1 History and development
1.1 Early optimal control foundations
Model predictive control builds on the broader idea of choosing future inputs to optimize a performance criterion subject to system dynamics. Early optimal control results established the use of mathematical models, optimization, and receding implementations, laying conceptual groundwork for later MPC formulations.
1.2 Emergence in industrial process control
MPC took practical form as industrial engineers sought controllers that could explicitly respect physical limits such as actuator bounds, safety envelopes, and quality constraints. Its ability to handle multivariable interactions made it attractive for chemical plants, refineries, and other processes where coupling between variables is pronounced.
1.3 Growth of computational methods
A key enabler for MPC was the maturation of numerical optimization algorithms and real-time computing. As solvers improved and hardware performance increased, it became feasible to repeatedly solve optimization problems at control-cycle rates, turning the predictive concept into a mainstream industrial technique.
1.4 Modern applications and variants
Over time, MPC expanded beyond classical linear constrained setups into nonlinear, robust, stochastic, and adaptive variants. Contemporary MPC platforms are used in domains ranging from manufacturing scheduling and robotics to energy management, often combined with state estimation and system identification workflows.
2 Basic principles
2.1 Prediction model
A prediction model forecasts how the controlled system evolves under candidate inputs. The MPC optimizer relies on this model to evaluate future behavior over a finite horizon.
2.1.1 State-space representation
In state-space MPC, the system is expressed with a state vector that captures internal dynamics. The prediction then propagates these states forward in time while applying candidate control actions.
2.1.2 Input-output models
In some applications, input-output descriptions are used instead of explicit states. Such models can simplify development when system identification yields transfer-function-like behavior or when observable quantities are more directly modeled.
2.2 Prediction horizon and control horizon
The prediction horizon defines the length of future time over which system responses are predicted and evaluated. The control horizon specifies how many future control moves are allowed to vary; beyond that, controls may be held constant or parameterized in a reduced form.
2.3 Receding horizon strategy
At each time step, MPC computes an optimal sequence over the horizon but applies only the first control move. The process then repeats at the next sampling instant using updated measurements (and possibly updated model parameters), creating a feedback loop driven by predictions.
2.4 Objective function
The objective function encodes trade-offs between desired performance, effort, and constraint-related behavior. It typically sums terms across the prediction horizon, reflecting system evolution and control decisions.
2.4.1 Tracking performance
Tracking terms penalize deviations between predicted outputs and reference trajectories. References may be constant setpoints or time-varying profiles.
2.4.2 Control effort
Control effort terms discourage unnecessarily aggressive actuation by penalizing input magnitudes or input changes. This can promote actuator longevity and reduce undesirable oscillations.
2.4.3 Constraint penalties
When constraints are handled softly, additional terms penalize violations. This allows the controller to trade constraint adherence against achievable tracking when strict feasibility is difficult.
3 Mathematical formulation
3.1 State variables and decision variables
MPC uses predicted state trajectories (and sometimes outputs) as variables that define how the system evolves. Decision variables usually include the control inputs over the horizon, and may also include slack variables for soft constraints or parameters for reduced control parameterizations.
3.2 Cost function design
A typical cost function aggregates stage costs and terminal costs. Stage costs commonly measure tracking errors and input-related penalties at each step, while a terminal term may reflect desired end-of-horizon behavior or stability considerations.
3.3 Equality and inequality constraints
Constraints enforce physical limits and safe operating regions. Equality constraints represent system dynamics or model relationships, whereas inequality constraints restrict inputs, outputs, or states within bounds.
3.4 Dynamic system equations
The predicted evolution is governed by discrete-time dynamics derived from the chosen model. These equations couple current state, applied input, and next state, forming the constraints that the optimizer must satisfy.
3.5 Optimization problem structure
The optimization problem can be posed as a quadratic program (QP), a nonlinear program (NLP), or other structured forms depending on model linearity and cost terms. The structure determines both solvability and computational load, influencing real-time feasibility.
4 Core algorithm
4.1 Model prediction step
Given the current state estimate and candidate control inputs, the controller propagates predictions across the horizon. This step uses the internal model to compute predicted outputs and states for each future time index.
4.2 Optimization step
An optimizer solves the finite-horizon problem, producing a control sequence that minimizes the chosen objective subject to constraints. Solver behavior—such as convergence criteria and warm-start capability—affects responsiveness in practice.
4.3 Control update step
Only the first element of the computed control sequence is applied to the plant. This implements the receding horizon principle and avoids reliance on errors that may accumulate over long predictions.
4.4 Feedback and repetition
At the next sampling instant, new measurements are incorporated, and the optimization problem is rebuilt or updated. State estimation may be run concurrently, especially when not all states are directly measured.
4.5 Stability and feasibility checks
Practical MPC implementations incorporate mechanisms to verify feasibility and maintain satisfactory closed-loop behavior. This may include fallback strategies when the solver cannot find a feasible solution under the current conditions.
5 Model types used in MPC
5.1 Linear MPC
Linear MPC assumes linear dynamics and typically uses quadratic cost functions. The result is often a convex optimization problem when constraints are linear, enabling efficient and reliable computation.
5.2 Nonlinear MPC
Nonlinear MPC allows nonlinear dynamics and nonlinear cost or constraint expressions. This can improve modeling accuracy for strongly nonlinear systems but usually increases computational complexity.
5.3 Robust MPC
Robust MPC is designed to maintain performance under model uncertainty. It incorporates worst-case or set-based assumptions about uncertainty, producing control actions that remain effective across a family of possible models or disturbances.
5.4 Adaptive MPC
Adaptive MPC updates model parameters during operation. As system behavior drifts due to aging, changing operating conditions, or unmodeled dynamics, parameter estimation helps keep predictions aligned with reality.
5.5 Stochastic MPC
Stochastic MPC accounts for uncertainty by modeling disturbances or noise as random variables. It aims to optimize performance in expectation or under probabilistic constraints, acknowledging that future evolution is not deterministic.
6 Constraints handling
6.1 Input constraints
Input constraints bound actuator commands such as throttles, motor torques, or valve positions. Enforcing these limits prevents unachievable commands and improves safety and hardware compatibility.
6.2 Output constraints
Output constraints restrict predicted trajectories of measured variables like temperatures, pressures, speeds, or product quality indicators. In many applications, output constraints are central to maintaining specification compliance.
6.3 State constraints
State constraints limit internal variables, including quantities that are not directly controlled but influence safe operation. They can be important when internal states correspond to physical safety margins or operational limits.
6.4 Soft constraints
Soft constraints allow controlled constraint violations by introducing slack variables. This is useful when strict feasibility may be lost due to disturbances, transient mismatch, or reference changes that temporarily exceed limits.
6.5 Constraint violation management
Beyond soft penalties, constraint violation management can use strategies such as priority levels, switching logic, or tightened constraints based on prediction error bounds. The goal is to avoid persistent constraint breaches while retaining solvability.
7 Design and tuning
7.1 Horizon selection
Choosing horizons involves balancing foresight against computational cost. Longer horizons can improve planning quality but require more prediction steps and potentially heavier optimization, affecting cycle time.
7.2 Weighting factors
Weighting factors in the objective function determine the emphasis on tracking, smoothness, and control intensity. Proper tuning aligns the controller’s behavior with performance priorities and avoids undesirable trade-offs such as excessive actuation or sluggish tracking.
7.3 Model calibration
Model calibration aligns the internal model with observed system behavior. This can involve parameter identification, model reduction, and validation steps to ensure that predictions are sufficiently accurate for constraint-aware control.
7.4 Disturbance modeling
Disturbance handling can be explicit (including disturbance states or measured disturbance inputs) or implicit (through robust or stochastic formulations). Accounting for predictable disturbances can substantially improve tracking and constraint adherence.
7.5 Performance tuning trade-offs
Tuning typically trades off speed of response, overshoot, control smoothness, and constraint tightness. It also interacts with solver settings and horizon length, meaning tuning must consider both control quality and real-time computation.
8 Stability and robustness
8.1 Closed-loop stability
Stability analysis for MPC depends on formulation details such as terminal constraints, terminal costs, and assumptions about system properties. For some linear settings, stability can be supported by designing terminal ingredients that create a controllable end region.
8.2 Robustness to model uncertainty
Robustness concerns how the controller behaves when the actual plant deviates from its model. Methods include robust MPC design, constraint tightening, and feedback mechanisms that reduce sensitivity to modeling errors.
8.3 Disturbance rejection
Disturbance rejection evaluates how effectively MPC counteracts external inputs or internal perturbations. The receding horizon nature often helps correct trajectory errors, while explicit disturbance models can further improve effectiveness.
8.4 Offset-free control
Offset-free control aims to eliminate steady-state tracking error caused by constant disturbances or model bias. Approaches often involve incorporating disturbance estimates, augmenting the model with integrator-like structures, or using observer-based corrections.
9 Computational aspects
9.1 Real-time implementation
MPC is inherently iterative and must complete the optimization within the sampling interval. Real-time implementation requires careful system integration of sensing, state estimation, solver execution, and actuator command timing.
9.2 Solver selection
The choice of solver depends on problem structure: linear MPC frequently benefits from QP solvers, while nonlinear MPC may require NLP methods. Solver features such as warm-starting, constraint handling, and detection of infeasibility are important in practice.
9.3 Numerical efficiency
Numerical efficiency relates to how quickly and reliably the solver computes solutions. Scaling, regularization, and careful formulation of cost and constraints can reduce ill-conditioning and improve convergence.
9.4 Embedded and industrial hardware
Deployments range from laboratory PCs to embedded controllers. Hardware constraints influence sampling rates and allowable model complexity, motivating simplified formulations, reduced horizons, or tailored solvers.
9.5 Complexity considerations
Complexity grows with horizon length, number of decision variables, and constraint count. Multi-input multi-output systems and nonlinear models can significantly increase the optimization burden, prompting strategies such as model simplification and sparse computation.
10 Industrial applications
10.1 Chemical process control
In chemical processing, MPC manages coupled dynamics such as temperature, pressure, and compositions while respecting constraints on equipment and product specifications. The approach helps coordinate multiple control loops within a single optimization framework.
10.2 Manufacturing systems
MPC can support coordinated control in manufacturing lines by optimizing trajectories of actuators, managing process quality variables, and anticipating constraint interactions. It is particularly relevant when delays and multivariable coupling affect throughput and stability.
10.3 Robotics and motion control
In robotics, MPC plans motion while considering actuator limits and collision avoidance constraints formulated in state or output terms. Depending on the model and constraints, it can produce smooth, feasible trajectories under changing conditions.
10.4 Power and energy systems
Energy applications often require constraint-aware control under fluctuating demand and generation. MPC can manage variables such as power flows, battery charge/discharge rates, and thermal dynamics, balancing efficiency with safe operating bounds.
10.5 Automotive systems
Automotive systems may use MPC for tasks that require predictive behavior under constraints, including longitudinal control and coordinated multi-variable driving assistance. The method’s constraint-handling capabilities align well with actuator limits and safety envelopes.
11 Software and implementation
11.1 Modeling environments
MPC implementation typically starts with modeling in tools that support system representation, discretization, and parameter management. Model development may combine physics-based dynamics with data-driven identification.
11.2 Optimization toolchains
The optimization layer is implemented via solver interfaces or dedicated MPC frameworks. Toolchains commonly provide automatic problem construction, code generation, and routines for warm-start and constraint configuration.
11.3 Simulation and testing
Simulation validates closed-loop behavior before deployment. Testing often includes scenarios with disturbances, reference changes, and constraint saturation to verify that control actions remain feasible and performance remains acceptable.
11.4 Hardware-in-the-loop deployment
Hardware-in-the-loop testing connects the controller to simulated plant models while using real control hardware. This reduces integration risk by verifying timing, data flow, and numerical behavior under conditions close to deployment.
12 Evaluation and performance metrics
12.1 Tracking error
Tracking error quantifies how closely outputs follow reference trajectories. Metrics may include mean squared error, integral absolute error, or worst-case deviation, depending on operational priorities.
12.2 Constraint satisfaction
Constraint satisfaction measures the frequency, magnitude, and duration of constraint violations. In applications with strict limits, maintaining feasibility or minimizing violation penalties becomes a primary evaluation criterion.
12.3 Energy efficiency
Energy efficiency assesses how control actions influence energy use, losses, or operational cost. For instance, penalizing input activity can indirectly promote efficiency, but direct energy metrics can also be incorporated.
12.4 Robustness measures
Robustness metrics evaluate sensitivity to disturbances and model mismatch. Testing across ensembles of conditions or uncertainty samples provides a practical sense of how consistently the controller performs.
12.5 Comparative benchmarking
Benchmarking compares MPC against alternative controllers using consistent scenarios and evaluation metrics. Differences in computation time, stability margins, and constraint behavior help interpret performance beyond tracking alone.
13 Limitations and challenges
13.1 Model mismatch
If the prediction model poorly represents the plant, the computed optimization may recommend suboptimal or infeasible actions. Improving model fidelity and incorporating uncertainty handling are common responses.
13.2 Computation time constraints
MPC depends on completing optimization within strict timing. When models are complex or horizons are long, computation may become bottlenecked, requiring simplification or faster solvers.
13.3 Tuning complexity
Selecting horizons, weights, constraint relaxations, and solver parameters can be time-consuming. Poor tuning may lead to oscillatory control, slow convergence to references, or frequent constraint interactions.
13.4 Feasibility issues
Under strong disturbances or aggressive references, strict constrained MPC may fail to find a feasible solution. Soft constraints, constraint tightening, or fallback logic can mitigate this issue, but they change the behavior that must be interpreted.
13.5 Scaling to large systems
As the number of controlled variables grows, the optimization problem can become large and difficult to solve. Techniques such as decomposition, sparsity exploitation, and reduced-order models help address scaling.
14 Related control methods
14.1 PID control
PID control is a widely used baseline technique that adjusts inputs based on proportional, integral, and derivative terms of error. Compared with MPC, PID is simpler but generally cannot enforce multivariable constraints as directly.
14.2 Linear quadratic control
Linear quadratic control methods optimize a quadratic cost for linear systems without explicit constraint handling. MPC extends this optimization idea by incorporating constraints and using a receding horizon structure.
14.3 Adaptive control
Adaptive control modifies controller parameters in response to observed system changes. MPC can be combined with adaptive mechanisms, but the main difference is that MPC primarily optimizes predicted trajectories at each step.
14.4 Optimal control
Optimal control refers broadly to strategies that choose inputs to optimize performance over time. MPC is a particular implementation of optimal control that repeatedly solves finite-horizon problems with receding updates and explicit constraints.
14.5 Reinforcement learning approaches
Reinforcement learning (RL) learns control policies from interaction with an environment. While MPC uses explicit models and optimization, RL approaches may learn policies that approximate decision-making, sometimes reducing the need for a precise model but introducing distinct challenges around training and safety.