Model Predictive Control Fundamentals

Model Predictive Control (MPC) represents a powerful advanced control strategy that excels where traditional methods struggle: managing complex, multivariable systems with strict limitations on both control actions and process outputs. Unlike a simple thermostat that reacts to current error, MPC uses a dynamic model to peer into the future, calculating not just the next move but a sequence of optimal moves, all while respecting real-world constraints. This makes it indispensable in industries like chemical processing, autonomous vehicles, and energy management, where safety, efficiency, and performance are governed by hard physical and operational limits.

The Core Principle: Predict, Optimize, Repeat

At its heart, Model Predictive Control (MPC) is a control algorithm that uses an explicit internal process model to predict the system's future behavior over a defined time window, known as the prediction horizon. This predictive capability is its defining feature. At each control interval or sampling instant, the algorithm solves an online optimization problem. It computes a sequence of future control moves (over a control horizon) that minimizes a cost function—typically penalizing deviations from a setpoint and excessive control effort—while satisfying all specified constraints. Only the first control action of this optimized sequence is implemented. Then, the horizon shifts forward by one time step, new measurements are taken (providing crucial feedback to correct for model inaccuracy and disturbances), and the entire optimization is repeated. This cycle is known as the receding horizon strategy.

Mathematical Formulation: The Optimization Problem

The power of MPC is codified in a mathematical optimization problem solved at time step $k$. The core elements are the model, the cost function, and the constraints.

First, a discrete-time model of the system is required. This could be a state-space model: $x(k+1)=Ax(k)+Bu(k)$, $y(k)=Cx(k)$, where $x$ is the state, $u$ is the input, and $y$ is the output. The MPC controller uses this model to predict future states and outputs over the prediction horizon $N_p$.

The optimization problem is then stated as: $$\underset{u(k|k),...,u(k+N_c-1|k)}{\min }\sum _{i=1}^{N_p}||y(k+i|k)-r(k+i)|{|}_Q^2+\sum _{i=0}^{N_c-1}||\Delta u(k+i|k)|{|}_R^2$$ subject to: $$u_{min}\le u(k+i|k)\le u_{max}$$ $$y_{min}\le y(k+i|k)\le y_{max}$$ $$\Delta u_{min}\le \Delta u(k+i|k)\le \Delta u_{max}$$

Here, $y(k+i|k)$ is the predicted output at future time $k+i$ made at current time $k$, and $r$ is the reference trajectory. $N_c$ is the control horizon. The notation $||\cdot |{|}_Q^2$ denotes a weighted squared norm, where $Q$ and $R$ are tuning matrices that balance tracking performance against control effort. The term $\Delta u(k+i|k)=u(k+i|k)-u(k+i-1|k)$ represents the change in the control action, and penalizing it leads to smoother control.

The Receding Horizon: Feedback and Robustness

The receding horizon strategy is what transforms a purely predictive, open-loop optimization into a closed-loop feedback control law. After solving the optimization, the controller applies only $u^{\ast }(k|k)$, the first element of the optimal control sequence, to the actual plant. At the next time step $k+1$, a new measurement (or state estimate) is obtained. This measurement provides critical feedback that accounts for three key realities: unmeasured disturbances affecting the plant, inherent model inaccuracy, and any changes in the reference signal. The controller then shifts its entire prediction window forward, initializes the new optimization with the latest measured state, and repeats the process.

This repeated re-optimization is the source of MPC's robustness. Even with a moderately accurate model, the feedback introduced at each step constantly corrects the predicted trajectory, preventing the controller from blindly following a plan that diverges from reality. It turns a sequence of open-loop optimizations into a resilient closed-loop policy.

Explicit Constraint Handling: MPC's Superpower

A primary advantage of MPC is its ability to handle constraints on inputs and outputs directly and explicitly within the optimization problem. This is a fundamental departure from classical control design, where constraints are often dealt with ad hoc (e.g., through anti-windup schemes) after the controller is designed.

• Input Constraints: These are physical limits on the actuator, such as a valve that cannot open more than 100% or less than 0% ($u_{min}\le u\le u_{max}$), or a motor that has a maximum rate of change ($\Delta u_{min}\le \Delta u\le \Delta u_{max}$).
• Output (or State) Constraints: These are safety or quality limits on the process itself, such as a reactor temperature that must not exceed a safe limit ($T\le T_{max}$) or a product concentration that must stay within a purity specification.

The optimizer inherently finds the best possible control sequence that keeps the process within these "hard" or "soft" boundaries. For example, if a disturbance pushes a temperature towards its maximum limit, the MPC controller will proactively adjust other variables (like cooling flow) well in advance to avoid violating the constraint, even if it means temporarily allowing a small deviation in another, less critical output.

The On-Line Optimization Engine

The need to solve a constrained optimization problem in real-time, at every sampling instant, is the main computational demand of MPC. For linear models with quadratic cost functions and linear constraints (Linear Quadratic MPC), the problem simplifies to a Quadratic Program (QP), for which extremely efficient solvers exist. The QP takes the standard form: $$\underset{z}{\min }\frac{1}{2}{z}^{T}Hz+f^{T}z\text{subject to}{A}_{ineq}z\le b_{ineq},A_{eq}z=b_{eq}$$ where $z$ is a vector containing the sequence of future control moves. The matrices $H$, $f$, $A_{ineq}$, $b_{ineq}$, $A_{eq}$, and $b_{eq}$ are constructed at each time step from the model, current state, references, and constraints. The rapid, reliable solution of this QP within the available sampling time is what makes modern MPC feasible for fast processes.

Common Pitfalls

1. Ignoring Model Quality: MPC's performance is directly tied to the accuracy of its internal model. A poor model leads to poor predictions, forcing the feedback mechanism to work excessively hard and resulting in sluggish or oscillatory control. Correction: Invest in systematic model identification. Use plant test data to develop and validate the predictive model before controller implementation.

1. Overly Aggressive Tuning: Setting the cost function weights ($Q$ and $R$) to demand extremely fast setpoint tracking can push the controller against its constraints constantly, leading to highly aggressive control actions and marginal stability. Correction: Tune conservatively. Increase the penalty on control effort ($R$) to smooth actuator movement. Widen the prediction horizon ($N_p$) to give the controller a longer-term perspective, promoting more stable, planning-based behavior.

1. Infeasible Optimization Problems: Defining constraints that are too tight or contradictory can lead to a situation where the optimizer finds no solution that satisfies all limits. For example, demanding a setpoint change that is physically impossible without violating an output constraint. Correction: Implement constraint softening or prioritize constraints. Critical safety limits can be kept as "hard" constraints, while performance-related limits can be made "soft" (allowing small, penalized violations) to ensure the optimization problem always has a solution.

1. Mismatched Time Scales: Choosing a sampling time that is too slow misses important process dynamics, while one that is too fast creates an unnecessary computational burden and can amplify measurement noise. Correction: Select a sampling time based on the dominant dynamics of the process. A good rule of thumb is to sample 4-10 times per rise time of the fastest important dynamic.

Summary

• Model Predictive Control is a receding horizon optimization technique that uses a dynamic model to predict future process behavior and computes optimal control actions at each time step.
• Its defining feature is the explicit, direct handling of constraints on both inputs and outputs within the online optimization problem, allowing for safe and optimal operation at process limits.
• The receding horizon strategy—applying only the first control move and then re-optimizing with new feedback—provides robustness against model mismatch and unmeasured disturbances.
• For linear systems with quadratic costs, the core online calculation simplifies to solving a Quadratic Program (QP), a well-understood and computationally tractable optimization problem.
• Successful implementation hinges on a reasonably accurate model, careful tuning of horizons and weights, and a thoughtful approach to constraint management to avoid infeasibility.