J. Gmeinbauer, C. Ramakrishnan, A. Reichhold, TU Vienna
By means of multivariate statistics a FCC model was developed that allows for the prediction of product distributions, i.e. the yields of gas and gasoline with the help of feed parameters. To provide a practical approach and to reflect the underlying chemical and physical processes as detailed as possible, model calculations were based on data received from a pilot scale FCC plant with internal CFB-design. The methods applied include Multi Linear Regression (MLR), Principal Component Analysis (PCA) and Principal Component Regression (PCR), whereby the FCC unit was considered as a blackbox.
An algorithm was developed to reduce the number of feed features describing the complex hydrocarbon feedstock to a minimum and finally enabled to reflect the feed by only four parameters, i.e. BMCI Calculation, Aniline Point, Mono Aromatics and the Aromatics - Non-Aroma tics - Ratio. Consequently correlations between these latter four features and the relevant product distributions could be calculated.
As far as common (and commercially applied) FCC feedstocks are concerned calculated and measured product distributions correspond very well and enable the plant's operator to predict yields arising from changes in the feed composition or new operating conditions.
No other refining process, except for physical separation by distillation, has had a longer history, or more of an impact on the oil industry than cracking of heavy hydrocarbon molecules to lighter ones. The beginning of the "automobile century" quickly required larger amounts of fuel than could be met by the available straight-run gasoline, which pushed refiners to develop new petroleum upgrading processes.
The need for optimisation of the FCCU performance has been increasing steadily, due to tough competition in the petroleum markets, especially when it comes to the prediction of gas and gasoline yields
From a historical point of view [1-12] research has mainly focused on highly complex computer models, which often do not take into account experimental examinations and therefore don't result in the desired prediction data. Hence, within this work a practical approach was used, i.e. the prediction of product yields and distributions by means of simple and easily determinable feed parameters and the application of statistical methods to data received from a pilot-scale FCCU.
Pilot Plant
The test runs were conducted in a pilot-scale FCC-unit with internal CFB-design as developed by Reichhold and Hofbauer (1996) based on pre-examinations of Krobath [13] and Reichhold [14].
particle separator riser
return flow tube
regeneration zone
siphon
feed inlet zone
Fig. 1 Pilot-scale FCCU scheme
product gas
flue gas
heating system
inert gas air inert gas
tube furnace
Figure 1 gives a schematic description of the FCCU. In contrast to commercial units, where both riser and regenerator are constructed as separate items, the pilot-scale FCC-plant integrates both into a compact unit with the riser concentrically mounted inside the regenerator. This internal CFB-design improves the heat balance of the process considerably, since direct heat is provided by the regenerator to meet the energy requirements necessary for the endothermic cracking reactions in the riser reactor.
zone, riser reactor and return flow tube, and the regeneration zone which only includes the regenerator itself. Since cracking reactions only take place in absence of oxygen, the reaction zone is fluidised by means of nitrogen and the vaporised feed, whereas fluidisation in the regeneration zone is realised by air.
The continuous circulation of bed material between riser and regenerator is ensured with the implementation of a fluidised siphon and bottom section. Both are fluidised with nitrogen as an inert "separation" gas in order to avoid gas leakage from the reactor to the regenerator and vice versa.
The dimensions and basic operating data of the pilot plant are given in Table 1.
Height [mm] 2500 Riser length [mm] 1977 Riser diameter [mm] 25 Regenerator diameter [mm] 180 Catalyst shape selective Zeolite Catalyst mass [kg] 9 Riser temperature [°C] 450-550 Regenerator temperature [°C] 550 - 700 Pressure Ambient
Tab. 1 Dimensions and basic operating data
The subsequent model calculations are based on experimental examinations of three different vacuum gas oils each with five different conversion-levels from the hydrodesulfurisation-unit, provided by OMV AG. Since operation parameters during these runs were kept constant a focus could be put on the influence of feed features on the product distributions.
Model Calculation
- - r
Fig. 2 Lump Classification
To enable the prediction of product distributions the examined FCCU had to be described by a simple model as shown in Fig. 2. The product range was therefore classified into 4 lumps, i.e.
3 lumps according to their boiling ranges (Tab.2) and the coke lump. The FCC process itself was considered as a black box operation.
Boiling Range [°C]
Gas Lump Gasoline Lump LCO - Residue Lump
< 2 0 20 - 200
>200 Tab. 2 Product lumps according to their boiling ranges
Since the available feedstock property data included 23 features, the principal goal was to reduce this number and to find those parameters that bear maximum-problem-relevant data and consequently find an algorithm to predict the desired product information.
The list of features available on the applied feedstocks, inclusive descriptive statistics, is given in table 3.
Aniline point [°C]
Mono - Aromatics [m%]
Di - Aromatics [m%]
Tri-Aromatics [m%]
Poly - Aromatics [m%]
Total Aromatics [m%]
Non - Aromatics [m%]
Total Aromatics : Non - Aromatics - Ratio [-]
Bromine Number [gBr/100g]
BMCI - Calculation [-]
Viscosity 70°C [mm2/s]
Viscosity 100°C[mm2/s]
Pourpoint [°C]
Base Number (by Total N) [ppm]
Total Nitrogen [ppm]
Total Sulfur [ppm]
Carbon [m%]
Hydrogen [m%]
C:H ratio Feed [-]
Ostwald Boiling Point [°C]
Volume Average Boiling Point [°C]
Cubic Average Boiling Point [°C]
9,3 Tab. 3 Applied Feed Parameters
MLR, also called Ordinary Least Squares Regression (OLS), constitutes the standard method for modelling correlations between an X -Matrix and one y - variable. The principal goal is to build a linear model for the prediction of a response y* from the independent features xi, X2,
..., xp.
Initially a MLR method was applied to find correlations between feedstock parameters and product distributions. Since the available feed features are highly correlative the calculation did not result in the desired algorithm, as the feature covariance matrix became singular and could not be inverted, as would be necessary for the analysis, which is always the case if highly correlative features are used.
To solve this problem the feedstock features had to be transformed into principal components, a calculation processed by SPSS for Windows 10.0. Principal Component Analysis [15,16]
constitutes one of the most frequently applied methods for defining latent variables and is a
method for the transformation of the vector x into a set of orthogonal components. PC A scores are uncorrelated by definition, so the C matrix was regular and could be inverted.
Further on a criterion had to be defined to find the features carrying maximum information and to eliminate those which do not contribute to the solution of the problem, hi the course of this paper an iterative algorithm was applied, the criterion being the communalities of the PCA scores. The communality Cj of a feature j is defined as the sum of the squared factor loadings bjk (k = 1 ... m) and therefore denominates the percentage to which a feature contributes to the respective PCA score. Therefore a communality of 1 means that the total amount of information of the respective feature is included in the PCA score.
Consequently a PCR method was applied to find correlations between the PCs and the product lumps, eliminating one feature every step taking into account the respective communalities and eliminating that PC contributing at the lowest to the PCA scores.
In step 1 the "communality - algorithm" was over-ruled by eliminating the following features due to overt correlations between the features or due to chemical and physical reasons:
Feature
Density 15°C [g/cm3] Total Aromatics [m%]
Non - Aromatics [m%]
Viscosity 70°C [mm2/s]
Carbon [m%]
Hydrogen [m%]
Ostwald Boiling Point [°C]
Volume Average Boiling Point [°C]
Cubic Average Boiling Point [°C]
Reason for over-ruling information is included in the BMCI calculation information is included in the aromatics ratio information is included in the aromatics ratio correlates with Viscosity at 100°C information is included in the C:H ratio information is included in the C:H ratio information is included in the BMCI calculation information is included in the BMCI calculation information is included in the BMCI calculation Tab. 4 Over-ruling
Additionally R2, the quotient of SSR ('fitted sum of squares) and SST (total sum of squares), was calculated as a quality criterion for the established model and an F - Test was applied to find the PCs with the highest contribution to the PCR regression model.
Fig. 3 gives R2 of the product lump regression models according to the number of features included in the PCA. A severe decrease in R2 applying less than 6 features can be observed, so the model determined by this PCA and the features included in these PCs (Table 5) were chosen for the prediction model.
Due to operative reasons the coke level remains highly stable which results in a low coke variance. Therefore the prediction model cannot give any relevant information on coke formation in the FCC reactor. Hence, for statistical reasons the coke lump was excluded from the model. Fig. 3 underlines this exclusion since R2 of the coke lump does not exceed 0,6 even with all the 23 features applied to the calculation. Including less than 7 features makes it even impossible to calculate PCs for the coke lump since its variance converges on zero.
Aniline point [°C]
Mono - Aromatics [m%]
Di - Aromatics [m%]
Tot. Arom. : Non - Arom. - Ratio [-]
BMCI - Calculation Pourpoint [°C]
Tab.5 Feed Parameters included in the PCR model
Since those 6 remaining features showed no further correlation, now a MLR method could be applied to the reduced data in order to determine the desired regression model without a preceding principal component analysis.
The applied feed parameters could be reduced to 4 features:
• BMCI - Calculation (BMCI), which includes density and boiling range of the feedstock and therefore reflects the molecular weights and chemical structures of the chemical components
• Aniline Point (AP), which is a measure for the aromatic character of the applied Ä feedstock
• Mono - Aromatics (MA), which determines the amount of benzene - derivatives in the feed
• Total Aromatics : Non-Aromatics Ratio (AR), which is the quotient of aromatic components and non aromatic components
As a conclusion, it has to be pointed out that by means of a statistical method 4 feed properties could be isolated (2 for the gas and 2 for the gasoline lump, resp.), which can also be confirmed from a physical and chemical point of view, since they comprise most important feed data i.e. the boiling range and essential chemical composition data.
Equations (1) and (2) therefore give the correlations to determine the product distributions (gas (G) and gasoline (GL) respectively):
G = 53,97-0,383 BMCI-0,083• AP (1) GL = 39,31 + 0,839 MA-23,744 AR (2)
A graphical description is indicated in Fig.4 and Fig.5
Fig. 5 Gasoline Lump Model
Since the principal goal of the model was to find a practical approach, easy to use by an FCC operator, nomograms as given in Fig. 6 and 7 were constructed, to facilitate a quick determination of BMCI and the product prediction.
o
-Fig. 6 Nomogramm Gas - Lump
e n
Fig. 7 Nomogram Gasoline Lump