Optimizing Energy Consumption in Maritime Transportation with Machine Learning Methods (in collaboration with Cetasol)

Title Optimizing Energy Consumption in Maritime Transportation with Machine Learning Methods (in collaboration with Cetasol)
Summary Develop machine learning methods for forecasting fuel consumption, path, and motion planning, with historical data from furries operation.
Supervisor Slawomir Nowaczyk, Yuantao Fan, Hadi Fanaee, or Mohamed Abuella
Level Master
Status Draft

This project will explore different AI-based approaches to understanding ferry operations, recognizing their activity at different abstraction levels, and finding energy-efficient operational solutions using historical and real-time onboard sensor data. The aim is to develop AI-based systems that will assist, and guide, the ferry operator toward Green Operation, by providing decision support on energy-efficient operation solutions. The project is a collaboration with Cetasol, a data-driven startup active in sustainability and has extensive knowledge of marine applications.

Reducing the fuel of marine traffic is a complex task that could involve the perception of the operating conditions, motion control, path planning, etc. The end goal is to provide a full motion control system for the vessel traveling from point A to B. This would include the planning of a path (including heading-course over ground difference) as well as speed (including acceleration pedal position) over time. To optimize fuel consumption, a recommendation system can be trained with historical data, with the help of a cost function that takes multiple factors into consideration, e.g. cost of fuel, time, path constraint, etc. The problem can also be formulated as a graph search. The graph will be a multi-dimensional graph where longitude, latitude, speed, heading, wind, etc can be considered as dimensions. The graph should be sufficiently detailed in different ranges of dimensions. For example, longitude and latitude can take longer steps in the open sea and shorter close to harbors. Several topics can be selected for the thesis:

1) Fuel modeling and forecasting

The modeling can be performed in offline and online modes. The offline model will forecast fuel consumption in a future time without any real-time input from the vessel. This will be based only on the average behavior of the vessel, and may not be very accurate, depending on how special the route is taken. The online model would be using time series from onboard live data. This may yield a more accurate prediction as unmeasurable parameters such as weight and hull fouling could be implicitly included in the modeling because of their live effect on fuel consumption.

2) Graph construction

The goal is to construct a graph with a desired resolution or abstraction, based on a set of physical parameters given. A big part of the motion planning would lay inside the possible states of the vessel. A state would be the physical state of the vessel, including position, speed, rotation, etc. As those parameters are listed the complexity of the graph grows exponentially, nlog(n). Therefore the available steps through this grid could be heavily limited by the vessel dynamics as well as going for larger steps where accuracy can be reduced. This would for example be to not let the vessel go from speed=10m/s to 0 in one short physical step as it would be either impossible or insufficient.

3) Motion planning

Motion planning can be carried out by classical algorithms such as A* or RRT* to traverse the graph. This would have to be calculated live as the vessel travels to different destinations and gets off route every now and then. The computation needs to be efficient. The goal is to produce an optimal path from the algorithm, with an estimated bound on time and variation of the path, given a graph or a set of graphs and a cost function.