Hydrocarbon Processing Copying and distributing are prohibited without permission of the publisher
Email a friend
  • Please enter a maximum of 5 recipients. Use ; to separate more than one email address.

Update temperature control for a Claus unit with neural networks

11.01.2013  |  Schabort, C. J. ,  North-West University, School of Chemical and Minerals Engineering, Potchefstroom, South AfricaNeomagus, H. W. J. P.,  North-West University, School of Chemical and Minerals Engineering, Potchefstroom, South Africa

Keywords: [neural sensors] [neural networks] [Claus unit] [sulfur recovery]

The Natref oil refinery in Sasolburg, South Africa, had experienced refractory problems in the sulfur recovery unit’s (SRU’s) combustion chamber since startup (Fig. 1). Even though the SRU was at all times operated within the temperature range of 1,250°C to 1,750°C (as measured on the original thermocouples and the pyrometer), refractory failures occurred that could only be accounted for by temperatures in excess of 1,850°C. It was concluded that the available temperature indication was not an accurate reflection of the actual temperature in the combustion chamber.

  Fig. 1.  A refractory failure in the combustion
  chamber of the SRU.

Three new sets of temperature measurement were installed to better understand the temperature profile in the combustion chamber (each set consists of two thermocouples and one pyrometer). Plus, a simulation was used to set up a neural network to predict the combustion chamber temperature based on three input variables: the oxygen enrichment level, the ratio of sour water stripper (SWS) offgas to acid gas and the amount of acid gas bypassed to the second chamber of the reaction furnace.

The reaction furnace was found to be operating at an average temperature of 1,400°C during normal operation. This temperature is substantially higher than the normal recommended 1,250°C for Claus combustion chambers. The neural network was connected to the Natref distributed control system (DCS) and it proved that, during the SRU’s startup in 2006, a temperature in excess of 2,000°C was reached inside the combustion chamber, explaining the refractory failure and subsequent shutdown of the unit.

To establish which temperature measurement should be used for control purposes during normal operation, nine temperature measurements over the length of the combustion chamber were compared to the neural network temperatures. All of the other thermocouples (T1–T6) were found to under-predict the simulated temperature, and the response of these thermocouples was much slower than that of the pyrometers.

With the exception of the burner box pyrometer, P1, all the pyrometers also under-predicted the simulated temperature. The response of the pyrometers was, however, much faster and showed the strongest linear correlation with regard to the adiabatic flame temperature. The reason for the under-prediction can be attributed to heat losses in the combustion chamber. The middle pyrometer, P2, gave the smallest deviation between measured and predicted temperature values of only 28°C on average and should be used as the main control point with regard to temperature.

Long view

Crude oil refining consists mainly of three processes—namely separation, conversion and product upgrading.1 Hydrocracking and hydrotreating play an important role in upgrading distillation cuts to meet the specifications of final products. In both cases, hydrogen is used to remove sulfur and nitrogen compounds from the feed streams.2 The hydrogen gas that is used to remove sulfur and nitrogen compounds is recycled to the hydrotreating and hydrocracking reactors, using recycle gas compressors. As the level of hydrogen sulfide (H2S) in the recycle hydrogen stream increases, a purge stream from the unit is sent to an amine unit to ensure that the hydrogen purity in the recycle hydrogen stream remains constant. An alkanolamine in an aqueous solution reacts with the H2S to form a chemical complex. A stripping column is used to release the H2S from the amine and the regenerated solvent is reused.3

Ammonia (NH3) is removed from the recycle hydrogen stream by a different process. H2S and NH3 in the recycle hydrogen stream react to form ammonium sulfide (NH4HS). This is a highly soluble salt, which is removed by injecting water upstream of the recycle hydrogen gas fin-fan coolers. The resulting sour water accumulates in the water boots and is routed to the SWS. In the SWS, the salt undergoes hydrolysis to form H2S and NH3 again. The resulting product is known as SWS offgas.

The released H2S from the amine units, as well as the H2S and NH3 from the units, must be processed in an environmentally friendly way to limit sulfur oxide (SOx) and nitrogen oxide (NOx) emissions. To achieve this objective, a Claus SRU is used. The modified Claus process produces elemental sulfur by partial oxidation of H2S according to these main process steps:

3H2S + 1.5O2 → SO2 + 2H2S + H2O (∆h0 = –520 kJ/mol)      (1)

SO2 + 2H2S → (3/x)Sx + 2H2O (∆h0 = –93 kJ/mol)                   (2)

Due to stricter environmental regulations, the requirement for co-processing of ammonia-bearing streams (offgas) has become imperative. Incomplete destruction of the NH3 in the combustion chamber, however, can lead to the formation of ammonium salts in cooler downstream units. These salts can lead to an increase in pressure drop over the system, causing unplanned shutdowns to clean the fouled equipment.4

Close examination

The Claus unit at the Natref refinery consists of a thermal stage configured for oxygen enrichment and ammonia destruction with split flow of amine acid gas, and two Claus catalytic reaction stages. The main feed to the SRU is acid gas (consisting primarily of H2S) and offgas (consisting primarily of H2S, NH3 and H2O). The mixture of acid gas, offgas and air enters the furnace, which operates at 1,250°C–1,450°C.5

The products that are formed are mainly SO2 and water (steam). The NH3 and hydrocarbons that are present in the feed are converted in the burner muffle. Combustion is controlled in such a way that a ratio of unreacted H2S to SO2 of 2:1 is obtained.6 Subsequently, sulfur is formed and condensed in the downstream waste-heat boiler (WHB), where high-pressure steam is produced for the catalytic stages of the SRU. The remaining H2S from the WHB is reacted with SO2 at lower temperatures (200°C–350°C) over an alumina- or titanium dioxide catalyst to produce additional sulfur.

Two parameters of importance to assist in the destruction of NH3 are the flame temperature and the residence time in the combustion chamber. Most designers recommend a minimum flame temperature of 1,250°C and a residence time larger than 0.8 seconds as part of the performance guarantee for the Claus unit. This corresponds to a combustion chamber exit NH3 level below 30 ppm.6, 7

With a residence time greater than 1 second, a flame temperature higher than 1,200°C is required. With a residence time greater than 0.5 seconds, a flame temperature higher than 1,250°C is necessary. This corresponds to a combustion chamber exit NH3 level of less than 60 ppm.

The maximum allowable temperature is dependent on the type of refractory installed. The normal refractory limit is seen as 1,650°C, but some refractories are robust enough to handle temperatures of up to 1,800°C. The melting point of refractory that is 90% alumina with silica and 94% alumina with magnesium is about 1,870°C and 1,926°C, respectively.8

Based on the available literature, it can be concluded that a flame temperature higher than 1,250°C and a residence time of greater than 0.8 seconds should be sufficient to obtain adequate NH3 destruction to prevent unintended downtime due to blockages or loss of catalyst activity due to sulfation. To protect the refractory, however, the combustion chamber should not be operated above 1,750°C in cases where robust refractories are utilized.

To simulate the reaction furnace of the modified Claus plant, designers normally rely upon thermodynamic equilibrium calculations to determine the temperature and composition of the product stream. Even though equilibrium is generally not reached, the equilibrium approach is still used in simulations because reaction rate constants for a kinetically limited reactors are not readily available for all the reactions that occur in the combustion chamber. Equilibrium results are thus adjusted to reflect the kinetic limitation for the reaction furnace.

One of the key elements of automated temperature control in a SRU combustion chamber is reliable temperature measurement. Since the feed to the SRU is received from various sources, unexpected increases in hydrocarbons or NH3 in the feed can lead to excessive temperatures that can damage the refractory.9

Stainless steel sheathed thermocouples were initially used as measurement tools in Claus units. Over time, the process gas, mainly consisting of H2S, corroded these thermocouples. To protect the thermocouple, it was placed inside a ceramic thermo-well. Unfortunately, the thermo-well cracked due to thermal expansion of the refractory upon startup. Thus, the thickness of the ceramic thermo-well was increased as a solution, but H2S still managed to penetrate the porous ceramic and corrode the thermocouple over time. As a final solution an air-purge or nitrogen-purge was utilized inside the thermo-well. This protected the thermocouple sufficiently against H2S attack.

Even though these purged thermo-wells were able to protect the thermocouples against the highly corrosive H2S at elevated temperatures, the configuration does not allow for fast enough response to provide reliable temperature reporting. With sheathed, purged thermo-wells the response time can vary between 1 second and 15 seconds.

Over the last couple of years, pyrometers have become a very popular technique for measuring temperatures in the Claus combustion chamber.10 A pyrometer can be used to either measure the process gas temperature or the refractory temperature in the combustion chamber.11

The infrared thermometer is mounted outside the combustion chamber sighting into the furnace through a view port. This keeps the pyrometer outside of the severe environment. Not only can pyrometers provide accurate temperature measurements of the ceramic wall over a wide range, but the response time is also below one second and thus much faster than with thermocouples.


The difference in the temperature readings of a pyrometer and a thermocouple in a nitrogen-purged thermo-well were shown in a study done by sulfur experts at the Natref refinery.12 In this study, the adiabatic flame temperature was calculated for the front zone of the combustion chamber. The pyrometer temperatures were generally 13% to 20% (200°C to 400°C) lower than the actual temperature. The temperatures from the thermocouples, however, were 13% to 38% (230°C to 675°C) lower. It was concluded that these thermocouples were not accurate and should not be used for controlling the operation of the reaction furnace.

Due to the inaccuracy of the temperature measurement in a SRU, as well as the slow response time, a feed-forward-based control system with a temperature as a check was suggested as the preferred arrangement in the study.

The main purpose of this study was to propose an alternative operating strategy for the Claus combustion chamber used at Natref. To achieve this, a combustion chamber simulation was used to set up a neural network to simulate temperature in the combustion chamber. This was done to determine which temperature measurement should be used for the furnace control to prevent any subsequent refractory damage.

Experimental research

All the experiments were carried out in the SRU at the Natref refinery. Both the acid gas and SWS offgas streams were sampled six times over three days. These samples were taken while the plant was online and fully operational.

The acid-gas stream was sampled using proprietary analysis for bulk sour-gas process samples. The offgas stream was also sampled. These samples were analyzed water and sulfur free, and the composition was reported on a dry basis.

After all the refractory failures experienced in the combustion chamber, three new sets of temperature measurement (consisting of two nitrogen-purged thermocouples and one pyrometer) were installed to better understand the temperature profile in the combustion chamber. The location of all the temperature measurements is shown in Fig. 2. These temperatures were recorded in minute intervals over a 30-day period.

  Fig. 2.  Location of temperature measurements
  in the combustion chamber (side view).

The thermocouples are fitted inside a ceramic well. The materials used in the construction of the Type B thermocouples were tungsten, platinum and 5% rhenium. These thermocouples can accurately measure any temperature between 200°C and 1,680°C.

The pyrometers installed at Natref consisted of two types. The first pyrometer installed had the capability to accurately measure any temperature between 1,000°C and 2,600°C. Dual silicon cell detectors operating at 0.85 µm to 1.1 µm were used. The second installed was a modular signal processor, which provided a large digital display of temperature and emissivity, as well as alarm settings.

The error on the pyrometers can be as high as 32 K, which is much higher than the error of 10 K of the thermocouples. The advantage of the pyrometers over the thermocouples, however, is the much faster response time.

Results and discussion

The composition of the acid gas and SWS offgas streams for the six test samples analyzed during the test run is summarized in Table 1 and Table 2. To establish whether the compositions of the streams remain constant over time, the six test samples are compared to the Natref daily analyses of the acid gas and SWS offgas streams over a three year period.



The acid-gas composition and the SWS offgas is a crude slate function that is fed to the refinery. If the crude slate remains constant, the feed composition should remain the same. There have not been considerable changes to the crude slate to the Natref refinery over the last couple of years, so it was expected that the feed composition would remain constant until major changes were made to the crude slate selection.

The average acid gas and SWS offgas feed compositions were normalized at the average knockout temperatures of 32°C and 92°C and pressures of 54 kPa (g) and 142 kPa (g) to give a wet basis flow. These results are shown in Table 3.


To determine the adiabatic flame temperature, the various flowrates to the combustion chamber were recorded in minute intervals simultaneously. Software was used to simulate the SRU’s combustion chamber. Due to temperatures in the combustion chamber being in excess of 1,000°C and pressures being relatively low, the ideal property method was selected as the thermodynamic equation of state for the simulation. The analytical results obtained were used as input for the various feed streams in the simulations.

The general simulation results were used to set up a neural network and to predict the adiabatic temperature in the reaction furnace based on the three input variables: the molar ratio of SWS offgas to acid gas, the oxygen enrichment level and the acid gas bypass percentage. The network topology used was one with three inputs, two hidden layers and two output nodes. The objective function used for minimization was the sum of the absolute percentage errors. This model was used to compare the actual temperatures measured in the reaction furnace by means of the pyrometer and thermocouples with the temperature predicted by the temperature model. Fig. 3 shows how well the neural network performed on the training data in relation to the adiabatic front temperature.

  Fig. 3.  A comparison of the actual simulated
  temperatures and the neural network
  predicted temperatures.

The training data could be incorporated successfully into the model with an accuracy of 98.5%. The model was further verified by using simulated data not used during the training and comparing that to the temperature results from the neural network model. This comparison is shown in Fig. 4. 

  Fig. 4.  Validation of the neural network
  predicted temperatures with simulated

For the adiabatic front temperature, a correlation of 99.29% was obtained with the proposed topology. The maximum relative error for the adiabatic flame temperature was 2.25%. The neural network was used to predict the actual temperature in the combustion chamber during the startup period in 2006 at the Natref facility, with data from the DCS plant history. It can be clearly seen from Fig. 5 that temperatures in excess of 1,750°C were reached inside the combustion chamber, thus explaining the refractory collapsed and a subsequent SRU shutdown. It can further be seen that the pyrometer read, on average, slightly higher than 1,500°C, while the two thermocouples only recorded temperatures between 1,250°C and 1,350°C.

  Fig. 5.  Measured and predicted temperature
  in the combustion chamber during the
  2006 Claus unit startup.

To better understand the temperature profile in the combustion chamber, the neural network was used to determine the accuracy of the nine installed temperature measurements by comparing the predicted temperatures with actual temperature measurements during normal operation. In Table 4, both the linear and Spearman rank correlation coefficients are shown, as well as the average deviation between the actual and predicted temperature.


The burner box thermocouples, T1 and T2, were damaged due to excessively high temperatures. It was concluded that these thermocouples should not be used for control purposes. None of the other thermocouples (T3–T6) showed a strong relationship with the neural network temperatures. The poor relationship can be explained by the location of the thermocouples inside the refractory bricks, leading to a slow response to changes inside the combustion chamber. It is not possible to determine the instantaneous temperature of the furnace wall by using thermocouples.

It was expected that the thermocouples should thus read a lower temperature at the same axial position if compared to the pyrometers. This was the case if the middle thermocouples, T3 and T4, are compared to the middle pyrometer, P2. The same trend was found when the back pyrometer, P3, is compared to the back thermocouples, T5 and T6.

A poor relationship is seen between the front pyrometer and the neural network temperature. The readings of the pyrometer can be erratic at times. The aim of the pyrometer was to measure the temperature of the refractory wall on the opposite side of the sight glass. If conditions occur where feed burns in the sight path of the pyrometer, this will have an influence on the pyrometer reading. It can be concluded that the front pyrometer was too close to the flame.

A strong relationship existed between the neural network temperature and the middle and back thermocouples. An average temperature difference of 28°C and 88°C between the simulated adiabatic temperature and the measured middle and back temperatures, respectively, can be explained by heat losses experienced over the axial length of the combustion chamber, as well as the temperature profile over the length of the combustion chamber. Taking the heat losses into account, these two pyrometers correlate the best with the actual temperature inside the combustion chamber.

To verify the findings of the neural network comparison, the linear correlation coefficients were calculated using plant data only to determine the linear relationship between all measured inputs to the combustion chamber and the measured temperatures on the pyrometers and the thermocouples. The matrix of the linear correlation coefficients is shown in Table 5.


The correlation between the acid gas feed stream and the oxygen feed stream was very strong (0.93). As more acid gas was fed to the unit, more oxygen was required for partial combustion of the H2S. There was also a strong correlation between the offgas feed stream and the oxygen (0.73), but to a lesser extent. This can be explained in that the oxygen demand is linked to the acid gas and offgas streams, but, due to the higher H2S content in the acid gas than in the offgas, the correlation was stronger with the acid gas than with the offgas.

The correlations between the various feed streams and the total flow was also significant. The total feed flow was the summation of the various feed streams and the strong correlation.

Another factor that can be evaluated from this data is the residence time. The correlation between the total flow and the measured temperature on the pyrometers, especially P2 (0.07) and P3 (0.08), was very poor. Residence time had a negligible influence on the measured temperatures on the pyrometers. The same poor correlation was found between the total flow and the temperature measured by the middle thermocouples, T3 (0.20) and T4 (0.16). This indicates that residence time did not have a significant influence on the measured temperatures on the middle thermocouples.

The only correlation that showed significance was the correlation between the total flow and the temperature measured by the back thermocouples, T5 (0.43) and T6 (0.40). The change in flowrate thus leads to a change in flow pattern close to the choke ring. These thermocouples should thus not be used for temperature control, as they are strongly influenced by the flow pattern inside the combustion chamber. This explains why an actual temperature of above 1,750°C was achieved during startup in 2006, while these two thermocouples never indicated a similar high temperature.

The last factor that is evaluated is the correlation between the individual temperature measurements. There is a poor correlation between the front pyrometer, P1, and the other temperature measurements. A very strong correlation of 0.99 is found between the middle and back pyrometers, P2 and P3. This shows that these two pyrometers follow a similar trend.

The correlation between P2 and P3 with middle thermocouples, T3 and T4, are strongly correlated (0.75 to 0.81). This is, however, not the case with the back thermocouples, T5 and T6. Furthermore, the four thermocouples are strongly correlated among themselves with correlations ranging from 0.73 to 0.92.

Wrapping up

Based on the neural network results, it was found that the reaction furnace was operated at an average temperature of 1,400°C, which is substantially higher than normally recommended for Claus combustion chambers.

The previous refractory failures in the SRU combustion chamber were due to temperatures in excess of 2,000°C during startup. The original temperature measurements (P3, T5 and T6) under-predicted the actual temperature in the combustion chamber. This can be attributed to the proximity to the choke ring, which has been proven to have a significant influence on the temperature measurement of the thermocouples (T5 and T6).

In the process of establishing which temperature measurement should be used for control purposes, the front pyrometer, P1, was found to be erratic and the correlation between the pyrometer and the neural network poor. The middle and back pyrometers, P2 and P3, measured, on average, 28°C and 88°C lower than the adiabatic temperature, with a strong correlation with the neural network temperature. Both these pyrometers can be considered for temperature control inside the furnace. There was a strong correlation between the middle and back pyrometers, P2 and P3, of 0.99. Even though there was a temperature difference between these two pyrometers of 60°C, this difference remained constant. The ratio between these two pyrometers can thus be monitored to determine whether calibration is required.

It was also found that there is a strong correlation between the two middle thermocouples, T3 and T4, of 0.97. The difference between these two thermocouples is 6°C. Furthermore, the correlation between these thermocouples and the pyrometers is 0.75 to 0.81, and it can thus be used as an additional measurement for calibration purposes. If compared to the pyrometers, the thermocouples take longer to respond to operational changes in the furnace. The pyrometers have an instantaneous indication of the wall temperature inside the combustion chamber.

The middle pyrometer, P2, shows the smallest average deviation of 28°C and should therefore be used for temperature control instead of P3, which was used for temperature control in the combustion chamber. Not only does the middle pyrometer have a shorter response time than all thermocouples, but the average temperature difference between the adiabatic flame temperature and P2 was smaller than the difference between the adiabatic flame temperature and P3. HP


1 Riazi, M. R., “Characterization and Properties of Petroleum Fractions,” ASTM International, West Conshohocken, Pennsylvania, 2005.
2 Gary, J. H. and G. E. Handwerk, Petroleum Refining: Technology and economics, Dekker, New York, 1975.
3 Sheilan, M. H. et. al., Amine Treating and Sour Water Stripping, 2nd edition, Amine Experts, Canada, 2006.
4 Monnet, F. et al., “Cobalt, A New Primary Burner for SRUs,” NPRA annual meeting, San Antonio, Texas, 2003.
5 Gupta, A. K. and M. Sassi, “Sulfur Recovery from Acid Gas Using the Claus Process and High Temperature Air Combustion (HiTAC) Technology,” American Journal of Environmental Sciences, 2008.
6 Paskall, H. G. and J. A. Sames, Sulfur Recovery, 8th edition, Sulfur Experts, Inc., Canada, 2003.
7 Monnery, W. D. et al., “Ammonia Pyrolysis and Oxidation in the Claus Furnace,” Industrial Engineering Chemical Resources, 2001.
8 Rameshni, P. E., “Challenges for SRU Expansion with Oxygen,” Sulfur Recovery Symposium, Brimstone Engineering Services, Vail, Colorado, 2002.
9 Emery, E., “The ultimate temperature measurement instrument for the sulfur reactor,” Annual European Sulfur Recovery Seminar, Bernkastel, Germany, 2006.
10 Kral, J. and E. K. Matthews, “Pyrolaser and Pyrofiber Infrared Temperature Measurement with Automatic Emissivity Correction,” International Fair of Metallurgy, Ostrava, Czech Republic, 1996.
11 Emery, E., “Practical Experiences with Temperature Measurement in the Modified Claus Sulfur Reactor,” Laurance Reid Gas Conditioning Conference, University of Oklahoma, Norman, Oklahoma, 1998.
12 Derakhshan, F., “Sulfur plant evaluation at the Natref Sasolburg refinery,” 2007.

Have your say
  • All comments are subject to editorial review.
    All fields are compulsory.

Related articles


Sign-up for the Free Daily HP Enewsletter!

Boxscore Database

A searchable database of project activity in the global hydrocarbon processing industry


Is 2016 the peak for US gasoline demand?




View previous results

Popular Searches

Please read our Term and Conditions and Privacy Policy before using the site. All material subject to strictly enforced copyright laws.
© 2016 Hydrocarbon Processing. © 2016 Gulf Publishing Company.