Apr-2014

Predicting severity, viscosity and yields 
in fuel oil visbreaking

Neuro-fuzzy technique is used to predict product properties in a 
commercial visbreaker

SEPEHR SADIGHI and REZA SEIF MOHADDECY
Research Institute of Petroleum Industry

Article Summary

Crude oils contain a large fraction of heavy products for which only few useful outlets exist. Indeed, world demand for light and middle distillate continually increases, while at the same time the available crude oil becomes heavier.¹ Visbreaking is a non-catalytic thermal process that converts atmospheric or vacuum residues via thermal cracking to gas, naphtha, distillates, and visbroken residue. A visbreaker thermally cracks large hydrocarbon molecules in oil by heating them in a furnace to reduce their viscosity and to produce small quantities of light hydrocarbons (LPG and gasoline). The process name of ‘visbreaker’ refers to the fact that the process reduces the viscosity of the residual oil._2,3
Two types of visbreaking technology are commercially available: the ‘coil’ or ‘furnace’ type and the ‘soaker’ process. In the coil process, conversion is achieved by high temperature cracking for a predetermined, relatively short, period of time in the heater. In the soaker process, a low temperature/high residence time process, the majority of conversion occurs in a reaction vessel or soaker drum, where the two-phase heated effluent is held at a lower temperature for a longer period of time. Therefore its heater duty and, in turn, its fuel consumption is only 70% of that for the coil-visbreaking process.⁴

Worldwide, about 200 visbreaking units are in operation; Europe alone accounts for about 55% of total visbreaking capacity.⁴ To have an effective design and perfect control of any process, a model is needed to predict product yields and qualities versus variables such as space velocity and temperature. However, the complexity of visbreaking feed and product makes it extremely difficult to characterise and describe its kinetics at a molecular level. Modern day rigorous simulators such as Aspen Plus or Hysys from Aspen Technology do not have such restrictive limits on the total number of components; therefore, it is possible to use a unique set of pseudo components for every petroleum assay stream.⁵ This approach increases the calculation time, and characterisation of the streams and subsequent reports become unnecessarily complicated.⁵
One approach to simplify the problem is to consider the partition of the species into a few equivalent classes, the so-called lumps or lumping technique, and then assume each class as an independent entity. Developing simple kinetic models (for instance, power-law model) for complex catalytic reactions is a common approach, and it can give basic information for reactor design and optimisation. In this field, many investigations were reported in which the visbreaking process was modelled with discrete two-lump,^6-8 three-lump,⁹ four-lump,^10,11 five-lump1^2,13 and seven-lump_¹⁴ approaches. In all of these investigation, the experiments were carried out in a micro or pilot scale reactor. However, over the last two decades, soft computing methods such as artificial neural networks and fuzzy logic were migrated into the realm of refinery processes for modelling, control, optimisation and kinetic estimation. The main advantage of this approach, named black box modelling, is its independence of any assumption or equation about kinetic study. Moreover, it is not needed to ignore the effect of some parameters to obtain a simpler mathematical expression.

The aim of this research is to use the neuro-fuzzy methodology for predicting the severity, product yields and viscosity of fuel oil for an industrial scale visbreaking unit equipped with a soaker drum. The performance of the neuro-fuzzy model with different membership functions is compared with industrial data obtained from an Iranian refinery.

Feed characterisation
An industrial soaker-visbreaker unit was chosen as a case study. This unit was designed to visbreak 20 000 b/d of a mixture of vacuum residuum and slop vacuum gas oil, both taken from a vacuum tower. The composition of the fresh feed can vary slightly with time from start of run (SOR) to end of run (EOR). The specification of the combined feed, which was obtained during this research, is shown in Table 1.

Process description
The visbreaking feed is charged to the coil furnace at a temperature of about 340°C. The visbreaking furnace is constructed in two sections, fired independently. After the coil furnace, the two hot streams are drained into a transfer line and the mixed product enters the soaker drum. The specifications of cells and the soaker drum are presented in Table 2. The output product from the soaker drum is quenched by the cold recycle stream to stop cracking reactions and inhibit coke formation. Finally, the combined stream is transferred to the fractionation tower and side strippers to separate the visbreaking products. A simplified process flow diagram of the described unit is shown in Figure 1.

During one year of data gathering, 46 sets of data, including product flow rates, feed inlet temperature and soaker outlet temperature, were gathered from the target visbreaking process. 
As Figure 2 shows, light gases including C₁, C₂ and LPG, gasoline and tar are the output streams from the visbreaking plant. Performing mass balance around the unit proved that the error for all experiments was less than 5%, mainly related to the gross error for the measuring of the gaseous products and maybe related to the coke formation. The boiling range of VGO feed, fuel and gas oil samples were analysed according to the ASTM D1160 standard procedure whilst the one for the gasoline sample was analysed according to the ASTM D86 method. Additionally, the viscosity of the visbreaking product (fuel oil) was measured according to the ASTM D2170 method.

Neuro-fuzzy modelling approach
The connection of fuzzy systems with an artificial neural network (ANN) is called a neuro-fuzzy (NF) system. Similar to neural networks where knowledge is saved in connection weights, it is interpreted as fuzzy if-then rules in NF systems. The most frequently used neural network in NF systems is radial basis function neural network (RBFNN) in which each node has radial basis function such as Gaussian and ellipsoidal.

The popularity of NF systems is due to the simplicity of structure, well-established theoretical basis and faster learning than in other types of neural networks. There are many developed fuzzy neural networks (FNN) as NF algorithms in the literature, but the adaptive network based fuzzy inference system (ANFIS) is one of the best known. In ANFIS, Takagi-Sugeno type fuzzy inference system is used. The output of each rule can be a linear combination of input variables plus a constant term or can be only a constant term. The final output is the weighted average of each rule’s output. In the forward pass, the ANFIS uses the least squares method to identify the parameters in the network layers. In the backward pass, the errors are propagated backward and related parameters are updated using the gradient descent method. The ANFIS can be also trained by a hybrid algorithm to identify the membership function parameters.¹⁵

To create a neuro-fuzzy inference system for the understudy visbreaking plant, Matlab Fuzzy Logic Toolbox version 2010a from MathWorks and ANFIS syntax were used. This syntax is the major training routine for Sugeno-type fuzzy inference systems. ANFIS uses a hybrid learning algorithm to identify the parameters of Sugeno-type fuzzy inference systems. It applies a combination of the least-squares method and the back-propagation gradient descent method for training fuzzy inference system membership function parameters to emulate a given training data set. The type of membership functions associated with a visbreaking unit was selected from all supported types in Matlab. To train the fuzzy model in an optimised time, two fuzzy rules were selected from the ANFIS toolbox, and the training process was stopped whenever the designated epoch number (100) was reached.