Application of artificial neural networks (ANN) for modeling of industrial hydrogen plant
Highlights
► The feed forward back propagation ANN is applied for modeling of hydrogen plant. ► The hidden neurons number is optimized by minimization of MSE. ► The ANN can successfully model a highly nonlinear process, e.g. hydrogen plant. ► The ANN can cover whole hydrogen plant with no need to detail modeling.
Introduction
Hydrogen is the one of the key materials in petroleum and petrochemical industries. Recently hydrogen is more important due to new fuel cell application. Hydrogen is produced industrially by several methods such as steam reforming, partial oxidation, electrolysis, gasification and ammonia dissociation. Also hydrogen is obtained as a by-product of some refining and chemical production processes. Catalytic steam reforming of methane is a well-known, commercially available and the most economic process for hydrogen production [1], [2].
The process scheme comprises reforming, water–gas shift, CO2 removal by amine solution and methanation units. Fig. 1 shows the block flow diagram of aforementioned units [3]. At the steam reforming unit synthesis gas, i.e. mixture hydrogen and carbon monoxide, is produced. The water–gas shift unit is designed to convert the bulk of carbon monoxide to carbon dioxide besides hydrogen generating. The water–gas shift reaction is run in an adiabatic fixed bed reactor that consists of two high and low temperature stages. In amine unit, to improve the product purity and elimination of some undesired property of CO2, such as corrosiveness and lack of heating value, CO2 is removed using amine solution. As carbon monoxide acts as catalyst poisons in the most applications especially in the fuel cell, its concentration of must be less than 5 ppm using methanation unit by following catalytic reaction:
In reformer unit, firstly the feedstock is desulphurized due to the catalysts used in the steam methane reformer and the shift reactor are extremely vulnerable to sulphur poisoning. The reformer provides the principle step of the process with following reaction system:
The reactions take place in the reformer tubes filled with a (usually nickel-based) catalyst. As the first two reactions dominate, the overall reaction is highly endothermic and high temperatures in the range of 750–1000 °C are favored.
Since 1968, various models have been developed for modeling of reformer reactors:
- 1.
One dimensional homogeneous model or thermodynamic simulation [4], [5], [6], [7]. Though this model has agreed with some reformer plant data, it cannot account the comprehensive phenomena in the reformer reactor due to consideration of homogeneous phases and no mass transfer limitations.
- 2.
One dimensional (1-D) heterogeneous model [8], [9]. In this model, mass transfer limitation is considered. This model is valid for reactor performance prediction in one dimension (axial direction).
- 3.
Two dimensional (2-D) heterogeneous model [10], [11]. This model has been developed for reactor performance prediction in both axial and radial directions. Mass transfer limitations are also considered.
In shift converter unit, the resulting synthesis gas is sent to one or more shift reactors, where the water–gas shift reaction takes place:
The favored reaction temperature is less than 600 °C and can take place as low as 200 °C with sufficiently active catalysts. The gas exiting the shift reactor contains mostly H2 (70–80%) plus CO2, CH4, H2O and small quantities of CO. To increase hydrogen content, the shift reaction is often accomplished in two stages. A high temperature shift reactor operating at 350–475 °C followed by a lower one at 200–250 °C.
Various models have been developed for modeling of shift converter. Barrio et al. [12], besides development of an algorithm for the simulation of an adiabatic fixed bed reactor, studied the kinetics of the water–gas shift reaction. Elnashaie et al. [13] provided mathematical modeling and computer simulation of industrial water–gas shift converters and checked the model performance against four industrial reactors. Ding et al. [14] used 2-D unsteady model with kinetic rates obtained from the experiments. Barrio et al. [15] modeled and simulated catalytic partial oxidation and steam reforming of methane reactors to compare the temperature profiles and hot spot minimization.
In amine unit, aqueous solution of alkanolamine reacts reversibly with acid gases. Well-known technology and negligible hydrocarbon loss are the major advantages of the amine treatment process. However, the operating and capital costs shoot up very rapidly as the concentration of acid gas in the feed gas increases [16]. Recently, mixed amine solutions have received increased attention. Traditional process simulator such as HYSYS and Aspen plus can simulate successfully the amine process.
There are many adjustable parameters in the hydrogen plant (e.g. T, P, x, F, …), so it has complication relations among parameters that may be unknowns. Therefore, it would be interesting and useful to utilize the model to simulate the whole plant in the wide range of operating conditions. It must be noted that almost all papers in literature have focused on individual units and there are few papers on simulation of the whole hydrogen production process.
The neural network methodology is widely used in the chemical processes such as modeling [17], fault detection and diagnosis [18], control [19], [20] and so on. It enables user to model and design nonlinear systems based solely on instances of input–output relationships. Moreover, the neural network is adapted easily to new environments by learning, and can deal with information that is noisy, inconsistent, vague, or probabilistic.
This work focuses on modeling the whole hydrogen production plant using an artificial neural network (ANN). Using a gradient descent algorithm, a three-layer ANN is used to identify a nonlinear model from input–output data. The results of designed network show that ANN is applicable for complicated nonlinear systems (e.g. hydrogen production plant) and it implements robust predicted model. The result can be used to gain better knowledge and to optimize hydrogen production plants.
Section snippets
Artificial neural network
Artificial neural network is a computational model, which replicate the simple function of a biological network and is used to solve complex nonlinear functions [21]. The basic element of a neural network is called neuron. The neurons are located in the network layers. It has the weight factors equal to the number of neuron connections and a bias variable. Each neuron may implement transfer function of different type. The layers are defined as the input layer, the output layer and the hidden
Industrial hydrogen plant
The required data for ANN training of hydrogen production plant are not available in the open literature due to proprietary reason. Therefore, we decided to generate process data by using modeling and simulation of hydrogen plant. The model is validated against the data obtained from industrial plant.
An industrial well-known hydrogen plant in the refinery has been simulated here. Each unit (reformer with its furnace, high and low shift converter, amine, and methanator) was modeled separately
Simulation and modeling of hydrogen plant
As mentioned before, an industrial hydrogen plant in Iran is selected for simulation. The plant uses the off gas of condensate recovery (CRU) and Platformer units and propane as feed and produces hydrogen by aforementioned process. The feed characteristics of the reformer are presented in Table 3.
The calculated results (based on the modeling and simulation) as compared with plant data, of the outlet unit composition for all units (reformer, amine, shift converters and methanator), are presented
Conclusion
In this paper, the feed forward neural network is designed for modeling the hydrogen production plant. The ANN model is identified by input–output data using a three-layer feed forward network and trained with a gradient descent algorithm. The number of hidden neurons is optimized by minimization of MSE of the neural network. ANN models temperature and the mole fraction of carbon monoxide and hydrogen in the product of hydrogen production plant. The present study shows ANN can successfully
Glossary
Acronyms
- ANN
- Artificial neural network
- CRU
- Condensate recovery unit
- DEA
- Di ethanol amine
- MSE
- Mean square error
- PSA
- Pressure swing adsorption
- WGS
- Water–gas shift
Latin
- A
- Network output
- Across
- Cross section area of reactor, m2
- Cp
- Heat capacity, kJ kg−1 K−1
- dp
- particle diameter, m
- E
- Performance function
- f
- Friction factor
- F(n)
- Activation function
- F
- Mole flow, kmol/h
- °C
- Centigrade
- k
- Number of learning epoch
- ki
- Equilibrium constant of reactions I, II, III, bar2, –, bar2
- Kj
- Adsorption constant for component j
- lr
- Learning rate
- mc
- Momentum rate
- n
- Number of
References (32)
Sulfur-passivated nickel catalysts for carbon-free steam reforming of methane
J Catal
(1984)- et al.
Optimizing a steam-methane reformer for hydrogen production
Int J Hydrogen Energy
(2009) Thermodynamic comparison of hydrogen production processes
Int J Hydrogen Energy
(1996)Thermodynamic investigation of hydrogen production by steam-methane reforming
Int J Hydrogen Energy
(1991)- et al.
Use of a heterogeneous two-dimensional model to improve the primary steam reformer performance
Chem Eng J
(2003) - et al.
Hydrogen production from the low-temperature water–gas shift reaction: kinetics and simulation of the industrial reactor
Int J Hydrogen Energy
(1995) - et al.
Mathematical modelling and computer simulation of industrial water–gas shift converters
Math Comput Model
(1989) - et al.
Water–gas shift reaction – a 2-D modeling approach
Int J Hydrogen Energy
(2008) - et al.
Reactor modeling to simulate catalytic partial oxidation and steam reforming of methane. Comparison of temperature profiles and strategies for hot spot minimization
Int J Hydrogen Energy
(2007) - et al.
Modelling the SOFC behaviors by artificial neural network
Int J Hydrogen Energy
(2009)
Sensitivity analysis and faults diagnosis using artificial neural networks in natural gas TEG-dehydration plants
Chem Eng J
A reliable multi-objective control strategy for batch processes based on bootstrap aggregated neural network models
J Process Control
Model based control of a yeast fermentation bioreactor using optimally designed artificial neural networks
Chem Eng J
A unified model for top fired methane steam reformers using three dimensional zonal analysis
Chem Eng Process
Complete removal of carbon monoxide in hydrogen-rich gas stream through methanation over supported metal catalysts
Int J Hydrogen Energy
Multi-layer feedforward networks are universal approximations
Neural Netw
Cited by (76)
Syngas production by ultrarich oxy-natural gas combustion in a pilot-scale porous burner
2024, International Journal of Hydrogen EnergyRecent advances in artificial neural network research for modeling hydrogen production processes
2023, International Journal of Hydrogen EnergyMachine learning in fermentative biohydrogen production: Advantages, challenges, and applications
2023, Bioresource TechnologyCitation Excerpt :Among the experimental runs, 70 % of the experimental data was utilized for training, while the remaining 30 % was used for testing and validation (Hossain et al., 2016). Zamaniyan et al., (2013) demonstrate the modeling and simulation of an industrial hydrogen plant using a forward artificial neural network. Each unit had been modeled separately and linked with each other using a single hidden layer.
Data-driven modelling and optimization of hydrogen adsorption on carbon nanostructures
2022, International Journal of Hydrogen EnergyProgress of artificial neural networks applications in hydrogen production
2022, Chemical Engineering Research and Design