Deep Learning-based Fuel Oil Consumption Prediction for Twin-screw Ship

Abstract

With growing pressure from the International Maritime Organization (IMO) to reach net-zero emissions by 2050, accurate prediction of ship Fuel Oil Consumption (FOC) has become increasingly important. Although numerous deep learning-based FOC prediction models have been proposed, studies targeting twin-screw ships remain scarce. In this paper, we propose a dual-branch deep learning architecture for multi-step FOC prediction of twin-screw ships, designed to capture the distinct operational characteristics of the port and starboard propulsion shafts. Each branch independently encodes feature-axis relationships using a Multi-Layer Perceptron (MLP) encoder and temporal patterns using a Gated Recurrent Unit (GRU) encoder. A Squeeze-and-Excitation (SE) block is employed to adaptively fuse the encoded representations, and an autoregressive GRU decoder then generates multi-step FOC predictions. The model is trained on five months of operational data from a RO-RO passenger ship navigating the Gunsan–Shidao route, supplemented with wave data from the Copernicus Marine Environment Monitoring Service (CMEMS). Experiments under various prediction conditions demonstrate that the proposed model consistently outperforms nine baseline models, including LSTM, GRU, transformer variants, and recent hybrid architectures, across the RMSE, MAE, and MAPE metrics.