Data driven digital twins for the maritime domain
Digital twins are computational models that replicate the structure, behaviour and overall characteristics of a physical asset in the digital world. In the maritime domain, conventional approaches have relied on mathematical modeling (e.g., linearised equations of motion) and heavy computations for estimating ship resistance and propulsion, seakeeping and maneuverability and overall hull form optimization, treating the vessel as a point body. For instance, the ability to predict
a vessel's future track in confined or congested waters presents a significant challenge due to the fact that as time passes, these models often fall out of sync with their digital counterparts due to changes that happen to the ship (e.g., foulding
affecting maneuverability). In addition to this, mostly due to computational resources required, in real world deployments models are simplified, thus reducing their overall prediction accuracy. In our work, we implement AI-enabled coupled
abstractions of the asset-twin system, which rely on machine learning methods for constant learning of the evolving over time behavior of a vessel based on historical trip data and information related to vessel’s structure and loading capacity. The
evaluation results indicate that the inclusion of vessel and journey specific information is beneficial for the predictions.
This work explores the use of DTs in the maritime domain, focusing mainly on the vessel trajectory prediction problem. Though several works provide accurate results for short-term route forecasting, the presented work is dedicated to predicting the vessel’s future path until the destination port, regardless of the distance. Furthermore, based on clustering and classification techniques, the proposed method uses past movement patterns from historical data and vessel-specific characteristics to return better-suited predictions. Experimental results on real data show that our approach performs prediction of high accuracy even for complex trips, while further analysis indicates that the inclusion of additional static and dynamic information features can refine the prediction accuracy. Finally, the addition of dynamic information related to weather conditions may be considered as an extension of the current work.