TLDR: A new research paper introduces a transfer learning method to improve shaft power prediction in maritime vessels. By pre-training a model on high-frequency sensor data and fine-tuning it with low-frequency noon reports, the approach significantly reduces prediction errors across sister, similar, and different vessel types. This innovation helps bridge the accuracy gap between sensor data and noon reports, enabling better fuel efficiency planning and predictive maintenance even with limited data, and accurately forecasts future power consumption trends.
Optimizing energy consumption is a critical challenge in global maritime transportation, directly impacting operational costs and environmental emissions. A key factor in this optimization is accurately predicting a vessel’s shaft power, which is the mechanical power transmitted from the engine to the propeller shaft and directly influences fuel usage.
Traditionally, predicting shaft power relies on either high-frequency sensor data or low-frequency daily noon reports. High-frequency sensor data, recorded every few seconds, offers high accuracy but is often expensive to install and maintain, and not all vessels are equipped with such systems. Noon reports, manually generated daily by the captain or a logging system, are more accessible but less accurate due to being averaged over 24 hours and prone to human error.
A new research paper, “From high-frequency sensors to noon reports: Using transfer learning for shaft power prediction in maritime”, proposes an innovative solution to bridge this gap. The study introduces a transfer learning-based approach that leverages the strengths of both data types. Transfer learning involves training a model on a rich source of data (high-frequency sensor data from one vessel) and then fine-tuning it with a smaller, more accessible dataset (low-frequency daily noon reports from other vessels).
How the Approach Works
The methodology involves training a neural network initially on detailed sensor data from a ‘source’ vessel. This pre-trained model then has its knowledge adapted and refined using noon reports from ‘target’ vessels. The target vessels were categorized into sister vessels (identical dimensions and configurations), similar vessels (slightly different in size and engine), and different vessels (distinct dimensions and configurations). Key features used for prediction included speed through water, shaft rotational speed (RPM), draft amidships, wave height, swell height, and wave and wind directions.
Key Findings and Impact
The experiments demonstrated significant improvements in shaft power prediction accuracy. Compared to models trained solely on noon report data, the transfer learning approach reduced the mean absolute percentage error (MAPE) by 10.6% for sister vessels, 3.6% for similar vessels, and 5.3% for different vessels. This effectively narrowed the performance gap that typically occurs when switching from high-accuracy sensor data to less accurate noon reports.
Furthermore, the study showed that the method could accurately predict future power consumption trends for 2024 and 2025, even when the base model was trained on data up to 2023. This temporal validation highlights the robustness of the approach in forecasting future performance based on historical data.
Also Read:
- Optimizing Container Stowage: A Deep Dive into Reinforcement Learning Benchmarks
- ZEROSHOTOPT: A Pretrained Model for Efficient Black-Box Optimization Without Tuning
Broader Implications
This transfer learning method offers several benefits for the maritime industry. It enables more accurate fuel efficiency planning and facilitates predictive maintenance, even for vessels without expensive sensor installations. By integrating prior knowledge from high-frequency data, it can also assist in detecting outliers in sensor-based predictions and improve data quality management for noon reports. This means ship operators can predict future performance degradation with a limited number of noon reports, leading to better operational decisions and cost savings.
While promising, the study acknowledges limitations such as the number of vessels evaluated and the need to investigate optimal retraining frequencies and the individual effects of weather parameters. Future work aims to address these aspects to enhance scalability and generalizability across diverse vessel types and operational conditions.
