The Humble PID Controller That Can Do What Engineers Thought Required AI

At minus 40 degrees Celsius, the heater in a barn full of cows maxes out. The fan slows to keep warm air in. CO2 starts climbing past 1,000 parts per million — the threshold for good air quality. Then past 2,000. At 3,000 ppm, the control system makes a final, automatic decision: the cows are better off breathing stale air than freezing. No engineer intervenes. No optimization algorithm solves a differential equation. A handful of classical PID controllers, arranged with deliberate logic, navigate the entire crisis.

That scenario — drawn from real simulations in a new paper by Sigurd Skogestad of the Norwegian University of Science and Technology — is a provocation. For roughly 40 years, the dominant view in academic control engineering has been that managing shifting, conflicting constraints requires sophisticated model-based tools: model predictive control (MPC), which uses a dynamic mathematical model of a process to continuously solve an optimization problem, or real-time optimization (RTO), which does the same at a slower timescale. These tools are powerful. They are also expensive to build, brittle when the model drifts from reality, and opaque to the operators who run industrial plants day to day. Skogestad (2026) argues — with worked examples and simulations — that a much older, simpler approach has been hiding in plain sight.

The Science

The approach is called Advanced Regulatory Control (ARC), and it is built from elements that have existed in industrial plants for more than 60 years. The two key building blocks are selectors and split-parallel control.

A selector is a logic element that picks between competing controller outputs. A MAX-selector passes whichever of its inputs is largest to the actuator; a MIN-selector passes whichever is smallest. In practice, this means a single valve or motor can be "claimed" by whichever controller currently has the most urgent constraint to enforce — and released back to normal duty when the crisis passes. A split-parallel controller handles the opposite problem: two actuators competing to control the same variable, each given a slightly different setpoint so they hand off smoothly without fighting each other.

Neither concept is new. Maarleveld and Rijnsdorp described override schemes for industrial distillation columns as far back as 1970. What Skogestad's paper does is systematize them — articulating formal rules for when to use MAX versus MIN selectors, how to order them in a hierarchy when constraints conflict, and how to set the setpoint offsets in split-parallel schemes to avoid instability. These rules are then stress-tested against two case studies drawn from genuinely complex industrial situations: an offshore oil and gas separator handling a two-phase hydrocarbon stream, and a ventilated barn where cow welfare depends on balancing air quality against warmth.

The barn case is deliberately approachable — most people have intuitions about temperature and fresh air. But it is not a toy problem. It has three constraints, two manipulated variables, and conflict scenarios that can cascade in non-obvious ways. The oil and gas case is straightforwardly industrial: seven interacting constraints, five potential bottlenecks, and the economic imperative to maximize production throughput at all times.

What They Found

The gas-liquid separator: making the bottleneck find itself

In oil and gas production, a separator tank takes a mixed stream from a well and splits it into gas (compressed by a compressor and piped away) and liquid (pumped out through a valve). The throughput manipulator — the TPM, the single variable that sets the production rate for the whole system — is traditionally either the feed valve (how fast fluid comes from the well) or the compressor (how fast gas leaves). Operators set it manually.

The problem: if you're trying to maximize production, the real limit shifts constantly. Sometimes it's the well pressure (push too hard and you damage the reservoir — a minimum pressure constraint of 170 bar in this study). Sometimes it's the compressor hitting maximum speed. Sometimes it's the liquid outlet valve fully open. Whichever constraint is active at this moment is the bottleneck, and ideally the TPM should be located there — automatically.

Skogestad shows that combining MIN-selectors with split-parallel pressure controllers (PCA and PCB, with setpoints of 70 bar and 71 bar respectively — a deliberate 1-bar gap) achieves exactly this. When the compressor saturates at 100% speed, the separator pressure starts to rise past 70 bar. PCB wakes up and begins throttling the feed valve. The TPM has migrated, by feedback, to the feed. When the compressor constraint relaxes, the process reverses. No model. No optimizer. The architecture itself encodes the priority logic.

The full scheme, shown in

, handles seven simultaneous constraints: minimum reservoir pressure, maximum compressor speeds for two compressors, maximum liquid valve opening, minimum and maximum pressures in the gas export pipeline. Each constraint maps to exactly one controller; each controller can claim or release its actuator through a MIN-selector. The structure repeats — every valve and compressor has a MIN-selector fed by an upstream L-setpoint, a downstream H-setpoint, and a desired operating value. Once you see the pattern, the diagram becomes readable.

A key design insight: setting the split-parallel setpoint gap $\Delta$ larger gives the system an economic buffer. While separator pressure drifts between SPL and SPH (neither controller active), gas or liquid accumulates in the tank — buying time before the feedrate must be cut. A larger $\Delta$ delays that cut, which has direct production value.

The barn: hierarchical constraint management for happy cows

The barn problem has a single manipulated variable — fan speed $u_1$, ranging from 0% to 100% — and three constraints: CO2 concentration $c\le 1000$ ppm (satisfied by a faster fan, which flushes air), temperature $T \geq 5°$C (satisfied by a slower fan, which retains heat), and temperature $T \leq 20°$C (satisfied by a faster fan again).

The first two constraints can pull in opposite directions. On a cold day, the CO2 controller wants to speed up the fan; the low-temperature controller wants to slow it down. These are genuinely conflicting. Skogestad's Selector Rule S2 says: when conflict is possible, the higher-priority constraint must be "at the end" of the selector chain — physically closest to the actuator — so it always wins. Here, animal health (CO2) outranks comfort (temperature), so the MAX-selector for CO2 sits last in the chain.

The result is a clear priority ordering that works at steady state across a wide range of outdoor temperatures, confirmed by simulation. The data are striking.

Barn Control: Fan Speed & CO₂ vs. Outdoor Temperature

Label	Value
15°C	77.2 %
10°C	50 %
5°C	50 %
0°C	50 %
-2.5°C	47.6 %
-5°C	45.6 %
-10°C	45.6 %
-20°C	24.4 %

Source: Skogestad (2026), Table 2 simulation results. CO₂ limit: 1,000 ppm (comfort), 3,000 ppm (health ceiling).

At $T_{out} = 15°$C, the fan runs at 77.2% and CO2 sits comfortably at 765 ppm — well under the 1,000 ppm limit. As it gets colder, the fan throttles back to protect warmth; CO2 rises but stays controlled at 950 ppm with the fan at 50%. Below about $-2.5°$C, the fan drops below the CO2 constraint threshold: at $-5°$C, CO2 hits exactly 1,000 ppm (the active constraint) and the heater switches on. By $-10°$C, the heater is maxed out at 100%. Below $-20°$C, the CO2 limit must be sacrificed to prevent freezing — the system allows CO2 to rise to 1,492 ppm, then 2,487 ppm at $-30°$C, and finally the absolute ceiling of 3,000 ppm at $-40°$C, where temperature drops to $-6.4°$C.

Heater Activation vs. Outdoor Temperature

Label	Value
15°C	0 %
10°C	0 %
5°C	0 %
0°C	0 %
-2.5°C	0 %
-5°C	25.7 %
-10°C	100 %
-20°C	100 %

Source: Skogestad (2026), Table 2 simulation results. Heater rated at 50,000 W at 100%.

This is not a failure mode. It is correct behavior: the system has correctly identified that at extreme cold, keeping cows above 0°C is more important than keeping CO2 below 1,000 ppm, but keeping CO2 below 3,000 ppm is more important than temperature once the heater is maxed. The hierarchy, encoded in selector ordering, produces exactly the intended trade-offs.

The full Case Study IIB scheme uses six PI controllers and three selectors. Each controller handles exactly one constraint. Anti-windup (a technique to prevent integral action from accumulating when a controller is not in command) is applied to all controllers that might be dormant. The barn is modeled as a well-mixed volume of $V=3,000{\mathrm{m}}^3$ with 80 cows, each generating $G_{\mathrm{C}{\mathrm{O}}_2}=5\times 10^{-5}{\mathrm{m}}^3/s$ of CO2 and $Q_{\mathrm{cow}}=1,000\mathrm{W}$ of heat. The full dynamics are captured by two coupled differential equations governing CO2 concentration and temperature, driven by fan airflow $q_{\mathrm{air}}(u_1)$ and heater power $Q_{h,\max }\cdot u_2$.

Why This Changes Things

The significance here goes beyond barns and oil wells. MPC has become the default recommendation in academic control curricula for any problem involving constraints. It is genuinely powerful — when you have a good model. But building and maintaining a dynamic process model is expensive. Models go stale as equipment ages. MPC solvers can fail silently. And for operators in a control room at 3 a.m., a black-box optimizer offers little intuition when something behaves unexpectedly.

The ARC approach is transparent. Every controller has a clear job. Every selector has an explicit priority. An operator can trace why the system is doing what it is doing by following the selector chain. Industrial practitioners have known this for decades — Skogestad notes that selector-based schemes have been used reliably in industry for over 60 years (Maarleveld and Rijnsdorp, 1970; Smith, 2010) — but the academic literature has largely ignored them, treating them as unprincipled hacks rather than systematic engineering.

What this paper adds to that industrial knowledge is systematization. The formal rules — particularly Selector Rules S1, S2, and S3 for ordering MAX and MIN selectors, and the guidance on split-parallel setpoint gaps $\Delta$ — give engineers a principled framework for designing these schemes from scratch. The radiation rule for inventory control (

), which says that inventory loops must propagate outward from the TPM in both directions, provides the geometric intuition for why the bidirectional scheme works and what makes it consistent.

The economic stakes are real. In oil and gas production, maximizing throughput while respecting equipment constraints is worth significant money per hour. A system that automatically finds the current bottleneck and tightly controls it — with minimum "back-off" from the constraint boundary — directly translates to production volume. The same logic applies to any process where capacity is the binding limit.

For climate and sustainability applications, the barn case hints at broader relevance. Agricultural buildings, data centers, commercial HVAC systems, and battery storage facilities all face analogous multi-constraint problems: energy efficiency versus air quality versus temperature, with constraints that shift by season, occupancy, and weather. ARC schemes could offer a path to intelligent, adaptive control of these systems without the model-development overhead that MPC demands.

What's Next

Skogestad is careful not to oversell. The paper explicitly acknowledges cases where ARC does not apply. The switching schemes assume a one-to-one pairing between each constraint and a single manipulated variable. When you have more constraints than actuators — and many industrial systems do — this pairing becomes impossible, or you must accept that some constraints will simply be violated. Bernardino and Skogestad (2024) have begun formalizing when decentralized switching can guarantee optimal steady-state operation, but only under the restrictive assumption that there are at least as many inputs as constraints.

There are also genuine theoretical loose ends in the dynamics of switching systems. Under certain conditions — noisy measurements, tight constraint boundaries, fast disturbances — selector-based schemes can exhibit chattering (rapid switching that wears out equipment and destabilizes control) or even chaotic behavior (Caponigro et al., 2018). Anti-windup design, particularly the tracking scheme with ${\tau }_T={\tau }_I$, suppresses chattering in practice, and the simulations in this paper show smooth transitions throughout. But a complete theoretical characterization of stability for hierarchical selector networks does not yet exist. This may be the core reason academia has been reluctant to engage with these schemes — they are hard to analyze formally, even though they work reliably in practice.

The paper's framing is ultimately a call for reconciliation. "Today advanced PID solution (ARC) and model-based control and optimization (MPC/RTO) are in parallel universes," Skogestad (2026) writes, "but both solutions are needed in the future control engineer's toolbox." These are not competing philosophies; they are complementary tools. ARC excels when constraints are known, measurable, and can be prioritized; MPC earns its keep when the interactions are dense, the model is reliable, and the optimization landscape is complex. The problem is not that one is better — it is that one has been systematically underrepresented in how engineers are trained to think.

The Matlab/Simulink simulation files for both case studies are publicly available from the author, which means the next generation of control engineers can test these ideas directly. That openness matters. Forty years of neglect is a long time to leave a useful tool in the drawer.