TLDR: This paper introduces FuelCast, a new dataset and standardized benchmark for predicting ship fuel consumption using tabular and time-series models. It evaluates various machine learning models, including the novel application of the TabPFN foundation model with in-context learning. Key findings show that TabPFN performs strongly, temporal context and environmental conditions significantly improve prediction accuracy, and the benchmark supports the development of data-driven fuel optimization strategies for the maritime industry.
The shipping industry faces immense pressure to optimize fuel consumption and reduce emissions, driven by both economic efficiency and environmental sustainability goals. Accurate prediction of ship fuel consumption is a critical step towards achieving these objectives, enabling better routing, operational planning, and emissions estimation. However, the field has been hampered by diverse methodologies and a scarcity of high-quality, standardized datasets, making it difficult to compare different modeling approaches effectively.
A new research paper, titled FuelCast: Benchmarking Tabular and Temporal Models for Ship Fuel Consumption, addresses these challenges head-on. Authored by Justus Vigaa, Penelope Muecka, Alexander Löser, and Torben Weis, the paper introduces a novel benchmark and dataset designed to standardize the evaluation of models for ship fuel consumption prediction.
Introducing FuelCast: A Comprehensive Dataset
One of the paper’s key contributions is the release of a new dataset, FuelCast, publicly available on Hugging Face. This dataset is unique because it compiles extensive operational and environmental data from three distinct ships: a small cruise passenger ship (CPS Triton), a large cruise passenger ship (CPS Poseidon), and an offshore supply ship (OSS Ceto). Unlike previous datasets, FuelCast offers long-term, high-resolution data, including detailed vessel-specific operational parameters (like speed, heading, and fuel consumption per consumer) and comprehensive environmental conditions (such as sea temperature, depth, currents, wind, waves, and air temperature). This rich contextual information is crucial for developing robust and generalizable fuel consumption models.
Standardized Benchmarking Tasks
The FuelCast benchmark defines two primary tasks for evaluating predictive models:
-
Tabular Regression: This task focuses on pointwise prediction, where models predict fuel consumption based on instantaneous operational and environmental features at a single moment in time. It helps assess how static parameters influence fuel use, particularly in steady cruising conditions.
-
Timeseries Regression: This more complex task incorporates temporal context, using current and past inputs to predict future fuel consumption. By considering historical patterns, models can learn dynamic behaviors like acceleration or maneuvering, which are vital for understanding fuel consumption across entire voyages.
Evaluating Diverse Modeling Approaches
The researchers evaluated a range of models, from simple baselines to advanced machine learning techniques. These included a third-order polynomial regression model (a speed-based baseline), CatBoost (a gradient boosting method), Multilayer Perceptrons (MLPs), Long Short-Term Memory Networks (LSTMs) for sequential data, and notably, TabPFN. TabPFN is a pretrained probabilistic transformer, a type of foundation model designed for tabular data, and its application in maritime fuel consumption prediction is a first in this domain. The study specifically investigated TabPFN’s potential for in-context learning, even with limited training data (500 or 1000 samples).
Key Findings and Insights
The results of the FuelCast benchmark revealed several important insights:
-
Strong Performance of TabPFN: TabPFN consistently demonstrated strong performance across all evaluated tasks and vessel types, often achieving the lowest Mean Absolute Error (MAE). This highlights the significant potential of foundation models with in-context learning capabilities for tabular prediction, especially in data-scarce scenarios common in maritime transport.
-
Importance of Temporal Context: Models that incorporated temporal information, either through lag-based features or sequential architectures like LSTMs, generally showed improved prediction accuracy. This confirms that dynamic patterns in vessel behavior and external conditions over time are crucial for accurate fuel consumption forecasting.
-
Environmental Conditions are Key: Simple polynomial baselines relying solely on vessel speed consistently underperformed. In contrast, models that included environmental variables like wind, waves, and ocean currents significantly improved prediction accuracy, underscoring their critical role in fuel consumption.
-
Vessel-Specific Variability: Model performance varied across the different vessels, with the offshore supply ship (OSS Ceto) presenting more complex operational profiles and higher variability in predictions compared to the cruise passenger ships. The small cruise passenger ship (CPS Triton) showed the lowest errors due to its more stable, fixed-route operations.
Also Read:
- TimePD: Enhancing Time Series Forecasting Across Domains Without Source Data
- Bridging the Gap in Observability Data: Introducing TelecomTS
Future Directions
While the FuelCast benchmark provides a robust foundation, the authors acknowledge limitations such as the dataset size and the fixed time context for time-series tasks. Future work will explore k-step-ahead prediction, investigate additional model architectures like Informer and TimesNet, and extend the approach to a wider range of vessel types and larger fleets. The framework’s applications include simulating ship models for fuel efficiency optimization, performing what-if scenarios, and analyzing past consumption to identify areas for improvement.
In conclusion, the FuelCast benchmark offers a standardized and reproducible basis for evaluating temporal regression methods in the maritime domain. It demonstrates the feasibility and strong potential of modern machine learning, particularly foundation models, for accurate onboard fuel estimation, even with limited data, paving the way for more efficient and sustainable shipping operations.
