Predicting VLCC Fuel Consumption with Machine Learning Using Operationally Available Sensor Data

Abstract

In the maritime industry, more accurate predictions of fuel oil consumption (FOC) could yield multidimensional results including more precise bunker calculations, emission reductions, more informed planning and limiting operational costs. However, models often require sophisticated data that may be partially unavailable to operators beforehand.

The present research aims to develop accurate main engine FOC forecasting models that utilize exclusively data from sensors and simple weather data readily available in operational practice. Commonly available sensor data from a Very Large Crude Oil Carrier (VLCC) were used, comprising speed through water, relative wind direction, relative wind speed, mean draft, trim, days since last drydock and laden or ballast vessel state.

Multivariate Polynomial Regression (MPR), Artificial Neural Networks (ANNs) and eXtreme Gradient Boosting (XGBoost) regression models were developed and evaluated based on their predictive accuracy for VLCC FOC. Results indicated that XGBoost had the best performance, yielding predictions within 5% of the true value in more than 86% of the total cases, followed by MPR and ANN. In addition, accurate aggregate FOC forecasting was conducted with XGBoost for a laden voyage and a ballast voyage of a VLCC.

Why it matters

The practical constraint in fleet analytics is rarely the algorithm — it is what data you can actually rely on having. This study deliberately restricts itself to operationally available inputs and still reaches accuracy levels useful for bunker planning, charter party assessment, and emissions estimation.