Estimation of moored ship motions using a combination of machine learning techniques

Abstract

The moored ship motions can cause problems for the efficiency of the operation, and for the people and equipment involved. Therefore, being able to predict movements and anticipate possible risk situations is of great interest to operators and the port community. This work presents a methodology applying different machine learning techniques that has allowed positive results to be obtained for this objective, with particular emphasis on the highest values (outliers), which are usually associated with problematic situations. The field campaigns carried out allowed 77 different vessels to be monitored in the outer port of A Coruña (Spain). The techniques used were gradient boosting (GBM), a neural network (DNN), a quantile regression (qReg) and several models generated by stacking (GBM*). The results indicate a lower root mean square error (RMSE) with the use of the latter technique (the validation on the swell is 0.13 m, while the DNN is twice as high), and a better performance on most motions in the outlier subset than those obtained with the individual models (the validation on the outlier subset for the pitch gives an RMSE of 0.12° and 0.2 for the GBM). Finally, the results show that this methodology can be extrapolated to other ports.