Fuel Consumption Prediction for a Ferry using ML and In-service Data
Published in Ocean Engineering (Elsevier 2023), this paper demonstrates that ferry fuel consumption can be accurately predicted from operational in-service data without expensive experimental calibration or dry-dock measurements. The machine learning pipeline was validated against real vessel telemetry from BC Ferries, establishing a baseline for AI-driven emissions monitoring in commercial shipping. It illustrates the practical value of building AI systems that work within the data constraints that actually exist in regulated industries.
Ferry operators face a practical dilemma: fuel efficiency depends on operating conditions (speed, loading, sea state, weather) in ways that are non-linear and difficult to model analytically. Traditional approaches use computational fluid dynamics (CFD) models that require detailed vessel geometry and expensive experimental calibration.
Data and Method
We worked with in-service operational data from a BC Ferries vessel, including engine telemetry, GPS and speed logs, draft measurements (load proxy), weather data, and fuel flow meter readings. This data is routinely collected but rarely used for optimization.
Feature engineering: Speed through water, engine RPM, pitch and roll from IMU, wind speed and direction relative to vessel heading, draft as a loading proxy, and derived features (power-speed relationship, wind load estimate).
Model selection: We evaluated multiple regression approaches - gradient boosted trees, ridge regression, and a simple LSTM for temporal dependencies. Gradient boosting performed best on the validation set; the LSTM was competitive on steady-state passages.
Validation: The final model was evaluated on a held-out set of voyage segments not used in any training or feature selection step. We report mean absolute percentage error separately for different operating regimes (port approach, open water, adverse weather), since failure in high-fuel-consumption regimes carries disproportionate cost.
Why This Matters
The result demonstrates that AI can deliver useful operational insights from data organizations already collect, without requiring new infrastructure or experimental downtime. This is a common constraint in regulated industries: you cannot run controlled experiments freely, so your AI systems must work with observational data and be validated under those same constraints.
The per-regime error breakdown is an example of disaggregated evaluation - measuring performance where it matters most rather than averaging over a distribution that may hide important failure modes.