An important step in determining the performance of a commercial ship is to determine the vessel fuel consumption in offdesign conditions. In previous studies, the vessel fuel consumption is obtained using machine learning algorithms which use navigational data and meteorological
...
An important step in determining the performance of a commercial ship is to determine the vessel fuel consumption in offdesign conditions. In previous studies, the vessel fuel consumption is obtained using machine learning algorithms which use navigational data and meteorological data to train the models. Due to the inaccuracy of the weather hindcasts, the accuracy of these machine learning techniques is not satisfactory for Compagnie Maritime Belge (CMB) which is the company for which this study is performed. Since the vessel motions are a direct result of the meteorological conditions an approach is investigated in this research to replace the meteorological data by motion data to train a machine learning algorithm and obtain accurate results. The vessel motions are captured by a motion sensor on board the vessel. Multiple machine learning techniques were tested in this research. The data used to train the best performing learning algorithm represent the motions by the first and second moment of the nonderivative motion and the second moment of the first derived motion. These moments are obtained from the motion distributions of a time window of 5 minutes. The obtained best parameters for representing the motions were obtained by testing if a spectrum must be provided, how many motion derivatives are required, which moments are needed and how large the time window must be. Although the accuracy of the developed learning algorithm is determined as satisfied no digital twin could be made and tested. Unfortunately, due to the limited amount of data available no separate algorithm could be made which uses the Extra Tree learning, as digital twin base, to determine the overconsumption of the vessel. Only data from one voyage was available. All this data was used to train the learning algorithm to have as much variance as possible in the data set such that a good learning algorithm could be developed.