Skip to main content

Results of Energy Management

  • The load profile for one voyage, including ferry departure, transit from A to B, and arrival at port, is explained. The total power demand from the propulsion system and onboard systems is emphasized.
  • The offline optimal case, in which the future power demand for the voyage is known, is explained. It is shown that most of the transit power is supplied by the fuel cell with relatively constant output, after which power is reduced and the fuel cell is shut down before port arrival while shore power is used to charge the battery.
  • How different reinforcement learning agents are used to distribute fuel-cell and battery power without knowledge of the next time step is explained. It is highlighted that discrete Q-learning follows the general trend but also causes frequent and unnecessary power fluctuations.
  • It is explained that deep reinforcement learning with a deep network yields more confident and stable control decisions. It is emphasized that the resulting power split is very close to the offline optimum despite uncertainty.
  • The comparisons of emissions, emission ratios, and costs with the offline optimum are explained. It is concluded that most agents perform only a few percent away from the optimum, and a generic energy management framework applied to an actual ship is introduced.
Updated on Jan 18, 2026