MPC

What is MPC?

MPC (Model Predictive Control) is an advanced control method that uses a dynamic process model to predict the future behaviour of a process and, in every control cycle, calculates the sequence of control moves that minimises an objective function within all process limits, after which only the first move is applied. MPC is therefore a form of control within process automation that makes industrial processes run more stably and efficiently. Where a classical PID controller only reacts to a deviation, MPC looks ahead and drives several variables at once. That makes MPC the core technique of Advanced Process Control and a standard layer above the base-layer control in the DCS of refineries, chemical plants and power stations.

🧠 How does MPC work?

An MPC controller runs through the same five steps every cycle, known as the receding horizon principle:

1. Measure - the current process values are read from the DCS or the historian
2. Estimate - the difference between the previous prediction and the measurement corrects the model (a bias update or a Kalman filter)
3. Predict - the model calculates how the process will behave over the prediction horizon
4. Optimise - a solver finds the sequence of future control moves that minimises the objective function without violating any limits
5. Apply - only the first control move is sent to the process; the next cycle starts afresh

Because each calculation starts again with fresh measurements, the result is a closed loop that absorbs disturbances and model errors. MPC works with three kinds of variables: manipulated variables (MVs, the handles MPC may move, usually setpoints of PID loops, see Manipulated Variable), controlled variables (CVs, which must stay on target or within limits) and disturbance variables (DVs, measured disturbances that MPC can anticipate but not control).

🧮 What mathematics lies behind MPC?

Linear MPC usually relies on a discrete state-space model:

$x_{k+1}=A{x}_k+B{u}_k,y_k=C{x}_k$

Here $x$ is the process state, $u$ the vector of manipulated variables (MVs) and $y$ the measured outputs (CVs). The controller then minimises an objective function over the horizon:

$J={\sum }_{i=1}^p\Vert y_{k+i}-r_{k+i}{\Vert }_Q^2+{\sum }_{i=0}^{m-1}\Vert \Delta u_{k+i}{\Vert }_R^2$

subject to $u_{min}\le u\le u_{max}$, $\Delta u_{min}\le \Delta u\le \Delta u_{max}$ and $y_{min}\le y\le y_{max}$. The first term penalises deviations from the setpoint $r$, the second penalises aggressive control moves. With a linear model and a quadratic objective this is a quadratic program (QP) that a modern solver handles within milliseconds at typical industrial sizes. Older industrial packages such as Dynamic Matrix Control use a step-response model instead of a state-space model, derived directly from step tests on the plant.

🔄 What is the difference between MPC and PID?

MPC does not replace PID; it directs it. The PID loops in the DCS keep pressure, flow and temperature on setpoint quickly, while MPC adjusts those setpoints every minute. Classical extensions such as cascade and feedforward control solve part of the problem; MPC pays off once many variables interact or the process has to run close to its limits.

🏗️ What types of MPC are there?

Models are increasingly supplemented with machine learning, for example as a soft sensor for a product quality that is hard to measure. A digital twin then serves as a test environment to validate the controller offline. This links MPC to Industrial AI, although the optimisation core usually remains a classical, explainable model.

📡 Where does MPC run in the OT architecture?

In the Purdue Model the MPC or APC server usually sits on level 2, or on the boundary with level 3, close to the DCS, SCADA and the historian. The server reads process data via OPC UA or older OPC interfaces and writes new setpoints back to the PID loops on level 1. Data for consumers outside the DCS can also be published via MQTT; transport runs over Industrial Ethernet, where QoS gives control traffic priority. In the ISA-95 architecture MPC belongs to process control, below the manufacturing operations layer on level 3. A watchdog in the DCS monitors the link: if the MPC server fails, the loops fall back to their last setpoint and the base-layer control carries on independently.

Not every MPC runs on a separate server. In PCS 7, Siemens offers the ModPreCon function block for small multivariable controllers with up to four MVs and CVs plus one measured disturbance, and since PCS 7 V8.1 an MPC block of up to 10×10. For fast applications such as motion control, drives and power electronics, MPC is embedded with cycle times in milliseconds or microseconds, on an RTOS or a soft PLC. Only there do real-time behaviour, latency, jitter and TSN really matter; process-industry MPC with a one-minute cycle is not time-critical.

🏭 Where is MPC used?

MPC pays off mainly in continuous processes with high throughput, long dead times and strong interaction between variables:

• Refining and petrochemicals - crude distillation, catalytic cracking and ethylene furnaces; this is where MPC originated (IDCOM in 1978 and Shell’s Dynamic Matrix Control in 1979)
• Chemicals and polymers - reactor temperature, product quality and transitions between grades
• Energy - boilers, steam networks, combined heat and power plants and battery storage
• Water management - pumping stations, discharge sluices and water treatment
• Buildings and HVAC - energy-efficient climate control using weather forecasts
• Maintenance - the same process models feed predictive maintenance, for instance by detecting heat-exchanger fouling from a changing model gain

In the Netherlands MPC is common in the refineries and chemical clusters around Rotterdam, where installations often fall under the Seveso regime (Brzo). A less well-known Dutch example is water management: Deltares developed RTC-Tools, an open-source package (now written in Python) for model predictive control of canals, polders and reservoirs. The Noorderzijlvest water board used it to build an MPC within Delft-FEWS, coupled to its SCADA system, for the pumping stations and tidal sluices in its district; it now supports the daily operation of the entire water system. There MPC pre-releases water ahead of forecast rainfall and shifts pumping to hours with lower energy prices. See also Water Sector OT Security.

🔧 How do you run an MPC project step by step?

1. Business case and baseline - define which KPI must improve (energy per tonne, throughput, quality variation) and measure the starting point
2. Fix the base layer - tune the PID loops and repair sticking control valves; MPC on top of poor base-layer control does not work
3. Process identification - run step tests on every MV and derive the dynamic models from them
4. Configuration and simulation - choose sample time, horizons, weights and limits, and test the controller offline against the model or a digital twin
5. Commissioning - first run the controller in prediction mode without acting, then switch on MV by MV
6. Sustainment - monitor the service factor (the share of time MPC is actually in control) and re-identify after process changes; every model change belongs in change management

🔐 How do you secure an MPC system?

An MPC server is allowed to write setpoints for an entire production process, which makes it an attractive target within OT cybersecurity. Anyone who manipulates the server or the process data can quietly push the process to an unfavourable or unsafe operating point.

Place the APC server in its own zone according to the zones and conduits model of IEC 62443, with network segmentation, an industrial firewall and monitoring with an IDS. Define hard limits in the DCS that MPC may never exceed, so that safety limits are enforced independently of the optimiser. The safety functions themselves belong in an independent SIS that stays out of MPC’s reach. For Seveso sites this falls under the supervision described in Seveso and Cybersecurity.

❓ Frequently asked questions

What does the abbreviation MPC stand for?

MPC stands for Model Predictive Control. The name describes how it works: a process model predicts future behaviour and the controller uses that prediction to choose the best control moves.

Is MPC the same as APC?

No, MPC is the most important technique within APC, but not the only one. Advanced Process Control also covers feedforward control, soft sensors, decoupling and real-time optimisation. In practice a multivariable MPC controller is often simply called “the APC”.

Can MPC run on a PLC?

Small MPC controllers can run in a PLC or DCS controller, for example with the MPC function blocks in Siemens PCS 7 or with explicit MPC, where the solution is precomputed as a table. Large multivariable controllers with dozens of variables usually run on a dedicated APC server next to the DCS.

What happens if the MPC server fails?

When the MPC server fails, a watchdog in the DCS notices that no new setpoints are arriving. The PID loops then fall back to local control at their last setpoint, so the process keeps running safely, just less optimised. This fallback behaviour should be tested during the commissioning of every MPC.

How much does MPC deliver?

The benefit of MPC comes from reduced process variation, which lets the plant run closer to limits such as capacity or product specification. How much that is worth depends strongly on the process and on the quality of the existing base-layer control. A baseline beforehand and monitoring of the service factor afterwards make the benefit measurable.

How often does an MPC model need updating?

An MPC model must be re-identified whenever the process changes significantly, for example after a revamp, a valve replacement or a different feedstock. Declining control performance and a low service factor are signs that the model no longer fits. Without this maintenance, operators tend to switch MPC off over time.

📌 In summary

MPC is a predictive, multivariable control method that uses a process model to calculate the optimal setpoints within all process limits every cycle and passes them to the PID base-layer control. MPC usually runs on level 2 of the Purdue model next to the DCS, delivers most value in continuous processes such as refining, chemicals and water management, and requires sound model management and a secure, segmented place in the OT Architecture of industrial automation. Within OT convergence and Industry 4.0, MPC is gaining ground as data, computing power and models become easier to access.