Improving prediction performance of machine learning models in maritime

Designing new architecture or enhancing existing ML models to improve predictive performance in maritime operations, such as estimating fuel consumption during voyage planning.

This project is part of the GASS (Green AI for Sustainable Shipping) initiative, funded by The Green Platform Initiative (https://www.gass-project.com/). Simula collaborates with Navtor (https://www.navtor.com/), a provider of ship operational data. The ship operational data spans multiple fleets, with historical records starting from 2020, and is updated in real time on a regular basis. The goal is to develop a chain of machine learning models to estimate RPM, shaft power, and fuel consumption for upcoming voyages with minimal errors. Various approaches have been explored, including heuristics, physics-based methods, traditional polynomial curve fitting, purely data-driven models, physics-guided models, and KAN models. However, each approach has limitations and tends to underperform for some vessels on different voyage scenarios provided by Navtor. We anticipate that advanced techniques, leveraging machine learning or data augmentation/fusion, can enhance accuracy or introduce a new paradigm of higher-performing models through this project.

Goal

The marine data is tabular, derived from sensor recordings, ship reports, and voyage logs, with certain fields augmented from public weather data sources. The objective is to enhance the accuracy of existing models or test new architectures under varying weather and sea conditions experienced during historical voyages. This includes creating an ensemble of models, exploring advanced data fusion techniques, and introducing or removing dependent variables based on analysis of existing models. Data-driven models, in particular, struggle to extrapolate to weather conditions or ship sailing profiles not represented in historical data and often fail to predict values beyond specific ranges, likely due to being struck in local minima. Models should address this limitation and seek to overcome it.

Learning outcome

1. Critical evaluation of ML models
2. Handling tabular & time-series data
3. Feature engineering & variable selection for marine domain
4. Work with real-world data and use cases in shipping industry

Qualifications

• Familiarity with machine learning frameworks
• Python programming

Supervisors

• Akriti Sharma

Collaboration partners

• Navtor
• GreigStar
• SinOceanic
• ScanReach
• Maritime Cleantech