Jul-2017

Modelling catalytic naphtha reforming

A data handling modelling approach is applied to predict the output performance of a heavy naphtha catalytic reforming unit

REZA SEIF MOHADDECY and SEPEHR SADIGHI Research Institute of Petroleum Industry
ERSHAD AMINI University of Tehran

Article Summary

Catalytic reforming is a chemical process used to convert heavy naphtha with a low octane number into a high octane product called reformate. This process increases the total amount of aromatic hydrocarbons and branched paraffins without changing their boiling point range. The antiknock characteristic of gasoline is an imperative property which leads to installing catalytic reforming.^1-3 However, the kinetic modelling of catalytic reforming has been limited due to the complexity of the process. Moreover, there is a large gap between fundamental studies and practical kinetic model reactions.^4-6 Even if an accurate model is obtained, it is highly complex, and it requires many simplifying assumptions to find a tangible solution.^7-10

On the other hand, developing a black box model, which is exclusively obtained from experimental data, can provide other practical methods in the field of process modelling. These models provide a dynamic relationship between input and output variables and bypass underlying complexity inside the system. Most of these common approaches rely on linear system identification models. The major processes found in chemical engineering are unfortunately non-linear processes, and previously mentioned approaches fail to respond regarding process nonlinearity. As an alternative to fundamental models, artificial neural networks (ANN) are a valuable estimate tool, and up to now numerous applications of ANN models in the engineering area have been reported.¹¹ ANN can perform better than regression models, and is tolerant to noise in data.^12-15 The increased importance of ANNs arises from their ability to parallel process data despite their components being independent of each other.¹⁶ On the other hand, straightforward theories do not offer adequate precision for the estimation of experimental data.

However, an ANN’s structure contains a massive complication of equations within its nodes and layers. Furthermore, the arrangement of networks is chosen manually or randomly which does not assure the best possible network. As a better alternative, the group method of data handling (GMDH) provides a self-organising neural network to express the genome of a system as well as using the most suitable configuration by means of the minimisation process. In other words, the GMDH utilises a feed-forward network whose coefficients are determined using regression together with imitation of self-organising activity.¹⁷ The algorithm chooses the most suitable polynomial expressions built by a combination of two independent variables at a time.

Some artificial neural network models have been developed in the literature to predict and control parameters in industrial catalytic reforming units.^18-20 But, based on our literature review, there is no study on using GMDH to model the heavy naphtha catalytic reforming process. Therefore, the present study is devoted to simulating the yield and RON of the product and the outlet temperature of reactors using GMDH for a commercial scale naphtha reforming unit. To validate the proposed model, 97 data points were gathered from an Iranian catalytic naphtha reforming plant during the complete life cycle of the catalytic bed (about 877 days).

Process description
A commercial fixed-bed catalytic naphtha reforming unit, a Platformer design licensed by Chevron Research Corporation, was chosen as a case study. The feed of the plant prior to entering the catalytic reformer should undergo a hydrodesulphurisation (HDS) reaction in the hydrotreatment unit. Then the produced naphtha, called Platcharge, is introduced to the catalytic reforming process.

The catalyst of the semi-regenerative catalytic reformer is regenerated during routine shutdowns of the process once every 18 to 24 months. Normally, the catalyst can be regenerated three or four times, then it must be returned to the manufacturer for reclamation of valuable platinum and/or rhenium elements.

As Figure 1 shows, Platcharge is first preheated by the first furnace (H-1), then enters the first reactor (R-1) where naphthenes are dehydrogenated to aromatics. Then the product stream from the first reactor passes through the second reactor (R-2), and the outlet stream of that enters the third reactor (R-3). Similarly, the product stream from the third reactor enters the fourth reactor (R-4). The overall reforming reactions are endothermic. Therefore, a preheater (H-1, H-2, H-3 and H-4) should essentially be provided before each reforming reactor.

Next, the product stream from the fourth reactor enters a separator, V-1, wherein the hydrogen produced during the reforming process (the gas stream) is recycled and then mixes with the Platcharge. Finally, the liquid product leaving the separator is introduced to the gasoline stabiliser in which LPG and light gases are separated from the gasoline (reformate). Thus, the vapour pressure of the gasoline can be set according to market requirements. The catalyst distribution in the reactors and the unit’s normal operating conditions are shown in Table 1.

Modelling heavy naphtha reforming
The basic structure of the brain has been widely employed for various fields such as modelling, control and pattern recognition. The GMDH, introduced by Ivakhnenko,21 is a hierarchical and learning network structure that provides an effective approach to identifying higher order non- linear systems. Its main purpose is the identification of relations in large, complex non-linear multidimensional systems as well as their approximation and prediction. In the GMDH network, the part which corresponds to the neuron of a neural network is called the “N-Adaline”, and it is generally expressed by a polynomial. The N-Adaline is composed of two inputs and one output, and the latter is generated by combinations of two inputs.22 Inputs xi and xj are then combined to produce a partial descriptor based on a simple quadratic transfer function:

                                                    (1)  

where yˆn is determined using the least squares method, and coefficients a0 to a5 are determined statistically and are unique for each transfer function. These coefficients can be thought of as analogous to weights found in other types of neural networks.

The GMDH topology is usually determined using a layer by layer pruning process based on a pre- selected criterion of what constitutes the best nodes at each level. The traditional GMDH method is based on an underlying assumption that data can be modelled using an approximation of the Volterra series or Kolmorgorov-Gabor polynomial: