Data driven analysis of ship performance in various ice conditions.

Abstract

Maritime activity in polar regions is increasing due to climate change resulting in opening of new Arctic Sea routes for commercial shipping. This increases the importance of reliable ship performance methods to ensure operational safety and economic viability. Classical methods such as Lindqvist and Malmberg are conservative and are developed based on assumptions. The increasing volume of actual ship machinery data and ice data provides a good opportunity to develop data driven methods. Existing data driven studies are based on coarse AIS datasets with limited machinery information. Additionally, they rarely provide a comparison with h-v curves developed using classical methods.

This thesis investigates data driven methods for predicting the performance of a ship navigating in ice using machinery aware full-scale sea trial data. It compares the data driven methods with classical semi empirical models by constructing h-v curves for the PSRV S.A. Agulhas II and Aranda using various Machine Learning techniques. Different machine learning regressors were trained on full-scale data by segregating the data based on various ice conditions (level and ridge ice).

In case of level ice, the raw sensor logs were time-synchronized and filtered to select only level-ice data while the ship is moving straight ahead. Outliers in the (h,v) data were removed using a Mahalanobis-distance filter, and a correlation-based feature-selection to retain the most predictive variables. The Extra-Trees ensemble achieved the highest predictive accuracy on in-sample test data. The resulting data driven h-v curves lie above the Lindqvist based h-v curve indicating its conservative nature.

In case of ridge ice due to non-continuous nature of loads and resistance 30s bins were created, and a 33s delay was rectified in ice thickness and speed. Polynomial and Extra-Tree models were used for the creation of h-v curve using the mean values of the bins. Both of these models indicated a monotonic loss of speed with increasing ridge thickness. The data driven h-v curves and the Malmberg based curve agree on the overall trend, but the Malmberg tends to predict slightly higher speeds at very thin ridges than those implied by the measured and learned curves.

The polynomial model struggled to fit the scattered data in both level and ridge ice. To conclude, machine learning regressors can be used for performance predictions if ample high-quality data is available.