Predicting CO2 Emissions from Maritime Transport with Feature-Oriented Methods

Prediction of CO2 emissions from maritime transportation is both a strategic and mandatory activity for shipping companies, which, from January 2018, are forced by the European regulations to set up a system for daily Monitoring, Reporting and Verification (MRV) of emissions from their fleet. Today’s multi-sensor systems are able to acquire measurements for a massive number of variables, often collinear and with a non-stationary behaviour, that are stored every five minutes, taking the form of profiles. In this setting, multivariate statistical process monitoring methods commonly used for monitoring batch processes, can be applied. However, the use of those methods is sometimes not straightforward and requires considerable additional data pre-processing, such as, for example, unfolding the data to handle the three-way structure (which leads to a very large number of pseudovariables), aligning and synchronizing the data (because each voyage has a different length and therefore different number of observations need to be warped into the same domain as well as the main events of the profiles should be aligned). The results can be, indeed, influenced by pre-processing techniques as well as the order in which they are applied. For those reasons, feature-oriented methods were proposed, that reduce the modelling complexity and lead to more parsimonious modelling. Examples include statistical pattern analysis and profile-driven techniques. In particular, statistical pattern analysis make use of four statistical features computed for the measured data, corresponding to the first four moments. On the other hand, profile-driven features claim better performance by identifying a dictionary of features that are characteristic of each variable. In this work, with the goal of predicting CO2 emissions, four classes of regression methods, namely variable selection, latent variable, penalized regression, and tree-based ensembles, are compared by means of the root mean squared error resulting from 50 iterations of double cross-validation. A real case study on shipping data acquired on board of a roll-on/roll-off passenger cruise ship owned by the shipping company Grimaldi Group is presented to illustrate both the satisfactory predictive performances of the proposed methods and the ease of use, which do not require neither pre-processing nor data alignment.