The maritime industry, a vital artery of global trade, is confronting a pressing challenge: minimising its environmental footprint. One strategic approach is the optimisation of fuel consumption by ships, which represents a significant financial burden and a major source of greenhouse gas emissions.
At the forefront of efforts to refine robust fuel consumption prediction (FCP) is a team led by Xiuju Fu, Director of the Maritime Research Programme and a Senior Principal Scientist at A*STAR’s Institute of High Performance Computing (IHPC).
“Accurate FCP enables stakeholders to devise effective decarbonisation strategies, including technical interventions such as alternative energy sources, retrofitting and engine derating, as well as operational optimisations such as weather routing and fleet operation strategies,” stated Fu.
Nevertheless, developing accurate FCP models presents its challenges. Existing models are mostly tailored to specific ships or routes and often neglect real-world variables such as incomplete ship data. Together with the Monohakobi Technology Institute, Fu’s team collaborated with a world-leading transportation company to use their shipping data in developing two multi-ship machine learning models.
These models were based on distinct methodologies: one on a tree-based gradient boosting algorithm (XGB) and the other on an artificial neural network (ANN).
Fu explained, “It was critical for us to work closely with our industry partners to understand the influencing factors that can affect the fuel consumption rate and any potential data quality issues.” The team proposed an application-oriented testing regime, developed with domain experts from the transportation company, which assessed the robustness of these FCP models under challenging but realistic conditions.
While previous studies often proved one model’s superiority over another, the IHPC team’s findings were less conclusive, reflecting the complex nature of real-life scenarios. “The performance comparison between ANN and XGB yielded a more complex outcome, with no definitive dominance of one algorithm over the other,” reported Fu.
Looking ahead, Fu advocated for a cooperative approach in model development rather than seeking a singular best model. To this end, Fu proposed focusing on enhancing data quality, broadening data sources across varied vessel fleets, and leveraging multiple machine learning models to effectively address diverse decarbonisation needs. This enhanced testing regime could allow researchers to better understand the strengths of various models and leverage this knowledge to develop more effective and robust FCP models.
The team’s current focus is on expanding the bank of testing scenarios to encompass a broader spectrum of artificial intelligence (AI) quality, aiming to cover more aspects of maritime AI models.
The A*STAR-affiliated researchers contributing to this research are from the Institute of High Performance Computing (IHPC).
