PTQ Q3 2024 Issue

Fuel gas consumption

Natural gas consumption

1.0

METHANE_SLIP

-0.11 1 -0.43 0.22 0.3 -0.31 0.0028

0.26

0.17 -0.017

NG_FEED

-0.096

1

-0.11

0.92

0.27

KF

-0.43 1 -0.37 0.055 0.33 0.47

0.0045 -0.37

0.24 0.63

0.8

0.22 -0.37 1 -0.1 -0.17 -0.28

0.35

0.15

0.13 -0.49

NG_LHV

-0.055

1

0.13

-0.086

-0.11

SC

AIR_TO_ FURNACE

0.3 0.055 -0.1 1 0.072 0.47

-0.31

0.35

0.55 0.55

0.6

0.063

-0.21 0.45

-0.31 0.33 -0.17 0.072 1

-0.13 -0.082

SC_FFI

H2_DEMAND

-0.1

0.92

0.13

1

0.1

HP_STEAM_ TEMP

0.4

0.0028 0.47 -0.28 0.47 0.063 1

-0.6

-0.22

0.75 0.69

-0.11 -0.37 0.35 -0.31 -0.082 -0.6

1 0.11

-0.4 -0.74

FG_LHV

NG_MW

0.27

-0.086

0.1

1

0.055

0.2

NG_FEED

0.26 0.0045 0.15 0.35 -0.13 -0.22

0.11

1

-0.11 0.049

REFORMER_ ENTRANCE_TEMP FG_ CONSUMPTION

0.17 0.24 0.13 0.55 -0.21 0.75

-0.4

-0.11

1 0.42

REFORMER_ INLET_TEMP

0.0

-0.096

-0.055

-0.1

0.055

1

0.017 0.63 -0.49 0.55 0.45 0.69

0.049 -0.74

0.42 1

Figure 3 Fuel gas and natural gas consumption heatmap

understanding of variable interactions emerged. The var- iables representing the temperature of the reformer inlet – referred to in the model as ‘REFORMER_INLET_TEMP’ and ‘NG_MW’ – showed a counterintuitive pattern. Despite a low direct correlation with the target variable, includ- ing these variables significantly improved the model’s R-squared (R²) value and lowered the mean absolute per- centage error (MAPE). The development process consistently considered the trade-off between multicollinearity and the enhancement of the model’s predictive accuracy, ensuring that improve - ments in model performance were carefully weighed against the fidelity of the variable relationships. The modelling process encompassed a battery of algo - rithms, including various regression techniques – such as linear regression, ridge regression, lasso regression, and Huber regressor – and tree-based methodologies like LightGBM, XGBoost, and RandomForest. A time-series val- idation approach was employed to validate the efficacy and robustness of these models. This was complemented by a sliding window method, which facilitated the assessment of model predictions over diverse periods. These model val- idation techniques ensured the predictive performance was accurate and consistent over time.

In the model development process, the performance met- rics, specifically R² and MAPE, were meticulously tracked and compiled into trends and tables for comprehensive analysis. A model was considered final and selected for deployment only if it consistently achieved a predefined R² and MAPE values threshold during both holdout and time- series validation processes. This rigorous evaluation criterion ensured the chosen model demonstrated robust predictive capabilities and reli- ability across validation frameworks. The results of these evaluations, detailing the performance of various model iterations, are summarised in Table 1 . The outcomes of this model serve as crucial inputs for the natural gas, fuel gas, and methane slip models. Therefore, the accuracy of this model’s results is paramount for the project’s success. Fuel gas consumption model Fuel gas constitutes a significant component of the energy cost profile for an HGU. Its consumption within the unit varies according to a range of operational parameters. The purpose of developing this model is to understand how fuel gas consumption fluctuates with changes in the S/C ratio and the amount of natural gas utilised. In the progression of our fuel gas consumption model for

Model performance table

NG

FG

Steam

Methane slip

R2

MAPE

R2

MAPE

MAPE

R2

MAPE

Model 1 Model 2 Model 3 Model 4

0.98 0.94 0.93 0.94

0.35 0.76 0.95 0.75

0.79 0.75 0.73 0.67

3.77 3.03 2.02 2.43

0.76 0.84 0.57 0.81

1.24 0.84 1.54 1.06

0.68 0.78 0.71 0.78

1.42 2.56 2.31 0.98

Table 1

66

PTQ Q3 2024

www.digitalrefining.com

Powered by