Model Predictive Control (MPC)

The system model is the mathematical representation of the process to be controlled. It predicts how the system's state 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 data-driven.
• Purpose: Provides the internal simulation used for prediction. The accuracy of this model directly limits the performance of the MPC controller.
• Example: For a chemical reactor, the model would include equations for mass balance, energy balance, and reaction kinetics.

The prediction horizon (N) is the finite number of future time steps over which the controller predicts the system's behavior. It defines the look-ahead window for optimization.

• Trade-off: A longer horizon improves stability and accounts for long-term effects but increases computational cost. A shorter horizon is faster but can lead to myopic, unstable control.
• Role: The controller solves for a sequence of optimal control actions over this horizon but typically only implements the first step before re-solving (receding horizon control).

The cost function is a scalar measure of performance that the MPC controller seeks to minimize over the prediction horizon. It mathematically encodes the control objectives.

• Typical Terms: Includes tracking error (deviation from a setpoint), control effort (magnitude of actuator movement), and economic objectives.
• Formulation: Often a quadratic function (e.g., sum of squared errors) for convex optimization, but can be any function suitable for the solver.
• Example: J = Σ (y(t) - y_desired)² + Σ Δu(t)² penalizes both tracking error and aggressive control moves.

Constraints are hard or soft limits imposed on system variables within the optimization problem. They are a defining feature of MPC, enabling safe and practical operation.

• Types:
  • Input Constraints: Limits on actuator range (e.g., valve position 0-100%) and rate of change.
  • State Constraints: Safety or operational limits on the process itself (e.g., maximum temperature, tank level).
  • Output Constraints: Limits on the measured process variables.
• Handling: The optimization solver finds a solution that satisfies all constraints, if feasible.

The optimization solver is the computational engine that, at each control interval, solves the mathematical problem: minimize the cost function subject to the system model dynamics and constraints.

• Real-time Requirement: Must find a solution within the sampling period of the controller.
• Algorithm Types: For linear MPC with quadratic costs, Quadratic Programming (QP) solvers are standard. For nonlinear MPC, Nonlinear Programming (NLP) solvers (e.g., interior-point, SQP) are used.
• Output: Produces the optimal sequence of future control moves. Only the first move is applied.

A state estimator or observer is required when not all system states are directly measured. It uses the system model and available sensor outputs to reconstruct the full internal state.

• Necessity: The optimization problem is formulated in terms of system states. Accurate state information is critical for correct prediction.
• Common Technique: The Kalman Filter (for linear systems) or Extended Kalman Filter (for nonlinear systems) is frequently used to provide optimal state estimates while filtering measurement noise.
• Role: Closes the loop by feeding the estimated current state into the MPC optimization at each step.

MPC is the dominant advanced control strategy in the chemical and petrochemical industries. It manages complex, multi-variable processes like distillation columns, catalytic crackers, and polymerization reactors.

• Key Function: Simultaneously controls temperature, pressure, flow rates, and product composition.
• Primary Benefit: Maximizes yield and product quality while strictly adhering to safety constraints (e.g., pressure limits) and operational bounds.
• Example: In a refinery, MPC adjusts multiple inputs to an oil distillation unit to maintain specified fractions of gasoline, diesel, and kerosene despite feed stock variations.

MPC is used for local path planning and motion control in self-driving cars and robots. It computes optimal steering, throttle, and brake commands over a short-term horizon.

• Key Function: Solves for the vehicle's future trajectory that minimizes deviation from a reference path while respecting dynamic constraints (e.g., tire friction limits) and obstacle avoidance.
• Primary Benefit: Provides smooth, anticipatory control that can react in real-time to other vehicles, pedestrians, and road conditions.
• Example: An autonomous car uses MPC to plan a smooth lane-change maneuver, predicting the states of surrounding vehicles and ensuring it stays within lane boundaries.

MPC is applied to aircraft autopilots, missile guidance, and satellite attitude control. It handles systems with fast dynamics and stringent performance requirements.

• Key Function: Computes optimal control surfaces (ailerons, elevators) or thruster firings to follow a desired flight path or maintain orientation.
• Primary Benefit: Explicitly manages actuator limits (e.g., maximum deflection angle) and state constraints (e.g., angle of attack to prevent stall).
• Example: A quadcopter drone uses MPC to stabilize its attitude and navigate to waypoints, accounting for wind gusts and its own battery-drain dynamics.

MPC is critical for smart grid operation, renewable energy integration, and building climate control. It balances supply, demand, and storage across dynamic, uncertain systems.

• Key Function: In a power grid, it dispatches generators and manages load to maintain frequency stability. In a building, it schedules HVAC systems to minimize energy cost.
• Primary Benefit: Optimizes economic performance while satisfying power flow constraints and thermal comfort bounds.
• Example: A microgrid with solar panels and a battery uses MPC to decide when to store, use, or sell energy to the main grid, predicting solar generation and electricity prices.

Industrial robots use MPC for precise, high-speed motion control of manipulator arms, especially when handling flexible or delicate objects.

• Key Function: Calculates optimal joint torques to move an end-effector along a desired trajectory, considering the full non-linear dynamics of the arm and its payload.
• Primary Benefit: Enforces collision avoidance and joint limit constraints in real-time, preventing damage. Improves performance in tasks like pick-and-place and assembly.
• Example: A robotic arm assembling electronics uses MPC to place a component with precise force, compensating for the arm's inertia and the component's fragility.

MPC is employed in automated drug delivery systems (e.g., artificial pancreas) and bioreactor control for pharmaceutical production.

• Key Function: In diabetes management, it computes optimal insulin infusion rates to regulate a patient's blood glucose levels based on continuous monitor readings and meal predictions.
• Primary Benefit: Maintains critical physiological variables within a safe, healthy range despite unmeasured disturbances and patient-specific dynamics.
• Example: A closed-loop insulin pump uses MPC to autonomously adjust basal rates, preventing both hyperglycemia and dangerous hypoglycemia.

The process of computing a sequence of control inputs and corresponding state trajectories that minimize a cost function while satisfying system dynamics and constraints. It is the core mathematical problem solved at each step of MPC.

• Offline vs. Online: MPC performs online trajectory optimization in real-time, re-planning at each control step.
• Direct vs. Indirect Methods: MPC typically uses direct methods (e.g., direct transcription) that discretize the problem into a nonlinear program (NLP).
• Applications: Found in robotics (manipulator motion), aerospace (rocket landing), and autonomous vehicles (lane changes).

A reinforcement learning (RL) paradigm where an agent learns an explicit model of the environment's dynamics and uses it for planning. MPC is a specific, often non-learning, instance of model-based planning.

• Key Distinction: In pure MPC, the model is given (e.g., physics equations). In model-based RL, the model is learned from interaction data.
• Dyna Architecture: A classic framework that alternates between learning a model, using it for planning (like MPC), and taking real actions.
• Sample Efficiency: Using a model for internal simulation, as MPC does, is typically more sample-efficient than model-free RL.

The defining operational principle of MPC. At each time step, the controller solves an optimization problem over a finite future horizon, implements only the first control action, and then repeats the process at the next time step.

• Horizon Management: The prediction horizon slides forward in time, providing continuous feedback correction.
• Feedback Integration: This re-planning inherently incorporates feedback, making it robust to model inaccuracies and disturbances.
• Contrast with Optimal Control: Traditional optimal control computes a full trajectory from start to finish; receding horizon control is inherently online and adaptive.

The mathematical field of optimizing an objective function subject to a set of constraints. MPC formulates the control problem as a constrained optimization problem solved at each time step.

• Constraint Types: MPC handles hard constraints (e.g., actuator limits, safety boundaries) and soft constraints (penalized in the cost function).
• Solvers: Relies on numerical solvers for Quadratic Programs (QP) (linear MPC) or Nonlinear Programs (NLP) (nonlinear MPC).
• Feasibility: A critical concern is ensuring the optimization problem has a feasible solution at every step, often managed via constraint softening or robust MPC formulations.

An optimal control solution for linear systems with quadratic cost functions and no constraints. LQR provides the optimal feedback gain matrix and is a foundational element often used within or alongside MPC.

• Unconstrained Optimum: LQR gives the globally optimal solution for its problem class. MPC with constraints reduces to LQR if constraints are inactive.
• Terminal Cost: In MPC, the LQR cost-to-go is often used as the terminal cost to ensure stability (infinite horizon approximation).
• Computational Simplicity: LQR has a closed-form solution (Riccati equation), making it extremely fast, whereas MPC requires online optimization.

A mathematical framework for planning under uncertainty where the agent cannot directly observe the true state. Advanced MPC variants address similar challenges through state estimation and robust formulations.

• State Estimation: MPC typically assumes full state observation. In practice, a state estimator (like a Kalman Filter) is used to feed the MPC, creating an implicit POMDP structure.
• Robust MPC: Explicitly accounts for bounded disturbances or model uncertainty, akin to solving a min-max optimization for worst-case scenarios.
• Observations vs. States: While POMDPs reason over belief states, MPC reasons over estimated states, blending estimation and control.