Traditional control engineering relies on a two-step process: first, build a mathematical model of a system (system identification), and second, use that model to predict and optimize future behavior (Model Predictive Control). However, for complex systems like urban traffic, biological gene circuits, or underwater robots, creating an accurate model is often the most difficult and error-prone part of the process.
Data-Enabled Predictive Control (DeePC) bypasses this modeling step entirely. Instead of using equations to represent a system, DeePC uses a library of raw historical trajectories to predict the future.
The Intuition: Data as the Model
The core insight of DeePC is rooted in Behavioral Systems Theory. For a linear time-invariant system, any possible trajectory the system can take is simply a linear combination of its past behaviors, provided those past behaviors are "rich" enough.
Imagine you are trying to predict the path of a drone. If you have a large enough database of how that drone has moved in the past—how it responded to wind, how it accelerated, how it turned—you don't need to know the physics of its rotors. If the drone is currently in a specific state, you can look at your database, find similar past sequences, and combine them to see what happens next.
This principle is formalized through Willems’ Fundamental Lemma. It states that if the historical input data is "persistently exciting"—meaning it has explored the system's dynamics thoroughly—then the subspace spanned by those historical trajectories contains all possible future trajectories of the system.
The Mechanism: The Hankel Matrix
To turn this intuition into a controller, researchers like John Lygeros and Florian Dorfler organize historical input and output data into a structured format called a Hankel matrix.
A Hankel matrix stacks time-shifted versions of past trajectories. By solving an optimization problem over this matrix, the controller finds a set of weights that combine these past trajectories to match the current observed state and project the most "optimal" future path. As Alessio Rimoldi and colleagues demonstrate in Urban traffic congestion control: a DeePC change, this allows a controller to manage complex urban traffic flows by looking directly at historical sensor data rather than trying to model the unpredictable behavior of thousands of human drivers.
Bridging Theory and Robustness
Raw data is rarely perfect; it is often plagued by noise and disturbances. In traditional MPC, noise is handled by the model's inherent structure. In the model-free world of DeePC, noise can lead to "overfitting" to the historical data.
To solve this, Linbin Huang and Florian Dorfler proved in Quadratic Regularization of Data-Enabled Predictive Control that adding a quadratic regularization term to the DeePC objective function is mathematically equivalent to solving a "min-max" robust optimization problem. This provides a formal bridge between heuristic data-driven methods and rigorous control theory, ensuring the controller remains stable even when the data is messy.
Versatility Across Domains
Because DeePC doesn't care about the underlying physics, only the patterns in the data, it has been applied to a remarkably diverse set of problems:
- Energy and Buildings: Jicheng Shi and Colin Jones used an adaptive DeePC framework to manage building climate control in Switzerland, achieving nearly 30% energy savings by learning from weather and occupancy patterns directly.
- Robotics and Vehicles: For autonomous underwater vehicles (AUVs), where hydrodynamic modeling is notoriously difficult, Sebastian Zieglmeier and others showed that DeePC can outperform traditional PID controllers by adapting to ocean currents in real-time.
- Synthetic Biology: In Host-Aware Control of Gene Expression, Liam Perreault and Idris Kempf applied DeePC to regulate bacterial growth and gene expression, a domain where mathematical models are often too simple to capture the chaotic nature of living cells.
- Traffic Smoothing: Jiawei Wang and Colin Jones developed Distributed DeeP-LCC, which uses a small number of automated vehicles to smooth out traffic jams without needing to model human driving behavior.
By treating data as the fundamental representation of a system's behavior, DeePC offers a "plug-and-play" alternative to traditional control, enabling high-performance optimization in environments where models are either too expensive to build or too complex to maintain.