Including crude type
Excluding crude type
30
6 7 8
25
20
5
15
3 1 2 4
10
5
0
0
Feature
Feature
Figure 2a (left) Feature importance graph including crude type; 2b (right) Feature importance graph excluding crude type
model was screened statistically based on its importance in regression learners using an F -test. A small p -value (<0.01) or large value in terms of -log( p ) indicates that the corre- sponding feature is important. The feature ranking was done based on score value, which is shown in Figure 2a and five input variables – crude type, throughput, feed density, reactor inlet pressure, and H 2 /HC, were screened as key features. Even though the score value of ‘crude type’ is high, it is excluded during model devel- opment (see Figure 2b ), as only limited data on a monthly average basis is available. Instead, its effect is incorporated as the correction factor using WABT pred obtained from the ML model:
During training, the network is learned by creating con- nections and adjusting the weights associated with the con- nections between neurons to improve model performance. There are different iterative techniques to refine weights until a desired accuracy level is achieved. Gaussian process regression (GPR) is based on the concept of a Gaussian pro- cess. It uses training data, similar to the k-Nearest Neighbors algorithm, incorporating prior knowledge through kernels. LR is a technique that provides a linear relationship between an independent variable and a dependent varia- ble. The model is trained for the line or plane of optimal fit (regression line with slope and intercept) based on the ‘least squares regression’ method. Also, model training in ML algorithms involves a selection of optimal hyperparameters. Here, we have used a grid search technique, where users specify the hyperparameters to be tuned within a possible range. This is used to control the complexity and learning rate. For example, in case of bagging and boosting trees, users can provide a number of trees, including the depth of the trees or the minimum leaf size, as hyperparameters to be optimised while specifying the range for each hyper- parameter. Similarly, the number of layers and nodes, the learning rate, and the batch size can be chosen for NN. The GPR model has hyperparameters in terms of basis function (zero, constant, and linear) and kernel function (noniso- tropic rational quadratic and isotropic rational quadratic). Figure 3 shows the plots of the results of a grid search for the hyperparameter tuning of bagging and boosting trees, NN, and GPR algorithms, respectively. The optimal hyper- parameter is the one that produces a lower RMSE value. As shown in Figure 3, the best combination of hyperpa- rameters yields the lowest RMSE values, which are shown in Table 1 . The results showed that the GPR algorithm had the best performance for the prediction of deactivation pattern in terms of WABT and is an effective model for datasets used in this study, with RMSE and mean absolute error (MAE) being 2.82ºC and 1.52ºC, respectively, for test sets. The model sup- ported with AI techniques gets trained periodically to learn and update the effect of new feed streams, thereby displaying daily the remaining life of catalyst on the dashboard.
WABT corr = WABT pred + Δ WABTX (% crude mix)
Where Δ WABT is correction factor. The distillation of feed can be explained inherently in the data through feed density and crude type. Thus, four key features – throughput, feed density, reactor inlet pressure and H 2 /HC, are taken as independent variables for model development. ML algorithms In this study, five ML algorithms, including bagging trees, boosting trees, neural network (NN), Gaussian process regression (GPR), and linear regression (LR), have been evaluated based on Root Mean Squared Error (RMSE) per- formance metrics. Bagging and boosting are both ensem- ble modelling techniques that train multiple tree models. In bagging trees, models are trained independently in parallel, and predictions from all sampled models are weight aver- aged to calculate the overall prediction. In boosting trees, models are trained sequentially, with each model learning from the errors of the previous one. Weightage is then assigned based on the prediction accu- racy of each sample model to calculate the overall predic- tion. Bagging reduces variance while boosting reduces bias. Neural Network algorithms inspired from the networking of human brain neurons, is made up of layers of intercon- nected nodes called neurons.
81
PTQ Q2 2025
www.digitalrefining.com
Powered by FlippingBook