Arriving on time using uncertainty aware deep learning
conference paper
The Port of Rotterdam is an important hub in international maritime supply chains. Efficient functioning of the port requires accurate planning of the processes, thus reducing waiting times and costs for all parties. Therefore, predictions should be incorporated in the planning of all logistic processes. In particular, uncertainty information is fundamental to estimate the reliability of predictions and to adjust the planning. It is estimated that only the direct benefits of optimizing a port call can add up to $80,000 and CO2 can be reduced up to 240 ton per port call (MaritiemNieuws, 2017). In this research we combine the domain knowledge about logistics with data driven approaches to obtain an uncertainty aware Estimated Time of Arrival (ETA) prediction for incoming vessels by using deep learning.
Our proposed network provides better ETA predictions than the estimations currently in use, and provides an uncertainty estimation on the prediction. Our experiments show that the uncertainty cannot be reduced by adding more training data of the same type as the one available. Finally, our experiments show the importance of domain knowledge combined with data-driven techniques, to understand the behavior of the network; the lowest uncertainty is
anyway obtained by combining all available parameters.
Our proposed network provides better ETA predictions than the estimations currently in use, and provides an uncertainty estimation on the prediction. Our experiments show that the uncertainty cannot be reduced by adding more training data of the same type as the one available. Finally, our experiments show the importance of domain knowledge combined with data-driven techniques, to understand the behavior of the network; the lowest uncertainty is
anyway obtained by combining all available parameters.
TNO Identifier
952836
Source title
IPIC 2018
Files
To receive the publication files, please send an e-mail request to TNO Repository.