Data-Driven Hull-Fouling Cleaning Schedule Optimization to Reduce Carbon Footprint of Vessels

In response to climate change, the International Maritime Organization has introduced regulatory frameworks to reduce greenhouse gas emissions from international shipping. Compliance with these regulations is increasingly expected from individual shipping companies, compelling vessel operators to lower the CO2 emissions of their fleets while maintaining economic viability. An important step towards achieving this is performing regular hull and propeller cleaning; however, this entails significant costs. As a result, assessing whether ship performance has declined sufficiently to warrant cleaning from an environmental and economic standpoint is a critical task to ensure both long-term viability and regulatory compliance. In this paper, we address this challenge by proposing a novel data-driven dynamic programming approach to optimize vessel cleaning schedules by balancing both environmental and economic considerations. In numerical experiments, we demonstrate the usefulness of our proposed methodology based on real-world sensor data from ten tramp trading vessels. The results confirm that over a four-year period, fuel consumption can be reduced by up to 5%, even when accounting for the costs of one or two additional cleaning events.

💡 Research Summary

The paper tackles the pressing problem of determining optimal hull‑and‑propeller cleaning schedules for commercial vessels under increasingly stringent IMO and EU emissions regulations. Recognising that cleaning events are costly and that the degradation of hull performance due to bio‑fouling directly inflates fuel oil consumption (FOC), the authors develop a data‑driven decision‑support framework that simultaneously minimizes environmental impact and operational expenses.

First, a robust FOC prediction model is built using a suite of machine‑learning regressors (Support Vector Regression, Extremely Randomized Trees, XGBoost, Random Forest, etc.). The models are trained on four years of high‑frequency sensor data from ten tramp‑trading ships, incorporating variables such as engine load, speed, weather conditions, and operational states (anchoring, low‑speed transit, high‑speed cruise). Model selection is guided by cross‑validation metrics (R², MAE, RMSE). To overcome the “black‑box” nature of many algorithms, the authors apply SHAP (Shapley Additive Explanations) to quantify each feature’s contribution to predicted fuel consumption. The analysis reveals that “Days Since Cleaning” (DSC) – the elapsed time since the last recorded cleaning – is the most influential predictor, confirming its suitability as a proxy for hull fouling intensity. Additional derived variables (e.g., cumulative anchorage hours, low‑speed exposure) are introduced to capture differing fouling growth rates under various operating conditions.

Second, the paper formalises the cleaning cost structure. Two cleaning modalities are distinguished: (i) dry‑dock cleaning, which involves removing the vessel from water, high‑pressure water‑jet or sand‑blasting, and often re‑coating with anti‑fouling paint; this option incurs high direct costs (USD 5 000–50 000) and significant downtime. (ii) In‑water cleaning, performed by divers or specialised robots, is cheaper but less effective at removing fouling. The authors compile cost estimates from literature and internal records, also accounting for lost revenue during downtime.

Third, a dynamic programming (DP) optimisation model is constructed. The state space consists of discrete time steps (days) and the number of cleaning actions taken. Transition costs combine the predicted fuel consumption for the ensuing interval (derived from the ML model) and the cleaning cost if a cleaning action is scheduled at that step. The objective is to minimise total cost over a four‑year horizon while respecting regulatory constraints (e.g., maintaining a required Carbon Intensity Indicator rating) and operational constraints (e.g., limiting vessel unavailability). The DP algorithm efficiently identifies the schedule that yields the lowest cumulative cost.

Empirical results demonstrate the practical value of the approach. Applying the DP‑derived schedules to the ten‑ship dataset yields an average fuel consumption reduction of 3–5 % compared with the baseline practice of cleaning every three to five years. Even when one or two additional cleaning events are introduced, the overall cost (fuel plus cleaning) remains lower, confirming that the marginal savings in fuel outweigh the extra cleaning expense. The optimal schedules tend to cluster cleaning actions just before periods where the ML model predicts a sharp rise in fuel consumption, effectively “resetting” the DSC variable and curbing the fouling‑induced drag penalty.

The study’s contributions are threefold: (1) a validated, explainable ML model for ship fuel consumption that integrates an indirect fouling indicator (DSC); (2) a cost‑aware DP optimisation framework that balances environmental benefits with economic realities; (3) a real‑world case study showing up to 5 % fuel savings and cost reductions for tramp vessels. Limitations include reliance on accurate cleaning logs (DSC is undefined for missing records) and the absence of direct fouling measurements such as underwater imagery. Future work should incorporate low‑cost underwater cameras and deep‑learning image classification to enrich the fouling representation, extend the optimisation to stochastic DP formulations that capture fuel price volatility, and test scalability across larger, more diverse fleets.