FUELSIGHT: MODEL FOR PREDICTING SHIP FUEL CONSUMPTION AND PROVIDING EXPLANATIONS

Анотація

The ability to predict ship fuel consumption is fundamental to improving the efficiency and sustainability of maritime operations. Existing approaches predominantly focus on predictive accuracy while overlooking interpretability, thereby limiting their practical applicability for operational decision-making. This study proposes FuelSight, a unified model architecture that jointly performs multi-horizon fuel consumption forecasting and generates structured, human-readable insights reflecting seafarers’ best practices. The proposed framework leverages multivariate time-series data derived from onboard telemetry and environmental conditions to capture complex operational dynamics. A large language model backbone based on GPT-2 is employed to process sequential inputs and enable both numerical prediction and explanation generation within a single architecture, providing an efficient and coherent modeling approach. Empirical evaluation on the FuelCast benchmark demonstrates that the proposed method achieves competitive performance across multiple vessels and forecasting horizons. In addition to numerical accuracy, the model produces interpretable outputs, which achieve good results when evaluated as a structured classification task. The results indicate that integrating forecasting and explanation within an LLM-based framework offers a promising direction for developing interpretable decision-support systems in maritime applicatio