Fredrik Lorentzon and Jakob Bruchhausen, Electrical Engineering

​Title: Passenger vessel models for fuel efficient fixed-routes

Students: Fredrik Lorentzon (MPCAS),  Jakob Bruchhausen (MPCAS)

Industrial supervisors: Joel Odlund; Simon Johansson, Ethan Faghani, Cetasol
Supervisor/Examiner: Balazs Kulcsar

Minimizing fuel consumption of marine vessels has environmental, economical, and health related advantages. Today, many vessels lack software to help operators drive efficiently. We present a complete reinforcement learning framework that models the behaviour of a marine vessel and optimizes it with regard to its fuel consumption. The reinforcement learning algorithm consists of an environment and an agent. The environment is built using an LSTM neural network trained on real-life data. The data is analyzed to find relevant features, strongly correlated with fuel consumption, and to remove irrelevant ones. The agent is built using a deep Q-learning architecture. Moreover, a Hidden Markov Model was implemented to infer latent variables. It evaluates states at each time-step and feeds its output to both the environment model and the agent. The LSTM and Hidden Markov models are built on data from the archipelago passenger vessel Burö, operating in the Göteborg archipelago.
The model manages to describe the vessel's behavior relatively well, when evaluated on test data. Implementation of the Hidden Markov Model significantly improves this result. The agent manages to find a good but not optimal policy. In conclusion, the proposed reinforcement learning algorithm is not accurate enough to be implemented into real-life applications. However, we present many important insights to future works, such as the reinforcement learning architecture and the importance of estimating latent variables.

Fredrik Lorentzon and Jakob Bruchhausen
Category Student project presentation
Starts: 23 May, 2022, 13:00
Ends: 23 May, 2022, 14:00

Page manager Published: Tue 10 May 2022.