(28a) A Hybrid Modeling Approach for Catalyst Monitoring and Lifetime Prediction

In this work, we present a hybrid modeling approach combining a fundamental reactor model with a data-driven empirical model to predict the lifetime of a commercial heterogeneously catalyzed reactor system. The fundamental model connects the inlet conditions, temperature profile, and conversion of the reactor with the catalyst activity – the main KPI of the process. The empirical model, in turn, relates the catalyst activity with the process variables and catalyst age, allowing for prediction of future performance. The state of the model is updated automatically with daily new plant data using filter algorithms. The model provides a tool for catalyst health monitoring, process disruption monitoring, catalyst lifetime prediction, and turn-around planning.

Introduction

Catalyst deactivation is a major cause for loss of production and unplanned turnarounds in chemical plants. The capability to predict the performance metrics of the reactor, such as production rate, conversion, product yield, etc., at different conditions in the future using catalyst activity as a proxy is highly desired. However, the inability to measure the catalyst activity on a commercial plant scale, the numerous and complex fundamental mechanisms for catalyst deactivation, and the years-long time scale of relevant deactivation modes (compared to days-long typical lab experiments) render it difficult to construct a first principle model for catalyst deactivation that is relevant for commercial chemical plants.

Hybrid models combine engineering-based fundamental model with data-driven empirical model to improve the model’s ability to explain real data while preserve its ability to make meaningful and explainable predictions for hypothetical, unobserved inputs. For an in-series hybrid model, the empirical portion calculates internal parameters for the fundamental model which, in turn, solves engineering equations to compute the outputs of the system. Kinetic parameters [1] or reaction rates [2] have been calculated from neural networks before being used in mass balance equations to solve for the conversion and yields of chemical reacting systems. In this work, we demonstrated the methodologies to construct, validate, and deploy an in-series hybrid model to estimate historical catalyst activity and predict catalyst lifetime and future performance of commercial reactors [3].

Methods

The first half of the constructed in-series hybrid model was a reactor model that contained the fundamental mass and energy balance equations, thermodynamics, and kinetics of the reacting system and used the historical inlet conditions and indicators of the catalyst activity, such as conversion, yield, temperature profile, etc., to calculate the historical catalyst activity. The partial-least-squares model then found an empirical correlation between the historical catalyst activity and the process variables in the chemical plant and allowed prediction of the catalyst activity in hypothetical reaction conditions. In tandem, the two fundamental-empirical models enabled the projection of the reactor performance in the future. The model was calibrated with process data from a commercial plant with 24 catalyst charges and 18 years of continuous operation, resulting in 370 variables and over 18,000 data points.

Results and Discussions

The calculated historical catalyst activity showed the expected decreasing trend with operation time and cumulative reactant feed. The empirical model successfully captured the catalyst activity as a function of seven process variables and achieved a goodness-of-fit of 87% and goodness-of-prediction of 61%, demonstrating its ability to predict the catalyst activity for unseen data. A 100-day look-ahead validation scheme was employed in which the empirical model predicts the catalyst activity 100 days from the date-of-interest and the fundamental model predicts the reactor conversion. Quantitative agreement was observed between the historical data and the prediction from this validation scheme, further highlights the ability of the hybrid model to predict key performance metrics of the commercial reactor. A projection tool from this model allowed the user to monitor the historical catalyst activity, predict the future activity, and estimate the timing of the next required turn-around and the remaining production.

Conclusion

This work demonstrated a framework for constructing, validating, and deploying hybrid fundamental-empirical models to monitor and predict the catalyst activity of commercial reactors, from there improved the reactor’s performance, optimized the turn-around planning of the chemical plant, and increased overall production and efficiency.

References

1. Azarpoura et al. Chemical Engineering Research and Design, vol. 117, pp. 149-167, 2017.
2. Saraceno, S. Curcio, V. Calabrò and G. Iorio, Computers and Chemical Engineering, vol. 34, pp. 1590-1596, 2010.
3. Bui et al. ACS Eng. Au, vol. 1, pp. 17-26, 2022