Power prediction for a vessel without recorded data using data fusion from a fleet of vessels

Recent legislation in shipping applies additional pressure to reducing fuel consumption. However, this is impossible without accurate power prediction, as it is required to allow comparisons between novel efficiency improving advancements and to have confidence in route optimisation. This prediction is particularly difficult in rough weather, which the traditional prediction methods struggle to account for. Neural networks trained on an operational dataset from the vessel are a potential solution to this problem, as they have been shown to predict powering to a mean error of 2% across all weather conditions. However, the gathering of these data is expensive and time consuming. There is currently no literature looking at how data from one vessel can be used to make predictions about another, reducing the cost and allowing prediction of the performance of new vessels. This paper investigates the accuracy in predicting powering for an unseen vessel, using a neural network trained on a fusion of data, from a range of sensors located on other vessels in a fleet. It demonstrates the level of extrapolation that can be achieved from the use of multiple datasets on a real application and suggests that, for the fleet of vessels used, ship parameters are less important for accurate power prediction than having sufficient data across the desired prediction domain. It concludes that prediction of around 4% error can be achieved for most ships in the fleet and discusses the cause of the higher errors seen for a minority of other vessels.

Automated summary

Recent legislation in the shipping industry has increased pressure to reduce fuel consumption, requiring accurate power prediction for comparison of efficiency advancements and route optimisation. Neural networks trained on operational data from vessels show promise in predicting powering with a mean error of 2%, but data gathering is costly. Fusion of data from sensors on multiple vessels in a fleet can achieve around 4% prediction accuracy for most ships, showing the potential for extrapolation and highlighting the importance of sufficient data across the desired prediction domain.