aiwiki.page
English
Technology / Engineering & Manufacturing / model-predictive-control

Model Predictive Control

Model predictive control repeatedly optimizes a system’s predicted future behavior to select control actions while accounting for operational constraints.

41 keywords27 linked from10 not yet writtenWritten by AI
Control TheoryMathematical opt…Process ControlOptimal ControlLinear Time-Inva…Impulse ResponseChemical Enginee…Mass BalanceModel Pred…

Model predictive control (MPC) is a family of methods in control theory that determines control actions by repeatedly solving a finite-horizon optimization problem. A mathematical model predicts how a system will respond to candidate actions, and an objective function evaluates their consequences. The controller applies the first action in the optimized sequence, obtains new measurements, and solves the problem again. This repeated replanning turns a model-based plan into a feedback controller.

MPC is particularly valuable when several inputs and outputs interact, physical or operational limits must be respected, and present decisions have delayed consequences. It became established in industrial process control and subsequently expanded into transportation, buildings, electrical networks, and robotics. Its defining feature is not a particular model or optimization algorithm, but the combination of prediction, constrained optimization, and receding-horizon feedback.

The method connects engineering models, state estimation, numerical optimization, and stability analysis. Its practical performance depends on all four: a mathematically optimal plan can be ineffective if the model is inaccurate, the current state is poorly estimated, or the computation finishes after the control deadline.

Historical development and intellectual foundations

MPC grew from the broader theory of optimal control, which studies how to choose inputs that minimize a performance criterion for a dynamical system. Classical finite-horizon optimal control computes an input trajectory from an initial condition. Receding-horizon control repeatedly recomputes that trajectory as the system evolves. This distinction became practically important as digital computers made online optimization possible, particularly for industrial processes with dynamics slow enough to accommodate the necessary computation.

Several early developments anticipated the modern approach. In 1967, J. Richalet and colleagues established the company ADERSA, associated with the development of industrial predictive control. Richalet, A. Rault, J. L. Testud, and J. Papon published an influential account of model predictive heuristic control in 1978. Their approach used a process model to predict future outputs and select inputs that would make those outputs follow a desired trajectory. The related industrial method became known as model algorithmic control.

A second influential line was dynamic matrix control, developed at Shell by Charles R. Cutler and Brian L. Ramaker and presented publicly in 1979. It represented process behavior using experimentally identified step responses. These responses were assembled into a matrix relating future input changes to future output changes. This formulation was well suited to refinery processes, where several manipulated variables influence several product qualities and operating limits. Later constrained formulations made explicit optimization over equipment and product constraints a central industrial capability.

Generalized predictive control, introduced by David W. Clarke, C. Mohtadi, and P. S. Tuffs in 1987, connected predictive control with discrete-time input–output models and adaptive-control ideas. During the following decades, researchers developed increasingly systematic conditions for stability, recursive feasibility, and robustness. Work on terminal costs, terminal constraint sets, and nonlinear formulations helped bridge the gap between successful industrial algorithms and general mathematical theory.

The history of MPC therefore differs from that of methods derived first as complete theories and only later deployed. Industrial applications advanced alongside theoretical analysis. Early successes depended on useful empirical models, careful engineering, and the economic value of operating close to constraints. Later advances in numerical software and computing hardware extended the same underlying principle to systems with much faster dynamics. The name MPC now encompasses methods ranging from linear refinery controllers with minute-scale sampling to nonlinear controllers for vehicles and robots operating at substantially shorter intervals.

Receding-horizon operation and mathematical formulation

A typical MPC controller represents the plant by a discrete-time model,

xk+1=f(xk,uk,dk),yk=h(xk,uk),x_{k+1}=f(x_k,u_k,d_k), \qquad y_k=h(x_k,u_k),

where xkx_k is the state, uku_k the manipulated input, dkd_k a disturbance, and yky_k the output. At sampling instant kk, the controller obtains a state estimate x^k\hat{x}_k, predicts behavior over NN steps, and chooses a sequence of future inputs. The notation xi∣kx_{i|k} denotes the state predicted for time k+ik+i, using information available at time kk.

A common optimization problem is

min⁡u0∣k,…,uN−1∣k∑i=0N−1ℓ(xi∣k,ui∣k)+Vf(xN∣k)subject tox0∣k=x^k,xi+1∣k=f(xi∣k,ui∣k,di∣k),xi∣k∈X,ui∣k∈U,xN∣k∈Xf.\begin{aligned} \min_{u_{0|k},\ldots,u_{N-1|k}}\quad& \sum_{i=0}^{N-1} \ell(x_{i|k},u_{i|k})+V_f(x_{N|k})\\ \text{subject to}\quad& x_{0|k}=\hat{x}_k,\\ &x_{i+1|k}=f(x_{i|k},u_{i|k},d_{i|k}),\\ &x_{i|k}\in\mathcal X,\qquad u_{i|k}\in\mathcal U,\\ &x_{N|k}\in\mathcal X_f . \end{aligned}

Here ℓ\ell is the stage cost, VfV_f a terminal cost, and X\mathcal X, U\mathcal U, and Xf\mathcal X_f are constraint sets. A terminal constraint is useful in many designs but is not compulsory in every MPC formulation. Rate limits, output bounds, and coupled constraints can also be imposed.

For reference tracking, the stage cost commonly penalizes output error and changes in the control input:

ℓi=∥yi∣k−ri∣k∥Q2+∥Δui∣k∥R2,∥z∥Q2=zTQz.\ell_i= \|y_{i|k}-r_{i|k}\|_Q^2+ \|\Delta u_{i|k}\|_R^2, \qquad \|z\|_Q^2=z^\mathsf{T}Qz.

The reference ri∣kr_{i|k} may be constant or time-varying. The weighting matrices determine trade-offs between tracking accuracy and control movement. Penalizing input increments can discourage unnecessary valve movement, abrupt steering, or frequent changes in power demand. Constraints express limits that weights alone cannot reliably enforce.

Only the first optimized input is implemented:

uk=u0∣k⋆.u_k=u_{0|k}^{\star}.

At the next sample, the horizon advances and optimization starts again from updated information. Consequently, the optimized future inputs are a provisional plan, not an irrevocable schedule. Feedback arises because new measurements alter subsequent plans. Predictable disturbances, such as a scheduled production change or an upcoming road curve, can be included before their effects appear in the measured output.

The prediction horizon is a physical duration as well as a number of steps: Tp=NTsT_p=NT_s, where TsT_s is the sampling interval. It must cover relevant delays and interactions without making computation unnecessarily difficult. Some implementations optimize fewer independent input moves than predicted output samples. This distinction between prediction and control horizons reduces the number of decision variables while retaining a longer view of future behavior.

Prediction models and state estimation

The model determines what an MPC controller can anticipate. For a linear time-invariant system, a common representation is

xk+1=Axk+Buk+Edk,yk=Cxk+Duk.x_{k+1}=Ax_k+Bu_k+Ed_k,\qquad y_k=Cx_k+Du_k.

The matrices may come from physical equations, experimental identification, or linearization around an operating point. Linear models are attractive because linear constraints and a convex quadratic objective produce a tractable optimization problem. They can also describe multiple interacting inputs and outputs without requiring separate control loops for every interaction.

Industrial predictive controllers have frequently used finite step-response or impulse-response models. These predict output changes by adding the effects of past and future input changes. Such models are intuitive for stable processes that settle after an input step. Integrating or unstable processes require special treatment rather than an unmodified finite-response approximation. State-space models offer a compact alternative, especially when the system has many measured variables or long response times.

First-principles models use conservation laws and constitutive relations. In chemical engineering, these often combine mass balances, energy balances, reaction kinetics, and transport relationships. A reactor model may include concentrations and temperature; a vehicle model may include position, velocity, and tire forces. Nonlinear models can represent wider operating ranges, but their additional detail increases identification, estimation, and computational demands. A highly detailed simulator is not automatically the best online control model.

Most plants do not measure every state. State estimation therefore supplies an estimate consistent with the model and available sensor data. A Kalman filter is widely used for linear models with stochastic disturbances. Nonlinear alternatives include extended and unscented Kalman filters. Moving-horizon estimation instead reconstructs recent states and disturbances through an optimization problem over a moving window, allowing explicit state and disturbance constraints. Its structure closely resembles MPC, although estimation reconstructs past behavior while control selects future actions.

Persistent model error can produce steady tracking offsets. A common remedy augments the prediction model with slowly varying or integrating disturbance states and estimates them online. An associated target calculation finds a feasible steady state consistent with the estimated disturbance and desired output. Under suitable model and observability conditions, this architecture supports offset-free tracking. Simply adding a disturbance term does not guarantee offset-free behavior for every plant or every controlled output.

Model validation must address the conditions in which the controller will operate. A model can predict routine data accurately yet misrepresent responses near input saturation or during a large transition. Relevant properties include delays, dominant time constants, cross-coupling, disturbance sensitivity, and extrapolation behavior. The model’s useful accuracy is therefore judged in relation to the decisions and constraints it must support, not solely by a global prediction-error score.

Stability, feasibility, and constraint guarantees

Solving a finite-horizon problem does not by itself prove that the resulting feedback system is stable. A controller may choose actions that appear attractive within its horizon but create difficulties immediately beyond it. Short horizons, weak terminal penalties, and poorly represented unstable dynamics can all cause this behavior. Stability analysis concerns the closed-loop system generated by repeated optimization, rather than the quality of any single predicted trajectory.

A standard construction uses a terminal region Xf\mathcal X_f, a local terminal controller κf(x)\kappa_f(x), and a terminal cost Vf(x)V_f(x). The terminal region is chosen so that the local controller keeps the system inside it while respecting constraints. A useful condition is

Vf(f(x,κf(x)))−Vf(x)≤−ℓ(x,κf(x)),x∈Xf.V_f\bigl(f(x,\kappa_f(x))\bigr)-V_f(x) \leq -\ell(x,\kappa_f(x)), \qquad x\in\mathcal X_f.

If the optimization is initially feasible, the unused portion of its optimal sequence can be shifted forward and extended with the terminal controller. This constructs a feasible candidate for the next optimization, assuming the nominal model is exact.

This argument establishes recursive feasibility: feasibility now implies feasibility at subsequent steps under the stated assumptions. Combined with suitable positivity conditions on the cost, it also makes the optimal finite-horizon value function a Lyapunov function for the closed loop. For linear regulation problems, terminal costs and local feedback laws are often constructed from the linear-quadratic regulator. Alternative MPC theories obtain stability through sufficiently long horizons, contraction requirements, or other structural assumptions without an explicit terminal region.

Constraints introduce an important distinction between hard and soft limits. A hard constraint must be satisfied by every feasible solution. A soft constraint introduces a slack variable, for example

yi∣k≤ymax⁡+ϵi,ϵi≥0,y_{i|k}\leq y_{\max}+\epsilon_i, \qquad \epsilon_i\geq0,

and penalizes ϵi\epsilon_i in the objective. Softening selected performance bounds can prevent an optimization problem from becoming infeasible after a disturbance. It does not make a physical violation harmless. Equipment limits and safety requirements need treatment consistent with their actual significance, rather than indiscriminate softening.

Nominal feasibility is also different from real-world constraint satisfaction. Unexpected disturbances, estimation errors, numerical tolerances, and actuator behavior can make the actual trajectory differ from the predicted one. Formal guarantees must specify these assumptions explicitly. Even a recursively feasible nominal controller does not automatically provide robustness against unmodeled effects.

Infeasibility management is consequently part of controller architecture. Possible mechanisms include prioritized constraint relaxation, a feasibility-restoration problem, a reference governor, and a prevalidated backup controller. Stability and feasibility results are properties of the complete design—including estimation, uncertainty assumptions, solver behavior, and fallback logic—not of the MPC label alone.

Uncertainty, robust control, and stochastic formulations

Uncertainty enters MPC through disturbances, imperfect parameters, unmeasured states, and prediction errors. Nominal MPC commonly optimizes around one estimated state and one predicted disturbance sequence. This can work well when feedback corrections are frequent and errors are modest, but it does not generally guarantee constraint satisfaction. Robust control and stochastic control provide two different ways to make uncertainty explicit.

Robust MPC treats uncertainty as belonging to a specified set. A min–max formulation chooses a control strategy that minimizes cost against the worst allowed disturbance sequence. The distinction between a control strategy and a fixed sequence is fundamental: future inputs should be allowed to react to future information. Representing all such reactions directly can become computationally prohibitive. Practical formulations therefore use structured disturbance-feedback policies, scenario trees, invariant sets, or other restricted policy classes.

Tube MPC separates a nominal trajectory from a bounded tracking error. A common structure is

uk=vk+K(xk−zk),u_k=v_k+K(x_k-z_k),

where zkz_k and vkv_k are the nominal state and input, and KK provides ancillary feedback. If the error ek=xk−zke_k=x_k-z_k remains inside a robust invariant set E\mathcal E, nominal constraints can be tightened so that actual states and inputs remain admissible. For example,

zk∈X⊖E,vk∈U⊖KE.z_k\in\mathcal X\ominus\mathcal E,\qquad v_k\in\mathcal U\ominus K\mathcal E.

The set operation ⊖\ominus denotes Pontryagin difference. The resulting “tube” bounds possible trajectories around the nominal plan. Its guarantees depend on valid uncertainty bounds and suitable invariance properties.

Stochastic MPC instead uses a probabilistic uncertainty description. It may minimize an expected operating cost while enforcing a chance constraint such as

Pr⁡(xi∣k∈X)≥1−α.\Pr(x_{i|k}\in\mathcal X)\geq1-\alpha.

The parameter α\alpha represents a permitted violation probability under the assumed model. A chance constraint at each prediction step is not equivalent to a bound on the probability of any violation over an entire operating period. Joint, individual, conditional, and long-run probability requirements are distinct and require different analyses.

Scenario methods approximate uncertainty by sampled disturbance realizations. Other approaches exploit particular distributions or propagate approximate moments. Distributionally robust MPC optimizes against a collection of plausible probability distributions rather than a single specified distribution. This can address limited statistical knowledge, but the choice of distributional ambiguity set strongly affects conservatism and computation.

Neither robust nor stochastic MPC eliminates the consequences of incorrect uncertainty assumptions. A bounded-disturbance guarantee does not cover disturbances outside the stated set, while a probabilistic guarantee depends on distributional validity and the precise event being bounded. Estimation errors and temporal correlations may also matter. The engineering task is therefore to match the mathematical uncertainty description to the physical environment and to distinguish formal guarantees from empirical performance.

Numerical optimization and real-time computation

For linear dynamics, quadratic costs, and linear inequalities, MPC usually reduces to quadratic programming. After collecting the decision variables into a vector zz, the problem takes the form

min⁡z 12zTHz+gTzsubject toGz≤b.\min_z\ \frac12 z^\mathsf{T}Hz+g^\mathsf{T}z \quad\text{subject to}\quad Gz\leq b.

If HH is positive semidefinite, the problem is convex; suitable positive definiteness yields uniqueness of the optimizer. Current measurements and forecasts alter the problem data at each sampling instant, while much of its structure remains unchanged.

Two major formulations exploit this structure differently. A condensed formulation eliminates predicted states using the dynamics, leaving primarily future inputs as decision variables. This reduces the variable count but can produce dense matrices. A sparse formulation retains states and inputs and expresses the dynamics as equality constraints. Its larger system has a highly structured, often block-banded form. The preferable approach depends on horizon length, state and input dimensions, hardware, and solver implementation.

Active-set solvers identify which inequalities are binding at the optimum. They can benefit from the fact that consecutive MPC problems often have similar active constraints. Interior-point methods approach the solution through a sequence of barrier problems and exploit structured linear algebra. Operator-splitting and other first-order methods can offer inexpensive iterations and convenient warm starts, although the iterations needed for a given accuracy depend strongly on conditioning and tolerances. No solver class is uniformly best across all control problems.

Nonlinear MPC produces a nonlinear programming problem. Sequential quadratic programming solves a succession of local quadratic approximations; nonlinear interior-point methods handle constraints through barrier formulations. Direct multiple shooting introduces state variables at intermediate points and constrains them to agree with integrated trajectories. Direct collocation approximates state trajectories over intervals and imposes the differential equations at selected points. These transcription choices affect sparsity, numerical conditioning, and integration accuracy.

Real-time iteration methods perform a limited amount of optimization work per sample, often one structured update, rather than solving each nonlinear problem to high precision. Their usefulness depends on tracking a suitable local solution as conditions change. Warm-starting with the shifted previous solution can greatly reduce computation. Nevertheless, nonlinear objectives and constraints may have local minima, so numerical convergence does not necessarily imply a globally optimal trajectory.

Real-time feasibility involves more than average solve speed. The complete cycle includes measurement acquisition, estimation, model evaluation, optimization, communication, and actuation. A controller must also handle infeasibility reports, solver failures, and missed deadlines. Scaling state and input units, monitoring residuals, and selecting tolerances consistent with physical margins are important because a numerically accepted solution can still be unsuitable if the underlying problem is badly conditioned.

Specialized formulations and architectures

Several important MPC variants change the objective, the model class, or the location of computation while retaining receding-horizon feedback. Economic MPC directly optimizes operating objectives such as energy expenditure, production value, or resource use. Conventional tracking MPC normally penalizes deviation from a prescribed target; economic MPC need not have a stage cost minimized at that target. Its preferred operation can involve transitions or periodic behavior rather than a single steady state. Stability analysis therefore often uses dissipativity arguments and a transformed cost instead of ordinary tracking-error penalties.

Explicit MPC moves much of the optimization offline. For certain linear-quadratic constrained problems, multiparametric programming partitions a feasible parameter region into polyhedral subsets. Within each subset, the optimal control action is an affine function of the state or other parameters. Online execution consists mainly of finding the relevant region and evaluating that function. This can provide predictable, fast execution for small systems, but the number of regions may grow rapidly with problem dimension and constraint complexity.

Hybrid MPC handles systems containing both continuous dynamics and discrete decisions. Examples include switching equipment on or off, selecting gears, and choosing operating modes. Logical conditions can often be encoded with binary variables, producing mixed-integer optimization problems. Such models represent decisions that continuous approximations may miss, including minimum on-times or mutually exclusive modes. Their computational burden can be substantially greater than that of convex continuous MPC, especially when many discrete choices extend across a long horizon.

Distributed MPC divides a large system among interacting controllers. Each controller optimizes a subsystem while exchanging predicted trajectories, coordination variables, or prices with others. The challenge is to preserve compatibility between local decisions when subsystems share constraints or influence each other dynamically. Decentralized control without communication is not equivalent to distributed optimization. Communication delays, incomplete iterations, and inconsistent local models can alter both performance and guarantees.

Hierarchical architectures assign different time scales and responsibilities to different layers. A scheduling layer may choose production targets, an MPC layer coordinate process operation, and local PID controllers regulate individual actuators or fast loops. In other applications MPC directly commands actuators. The division depends on dynamics, computational resources, and the existing automation infrastructure. MPC is therefore not necessarily a replacement for conventional feedback control; it often coordinates and constrains it.

Adaptive MPC updates model parameters during operation, while learning-based MPC may use statistical models to represent uncertain dynamics or costs. These approaches introduce additional questions about identification quality and safe adaptation. Updating a model changes the prediction problem and can invalidate a previously computed feasible region or uncertainty bound. Successful adaptation consequently requires more than replacing old parameter estimates with new ones.

Industrial and cross-sector applications

The strongest early industrial adoption occurred in petroleum refining and related chemical processes. A distillation unit may have interacting reflux, reboiler duty, feed rate, and pressure controls, while product purity and equipment capacity impose limits. MPC predicts these interactions and coordinates manipulated variables rather than treating every controlled output independently. Its economic value often comes from operating closer to binding constraints with less variability, thereby increasing throughput, improving product consistency, or reducing utility consumption.

Comparable structures occur in papermaking, polymer production, cement manufacture, and wastewater treatment. A wastewater controller, for example, may coordinate aeration with predicted loading while accounting for dissolved oxygen and effluent objectives. In a batch process, the desired trajectory can change throughout a production cycle, and the relevant end condition may be final product quality rather than continuous tracking of a fixed setpoint. Process delays and imperfect quality measurements make model and estimator design especially important.

Buildings provide a different opportunity: thermal inertia allows a controller to shift heating and cooling in time. Predictions of outdoor temperature, occupancy, electricity prices, and available power can be combined with indoor comfort constraints. An MPC controller may precool a building before an expensive period or coordinate a heat pump with thermal storage. Outcomes depend on weather-forecast quality, building-model accuracy, equipment limitations, and the operational definition of comfort. Energy reduction and cost reduction are related but are not identical objectives.

In electrical systems, MPC coordinates energy storage, flexible demand, and variable generation. A microgrid controller can schedule battery charging subject to state-of-charge and power limits while balancing forecast generation and demand. Models may also include conversion losses and approximate degradation costs. At faster time scales, predictive control of power electronics can select switching actions or voltage references. Finite-control-set MPC explicitly evaluates a limited collection of converter switching states, although practical implementations vary in how they handle switching frequency, losses, and computational delay.

Vehicle applications include trajectory tracking, speed planning, and energy management. A road vehicle controller can predict how steering and acceleration affect future motion while constraining steering rate, tire-force demand, and separation from obstacles. In an electric vehicle, energy-oriented MPC can coordinate propulsion demand, battery operation, and thermal management. Predicted routes or road gradients provide information unavailable to a purely reactive controller. However, accurate treatment of friction, other road users, and uncertain perception remains essential; optimization alone does not establish driving safety.

In robotics, nonlinear MPC can coordinate motion under joint, contact, and actuator constraints. Legged robots may optimize contact forces or body trajectories, while manipulators may account for collision geometry and joint limits. Full rigid-body models are often simplified to meet computational deadlines, creating a trade-off between prediction fidelity and optimization speed. Irrigation and water-network applications exhibit slower dynamics but similar reasoning: present pumping or allocation choices influence future storage, pressure, and resource availability. Across these sectors, MPC is most useful when anticipating interactions and constraints has substantial value beyond immediate error correction.

Deployment, learning, and continuing research

An implemented MPC controller is a software and instrumentation system as well as an optimization formulation. Industrial deployments connect to a distributed control system, programmable logic controller, or supervisory platform. Measurements must be time-aligned, validated, and expressed in consistent units. Missing data, actuator saturation, maintenance states, and manual operation all affect the relationship between computed commands and actual plant behavior. Without explicit handling, these conditions can undermine assumptions used in both prediction and stability analysis.

Commissioning commonly involves model identification, offline simulation, closed-loop testing, and staged transfer of control authority. Simulation can expose constraint conflicts and unsuitable tuning, but its conclusions depend on how well the test model represents the plant. Hardware-in-the-loop testing includes actual control hardware and communication interfaces, revealing timing and integration problems that ideal simulations may omit. Monitoring after deployment is equally important because equipment wear, feedstock changes, and altered operating practices can gradually invalidate the original model.

Safety-critical use requires a distinction between operational optimization and independent protection. Functional safety functions and a safety instrumented system address specific hazardous conditions through separately specified protective actions. An MPC constraint can contribute to normal operation within a safe envelope, but it is not automatically equivalent to a certified protection function. Cybersecurity also matters because manipulated measurements, forecasts, or commands can change the optimization problem and the physical response.

Machine learning has expanded the range of predictive models available to MPC. Learned residual models can supplement physical equations; neural models can approximate nonlinear dynamics; and learned surrogates can accelerate repeated optimization. These approaches introduce a distinction between accurate average prediction and dependable behavior in control. Training data may provide little evidence about rare disturbances or operation close to constraints. Distribution shift, extrapolation error, and uncertainty calibration therefore have direct control consequences.

Connections with reinforcement learning include learning value functions, terminal penalties, system models, or controller parameters. MPC can also serve as a structured policy whose actions are obtained through optimization. Conversely, a learned policy may approximate an MPC controller to reduce online cost. Approximation changes the implemented control law, so guarantees for the original optimizer do not automatically transfer. Computational savings must be assessed together with approximation error and constraint-handling behavior.

Continuing research addresses reliable control with limited models, faster nonlinear optimization, distributed coordination, and the interaction between control and information gathering. In dual-control settings, an input affects both plant operation and what can be learned about the plant; ordinary certainty-equivalent prediction may neglect this second effect. Other work combines MPC with formal verification and runtime monitoring to identify when a nominal controller should yield to a backup mechanism. These problems concern not only finding better trajectories, but also establishing when a computed trajectory can be trusted and how control should proceed when its assumptions cease to hold.