⚡ electrical engineering

Artificial-reference tracking MPC with probabilistically validated performance on industrial embedded systems

This paper presents an efficient, feature-rich Model Predictive Control (MPC) implementation for industrial embedded systems that utilizes a structure-exploiting ADMM algorithm with offset-free tracking and constraint handling, validated through a probabilistic framework on a hardware-in-the-loop setup controlling a nonlinear reactor.

Original authors: Victor Gracia, Pablo Krupa, Filiberto Fele, Teodoro Alamo

Published 2026-06-09

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Victor Gracia, Pablo Krupa, Filiberto Fele, Teodoro Alamo

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Overworked Chef" in a Tiny Kitchen

Imagine a factory as a busy restaurant kitchen. The goal is to cook a perfect meal (control a machine) while following strict rules: don't burn the food (safety limits), don't run out of ingredients (resource limits), and make sure the dish tastes exactly as the customer ordered (tracking a target).

Usually, the "chefs" in these industrial kitchens are simple, reliable robots (like PID controllers) that follow basic recipes. They are safe, but they aren't very good at handling complex situations, like sudden changes in the order or when the oven temperature fluctuates unexpectedly.

Advanced chefs (called Model Predictive Control or MPC) are much smarter. They look ahead, simulate different cooking paths, and choose the best one. However, these advanced chefs usually need a supercomputer to do their math. Putting a supercomputer in a tiny, low-power industrial robot (like a PLC) is like trying to fit a full-sized industrial oven into a microwave; it just doesn't fit or work fast enough.

This paper solves that problem. The authors created a way to put a "smart chef" into a "microwave-sized" industrial robot. They did this by making the math super efficient and adding safety nets so the robot doesn't crash when things go wrong.

The Three Secret Ingredients

To make this smart chef work on a tiny computer, the authors added three special features to their recipe:

1. The "Magic Reference" (Artificial Reference)

The Problem: Sometimes, the customer orders a dish that is impossible to make with the current ingredients (e.g., "Make a steak, but we only have tofu"). A normal smart chef would get stuck, panic, and stop cooking because the goal is unreachable.
The Solution: The authors gave the chef a "Magic Reference." Instead of demanding the impossible dish immediately, the chef says, "Okay, I can't make that exact steak right now, but I will make the closest possible tofu dish that looks and tastes like a steak."
This allows the system to keep working smoothly even when the target is impossible to reach perfectly. It finds the best possible compromise without getting stuck.

2. The "Soft Safety Belts" (Soft Constraints)

The Problem: In a real factory, things are messy. Sensors lie, machines vibrate, and the weather changes. If a chef is told, "Never let the temperature go above 100 degrees," and a sudden gust of heat pushes it to 101, a rigid chef might freeze because it broke a rule.
The Solution: The authors introduced "Soft Safety Belts." Instead of a hard wall that says "STOP," they use a bouncy rubber band. If the temperature hits 101, the chef feels a little "pain" (a penalty in the math) but keeps cooking. This ensures the robot never stops working, even when the world gets messy. It keeps the system running smoothly rather than crashing.

3. The "Safety Buffer" (Back-off Parameters)

The Problem: Even with soft belts, you don't want to get too close to the edge of a cliff. If the cliff is 100 degrees, you don't want to aim for 99.9 degrees because a tiny error could push you over.
The Solution: The authors added a "Safety Buffer." They tell the chef, "The limit is 100, but pretend the limit is 95." This creates a cushion. If the machine wobbles, it stays within the safe zone. The paper explains how to calculate exactly how big this cushion needs to be so you aren't too cautious (wasting performance) but also not too reckless.

The "Crystal Ball" Test (Probabilistic Validation)

One of the biggest challenges in engineering is: "How do we know this will work in the real world?"

Usually, engineers test a controller a few times. If it works, they ship it. But what if the one time it fails is the day the factory goes bankrupt?

The authors created a "Crystal Ball" testing method.

Instead of testing the robot 10 times, they simulated it running 1,156 times in a computer simulation.
In every simulation, they changed the weather, the ingredients, and the customer orders randomly.
They used a special math trick to say: "We are 99.9999% sure that in the real world, this robot will break the safety rules less than 3% of the time, and it will finish its math calculations fast enough to keep up."

It's like testing a parachute by jumping out of a plane 1,000 times in a simulator. If it works 997 times, you can be statistically confident it will work for the real jump.

The Real-World Test: The Stirring Pot

To prove their idea works, the authors didn't just use theory. They built a real test:

The Machine: A "Continuous Stirred-Tank Reactor" (CSTR). Think of this as a giant, high-tech pot where chemicals are mixed and heated to create a product. It's non-linear and tricky to control (like trying to keep a pot of boiling soup from splashing over while adding ingredients).
The Brain: They put their new smart algorithm onto a Siemens S7-1500 PLC. This is a standard, rugged industrial computer used in factories worldwide. It has very limited memory and speed compared to a laptop.
The Result: The "microwave-sized" computer successfully controlled the "giant pot." It kept the temperature and chemical levels within safe limits, handled disturbances, and did all the math in under 2 seconds (while the system updates every 75 seconds).

Summary

This paper is about taking a very smart, complex control strategy and shrinking it down so it can fit inside a standard, low-cost industrial robot.

They made the math fast using a special algorithm (ADMM) that exploits the structure of the problem.
They made it robust by allowing the robot to handle impossible goals and messy real-world data without crashing.
They proved it works by running thousands of virtual tests to statistically guarantee that the robot will be safe and fast enough for real-world use.

The result is a bridge between high-level academic control theory and the gritty reality of factory floors, showing that advanced control is finally ready for the "microwave" computers that run our industries.

Technical Summary: Artificial-Reference Tracking MPC with Probabilistically Validated Performance on Industrial Embedded Systems

Problem Statement
Industrial embedded systems, such as Programmable Logic Controllers (PLCs), are typically constrained by limited computational and memory resources, leading to the widespread use of simple control laws like PID. While Model Predictive Control (MPC) offers superior performance in handling constraints and optimizing operations, its implementation on such hardware is challenging. Existing solutions often rely on simplified linear formulations or explicit controllers to ensure fast computation, but these lack robustness against real-world issues like plant-model mismatch, external disturbances, and the need to track non-admissible references. Furthermore, selecting controller parameters (e.g., constraint back-offs and soft constraint weights) to guarantee safety and performance in uncertain environments is non-trivial, as traditional robust methods can be overly conservative or computationally demanding.

Methodology
The paper proposes a comprehensive framework for deploying an advanced Model Predictive Control for Tracking (MPCT) scheme with an artificial reference on industrial embedded hardware. The methodology integrates three main components:

Enhanced MPCT Formulation:
- Artificial Reference: The controller utilizes an artificial reference as a decision variable, allowing the system to track references that are not admissible or steady states of the model, thereby enlarging the domain of attraction.
- Offset-Free Tracking: A state-disturbance estimator (Luenberger observer) is integrated to compensate for plant-model mismatch and constant additive disturbances, ensuring zero steady-state offset.
- Constraint Handling: To ensure recursive feasibility and robustness, the formulation incorporates:
  - Back-off Parameters: These tighten the model constraints to account for disturbances and estimation errors, ensuring the real system satisfies its physical limits.
  - Soft Constraints: Inequality constraints are relaxed using an exact penalty approach. This guarantees that the optimization problem remains feasible even under severe disturbances, while penalizing violations in the cost function.
- Joint Constraints: The formulation explicitly handles coupled state-input constraints.
Efficient Solver Implementation:
- The optimization problem is solved using a structure-exploiting Alternating Direction Method of Multipliers (ADMM) algorithm.
- The solver leverages the specific semi-banded structure of the MPCT problem to achieve low memory footprints and fast computation times, making it suitable for PLCs.
- A warm-start strategy is employed, initializing variables based on the previous time step's solution to accelerate convergence.
- The algorithm is implemented using standard IEC 61131-3 programming languages on a Siemens S7-1500 PLC.
Probabilistic Validation Framework:
- To address the difficulty of parameter tuning, the authors extend a probabilistic validation method (based on Karg et al., 2021).
- This framework allows for the selection of controller parameters (back-offs and soft constraint weights) by running a set of $N_s$ closed-loop experiments with random uncertainty realizations.
- It provides probabilistic guarantees (with confidence $1-\delta$ ) that the probability of performance metrics (e.g., constraint violations, solver iterations) exceeding a specific threshold is bounded by $\epsilon$ . This enables the selection of a controller that balances constraint satisfaction with computational efficiency without requiring explicit worst-case uncertainty bounds.

Key Contributions

Practical MPCT for Embedded Systems: The article presents a tailored MPCT formulation that includes essential industrial features (offset-free control, constraint tightening, and soft constraints) while maintaining computational efficiency suitable for resource-constrained hardware.
Probabilistic Parameter Validation: A novel framework is proposed to validate and select controller parameters based on empirical data, offering probabilistic guarantees on long-term closed-loop performance regarding constraint violations and solver iteration counts.
Industrial Implementation: The approach is successfully implemented on a Siemens S7-1500 PLC within a Hardware-in-the-Loop (HiL) setup controlling a nonlinear Continuous Stirred-Tank Reactor (CSTR).

Experimental Results
The proposed scheme was tested on a nonlinear CSTR model with coupled state-input constraints and external disturbances.

Performance: The controller successfully tracked piecewise constant references. The probabilistic validation identified a set of back-off parameters ( $\eta$ ) and soft constraint weights ( $\beta$ ) that minimized constraint violations while keeping the number of ADMM iterations low.
Validation: With a confidence of $1-\delta = 99.9999\%$ and a probability bound of $\epsilon = 0.03$ , the selected controller (C1) was shown to violate constraints by no more than a small threshold (measured by $\phi_1 \le 0.0045$ ) in at least 97% of new experiments. The maximum number of ADMM iterations was bounded at 246.
Computational Efficiency: The solver achieved a maximum execution time of 1.98 seconds on the PLC, which is significantly shorter than the 75-second sampling time. The total memory usage for the estimator, solver, and reference calculator was 0.73 MB (3% of available memory).
Trade-offs: The results highlighted the trade-off between constraint satisfaction and tracking performance. Controllers with aggressive back-offs (like C1) ensured constraint adherence but sometimes resulted in steady-state offsets when the reference was not admissible. Conversely, controllers without back-offs (C0) achieved better tracking but violated constraints more frequently.

Significance and Claims
The paper claims to bridge the gap between academic MPC research and industrial practice. It demonstrates that advanced, practical MPC formulations can be deployed on standard industrial embedded systems (PLCs) with solution times in the order of seconds. By combining an efficient structure-exploiting solver with a probabilistic validation framework, the work provides a systematic way to configure controllers that are both computationally efficient and robust against real-world uncertainties. The authors emphasize that their approach allows for the "probabilistic certification" of constraint satisfaction and computational bounds, which are critical for safe and reliable real-time control applications.