Greg: We last got Vivek R. Dabholkar to share his expertise in a May 2026 Control Talk Column “How to diagnose PID tuning problems”. How have you become an excellent solver of furnace control problems?

Vivek: Furnaces are widely used in the process industries such as heaters (crude tower) or as reactors (ethylene plants and other units). Decades ago, as a fresh Ph.D. engineer working in industry, I wondered how one can control a furnace without understanding quantitative details of firebox geometry, furnace design and reaction chemistry (in case of furnace as a reactor) with the very large number of coupled detailed differential equations. This is where one needs to make a drastic change in thinking.

First understand the process qualitatively in terms of multivariable effects (either through a mentor who is willing to share his/her hard-earned experience or by observing process trends and keeping notes of what you learned over the years) then use the plant data to gain quantitative understanding of gains and dynamics involved. Once you linearize the process (not as explained in process control textbooks, but through the nonlinear transformation over the entire range practiced by APC experts in the field) local gains work well over the full range along with the continuous process feedback.

It took me a while to make this hard transition since my Ph.D. work was highly theoretical involving esoteric “Bifurcation Theory” with symmetry math applied to chemical engineering “problem.” My head was in the clouds until reality check in the industry hit me like a rock. I am still amazed by quite a few industrial optimization engineers, when you ask them for qualitative explanation on some optimizer solution, they respond “that is why we have optimizer!” I believe an engineer in an industrial setting should be able to convincingly explain the solution in qualitative terms.

Greg: What is essential upfront knowledge?

Vivek: There are three critical elements in controlling any furnace:

1. Combustion control controlling O₂ to make sure that O₂ is healthy (>1.5%) at the arch. Measurements should include O₂ and CO/combustibles.
2. Draft control should be sufficiently negative to let outside air come into the furnace through thermo-siphon effect, so a cold air column pushes out less dense hot air inside the furnace via  the u-tube principle studied in high school.
3. Coil outlet temperature (COT) control is either a process requirement of a downstream column such as a crude tower or a conversion requirement in a cracking furnace.

Greg: Where should furnace oxygen (O₂) be measured?

Vivek: Furnace O₂ should preferably be measured and controlled at the arch to represent condition that closely resemble conditions at the firebox. Arch O₂ measurement poses some instrument shielding challenges due to elevated temperature of the firebox. Stack O₂ measurement can be quite different from at the arch due to air leaks through the convection section.

Sometimes only stack O₂ is available in this case historic data must be plotted for combustibles vs. stack O₂ in a x-y plot to set stack O₂ low limit so that combustible do not show up in the stack beyond baseline noise level. I have seen stack O₂ as high as 6% in case bad leaks.

Typically, a zirconium oxide-based probe is used to measure O₂. In modern cracker plants, there is a trend to measure O₂ through laser setup. One must be careful of even slightest ground movement that can move (change the angle) the beam especially if combustion chamber width is large causing a miss on the receiver end.

Another hurdle is if one is decoking into the firebox. Leftover particulates adversely affect reliability of the measurement. These were some of the challenges at a greenfield cracker with which I was associated.

Greg: What do we need to know about draft control?

Vivek: Draft is the difference in static pressure between inside minus the outside value at a given height. It is a negative number; maximum draft refers to large negative value (contrary to common math terminology). It is measured in inches or centimeters of a water column (wc), typically in the range of -0.03 to -0.2 in wc.

Furnaces can be on natural draft, induced draft or forced draft. In a pure natural draft, furnace stack would have to be very tall to ensure airflow through the furnace (despite congested convection section) due to thermo-siphon effect (based on density difference in outside colder air vs. firebox air).

Induced draft furnaces have ID fan to create suction (driving force for combustion products to escape into atmosphere) increasing the draft. In case of induced draft, draft is maximum (most negative) at the furnace bottom and near the ID fan.

Unlike forced draft furnace direct air flow measurement is not possible due to extremely low pressure drop available for measurement. As a result, no cross-limiting (lead-lag) combustion control is possible in case of induced/natural draft furnaces.

In case of forced-draft furnace air is added before the fuel and the fuel is reduced before the air is reduced to ensure safety of the furnace. Usually, draft is controlled by adjusting the damper OP which is often non-linear due to butterfly valve like action.

Alternatively, draft can be controlled by adjusting speed of variable speed fan at fixed damper output. Variable speed fan setup tends to be more expensive. An effort must be made to characterize damper output arch O₂ for fixed feed, so that draft is controlled effectively over the entire range of operation. This can be accomplished through careful air-to-close valve characterization or variable gain using piece-wise linear (PWLN) function. If you fail in mapping entire range, at least identify the knee point beyond which damper is as good as wide open (usually around 70%) and conjure up a nonlinear transformation/PWLN gains to be modified as more data becomes available on closed-loop behavior.

When draft is increased (made more negative) damper must open lowering the resistance for flue gas, this in turn increases arch O₂ but also carries heat from combustion chamber to higher level in convection section increasing cross-over temperature (XOT). Maintaining XOT below high limit is important due to markedly different tube-metallurgy in the combustion chamber vs convection section.

Sometimes you may observe O₂ may be healthy, but draft is low (close to zero). This usually occurs when excessive air enters through burner air-registers which need to be reduced. Sometimes, several furnace stacks are combined and are routed to waste heat boiler (WHB) for additional heat recovery. In this case there is inherent interaction among furnaces, i.e., furnace variables on one affect the other furnace which often is corrected on feedback as opposed to explicit model.

The effect is pronounced on neighboring furnace sharing a stack when one furnace goes down for decoke, making flue gas easier to pass through common duct, much like reduced traffic jam across multiple lanes on the highway, causing O₂ to rise if damper is manipulated variable within DMC. On the other hand, furnace coming up sharing a common duct causes flue gas traffic jam leading to O₂ drop or draft output to increase depending on whether damper OP is directly manipulated within DMC or the draft setpoint is manipulated.

Greg: How do we control coil outlet temperature?

Vivek: COT is controlled by adjusting fuel gas flow, either through cascaded fuel gas flow controller or fuel gas pressure controller. If O₂ is extremely low (causing combustibles to rise), then increasing fuel flow lowers COT. PV causing dangerous positive feedback increasing fuel flow putting a furnace in dangerous region, called furnace flooding. This must be avoided at all costs and hence periodic calibration of O₂ is mandatory.

Stay away from the temptation/operator request to control combustibles directly or by varying O₂ low limit within DMC, you are playing with safety trying to save few bucks in energy loss through the stack. If fuel gas quality (mole weight) is changing and culprit is hydrogen content then linear correlation of lower heating value (LHV, kcal/mole) correlation to mole weight (density) can be used to implement a duty controller above fuel gas flow controller adjusting the fuel flow in response to online density changes due to changes in hydrogen content of fuel gas.

If density meter is being calibrated, make sure PVSRC is set to manual or else all blocks in COT control will go bad due to propagation of bad value. Even more important, when new reading for density becomes available, if PV is much different from PVMAN slowly changes it towards new value or else you may cause big upset to furnaces.

Some textbooks preach use of Wobbe Index, but analyzer is much delayed and using it in COT control is like after your car windshield gets shattered by hailstorm and then you get a warning of hailstorm on the weather channel. If change in heating value is due to increased olefins in fuel gas, then LHV correlation does not entirely correct. If COT is cascaded to fuel gas flow, then there should be first low select block, selecting lower output between fuel gas flow controller and high fuel gas pressure override followed by high select block selecting higher of the outputs between previously selected output and the low fuel gas pressure override controller output. High fuel gas pressure override protects against flame lift off due to excessive fuel gas pressure often due to a smaller number of burners in service for processing larger than expected furnace load.

Low fuel gas pressure override protects against flame out due to low fuel gas pressure often due to substantial number of burners in service when feed is introduced to the furnace. It is tricky to implement duty controller without fuel gas flow measurement; it requires manual input on number of burners along with vendor provided burner characteristics along with fuel gas pressure. During unusual fuel gas composition disturbances, O₂ can drop rapidly below dangerous level, protection must be provided by implementing low-low O₂ override just above low fuel gas pressure override to put furnace in safe condition.

Get your subscription to Control's tri-weekly newsletter.

I witnessed one such episode while working at a ExxonMobil joint-venture site as an expat by examining incidence trends on the weekend. We were lucky to avoid explosion despite black smoke in the stack as communicated by the operations. Mixture was too rich to explode. Appropriate operator’s action is to reduce fuel and not to add air by opening the top damper which is a usual action to recover from low O₂. As a result, we implemented low-low O₂ protection for furnaces. In later years, ExxonMobil took corporate wide strong leadership position to implement this across all chemical plants and refineries for inherent safety and to save lives. This is clearly lacking in other major players like Shell Monaca (where I was unsuccessful in convincing management at highest level you can imagine within the corporation upsetting several folks in the management ladder to implement this safety related control prior to start-up due to inexperienced operators, where the middle management was regularly giving sermons on safety) and other minor players in the industry. Implementation while necessary can be onerous, if a furnace has multiple zones in the combustion chamber with COT to fuel gas pressure cascades.

There is less commonly used attractive scheme for COT control used by the leader in the industry. Controlling COT with Fuel gas flows involve zone interactions, firing one zone harder creates more heat flow to neighboring zones within same combustion chamber causing their COT controllers to cut back on respective fuel gas flows, if one has not accounted for cross interactions in the base level DCS controls then there can be significant sloshing around duties and COT process values across the furnace. This cannot be fixed at DMC level.

Alternatively, one can use master fuel gas duty controller with fan-out configuration to fuel gas flows with usual overrides. COT is controlled by adjusting coil feed flows. In this case zone interaction is eliminated. Coil feeds may be equalized using fuel gas bias stations if available. Another advantage if very same furnace can crack naphtha that cracks at lower temperature (due to lower C-C bond strength result of much longer chain length) in some coils and ethane that cracks at much high temperature (due to relatively stronger C-C bond)  in other coils as opposed to co-cracking across coils by pre-mixing different feed types ahead of cracking that often leads to excessive uncracked ethane and over cracking of naphtha in case of co-cracking with naphtha.

Another advantage of this scheme is that low-low O₂ protection can be enabled at one place at total duty controller output before it fans out. But there are unusual effects one needs to be aware of in this configuration. For example, if one lowers draft setpoint fuel flow is unchanged but feed rate is reduced due to COT to Feed cascade and arch O₂ will increase due to increased draft (more negative).

If the total duty is increased, then total furnace feed is increased automatically due to COT controller action and O₂ will drop due to increased fuel gas. In other words, total duty controller acts like total feed controller. Increasing COT setpoint has no model to arch O₂ or draft output but only coil feed reduced on the same coil. In this configuration, COT increase has bigger effect on conversion due to combination increased temperature and reduced feed at the same time. If configuration is used for Cracker Furnace and if one increases Steam/HC ratio, COT controller will cut back coil feed to hold COT at its setpoint. A combination of which could have much diminished gain between S/HC and coil conversion.

Another nuance is between total duty and draft effects on the controlled variables. If one increases duty or decreases draft (absolute value of draft) in both cases coil feed is increased through COT action albeit each by different amount for unit changes in the inputs. Therefore, any other controlled variable that is affected by coil feed increase such coil feed output, steam output (with S/HC ratio constant) must have models to both duty and draft, which is first level conclusion. Engineers must think through this logic mentally and not look for case-run button or magic AI button to figure it out.

Let us take it to next level in the thought process, where one makes unit increase in duty and makes a counter move in draft by some amount such that net change in feed is exactly zero (double-precision). Then logical inference is the net change in feed output and steam output must also be exactly zero for the same combination of moves. This holds true everywhere locally regardless of the choice of operating point. This is making problem exactly singular and eliminating non-existent degree of freedom.

Mathematically, one can spot this through so called relative gain array( RGA) analysis, but we offer simple logical thinking process that can be applied to any problem systematically. Now even if you were to model firebox and other furnace aspects in full dynamic details, there is no way around it than to obey this strictly to maintain model consistency. Otherwise, DMC would come up with unpalatable solution exploiting non-existent degree of freedom.

Similar effects can be reasoned out provided common effect is only via single controlled variable and not through other mechanisms in two-by-two case. General thinking process is if any pair of independent variables individually affect a selected controlled variable (intermediate variable) and some (not all) of the other controlled variables are changed only through the of change this intermediate controlled variable then a combination of moves in independent variables that produces net-zero effect on the intermediate controlled variable. Then the same combination must also produce net zero effect on other controlled variables with peculiar characterization described above.

I know this is mouth full but does not require knowledge of eigen-values, RGA, etc. This is much more enjoyable intellectually stimulating exercise than mechanically running RGA tool. Issue with using RGA tool blindly is, if the original gains are far away from being collinear, then RGA tool will never flag it and fix it. On the other hand, if the pair of gains are nearly collinear due to inadequate analysis, RGA will eliminate a valid degree of freedom. I prefer to think first and then run RGA tool to educate oneself what other things one missed, thereby increasing engineering judgement over the years. It will not happen overnight but a slow process until you develop good judgement.

In case of cracking furnaces, conversion/severity that is calculated by SPYRO program/regression is updated by online effluent analyzer. Conversion adjusts COT setpoint by the DMC or any other multivariable controller.

Traditional pass-balancing on furnaces involve differentially putting more feed into hotter coil while taking out feed from colder coils, balancing individual coil outlet temperatures within high/low limits. This can be easily achieved within DMC or any other multivariable controller. A variation of traditional pass balancing is using SPYRO coil conversion balancing while only single online furnace effluent analyzer is available.

In 99% of cases, coil feed flows are manipulated variables along with coil COTs and draft setpoint/damper output. In case when DMC is “off”, the operator needs to adjust total feed a separate DCS based total furnace feed must be in place in fan out configuration to adjust total feed to reduce fat-fingering by the operator trying to set individual coil flows due to substantial number of passes.

I encountered one odd case where coil flow biases (instead of coil feed flows) along with total furnace feed were manipulated variables (MV), along with coil COT setpoints and S/HC ratios, damper controller output (OP) as MV. In this since total feed is MV, coil flow biases individually should have net zero effect on stack O₂ and draft due to compensating effect of changes in other pass-flow to hold constant total furnace feed. Changing coil flow bias on say coil-1 will affect conversions, fuel gas Ops etc. on other passes since other pass flows would be moving in opposite direction to hold constant total feed. Utmost care must be taken in modeling such an highly integrated system resulting into dense model matrix.

Greg: What are your concluding remarks?

Vivek: In summary, one must think through each control configuration and its interaction with other DCS based controllers before testing and model building process for multivariable controller. This will eliminate bogus models and its implications for multivariable controller. Tendency these days is to crank the handle on software without first thinking through each model and internalizing what is means.

This and other articles are especially written for newcomers to the process control world. The world’s Top APC practitioners may already be familiar with it. Most universities or even graduate programs are focused too much on mathematical analysis without adequate introduction to how basic controls work on important pieces of process-equipment, as a result even several Ph.D. earners in process control with perfect GPA are ill-prepared for Industrial applications job, and most spend time on procedures, guidelines, useless meetings and towing the line of higher ups doing little of actual value to the company.

Greg: I have worked on furnace pressure controls in specialty chemical production units that required fast controller execution rates, measurements, and control valves. I had several production furnaces need controllers with a 0.05 sec execution rate, 0.1 second transmitter lag, and 1 second valve 86% response time (T86). To achieve T86 required putting a booster on the output of the valve positioner with its bypass valve barely open. Then there was an extreme case with an analog measurement and analog furnace pressure controller manipulating a fast variable speed drive. The pressure could go off-scale in less than 0.1 seconds due to disturbances. I could not safely migrate to a digital measurement or digital controller or use a control valve.