(89b) How to Develop Quickly New Kinetic Models By Including a Priori Information (Part II): Transfer Learning from Fossil Feedstocks to Waste Plastic Pyrolysis Oil

How to develop quickly new kinetic models by including a priori information (Part II): Transfer Learning from Fossil Feedstocks to Waste Plastic Pyrolysis Oil

Warumporn PEJPICHESTAKUL, Per Julian BECKER, Benoit CELSE

^aIFP Energies nouvelles, Rond-point de l'échangeur de Solaize, BP 3, 69360 Solaize, France

warumporn.pejpichestakul@ifpen.fr

Abstract

The need to address plastic waste management and reduce reliance on fossil feedstocks in the petrochemical industry has sparked sustainable solutions. One such solution is the upgrade of waste plastic pyrolysis oil (py-oil) through hydroprocessing, which can utilize the existing facilities in the refinery and provide a means to transform the low-quality and impure py-oil into more valuable products that can be used as chemical feedstocks for the steam cracker to produce new plastics. However, this process is challenging due to contaminants, varying polymer compositions in plastic waste, and pyrolysis conditions [1].

Compared to extensive datasets available for conventional fossil feedstocks, experimental studies on hydroprocessing py-oil are limited. Generating experimental data is extremely time-consuming and expensive. It would therefore be beneficial to transfer knowledge from fossil hydroprocessing into models for the new feedstock(s). One promising approach to bridge the domain gap between these feedstocks is the transfer learning, a data-driven technique that leverages existing data and predictive models from fossil feedstocks to fine-tune models for py-oil. Transfer learning has shown success in enhancing kinetic and product property models for hydroprocessing fossil fuels [2–4].

This work focuses on developing product property and kinetic models for the hydroprocessing of waste plastic py-oil by applying transfer learning based on the Monte-Carlo Markov-Chains (MCMC) approach. The product property models are generally linear correlations, while the kinetic model is a more complex continuous lumping model [5]. We compare the model obtained from transfer learning (fossil to py-oil) with those developed solely from py-oil data, both with the same set of features. Several criteria are used to evaluate the models, which are accuracy and robustness. The exercise of including a new feature from the new domain (py-oil feed) is carried out on the product property model to evaluate the adaptability of the current transfer learning method, which is developed for to same feature space (homogeneous transfer learning), to the different feature space (heterogeneous transfer learning). Both sets of models perform well on the relatively small calibration dataset, but transfer learning leads to more robust results. We particularly emphasize the models' ability to predict outcomes with different py-oil feedstocks (not seen during calibration), thus ensuring their applicability across a range of scenarios.

KEYWORDS

Monte-Carlo Markov Chains, Bayesian Statistics, Pyrolysis Oil; Waste Plastics; Hydroprocessing

References

[1] M. Kusenberg, A. Eschenbacher, M.R. Djokic, A. Zayoud, K. Ragaert, S. de Meester, K.M. van Geem, Opportunities and challenges for the application of post-consumer plastic waste pyrolysis oils as steam cracker feedstocks: To decontaminate or not to decontaminate?, Waste management (New York, N.Y.) 138 (2022) 83–115.

[2] P.J. Becker, L. Iapteff, B. Celse, in: A.C. Kokossis, M.C. Georgiadis, E. Pistikopoulos (Eds.), 33rd European Symposium on Computer Aided Process Engineering, Elsevier, 2023, pp. 1053–1058.

[3] L. Iapteff, J. Jacques, M. Rolland, B. Celse, Reducing the Number of Experiments Required for Modelling the Hydrocracking Process with Kriging Through Bayesian Transfer Learning, J. R. Stat. Soc. Ser. C. Appl. Stat. 70 (2021) 1344–1364.

[4] L. Iapteff, J. Jacques, Benoît Celse, V. Costa (Eds.), Bayesian Transfer Learning to Improve Predictive Performance of an ODE-Based Kinetic Model, 2022.

[5] P.J. Becker, B. Celse, D. Guillaume, V. Costa, L. Bertier, E. Guillon, G. Pirngruber, A continuous lumping model for hydrocracking on a zeolite catalysts: model development and parameter identification, Fuel 164 (2016) 73–82.