16. des. 2025
The Challenge: Data Scarcity in Maritime Operations
Leveraging knowledge from similar vessels with sensor connections to vessels without sensor connections.
Global shipping powers nearly 80% of world trade, but its environmental footprint needs attention and actions. The International Maritime Organization (IMO) has set ambitious targets: reduce greenhouse gas emissions by 40% by 2030 and 70% by 2050 compared to 2008 levels[1]. Achieving these goals requires smarter energy optimisation strategies, and one critical piece of the puzzle is accurate power prediction.
Although definitive statistics are scarce, the insight indicates that many vessels still lack accurate systems for monitoring fuel consumption via direct flowmeters—installation remains uneven across fleets. In the GASS project, all partner vessels are equipped with sensor data.
Power consumption directly impacts fuel usage and emissions. Traditionally, this prediction relies on high-frequency sensor data—recorded every few seconds—but installing and maintaining sensors across fleets can be costly and impractical. On the other hand, noon reports, which provide daily operational summaries, are universally available but lack granularity. This creates a data gap that limits predictive accuracy.
Can we bridge the gap?
The Data Paradox: Rich vs. Routine
High-frequency sensor data, collected in intervals, offers precision but demands installations for the majority of the world trade fleet. Conversely, noon reports, which are mandatory operational logs submitted daily, are readily available across fleets but lack the granularity for refined energy models. This mismatch creates a significant data gap, hindering the effectiveness of many optimisation efforts.
Our Approach: Transfer Learning Across Frequencies
The GASS Project proposes a novel strategy: transfer learning. This involves:
Training a neural network model on rich sensor data to capture complex relationships between operational parameters (e.g., speed, RPM, draft, wave height, wind direction) and shaft power.
Fine-tuning the model using noon reports from diverse vessel types—ranging from sister vessels to entirely different ship classes.
By freezing the initial layers of the model and retraining only the final output layer on noon report data, we effectively transfer knowledge learned from rich datasets to those with limited information.
Results: Quantified Gains
Using this method, power prediction errors decreased significantly compared to models trained solely on noon reports:
10.6% error reduction for sister vessels
3.6% for similar vessels
5.3% for different vessels
Initially, models were trained on vessel sensor data up to 2023 and evaluated on 2024 shaft power predictions, yielding error rates of 16.41% for sister vessels, 8.59% for similar vessels, and 13% for different vessels. When using noon reports, errors increased to 42.6%, 18%, and 28%, respectively. Applying transfer learning to noon-report–based shaft power prediction reduced the errors to 32.07% for sister vessels, 14.36% for similar vessels, and 22.70% for different vessels, resulting in a considerable reduction in error.
Additionally, the model successfully forecasted annual power consumption trends for the year 2024—even when trained on sparse data—demonstrating the robustness and generalisability of this approach. These results are documented in the paper.
We further explored transfer learning to transfer high-granularity weather–vessel response knowledge from sensor data to sparse noon reports, where vessel physical dimensions are not important. To evaluate this, models were fine-tuned using noon reports from several tanker vessels, the earlier vessels belonged to different categories of container ships.
Fine-tuning with only six months of noon-report data sampled across 2020–2023 reduced the 2024 shaft power prediction error from 23.85% to 19.17%. Further reducing the training data to a few weeks demonstrated that transfer learning became effective from the third week onward for both sister and different vessels.
Moreover, comparing to a traditional heuristic model that empirically models wind and wave resistance, the transfer learning approach consistently achieved better performance across vessels.
Why It Matters
Emission Reduction: More accurate power predictions enable better voyage planning, leading to lower fuel consumption and emissions.
Cost Efficiency: By augmenting noon-report-based predictions, fleets can avoid sensor installations without sacrificing performance.
Operational Scalability: The method is adaptable to under-instrumented vessels, making it practical for broad fleet application.
Future Directions
The next steps include:
Defining retraining cadence to maintain reliability over time.
Develop metrics for choosing the best sensor-equipped ‘base model’ to use for the new non-sensor-equipped vessel.
Conclusion
While exact numbers aren't publicly disclosed, evidence suggests a significant minority—likely 30–50%—of the global merchant fleet has installed fuel flowmeter sensors, with concentration in larger, trade-sensitive vessels. Adoption is increasing steadily due to regulatory and economic drivers, but full coverage across the ~50,000-ship fleet remains a future milestone.
Transfer learning bridges the gap between data-rich and data-sparse sources. By leveraging sensor-informed models and adapting them to routine operational data, shipping operators gain powerful tools for optimising performance and reducing emissions. In doing so, we move a step closer to a greener, more efficient maritime industry.
Read the full paper: From high-frequency sensors to noon reports: Using transfer learning for shaft power prediction in maritime (link to Cornell University)
The Science Paper Authors
Akriti Sharma, Dogan Altan, Dusica Marijan, Arnbjørn Maressa
References:
The GASS-project is made possible by The Green Platform Initiative (“Grønn Plattform”) funding scheme by Norges forskningsråd, Innovation Norway and Siva SF under the Grant Agreement No. 346603.