Model Predictive Control (MPC)

The dynamic process model is the mathematical heart of an MPC controller. It is an explicit representation—often a set of differential or difference equations—that predicts how the system's states will evolve over time in response to control inputs and disturbances.

• Types: Can be linear (e.g., state-space), nonlinear (e.g., neural network), or hybrid.
• Purpose: Used within the prediction horizon to simulate future system behavior for different candidate control sequences.
• Example: For a drone, the model predicts future position, velocity, and attitude based on proposed motor thrusts.

The cost function (or objective function) quantifies controller performance, defining what "optimal" means. The MPC solver minimizes this function over the prediction horizon.

• Typical Terms: Penalizes tracking error (deviation from setpoint), control effort (actuator movement), and rate of change.
• Constraints are hard or soft limits imposed on:
  • State Variables (e.g., maximum temperature, safe velocity).
  • Control Inputs (e.g., actuator saturation limits).
  • Output Variables (e.g., product purity in a chemical process).
• The optimization finds the control sequence that minimizes cost while satisfying all constraints.

The receding horizon principle is the defining operational mechanism of MPC. At each control time step, the controller:

1. Measures the current state of the system.
2. Solves an optimization problem over a finite prediction horizon (N steps into the future) to determine a sequence of optimal control inputs.
3. Applies only the first control input from this optimized sequence to the physical system.
4. Repeats at the next time step, incorporating new measurements.

This feedback mechanism provides inherent robustness to model inaccuracies and disturbances.

The optimization solver is the computational engine that performs the online minimization at each control step. Its speed and reliability are critical for real-time operation.

• For Linear/Quadratic Problems: Uses fast, dedicated algorithms like Quadratic Programming (QP) solvers (e.g., active-set, interior-point methods).
• For Nonlinear Problems: Employs more complex Nonlinear Programming (NLP) solvers (e.g., Sequential Quadratic Programming, interior-point).
• Real-Time Requirement: The solver must complete its calculation within the controller's sampling period, often in milliseconds. This drives the use of specialized, high-speed solvers.

A state estimator or observer (like a Kalman Filter) is frequently required because not all system states are directly measurable. It reconstructs the full internal state from available sensor measurements and the process model.

• Function: Provides the complete state vector needed to initialize the MPC's prediction at each step.
• Crucial for: Handling sensor noise, inferring unmeasured variables (e.g., concentration in a reactor), and compensating for model-process mismatch.
• In many formulations, the estimator and controller are designed separately but work in tandem.

These are the key tuning parameters that define the scope of the MPC's look-ahead optimization.

• Prediction Horizon (Np): The number of future time steps over which the system's behavior is predicted using the model. A longer horizon improves stability and long-term performance but increases computational cost.
• Control Horizon (Nc): The number of future control moves that are optimized. Beyond Nc, control inputs are typically held constant (e.g., at the last optimized value) for the remainder of the prediction horizon. This reduces the number of decision variables, simplifying the optimization.
• Tuning these horizons balances performance, robustness, and computational feasibility.

MPC is the de facto standard for advanced process control in refineries, petrochemical plants, and pharmaceutical manufacturing. It manages complex, interacting variables like temperature, pressure, and flow rates to optimize yield, quality, and energy consumption while strictly adhering to safety and operational constraints.

• Key Variables: Reactor temperatures, distillation column pressures, product purity.
• Objective: Maximize throughput, minimize energy use, ensure safe operation within constraints.

In autonomous driving, MPC is used for path tracking and motion planning. It predicts the vehicle's trajectory over a short horizon and computes optimal steering, throttle, and brake inputs to follow a desired path while avoiding collisions and maintaining passenger comfort.

• Key Variables: Vehicle pose (x, y, yaw), velocity, acceleration, distances to obstacles.
• Constraints: Actuator limits, friction circle (adhesion limits), traffic rules.
• Use Case: Lane keeping, adaptive cruise control, evasive maneuvering.

MPC provides robust control for aircraft, spacecraft, and drones, handling highly nonlinear dynamics and strict safety margins. Applications include:

• Attitude Control: Managing orientation (roll, pitch, yaw) of satellites and aircraft.
• Trajectory Optimization: Calculating fuel-efficient ascent paths for rockets or re-entry profiles.
• Quadcopter Stabilization: Precisely controlling multiple rotors for stable hover and agile flight under wind disturbances.

For robotic arms and mobile manipulators, MPC enables dynamic, torque-controlled motion that respects physical limits. It is crucial for tasks requiring contact with the environment.

• Key Variables: Joint angles, velocities, torques, end-effector pose, contact forces.
• Constraints: Joint limits, torque saturation, self-collision avoidance.
• Applications: Dexterous manipulation (e.g., handling fragile objects), locomotion for legged robots (Boston Dynamics-style walking), and human-robot collaboration where safety constraints are paramount.

MPC optimizes the operation of complex energy systems, balancing supply, demand, and storage in real-time.

• Power Plant Control: Manages boiler-turbine-generator systems for efficient load following.
• Microgrid Management: Coordinates distributed energy resources (DERs) like solar panels, wind turbines, and batteries to maintain grid stability and minimize cost.
• Building HVAC: Optimizes heating, ventilation, and air conditioning across a building to minimize energy use while maintaining comfort constraints.

MPC delivers personalized, automated therapy by modeling human physiology.

• Artificial Pancreas: For Type 1 diabetes, MPC uses a model of glucose-insulin dynamics to compute optimal insulin infusion rates from a continuous glucose monitor (CGM), maintaining blood sugar within a safe range.
• Anesthesia Delivery: Automatically adjusts the infusion rate of anesthetic drugs (e.g., propofol) based on measured physiological signals like the Bispectral Index (BIS) to maintain a target depth of sedation.
• Key Challenge: High inter-patient variability and safety-critical constraints require robust, adaptive models.

Optimal control is the broader mathematical field from which MPC is derived. It seeks to find a control law for a dynamical system that minimizes a specified cost function over time. Unlike MPC's receding-horizon approach, classical optimal control (like Linear Quadratic Regulator - LQR) often solves for an optimal policy over an infinite horizon offline.

• Key Distinction: MPC is an online approximation of optimal control, solving a finite-horizon problem repeatedly.
• Foundation: MPC's optimization problem is a direct application of optimal control theory.
• Use Case: LQR provides a globally optimal, closed-form solution for linear systems, while MPC handles nonlinearities and constraints.

Reinforcement Learning is a machine learning paradigm where an agent learns a control policy through trial-and-error interaction with an environment to maximize cumulative reward. It represents a data-driven alternative to model-based control like MPC.

• Model-Based vs. Model-Free: MPC uses an explicit dynamics model; many RL algorithms are model-free, learning a policy directly from experience.
• Hybrid Approaches: Algorithms like Model-Based RL (MBRL) learn a dynamics model from data and then use it for planning, blurring the line with MPC.
• Trade-off: RL can discover novel policies in complex environments but typically requires vast amounts of data and lacks MPC's inherent constraint handling and stability guarantees.

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

• Relation to MPC: For an unconstrained linear system, an infinite-horizon MPC formulation reduces exactly to an LQR controller.
• Practical Use: LQR is often used as the underlying regulator within an MPC scheme for the final cost-to-go approximation (i.e., the terminal cost).
• Limitation: LQR cannot handle state or input constraints directly, which is a primary motivation for using MPC.

State-space representation is the standard mathematical model used in MPC and modern control theory. It describes a dynamical system using a set of input, output, and state variables related by first-order differential (or difference) equations.

• Core Model: MPC's predictive model is almost always formulated in state-space: x_{k+1} = f(x_k, u_k).
• States (x): Variables that fully define the system's condition (e.g., position, velocity, temperature).
• Inputs (u): Manipulated variables the controller commands.
• Outputs (y): Measured variables, which may not be the full state, necessitating a state estimator.

A state estimator, such as a Kalman Filter or Moving Horizon Estimator (MHE), is a critical companion to MPC. It reconstructs the full internal state of the system from noisy and partial sensor measurements (outputs).

• Necessity: MPC requires knowledge of the current state x_k. Sensors often measure outputs y_k, not states.
• Kalman Filter: The optimal linear estimator. Extended Kalman Filters (EKF) and Unscented Kalman Filters (UKF) handle nonlinear systems.
• Dual of MPC: Moving Horizon Estimation (MHE) solves an optimization problem to estimate the state, making it the estimation counterpart to MPC's control optimization.

Hardware-in-the-Loop simulation is a testing methodology where real controller hardware (like an MPC algorithm running on an embedded computer) is connected to a real-time simulated model of the plant and environment.

• MPC Validation: Essential for safely and rigorously testing MPC controllers before physical deployment, especially for safety-critical systems like autonomous vehicles or robotics.
• Workflow: The MPC controller sends control signals u to the high-fidelity physics simulator, which returns simulated sensor measurements y.
• Benefit: Allows for stress-testing under failure modes and edge cases that would be dangerous or expensive to replicate in the real world.