Model Predictive Control (MPC)

MPC operates on a receding or moving horizon. At each control interval, the algorithm:

• Solves an optimization problem over a finite prediction horizon.
• Applies only the first control action from the computed optimal sequence.
• At the next step, the horizon shifts forward, and the optimization is repeated with new state measurements. This feedback mechanism provides inherent robustness to model inaccuracies and disturbances, as the plan is constantly re-evaluated.

A defining strength of MPC is its ability to directly incorporate hard and soft constraints into the online optimization problem. Common constraints include:

• Actuator Limits: Physical bounds on control inputs (e.g., valve position, motor torque).
• State Constraints: Safety or operational limits on process variables (e.g., temperature, pressure).
• Output Constraints: Limits on the measured system outputs. The optimizer finds the best control sequence that satisfies these constraints, making MPC ideal for safety-critical applications where operating near limits is necessary.

Unlike traditional PID controllers with fixed gains, MPC solves a numerical optimization problem in real-time at every control step. The standard formulation is a Quadratic Program (QP) minimizing a cost function, typically: J = Σ (output error)² + Σ (control effort)² The solution requires:

• A process model to predict future states.
• The current measured or estimated state.
• The desired reference trajectory. Advances in solver algorithms (e.g., interior-point, active-set) and dedicated hardware enable this computation within stringent sampling times, even for large systems.

MPC naturally handles Multiple-Input, Multiple-Output (MIMO) systems. The optimization framework simultaneously considers all manipulated inputs and all controlled outputs, along with their complex interactions. This allows it to:

• Manage coupling between different loops inherently.
• Optimize trade-offs between competing objectives across the entire plant.
• Coordinate the actions of multiple actuators to achieve a unified goal. This makes MPC superior to decentralized PID loops for integrated processes like chemical reactors, distillation columns, and autonomous vehicle trajectory tracking.

MPC can proactively compensate for measurable disturbances and known future changes in references through feedforward action. The optimizer uses:

• A disturbance model to predict the impact of measured upsets.
• Future reference trajectories (e.g., a planned setpoint change) within the prediction horizon. By optimizing with this foresight, MPC can initiate corrective action before the disturbance affects the primary outputs, significantly improving performance over purely feedback-based controllers.

The power of MPC comes with inherent trade-offs that dictate its application:

• High Computational Demand: Solving an optimization online requires significant processing power, limiting use in very high-frequency systems (<1ms).
• Model Dependency: Performance is directly tied to the accuracy of the internal model. Poor models lead to poor control.
• Tuning Parameters: While more intuitive than PID loops, MPC requires tuning the prediction horizon, control horizon, and cost function weights. Thus, MPC is best suited for processes with slow-to-moderate dynamics (e.g., chemical, energy, aerospace) where its optimization benefits outweigh the computational cost.

MPC is the de facto standard for advanced process control in refineries and chemical plants. It optimizes complex, interacting units like distillation columns, catalytic crackers, and reactors by:

• Maximizing yield of high-value products while respecting quality specifications.
• Minimizing energy consumption by coordinating heaters, compressors, and heat exchangers.
• Smoothly handling feedstock changes and product grade transitions.
• Enforcing hard safety constraints on pressures, temperatures, and flow rates. The economic impact is measured in percentage-point increases in throughput and significant reductions in utility costs.

In automotive systems, MPC provides high-frequency, constraint-aware control for safety-critical functions. Key applications include:

• Adaptive Cruise Control (ACC) & Cooperative Driving: Computes optimal acceleration/deceleration to maintain safe following distances and smooth traffic flow.
• Path Tracking & Lane Keeping: Solves for optimal steering angles to follow a reference trajectory while accounting for vehicle dynamics and tire friction limits.
• Torque Vectoring & Stability Control: Allocates drive and brake torque to individual wheels to maintain vehicle stability during aggressive maneuvers.
• Predictive Energy Management in Hybrid/Electric Vehicles: Optimizes the split between engine and battery power over a predicted route to maximize efficiency. MPC's ability to handle nonlinear dynamics and state/input constraints in real-time (< 50 ms) makes it ideal for these applications.

MPC is used for trajectory optimization and attitude control of aircraft and spacecraft, where violating constraints can be catastrophic.

• UAV/Drone Navigation: Plans optimal paths through obstacle fields while respecting dynamic limits on velocity and acceleration.
• Rocket Landing Guidance (e.g., SpaceX): Solves fuel-optimal descent trajectories in real-time, adjusting for wind and other disturbances.
• Aircraft Energy Management: Manages the complex interplay of engines, batteries, and auxiliary systems in more-electric aircraft architectures.
• Satellite Attitude Control: Computes torque commands for reaction wheels or thrusters to achieve desired orientation while minimizing fuel use. These applications leverage MPC's strength in multi-input, multi-output (MIMO) control under strict actuator saturation limits.

For robotic systems, MPC enables dynamic, reactive motion planning that respects the robot's physical limits and environmental constraints.

• Dexterous Manipulation: For robotic hands and arms, MPC computes joint torques to perform contact-rich tasks (e.g., assembly, grasping) while ensuring forces stay within safe bounds.
• Bipedal & Quadrupedal Locomotion: Generates stable walking and running gaits by solving for optimal foot placements and body motions over a preview horizon, adapting to uneven terrain.
• Mobile Robot Navigation: Plans collision-free velocity commands for wheeled robots in crowded, dynamic environments like warehouses.
• Human-Robot Collaboration: Predicts human motion and plans robot trajectories to maintain safe separation distances. The core challenge is solving the nonlinear optimization problem at control rates (100-1000 Hz), often requiring simplified models or specialized solvers.

MPC is crucial for managing modern, decentralized energy grids with high penetration of intermittent renewable sources.

• Smart Grid Management: Balances electricity supply and demand across the grid by coordinating generators, storage (batteries), and flexible loads, ensuring frequency and voltage stability.
• Building Climate Control (HVAC): Optimizes heating, cooling, and ventilation for energy efficiency and occupant comfort across an entire building, using weather forecasts.
• Wind Turbine Control: Maximizes power capture while minimizing mechanical loads on the turbine structure by optimally adjusting blade pitch and generator torque.
• Microgrid Operation: Manages local generation, storage, and consumption in islanded or grid-connected mode to minimize cost and maximize reliability. These applications typically involve large-scale, distributed systems where MPC coordinates numerous subsystems over long prediction horizons (hours to days).

In biomedical engineering, MPC provides personalized, adaptive therapy by treating the human body as a complex dynamical system.

• Artificial Pancreas for Diabetes: Automatically regulates insulin (and potentially glucagon) infusion based on continuous glucose monitor readings, meal announcements, and physical activity predictions. It must strictly avoid life-threatening hypoglycemia.
• Anesthesia Delivery: Controls the infusion rate of anesthetic drugs to maintain a desired depth of anesthesia (measured by EEG signals like BIS) while accounting for patient-specific pharmacokinetics and pharmacodynamics.
• Cardiac Assist Devices: Optimizes the operation of ventricular assist devices (VADs) to support heart function while preventing suction events.
• Drug Dosage Optimization: Personalizes chemotherapy or other drug regimens to hit target concentration levels while minimizing toxic side effects. These are safety-critical applications where MPC's constraint-handling is non-negotiable, and models are often low-order, linear approximations of complex physiology.

Optimal Control is the foundational mathematical framework for determining a sequence of control inputs that minimizes a cost function (or maximizes a performance index) for a dynamical system over time. It provides the theoretical basis for MPC.

• Key Principle: Solves for an optimal control trajectory, often using calculus of variations or dynamic programming.
• Relation to MPC: MPC is a receding horizon implementation of optimal control, solving a finite-horizon problem online at each step.
• Example: The Linear Quadratic Regulator (LQR) is a classic optimal control solution for linear systems with quadratic costs.

Trajectory Optimization is the computational process of finding a sequence of states and control inputs that minimizes a cost function while satisfying system dynamics and constraints over a fixed time horizon.

• Core Activity: It is the numerical optimization problem solved at the heart of each MPC iteration.
• Methods: Includes direct (e.g., direct transcription) and indirect methods, often using nonlinear programming solvers like IPOPT or SNOPT.
• Application: Used in robotics for motion planning (e.g., computing a robot arm's path) and is the central planning engine within an MPC controller.

Model-Based Reinforcement Learning (MBRL) is an RL paradigm where an agent learns an explicit model of the environment's dynamics and uses it for planning or to improve policy learning efficiency.

• Shared Foundation with MPC: Both use an internal model for prediction. MPC typically uses an engineered model (e.g., physics-based), while MBRL uses a learned model (e.g., a neural network).
• Key Difference: MBRL aims to learn a policy from experience, while MPC solves an online optimization problem at each step without storing a general policy.
• Hybrid Approaches: Modern methods like PETS or MuZero blend learned models with MPC-style planning.

The Linear Quadratic Regulator (LQR) is a foundational optimal control solution for linear systems with quadratic cost functions. It provides a closed-form, optimal feedback control law.

• Mechanism: Solves the Algebraic Riccati Equation offline to compute a constant gain matrix K. The control is u = -Kx.
• Contrast with MPC: LQR is optimal over an infinite horizon with no constraints. MPC handles nonlinearities and explicit constraints but is computationally heavier.
• Use Case: LQR is often used as the underlying regulator in simpler, unconstrained cases or as a local controller within a larger architecture.

Receding Horizon Control (RHC) is the operational principle that defines MPC: at each control interval, an optimal control problem is solved over a finite future horizon, but only the first control action is executed before the horizon recedes forward.

• Core Loop: 1) Measure current state. 2) Solve optimization over horizon N. 3) Apply first control input. 4) Repeat at next time step.
• Benefit: Provides implicit feedback and disturbance rejection by constantly re-planning based on new measurements.
• Synonym: Often used interchangeably with MPC, though MPC explicitly emphasizes the use of a model for prediction.