Process Control

Process control is the engineering of keeping a plant at the conditions it was designed for.

11 frameworks, from 1868 to today · Chemical Engineering · Engineering

still a working view ended · each leads to its entry below

Overview

Process control is the engineering of keeping a plant at the conditions it was designed for1. A chemical plant is a web of coupled vessels, pipes, and reactors in which temperature, pressure, flow, level, and composition must be held within narrow ranges while disturbances—a feed that changes composition, a pump that drifts, a demand that rises—keep pushing them off. The controller is the part of the plant that watches and corrects12. It is what lets a refinery run unattended for hours, a reactor stay below the temperature at which it would run away, and a product meet a specification. The history of process control is the history of how engineers learned to make that correction automatic, and then how they learned to make it smart: to predict, to optimize, and to learn.

Regulators and the Birth of Feedback

The oldest idea in the field is feedback: measure the output, compare it to the target, and act on the difference. The centrifugal flyball governor, which uses spinning weights to open and close a valve as speed changes, was applied to windmills in the 18th century and made famous by James Watt's steam engine in the 1780s. The problem it solved was that a steam engine under varying load would otherwise run at varying speed; the governor made the engine self-correcting3.

The trouble was that feedback can also destabilize24. In 1868 James Clerk Maxwell published "On Governors," an analysis of why some governor arrangements oscillate or run away, and in the 1870s Edward John Routh and Adolf Hurwitz independently developed algebraic tests for whether a feedback system is stable. These were the first mathematical treatments of the control problem, and they established that a feedback loop is a dynamical system whose stability must be proved, not assumed.

The industrial form that dominated for most of the 20th century was the proportional-integral-derivative (PID) controller. Its three terms are easy to state: the proportional term pushes in proportion to the current error; the integral term pushes in proportion to the accumulated error, which removes steady offset; the derivative term pushes against the rate of change, which damps overshoot3. The three-term controller was developed in the 1920s and 1930s—Nicolas Minorsky's 1922 work on ship steering is a standard attribution for the full three-term form—and it spread through the process industries because it was simple, needed no model, and worked well enough on the single loops that made up most plants. The pneumatic PID controller, running on compressed air, became the standard panel instrument; its tuning was a craft, and the Ziegler–Nichols rules of the early 1940s gave a systematic way to set the three constants from a plant experiment.

Frequency Response and the Loop-Shaping View

A parallel tradition came out of communications engineering. Harry Nyquist's 1932 criterion showed how to judge closed-loop stability from the open-loop frequency response, and Hendrik Bode's 1940s work on gain and phase margins gave engineers a graphical language for design. Applied to process loops, this "frequency-response analysis" let an engineer see how a loop would respond to fast and slow disturbances and how close it was to instability. The classical tuning rules and the frequency-response view coexisted: PID was what you installed, and frequency response was how you understood why it worked or didn't.

The Model-Based Turn

The limits of single-loop PID became acute in the 1950s and 1960s, when refineries and petrochemical plants grew into tightly coupled networks. A change in one unit's duty rippled through many others, and a plant full of independent loops could fight itself. The response was to bring models into control5.

Model predictive control (MPC) was the decisive step5. Instead of reacting to an error after it appears, MPC uses a dynamic model of the plant to predict the future trajectory of the outputs and computes the input sequence that keeps them on target, re-solving the problem at every sampling instant5. The idea took root in the late 1970s and early 1980s in the process industries, where Shell engineers developed the approach that became known as dynamic matrix control, and it was commercialized for refinery and petrochemical applications. MPC's appeal was that it could handle constraints explicitly—maximum valve opening, maximum reactor temperature—which PID cannot5. It became the standard for the hard, interacting, constrained problems at the top of the control hierarchy.

Internal model control (IMC), formulated by Manfred Morari and co-workers in the early 1980s, gave a clean theoretical frame: put a model of the plant in parallel with the plant, feed back the difference between model and reality, and design the controller around the model. IMC made the role of the model explicit and produced a simple, tunable design procedure; it also clarified the trade-off between performance and robustness.

System identification supplied the models6. The question was how to get a dynamic model of a plant too complex or too dangerous to derive from first principles. The answer, developed from the 1960s onward—with major contributions from Karl Åström and Torsten Bohlin, and the prediction-error and subspace methods that followed—was to perturb the plant with test signals and fit a model to the input-output data. Identification turned model-based control from a theoretical possibility into routine practice.

State estimation and inferential control addressed a different gap: many of the quantities you most want to control—composition, product quality—cannot be measured online, or are measured only by slow laboratory analysis. A state estimator (the Kalman filter, introduced in 1960, is the canonical form) combines a process model with whatever measurements are available to infer the unmeasured states7. Inferential control then regulates those inferred values. The practical effect was that a column could be controlled on an estimated composition rather than waiting hours for a lab result.

Multivariable Structure and Robustness

As plants became more integrated, the question shifted from "how do I tune this loop" to "which loops should exist at all." Multivariable control-structure design—the analysis of pairing, interaction, and relative gain, associated with the work of Edgar Bristol in the 1960s and the singular-value tools that followed—asked which manipulated variable should control which measured variable, and showed that a bad pairing could make a plant unstable even with perfectly tuned individual loops. The relative gain array became a standard screening tool8.

Robust control confronted the fact that models are always wrong9. The question it asked was: if the real plant differs from the model by some bounded amount, will the controller still be stable and perform acceptably9? The H-infinity methods developed by George Zames and others in the 1980s, and the structured singular value framework of John Doyle, gave a rigorous answer for linear systems. Robust control was more influential in aerospace and academia than in the day-to-day process plant, but it reshaped how process control engineers thought about model uncertainty and set the standard against which simpler designs are judged.

The Hierarchy and the Optimizer

By the 1980s the plant was understood as a layered system. At the bottom, regulatory control (PID loops) held flows, levels, and temperatures. Above it, MPC handled the interacting unit. Above that, real-time optimization adjusted the setpoints to maximize profit given current prices, feed quality, and constraints, and above that, planning and scheduling decided what to make. This control hierarchy—regulatory, supervisory, optimization, planning—became the organizing architecture of the field, and the interfaces between layers (how often the optimizer updates, how the MPC handles infeasible setpoints) became problems in their own right.

Data-Driven and Learning-Based Control

The newest turn is the move away from first-principles models toward learning from data1011. The idea is not new—identification has always been data-driven—but the scale changed6. Machine learning methods, including neural networks and, more recently, reinforcement learning, are being applied to problems where a mechanistic model is unavailable or too expensive: nonlinear processes, grade transitions, and controller tuning itself. The open questions are the ones the field has always asked of any controller: will it be stable, will it respect constraints, and can its behavior be guaranteed when it encounters conditions it has not seen? The attraction is that these methods can capture behavior that no tractable model can; the caution is that a controller that learns online can also learn the wrong thing.

The Present

Process control today is a layered practice1. PID remains the workhorse at the bottom of nearly every plant, tuned by rules and experience. MPC is the standard for constrained, interacting, high-value units, and its industrial implementations are mature. System identification and state estimation are routine engineering tools67. Robust control provides the language for uncertainty, and data-driven methods are moving from research into pilot applications. The questions still open are the old ones in new form: how to guarantee safety and stability when the model is learned rather than derived; how to make the layers of the hierarchy agree; and how to control plants that are increasingly integrated, increasingly variable in feed and energy supply, and increasingly expected to run themselves.

References

  1. Industrial process control (introduction, Hierarchy), Wikipedia
  2. Control theory (introduction, History), Wikipedia
  3. PID controller (Origins, introduction), Wikipedia
  4. Control engineering (History), Wikipedia
  5. Model predictive control, Wikipedia
  6. System identification, Wikipedia
  7. Kalman filter (introduction, History), Wikipedia
  8. Relative gain array, Wikipedia
  9. Robust control, Wikipedia
  10. Computational intelligence (History), Wikipedia
  11. Machine learning control, Wikipedia

The frameworks of Process Control

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1868-Present

    Classical Feedback Control Theory

    When does a feedback loop become unstable?

    Fly to its tower in Nousfarer

  2. 1922-Present

    PID controller

    How should the three control terms be set?

    Fly to its tower in Nousfarer

  3. 1950-1990

    Frequency-Response Analysis

    How does a loop respond to fast and slow disturbances?

    Fly to its tower in Nousfarer

  4. 1960-Present

    State estimation and inferential control

    How can unmeasured quantities be inferred and controlled?

    Fly to its tower in Nousfarer

  5. 1965-Present

    System identification

    How can a dynamic model be fitted from plant data?

    Fly to its tower in Nousfarer

  6. 1966-Present
  7. 1978-Present

    Model predictive control

    What inputs keep future outputs on target under constraints?

    The 13 concepts its tower teaches
    • Constraint handling
    • Control horizon
    • Cost function
    • Disturbance model
    • Feedback correction
    • Feedforward
    • Optimization problem
    • Prediction horizon
    • Prediction model
    • Receding horizon
    • Setpoint tracking
    • System identification
    • Tuning

    Fly to its tower in Nousfarer

  8. 1980-Present

    Control hierarchy

    How should control tasks be layered from regulation to planning?

    Fly to its tower in Nousfarer

  9. 1981-Present

    Robust control

    Will a controller stay stable when the model is wrong?

    Fly to its tower in Nousfarer

  10. 1982-2000

    Internal model control

    How does a plant model set the performance-robustness trade-off?

    Fly to its tower in Nousfarer

  11. 2010-Present

    Data-driven and learning-based control

    Can controllers learn behavior that models cannot capture?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Classical Feedback Control TheoryPID controllerFrequency-Response AnalysisState estimation and inferential cont…System identificationMultivariable control-structure designModel predictive controlControl hierarchyRobust controlInternal model controlData-driven and learning-based control

Beyond the field

Where Process Control's ideas came from in other fields, and what they led to there: the strongest 12 of 56 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Process Control is a planet in Nousfarer

Its city has a tower for each of these 11 frameworks. Land, walk in, and climb one concept at a time.

Land on Process Control

Related fields

Fields its ideas are linked to

More in Chemical Engineering

See where Process Control sits on the Map of Human Knowledge