(61b) Neural Network Models for Predicting Impurity Removal Amount in ARDS Process and Operation Condition Optimization Using Genetic Algorithm

With the development of the fourth industrial revolution, technological growth progressed rapidly, and the development of these technologies served as an accelerant for various industries to enter the digital era. The oil refining industry also introduces digital twin and artificial intelligence (AI) technologies in multiple areas, such as process simulation, optimization, monitoring, and anomaly detection systems [1].

One of the refining processes, the Atmospheric Residue Desulfurization (ARDS) process, is a pretreatment process that removes impurities from the High-sulfur Atmospheric Residue (HS AR) generated by the Crude Distillation Unit (CDU). The ARDS process produces various products such as Liquified Natural Gas (LNG), Naphtha, Diesel, and Treated-Atmospheric Residue (T-AR) through hydrogenation and catalytic reactions [2]. The reactor of the ARDS process is divided into two trains, and raw materials are injected into each train at almost the same rate. In addition, since each train has three reactors, the ARDS process consists of six reactors. HS AR contains impurities such as sulfur, nitrogen, metals (nickel, vanadium), and Conradson Carbon Residue (CCR), causing environmental pollution problems and affecting the purity of the product. [3]. These impurities are removed through a reactor containing a catalyst layer. The metal component and CCR are deposited on the catalyst's surface while removing contaminants. This increases the inactivity of the catalyst and leads to a decrease in the lifespan of the catalyst [4]. For this reason, the catalyst needs to be replaced periodically. Since the catalyst replacement costs considerably, removing impurities as efficiently as possible within a limited life period is necessary. However, it is limited to accurately measuring the information on the catalyst and many variables and parameters required to calculate the amount of impurity removal during actual plant operation. In addition, the ARDS process includes catalytic reactions and hydrocracking to remove impurities. Due to the nature of the refining process, most of the processes occurring are nonlinear, and the reactions and their reaction rate expression are so complicated. The first-principle approach for modeling these complex characteristics is cumbersome and challenging [5].

We propose the data-based modeling approach for the reactor parts using neural networks and several training strategies for model reliability. Each neural network predicts the amount of removal of each five impurities (sulfur, nitrogen, nickel, vanadium, CCR) through the reactors. Considering the aging status, we can determine the smallest temperature conditions where the impurities are removed sufficiently. Finally, we suggest the optimal operation temperatures of the reactors by using the neural networks with a genetic algorithm (GA).

We use two fully connected neural networks (FCNN) to predict each impurity's removal amount from each train. The features for FCNNs are selected based on domain knowledge. The reactor temperatures, where the related catalyst is located, and the impurity feed flow rate are selected for the FCNNs features for each impurity. For example, for the FCNNs predicting the S removal amount, the S feed flow rate and the reactor temperatures at the HDS catalysts are used as the features. The outputs from FCNNs are multiplied by aging factor functions, then summed up to represent the total S removal amount from two trains. Because only the composition at the inlet and oultet of the two trains, the two FCNNs and the parameters in two aging factors should be trained at once. Because the removal amount from each train, the flow rates to each train, and the reactors' temperatures are operated similarly, we need to penalize the imbalance for the FCNNs representing each train. To this end, we add the penalizing term in the cost function for NN training along with the basic mean squared error terms. In addition, before considering the aging factor, the FCNNs outputs need to have a positive correlation with the features. In other words, before considering aging effects, the removal amount of the impurities should increase as the temperatures rise or the feed impurity flow rate increases. Thus, we clip the parameters in FCNNs to the positive values after a one-step update with the Adaptive Moment Estimation (ADAM) algorithm. With this strategy, we can guarantee a positive correlation between the features and the FCNNs outputs. Finally, every day, we train the model with the latest 28-day data to follow up on the catalyst aging effects. As a result, the proposed impurity removal amount prediction model shows that the error rate is less than 10% on average for all impurity components, and the performance of predicting the actual plant is excellent.

The operating temperatures for the reactors are determined using the NNs and genetic algorithm. The decision variables are the reactor temperatures. To calculate the objective function for the minimization formulation, the NNs with the explored decision variables predict the impurity removal amount. If the impurity amount remaining in the outlet from the reactors is smaller than the required specification, then the objective function is the sum of the reactor temperatures. If not, the objective function is set as a very high value, such as 1e10. This is because the required purity is not satisfied with the explored reaction temperatures, so the greater temperatures should be explored. The bounds for the reactor temperatures are set in the range of 350-400℃, the actual operating range. Through this optimization, we can determine the optimal operation conditions daily while updating the NNs with the latest 28 days' data.

References

[1] Wanasinghe, Thumeera R., et al. "Digital twin for the oil and gas industry: Overview, research trends, opportunities, and challenges." IEEE access 8 (2020): 104175-104197.

[2] Rana, Mohan S., et al. "A review of recent advances on process technologies for upgrading of heavy oils and residual." Fuel 86(9) (2007): 1216-1231.

[3] Nguyen, Thanh-Huong, et al. "Hydrodemetallization of heavy oil: Recent progress, challenge, and future prospects." Journal of Petroleum Science and Engineering (2022): 110762.

[4] Ali, Mohammad Farhat, and Saeed Abbas. "A review of methods for the demetallization of residual fuel oils." Fuel Processing Technology 87(7) (2006): 573-584.

[5] Gueddar, Taoufiq, and Vivek Dua. "Novel model reduction techniques for refinery-wide energy optimisation." Applied energy 89(1) (2012): 117-126.

Acknowledgements

This work was supported by the National Research Foundation of Korea(NRF) grant funded by the Korea government(MSIT) (NRF-2021R1C1C1004217), and by the Korea Agency for Infrastructure Technology Advancement(KAIA) grant funded by the Ministry of Land, Infrastructure and Transport (No. 21ATOG-C162087-01).