PTQ Q2 2025 Issue

based on the start-of-run (SOR) kinetic parameters. Hence it is required to update with catalyst deactivation profiles to predict the unit’s performance throughout the catalyst lifecycle. Deactivation mechanisms Catalyst deactivation is a complex phenomenon with differ- ent mechanisms of coke and nitrogen compound deposition, which is challenging to model kinetically. Several mathemat- ical models have been described in the literature explaining the single or multiple causes. These include active site/pore mouth blockage, diffusional resistance, and sintering of the active phase, and the rate of which depends on feed compo- sition (particularly basic nitrogen compounds and amount of coke precursors), catalyst characteristics (pore size, volume, and surface area), and reaction conditions (WABT, pressure, hydrogen-to-hydrocarbon ratio [H 2 /HC], and LHSV). The nature of the deactivation profile, evidenced by an increase in WABT over time, is well established as an S-shape curve in refinery units. In the DHDT process, the initial high rate of deactivation is due to rapid coke forma- tion. It then reaches a steady-state level before showing steep deactivation at the end of the cycle. The deactivation curve may also change in between, depending on intermittent refractory feed and severity of operation. Developing a robust model requires huge exper- iments at different reaction conditions with a wide range of feed and catalyst. Moreover, it is difficult to mimic an indus - trial catalyst deactivation profile at the lab scale through accelerated ageing experiments. In this study, ML algorithm applications have been lever- aged to identify WABT patterns associated with feed stream properties and process operations based on the unit’s his- torical performance data. Further, the model is automated through AI techniques to forecast the catalyst deactivation rate with time, enhancing the utility of the kinetic model throughout the catalyst lifecycle for optimisation. Data-driven AI/ML model Unlike conventional empirical regression models where linear WABT trends over different periods are weight aver- aged and linearly extrapolated to estimate deactivation rates, the ML model can accurately predict and forecast the non-linearity in the WABT pattern due to changes in feed quality, amount processed, and operating severity. Using an ML model approach to learn the behaviour of the deacti- vation process can increase the efficacy of predictability for remaining life assessment studies. The first step in building a model is data collection and pre-processing. The daily average operating data and corresponding lab analysis of one day are considered as one data set. The historical record of 649 data set points (approximately 1.5 years of data) is considered for model development, in which 70% of data is used for training the model and the rest for model testing. The data pre-processing includes the removal of outliers and estimating the missed values using either forward, backward or interpolation methods. The number of inde- pendent variables (features) to be provided as input to the

Olens saturation:

R CH

CH + H

R CH CH

HDS: HDN:

R SH + H R NH + H

R H + HS

R H +NH

k mono

k poly

k di

HDA:

Poly aromatic saturation

Mono aromatic saturation

Di aromatic saturation

‘R’ represents a hydrocarbon

feed ranging from thiophene to 4,6-dimethyldibenzothio- phene molecules. Suitable sulphur and nitrogen lumps have been considered in the model to capture accurate chemistry of desulphurisation and denitrogenation reactions, includ- ing inhibitions due to the presence of H 2 S and nitrogen compounds in the system. The aromatic mono, di, and poly compounds character- ised through HPLC ASTM D 6591 have been used for their saturation chemistry and accordingly incorporated into the model along with reversibility reactions considering both kinetic and thermodynamic regimes. As the olefin satura - tion reaction is very fast at DHDT process conditions, the single reaction for olefins saturation is considered in the model. Also, cracking reactions have been considered in the model to capture naphtha and gas formation from diesel. Considering the reactor as plug flow, the whole bed length is uniformly divided into small zones over which ordinary differential equations for mass and energy balance are applied for each reacting lump: Figure 1 Main chemical reactions associated with the hydrotreating process

_ mass balance for species i

_ Energy balance equation

These ordinary differential equations are solved simul- taneously to estimate the sulphur, aromatic, nitrogen, hydrogen sulphide (H 2 S), ammonia (NH 3) , H 2 consump- tion, diesel yield, and temperature across the length of each bed. The intrinsic kinetics are generated from pilot- scale experiments, and diesel product properties (density, CI, and distillation curve) are estimated as a function of feed properties and saturation level of different aromatics lumps. The process input variables considered in the model are the total pressure, hydrogen purity, recycle gas rate, reactor inlet temperature, and space velocity (liquid hourly space velocity [LHSV]). The LHSV and weighted average bed temperature (WABT) are the key hydrotreating unit parameters for unit optimisation and control of the catalyst lifecycle. However, the process kinetic model configured

PTQ Q2 2025