial intelligence, evaluating the Gated Recurrent Unit (GRU) model's performance for long-term monthly forecasts. Key predictive variables, including LNG prices, LNG inventory levels, sailing speeds of LNG carriers, port call indices, and charter rates, were selected for analysis. The study compares the performance of the Long Short-Term Memory (LSTM) model and the GRU model by incrementally extending the prediction period up to 16 weeks. The results demonstrate that the GRU model achieved approximately 70% lower Mean Squared Error (MSE) compared to the LSTM model, achieving superior accuracy even in long-term predictions. Moreover, the GRU model exhibited a smaller MSE increase over extended periods, demonstrating greater stability. These findings demonstrate the GRU model as an efficient and reliable tool for predicting LNG spot freight rate fluctuations. This study emphasizes the importance of selecting appropriate models to accurately capture the dynamics of highly volatile markets and provides valuable insights into LNG transportation planning and decision-making.

Crude oil holds strategic importance as the primary source for refined products such as bunker fuel. Accurate crude oil price forecasting is crucial for global trade and the international shipping industry. This study introduces a novel hybrid model for forecasting Brent crude oil and Rotterdam HSFO 380cst prices from June to November 2023. The proposed approach integrates deep neural networks with dimensionality reduction techniques (DRTs), using a dataset comprising various bunker prices, currency exchange rates, and other shipping market indicators. Unlike previous studies, which often fail to incorporate both feature selection and feature extraction as DRTs in forecasting Brent crude oil and Rotterdam HSFO 380cst prices, this study employs a combination of mean square error filter and principal component analysis alongside a deep learning (DL) model to enhance forecasting accuracy and efficiency. The experimental results demonstrate that the DL model with DRTs outperforms both the DL model without DRTs and the ARIMA (1,1,2) model in forecasting Brent crude oil prices, with the mean absolute percentage error decreasing from 31% (DL model without DRTs) to 14% (DL model with DRTs). Furthermore, the Diebold-Mariano test confirms a statistically significant difference between the forecasts generated by the two models.

Accurate forecasting of cargo throughput is vital for optimizing port operations and global logistics. This study focuses on major Australian coal-exporting ports, including Newcastle, to improve predictive accuracy by integrating the Automatic Identification System (AIS) data and statistical indicators. Seasonal-trend decomposition using locally estimated scatterplot smoothing and discrete wavelet transform reveals that long-term trends and seasonal patterns dominate throughput variability, while short-term fluctuations exert comparatively limited influence. Baseline forecasts are first established using univariate Prophet and long short-term memory (LSTM) models, demonstrating that LSTM outperforms Prophet according to the metrics. Building on these baselines, a multi-feature LSTM model initially integrates AIS-derived and economic features, with permutation importance pinpointing those most relevant. This approach significantly improves accuracy across shorter and longer forecast horizons, indicating that operational efficiency metrics drive near-term predictions while international demand factors grow in importance over extended windows. Additional testing at a more volatile port with a monthly forecast confirms that feature selection can flexibly adapt to varying contexts, underscoring the potential for data-driven decisions in port and logistics management.

Marine resources play a crucial role in sustaining human life, supporting biodiversity, and underpinning global fisheries, shipping, and aquaculture. Protecting marine ecosystems from illegal, unreported, and unregulated (IUU) fishing, as well as overfishing, is an urgent necessity. To achieve this goal, we propose a semi-supervised methodology that automatically detects fishing activities based on Automatic Identification System (AIS) data when there is a lack of labelled data. Instead of using raw data as input, we introduce an innovative mobility pattern algorithm based on convolutional auto-encoders (CAEs) to extract the hidden representation of vessel trajectory as input features, which can reduce computation complexity. To this end, we leveraged the unsupervised learning to infer the labels of vessel trajectories, which indicate fishing and sailing activities. The inferred labels obtained from the unsupervised learning are used to train the classifier for fishing activity detection. The proposed approach is then tested on a real vessel trajectory dataset. The experimental results show that our proposed approach can effectively detect fishing activity when the labelled data is scarce.