Physics-based shaft power prediction for large merchant ships using neural networks

There are currently over 100,000 merchant ships operating globally. To reduce emissions requires predicting and benchmarking the power they use. This is relatively straightforward for calm conditions but becomes almost impossible in larger waves. Design power predictions for ships in weather are typically derived by applying a 'margin' onto a reference 'calm water power'. This is of questionable accuracy as the techniques available to estimate these 'margins' are inaccurate. To improve the accuracy and flexibility of such predictions this paper investigates the use of neural networks. For this, 27 months of continuous monitoring data are used from 3 vessels of the same design, sampled every 5 min. Multiple network sizes are considered and evaluated to determine the quantity and quality of data required for predictions. A key aspect is determining network architectures optimised not just for accuracy, but that give close relationships between the input variables and shaft power. Predictions are compared to the results of a regression, the conventional tool to determine shaft power from measured full-scale data from ships. The predictions from this network are similar in accuracy to those of standard practices, with an error less than 10%, but the scope for further improvements is large.