Data-driven Prediction of Added Resistance on Ships in Waves

In combination, international regulations and digitalization have led to increasing acquisition of in-service data from ships, including navigational and engine-related recordings. However, associated analyses are challenging due to the limited availability of reference data and operation in an environment governed by uncertainty, e.g. related to wind and waves. The added wave resistance can take up a relatively large proportion of a ship’s required engine power. Still, its magnitude is notoriously difficult to determine – especially in short and oblique waves. Hence, the overarching goal of this thesis is to enhance the predictability and understanding of added resistance in real conditions using statistical and machine learning methods.

The first part of this thesis addresses the added resistance in regular waves and focuses on estimating the associated quadratic transfer function. Several machine learning methods are trained on the results of numerical methods for various hull shapes. The importance of data preprocessing and the generalization capability of neural networks stand out. A separate study investigates the uncertainty of the added resistance transfer function and its estimation via a semi-empirical formula. For this reason, parameter calibration of the underlying method is performed separately for both blunt-type and slender vessels. The methodology is based on experimental data, and a 90% prediction interval is implemented to improve the method’s transparency in an adapted version.
An estimate of the corresponding wave energy density spectrum at the exact spatiotemporal point of operation is needed to calculate added resistance in actual seaways characterized by irregular waves. Hence, in the second part of the work, in-service sensor data of a container vessel and different neural networks are utilized for sea state identification. Overall, neural networks can produce a satisfactory mapping from measured vessel responses to sea state parameters.
The third part of the work is about correlating theoretical and empirical estimates of the mean added resistance in actual conditions using in-service data from a fleet of more than 200 container vessels. Theoretical estimates are calculated in the spectral domain by combining the adapted semi-empirical procedure and historical wave data. The empirical predictions are determined using the measured shaft power combined with a resistance decomposition. It is confirmed that the actual added resistance is highly complex to determine and subject to significant uncertainty.
In the fourth and final part of the work, synthetic monitoring data of a standard tanker (KVLCC2) is simulated for varying operating conditions using a semi-empirical framework. It is shown that the performance data of ships is subject to a distributional shift, and thu neural networks are trained adaptively to pinpoint the added power due to biofouling. It turns out that methods for incremental learning are influenced by data quality and that the overall methodology may be applicable for determining the effect of energy-saving devices.

In the future, this work may serve the development of digital twins used to assess the safety and the energy efficiency of ship operations. Due to the sensitivity of machine learning methods to data quality and availability, it seems favorable to follow a hybrid approach in combination with established physical models. In a practical context, the findings of this work are applicable for voyage optimization or performance monitoring and may assist the maritime industry on its path to becoming sustainable