Machine learning in Maritime Technology: Energy storage size reduction using load forecasting for pipelaying vessels

Machine learning in Maritime Technology: Energy storage size reduction using load forecasting for pipelaying vessels

Menon, Akash (TU Delft Mechanical, Maritime and Materials Engineering; TU Delft Marine and Transport Technology)

Visser, K. (graduation committee)
Vrijdag, A. (mentor)
Polinder, H. (graduation committee)
Godjevac, Milinko (mentor)
Dijk, Marijn (graduation committee)

Delft University of Technology

2019-06-17

During dynamic positioning operations, vessels typically run with an extra generator (spinning reserve) for redundancy purposes such that no single fault will cause the vessel to lose its position or heading. As a result, all other engines on the grid share the load equally to ensure that if a failure were to occur, the remaining healthy generators would be sufficient to satisfy the total power demand. As DP vessels operate on a split-bus mode, the redundancy requirement must be satisfied across both switchboards on the vessel. Past research has shown that significant fuel and maintenance savings can be made by eliminating the spinning reserve with a battery energy storage system. However, as Lithium-ion batteries are relatively expensive, efforts have been made in the past to attempt to incorporate a single battery system than can be connected to either switchboard in the event of a failure, however the analysis was conducted only for deep-water pipelaying operations. Shallow water operations are characteristic of large power surges in contrast to deep-water operations which has been the limiting constraint in ESS design and architecture. In this research, a solution was developed as a split battery design which is arranged such that the power surges can be handled by two independent units connected to one switchboard while a tertiary unit from the secondary switchboard can be independently connected to the one switchboard wherein a failure has occurred, hence acting as a spinning reserve. This allows all the battery units to be maintained at a low state of charge, which maximises the battery life. To complement the design, two power management system integration methods were pursued. Firstly, a rule-based monitoring system was developed which allowed the power management system to make generator start-stop decisions based on a back-looking principle of measuring power demand and power surge characteristics experienced by the battery. This resulted in fuel savings of 3.14 tonnes and a running hour reduction of 52.5 hours. The average engine loads improved on average 10-20\% in contrast to the current situation on the vessel and milder weather scenarios showing nearly a 30-40\% improvement.
And secondly, a neural network based machine learning approach was used to forecast the vessel loads for the day ahead using weather and route plan parameters.The results suggest that in contrast to the rule-based system, the forecast model overestimates or underestimates the load which causes the fuel savings and running hour reduction to be less than the former control strategy. The neural network training was conducted on a data sample of approximately 2 million data points which was found to be insufficient to capture all the dynamics. In the scenario that the model can be trained on several years of data, it could be possible to forecast the load with enough fidelity to allow a 38\% reduction in ESS sizing compared to the base case design used with the rule-based control system.

energy storage system
machine learning
power management system
ESS architectures
system integration

http://resolver.tudelft.nl/uuid:9bca13b7-1256-4f02-9b6e-45d6a49c4cd4

2024-06-17

Student theses

master thesis

© 2019 Akash Menon