(16e) State Observer Design and Model Predictive Control of an Industrial-Scale Biochemical Fermenter

Knowledge of process states is crucial to implement control loops, perform troubleshooting, monitor critical tasks, and minimize the process cost. In many cases, not all measurements are available at the desired frequency or are measured online due to the unavailability of sensors, time delays, the high cost of sensor devices, or the environmental conditions where the measurement is taken [1]. For bioreactors, precise knowledge of the state variables is vital due to the high sensitivity of microorganisms to slight changes in the culture medium. In such reactor systems, many state variables like substrate and biomass are challenging to measure; hence, development of soft sensors is needed [2]. Motivated by this, we designed a state observer that accurately estimated the process states, including the substrate and the product concentrations. The observer was developed for both the batch and fed-batch operations [3].

For designing an observer, we need a mathematical model that can predict states' behaviors for both batch and fed-batch fermentation [4]. A hybrid model was developed for modeling the bioreactor [5], which combined the first principles model and neural networks technique using the available experimental data. This model was able to capture the uncertainties of the process and provide a reasonable prediction for the states. However, due to batch-to-batch variability and infrequent measurements being collected at different rates either online or offline, traditional Kalman filters [6] were unable to estimate well, thus requiring a multi-rate observer [7]. Hence, we designed a multi-rate observer where the process states were re-initialized with a new set of measurement values when they became available. The hybrid model estimates the states in the time-intervals between these measurements enabling the state observer to track even those internal states like biomass, substrate, and product which are either un-measured or are measured in irregular time-intervals. This method is computationally inexpensive, can account for batch-to-batch variation and for time-varying intervals of process measurements, and can give a good prediction for both the batch and fed-batch fermentation processes. This method is also helpful in tracking the concentration of the undesired product in the fermentation process. We then validated the developed observer against a new set of process data, and the results showed that the observer was able to track both the measured and un-measured process states reasonably well.

Once the observer was developed, optimal conditions were determined for the fermenter to maximize the product and profitability while maintaining the temperature and flow rate constraints and ensuring the safety of the operation. A model predictive control (MPC) algorithm was developed to maintain the process at optimal operating conditions via manipulation of temperature and catalyst concentration. The algorithm was able to effectively control and achieve the target states, considering all the practical constraints. As a case study, we implemented the observer model and optimal control algorithm on an industry-scale fermentation plant and demonstrated the proposed framework's performance with real-time monitoring.

References:

1. De Assis, Adilson José, and Rubens Maciel Filho. "Soft sensors development for on-line bioreactor state estimation." Computers & Chemical Engineering 24.2-7 (2000): 1099-1103.
2. Villeros, Pablo De, Héctor Botero, and Hernán Alvarez. "State observer design for biomass and ethanol estimation in bioreactors using cybernetic models." Dyna 83.198 (2016): 119-127.
3. Pan, Xinghua, et al. "Estimation of unmeasured states in a bioreactor under unknown disturbances." Industrial & Engineering Chemistry Research 58.6 (2019): 2235-2245.
4. Raftery, Jonathan Patrick. Continuous Biochemical Processing: Investigating Novel Strategies to Produce Sustainable Fuels and Pharmaceuticals. Diss. 2017.
5. Shah, P. J., Sheriff, M. Z., Kwon, J. S.-I., & Kravaris, C. (2020, November 16–20). Developing a Hybrid Model of a Biochemical Fermentation Process [Oral Presentation]. 2020 Virtual AIChE Annual Meeting.
6. Grewal, Mohinder S., and Angus P. Andrews. Kalman filtering: Theory and Practice with MATLAB. John Wiley & Sons, 2014.
7. Ling, Chen, and Costas Kravaris. "Multirate sampled-data observer design based on a continuous-time design." IEEE Transactions on Automatic Control 64.12 (2019): 5265-5272.