Model predictive control, widely known as MPC, is a control strategy that uses a mathematical model of a system to predict its future behavior and then solves an optimization problem at every time step to determine the best course of action. Unlike simpler control approaches that react only to what has already happened, MPC looks ahead over a rolling time window, adjusting its plan continuously as new measurements come in. The concept traces back to the 1960s, but it gained real traction after papers on dynamic matrix control and related methods appeared in the late 1970s and 1980s, initially inside petroleum refineries and chemical plants.1Computers & Chemical Engineering. Model predictive control: past, present and future – Section: The past Today MPC has spread far beyond those roots, turning up in autonomous vehicles, power grids, humanoid robots, and even artificial pancreas systems for people with diabetes.
How the Receding Horizon Works
The core idea behind MPC is sometimes called the “receding horizon.” At each moment in time, the controller takes a snapshot of the system’s current state, then uses a model to simulate what will happen over some future window, say the next several seconds or minutes. Within that window it searches for the sequence of control moves that best meets a goal, whether that goal is keeping a temperature at a setpoint, steering a car along a lane, or minimizing energy cost, all while respecting hard constraints like pressure limits or actuator speed. Once it finds the optimal sequence, it applies only the very first move, throws the rest away, takes a fresh measurement, and solves the whole problem again from scratch.
This “plan a lot, execute a little, replan” cycle gives MPC a distinctive advantage: it can anticipate problems before they arrive. If a disturbance is heading toward a constraint boundary, MPC starts correcting early rather than waiting until the limit is breached. That anticipation is what makes MPC especially valuable for systems with multiple interacting variables and hard safety limits. A distillation column, for instance, might have to keep dozens of tray temperatures within range while simultaneously meeting product purity specs and respecting valve travel limits. Traditional single-loop controllers struggle to coordinate all of those requirements; MPC handles them in a single optimization at each step.
The trade-off is computational cost. Solving an optimization problem in real time, sometimes hundreds of times per second, demands serious processing power, especially when the model is complex. Much of the research over the past two decades has focused on making that optimization faster and more reliable. High-performance quadratic programming frameworks, for example, are designed specifically to exploit the structure of MPC problems so the solver can finish within tight deadlines.2IFAC-PapersOnLine. HPIPM: a high-performance quadratic programming framework for model predictive control
Industrial Roots in Chemical Processing
MPC’s origin story is firmly planted in the petrochemical industry. In the late 1970s, engineers at Shell Oil and Adersa (a French company) independently developed control algorithms, dynamic matrix control and IDCOM respectively, that used step-response models to predict process behavior over a future horizon. These weren’t academic exercises: refineries needed to push throughput closer to constraint limits without violating safety boundaries, and MPC let them do exactly that. By the late 1980s, generalized predictive control had joined the family, and thousands of MPC installations were running in plants worldwide.3Computers & Chemical Engineering. Model predictive control: past, present and future – Section: The past
Chemical processing remains a stronghold. Distillation columns, reactors, and heat-exchanger networks are natural fits because they involve slow dynamics, tight purity constraints, and strong interactions between variables. Experimental work on pilot-plant distillation columns has shown that even a robust version of MPC, one designed to tolerate model errors and uncertainty, can maintain stability and performance across the operating range of the column.4Control Engineering Practice. Robust model predictive control of a pilot plant distillation column More recent approaches use data-driven linearized models built from operating data rather than first-principles equations. In case studies on both binary and four-product dividing-wall columns, MPC controllers built this way achieved good dynamic performance with small deviations and effective disturbance rejection.5Industrial & Engineering Chemistry Research. Operation of Distillation Columns Using Model Predictive Control Based on Dynamic Mode Decomposition Method
Energy Systems and Microgrids
As renewable energy has grown, so has the need for smarter grid management. Solar panels and wind turbines don’t produce power on demand; their output fluctuates with weather. That stochastic nature, combined with variable consumer demand, creates voltage and frequency swings that conventional controllers handle clumsily. MPC is increasingly used at multiple layers of microgrid control, from switching individual power converters up to managing power flow between an entire microgrid and the main utility grid. A comprehensive review found that MPC is emerging as a competitive alternative to conventional methods in voltage regulation, frequency control, power flow management, and economic optimization of microgrid operation.6Renewable and Sustainable Energy Reviews. Model predictive control of microgrids – An overview
One reason MPC fits energy systems so well is its natural ability to incorporate economic objectives directly into the optimization. Instead of tracking a fixed setpoint, the controller can minimize electricity cost or maximize revenue from selling power back to the grid, all while keeping voltages and currents within safe bounds. This economic-style MPC has gained substantial attention because it optimizes process operations in a time-varying fashion rather than just holding variables at steady-state targets. For a microgrid operator deciding when to charge a battery, when to sell surplus solar power, and when to buy from the grid, that kind of dynamic economic reasoning is exactly what’s needed.
Autonomous Vehicles and Steering
Self-driving cars and advanced driver-assistance systems have become one of MPC’s highest-profile applications. The problem fits the framework well: the vehicle has a dynamic model (how steering angle and throttle translate into motion), a reference trajectory (the desired path), and hard constraints (don’t leave the lane, don’t exceed maximum steering rate). MPC plans a short sequence of steering and throttle commands, executes the first one, and replans.
A data-driven MPC approach for autonomous vehicle steering demonstrated tracking errors generally between negative 0.1 meters and 0.2 meters during lane-change maneuvers, with smooth steering-angle changes staying within a range of about negative five to positive five degrees. The algorithm handled curve-to-straight transitions better than both a simpler PID controller and a kinematics-based MPC, which tended to let the vehicle drift off track during those transitions. Computation time was roughly half that of the kinematics-based MPC, since the data-driven version’s solve time depends on the data volume and prediction horizon rather than on the complexity of the vehicle model itself.7arXiv. Data-Driven Model Predictive Control for Autonomous Vehicle Steering
Speed matters here. A car traveling at highway speed covers many meters each second, so the controller has to finish its optimization in a few milliseconds. That pressure has driven innovations in both solver algorithms and the way vehicle models are formulated. Some teams use simplified linear models for speed, others use neural-network surrogates, and still others lean on the data-driven approach described above to sidestep the need for a physics model altogether.
Keeping Robots on Their Feet
Balancing a humanoid robot is a notoriously hard control problem. The robot’s center of mass sits high above a small support polygon (its feet), and even a gentle push can topple it. MPC helps by planning footstep adjustments and body motions several steps ahead, accounting for the robot’s dynamics and the constraint that the pressure center must stay within the footprint.
One approach formulates the balance-recovery problem as a quadratic program inside MPC. When a humanoid robot receives an external push, the controller measures the resulting error in what’s called the “capture point,” the spot on the ground where the robot needs to step to avoid falling, and adjusts both the desired pressure distribution and the next footstep location through MPC. This was tested on a full-size humanoid robot (DRC-HUBO+), which regained balance after strong pushes and after stepping on unexpected obstacles.8Robotics and Autonomous Systems. Balance recovery through model predictive control based on capture point dynamics for biped walking robot
Another study used MPC together with a whole-body controller for one-legged standing balance. The MPC-generated reference trajectories for the swing foot nearly coincided with the actual trajectories, and the scheme produced larger swing-foot motions than a controller without MPC, confirming that actively swinging the free leg plays a vital role in keeping a biped upright.9PubMed Central. Balanced Standing on One Foot of Biped Robot Based on Three-Particle Model Predictive Control The broader takeaway is that MPC lets robots plan their recovery motions before they’ve finished falling, much like a person sticking an arm out before actually losing balance.
The Artificial Pancreas
One of MPC’s most life-changing applications is in closed-loop insulin delivery for people with type 1 diabetes. In a healthy body, the pancreas continuously adjusts insulin release based on blood glucose levels. An artificial pancreas mimics this by reading glucose from a continuous sensor, running an MPC algorithm, and commanding an insulin pump to deliver the right dose. The controller’s model predicts where glucose is heading over the next hour or two, and it calculates an insulin dose that steers the predicted trajectory into a safe range while avoiding the dangerous low end (hypoglycemia).
Early clinical work at the University of Virginia demonstrated that MPC-based closed-loop control could manage blood glucose overnight and after a standardized breakfast. The system controlled glucose as effectively as patient-directed manual dosing after a meal, but was significantly superior to manual control in preventing overnight hypoglycemia, a period when low blood sugar can go undetected during sleep.10PubMed Central. Closed-loop artificial pancreas using subcutaneous glucose sensing and insulin delivery and a model predictive control algorithm: the Virginia experience
Subsequent refinements have tackled one of the hardest practical challenges: unannounced meals, when a person eats without telling the system in advance. An adaptive zone MPC approach that adjusts its aggressiveness based on current glucose level and its rate of change achieved statistically significant improvements in both average glucose and time spent in the target range of 70 to 180 mg/dL compared to a non-adaptive version, without increasing the risk of hypoglycemia.11PubMed Central. Adaptive Zone Model Predictive Control of Artificial Pancreas Based on Glucose- and Velocity-Dependent Control Penalties Commercial hybrid closed-loop systems are now widely available and use MPC or MPC-like algorithms at their core, representing one of the clearest cases where this control strategy directly improves daily quality of life.
Nonlinear MPC and Its Computational Cost
The earliest industrial MPC systems used linear models: simple representations where doubling an input doubles the output. Linear MPC is well understood, computationally cheap, and works fine when the system stays near a single operating point. But many systems are nonlinear. A magnetic levitation platform, for instance, has force-distance relationships that are sharply curved; a chemical reactor’s kinetics change with temperature in an exponential way. For these systems, a linear model only tells the truth in a narrow neighborhood.
Nonlinear MPC replaces the linear model with one that captures these curved relationships, at the cost of solving a harder optimization problem at every step. A fast nonlinear MPC scheme applied to a magnetic levitation system demonstrated that capturing the inductor current dynamics of the electromagnet, in addition to the position dynamics, was necessary to achieve high performance for both stabilization and rapid setpoint changes.12Control Engineering Practice. Nonlinear model predictive control of a magnetic levitation system Getting the model right matters: a model that’s too simple misses the physics, while one that’s too detailed takes too long to solve.
One way to guarantee that a nonlinear MPC controller will remain stable is the “terminal cost and terminal set” method, where the optimization includes a penalty and a safe region at the end of the prediction horizon. This is theoretically appealing but often impractical in real systems because computing those terminal ingredients for a nonlinear model is itself a difficult problem.13IFAC-PapersOnLine. Computation of terminal costs and sets for discrete–time nonlinear MPC Researchers continue to look for workarounds, including learning the terminal cost from data or using longer prediction horizons that make the terminal penalty less critical.
Another approach avoids some of the computational burden by learning a value function offline, essentially pre-computing an approximation of the cost-to-go, and then using it to warm-start or shorten the online optimization. This combines the safety guarantees of trajectory optimization with the speed of a pre-trained approximator.14Optimal Control Applications and Methods. Fast Nonlinear Model Predictive Control Combining Online Trajectory Optimization and Value Function Regression
When Machine Learning Enters the Loop
Machine learning and MPC have been converging from both sides. On one hand, MPC researchers want better models but don’t always have the physics to build them from scratch. On the other, machine-learning researchers have noticed that MPC’s constraint-handling and planning abilities complement the pattern-recognition strength of neural networks. Most of the work in learning-based MPC focuses on automatically improving the prediction model from recorded operating data.15Annual Review of Control, Robotics, and Autonomous Systems. Learning-Based Model Predictive Control: Toward Safe Learning in Control
The appeal is straightforward. Building a first-principles model of, say, a building’s thermal behavior across all seasons, occupancy patterns, and HVAC configurations is tedious and error-prone. A data-driven model trained on a year of sensor data can capture those dynamics without anyone writing down heat-transfer equations. MPC then wraps that learned model in an optimization that enforces comfort constraints and minimizes energy cost. The risk, of course, is that the learned model can be wrong in regions it hasn’t seen. Safety-aware learning-based MPC methods try to quantify that model uncertainty and tighten constraints accordingly, so the controller stays cautious in unfamiliar territory.
The autonomous-vehicle steering work mentioned earlier is a concrete example of this trend. Rather than modeling tire friction, suspension geometry, and road surface, the controller learned the relationship between inputs and outputs directly from driving data, cutting computation time while maintaining tracking accuracy.
State Estimation and the Problem of Missing Information
MPC assumes it knows the current state of the system, but in practice you rarely measure everything. A pipeline network might have pressure sensors at a few locations, not along every meter of pipe. A chemical reactor might measure outlet temperature but not the concentration of an intermediate species. The controller needs those unmeasured states to make good predictions.
The standard companion to MPC is a state estimator, often a Kalman filter for linear systems or a moving-horizon estimator for nonlinear ones. Moving horizon estimation (MHE) works on a similar principle to MPC itself: it looks at a window of past measurements and finds the state trajectory that best explains those measurements, subject to the model equations and any known bounds on the states. For distributed-parameter systems like pipeline networks, MHE has been paired with tracking MPC to estimate unknown states and uncertain parameters, enabling control that satisfies physical constraints even without full state information across the entire system.16Computers & Chemical Engineering. Tracking model predictive control and moving horizon estimation design of distributed parameter pipeline systems
The quality of the state estimate directly affects MPC performance. Feed the controller a bad estimate and it will optimize around a phantom version of reality, potentially driving the real system into trouble. This is one reason commissioning an MPC system involves more than just tuning the controller; getting the estimator right is equally important, and often harder.
Hybrid Systems and Mixed-Integer MPC
Some systems don’t fit neatly into the continuous-math world that standard MPC inhabits. Think of a heating system with an on/off valve, a gearbox that shifts between discrete ratios, or a power converter that switches between circuit topologies. These hybrid systems combine continuous dynamics (temperatures, voltages, flows) with discrete decisions (on or off, gear 1 or gear 2). Standard MPC, which solves a smooth optimization, can’t handle those binary choices directly.
Mixed-integer MPC addresses this by formulating the system as a mixed-logical dynamical model and solving a mixed-integer programming problem that contains both continuous and discrete decision variables.17IFAC-PapersOnLine. Mixed-Integer Model Predictive Control of Hybrid Impulsive Linear Systems The optimization simultaneously decides, for example, which valves to open and how far to open them. The downside is computational: mixed-integer problems are inherently harder than their continuous cousins, and solve times can jump by orders of magnitude. For slow processes like building climate control, where decisions happen every few minutes, this is manageable. For fast-switching power electronics operating at tens of kilohertz, it remains a research frontier.
Practical implementations often sidestep the worst of the computational burden through heuristics, warm-starting from the previous solution, or relaxing integer variables to continuous ones and rounding. The results are suboptimal in theory but fast enough to be useful, which in engineering is usually the more important metric.