3.3
4.5
Min leaf size
4.4
2 6 10
3.2
4.3
4.2
3.1
Min leaf size
4.1
5 6 9
3.0
4.0
0
20
40
60
80
100
0
20
40
60
80
100
# Bagging iterations
# Bagging iterations
6.0 5.0 4.0 7.0 8.0 9.0
3.5
3.3
3.1
2.9
2.7
3.0
2.5
1.0
1.5
2.5
4.0
10
20
30
40
50
60
2.0
3.0
3.5
# no. of nodes in 1st hidden layer
# Kernel function
No. of nodes in 1st hidden layer
Basis function
35 55
PureQuadratic
5
Constant
Linear
Figure 3 Grid search results
Hybrid digital twin The data-driven AI/ML model and first principles kinetic model described in the previous sections are integrated in a block-oriented manner, resulting in a hybrid model (see Figure 4 ), which is suitable, especially in this case, to strengthen the fundamental model for capturing the underlying deactivation mechanisms from available com - mercial process time-on-stream (TOS) data – to deduce the unknown patterns and insights. Here, the historical daily average data consisting of feed - stock processed along with feed density and reactor operat - ing conditions (H2/HC and inlet pressure) are used as inputs to the data-driven model to capture the WABT trend with time. And subsequently to quantify the catalyst deactivation rate along with the catalyst remaining life prediction. Increasing the WABT pattern with time results in decreased catalyst activity, which is passed to the first principles IOCL proprietary DHDT model to recalibrate kinetic parameters automatically with time. This ensures
model predictions of product properties, including density, T95, CI, and hydrogen consumption, are robust throughout the catalyst life cycle. Also, simulated WABT data can be generated from the kinetic model with future expected feed streams to be processed and passed to the data-driven AI/ ML model for forecasting catalyst life. In this way, the first principles kinetic model represents the balance equations and the inner fundamental relation - ships of the system, while the data-driven part describes the deactivation phenomena, including non-idealities, improving the hybrid model’s application for quick deci - sion-making, such as reducing throughput or cutting down refractory feed streams and accordingly scheduling pro - curement for the next charge of catalyst. The DHDT Digital Twin Dashboard developed using MATLAB software has been implemented at two IOCL refineries, which takes operating data from DCS and lab data from SAP and updates daily to predict product quality and track catalyst life.
Model Performance for prediction of deactivation rate in terms of WABT increase rate
Models
Selected hyperparameters
RMSE
MAE
Training
Testing
Training
Testing
Bagging trees Boosting trees Neural network
ntrees=35; min leaf size=5 ntrees=25; min leaf size=2
1.94 2.61
3.05 4.03
1.0.7 2.01
2.51 3.12
No. of nodes in 1st hidden layer = 50; No. of nodes in 2nd hidden layer = 5 Kernel function = Exponential; Basis function = Constant intercept = TRUE; independent variables considered = Density, Tpt, H2/HC, (Density * tpt), (Tpt * H2/HC), Tpt^2
2.71
4.15
1.82
3.14
Gaussian process
regression
1.02
2.82
0.74
1.52
Linear regression
2.73
4.63
2.06
3.66
Table 1
82
PTQ Q2 2025
www.digitalrefining.com
Powered by FlippingBook