Heterogeneous transfer learning for highly non-linear regression tasks with application to the hydrotreatment of tire pyrolysis feedstocks
Résumé
Hydrotreatment is a crucial step in removing impurities, such as nitrogen, from feedstocks in order to improve the hydrocracking reaction and avoid early catalyst deactivation. The objective of this work is to predict the nitrogen concentration after the hydrotreatment step under a scarce data regime. In particular, for renewable feedstocks, the available data is limited, whereas rich data sets are often available in the fossil domain, where they can be leveraged to improve the prediction task. This motivates the use of transfer learning, especially heterogeneous transfer learning, since the feature spaces of the two domains differ. Three new heterogeneous transfer learning methods for regression tasks have been developed, achieving substantially lower prediction errors than classical methods both with and without transfer learning on simulated and real data sets.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |