Traditional control design follows a two-step pipeline: first, collect data to identify a parametric model (such as a state-space model $x_{k+1} = Ax_k + Bu_k$), and second, use this model to synthesize a controller like Model Predictive Control (MPC). However, as noted by Xiong et al., this indirect approach can introduce and propagate identification errors. If the model is slightly wrong, the controller's actions may be suboptimal or even unstable.
Data-Enabled Predictive Control (DeePC) and data-driven MPC bypass this modeling bottleneck entirely. Rooted in behavioral systems theory Markovsky et al., these frameworks treat a system not as a set of hidden internal states, but simply as the collection of all its possible observable trajectories Coulson et al.. By using raw historical data directly inside the control loop, they achieve optimal predictive control without ever identifying a parametric model.
The Core Intuition: Trajectories as a Subspace
How can raw data predict the future without a mathematical model? The foundational insight comes from Willems’ Fundamental Lemma. For a linear time-invariant (LTI) system, the set of all valid input-output trajectories of a fixed length forms a linear subspace.
If we run an initial experiment and apply an input sequence that is persistently exciting—meaning it excites all modes of the system—the resulting trajectory contains complete information about the system's dynamics Alsalti et al.. We can slice this long historical trajectory into overlapping windows and arrange them as columns in a structured matrix called a Hankel matrix Zagorowska et al..
Because the valid behaviors of the system form a linear subspace, any new, unseen trajectory of the system must be a linear combination of the columns of this Hankel matrix Berberich et al..
$$\begin{bmatrix} U_p \\ U_f \\ Y_p \\ Y_f \end{bmatrix} g = \begin{bmatrix} u_{\text{past}} \\ u_{\text{future}} \\ y_{\text{past}} \\ y_{\text{future}} \end{bmatrix}$$
Here, the Hankel matrix is partitioned into past ($p$) and future ($f$) inputs ($U$) and outputs ($Y$). To predict the future, we look at our immediate past measurements ($u_{\text{past}}, y_{\text{past}}$) and find a compounding vector $g$ that aligns the upper half of the matrix with our current reality. Once $g$ is found, multiplying it by the bottom half ($U_f, Y_f$) gives us a mathematically guaranteed, valid prediction of future trajectories ($u_{\text{future}}, y_{\text{future}}$) Coulson et al.. The past data acts as an implicit state estimator, while the future data acts as the predictive model.
From Representation to Control
In DeePC, this non-parametric trajectory representation is embedded directly into an optimal control problem Huang et al.. Instead of optimizing over state equations, the controller optimizes over the vector $g$ and the future inputs $u_{\text{future}}$:
$$\min_{g, u_{\text{future}}} \mathcal{J}(u_{\text{future}}, y_{\text{future}}) \quad \text{subject to Hankel constraints}$$
At each sampling time, the system solves this optimization problem in a receding horizon fashion, applies the first step of $u_{\text{future}}$, and updates the past measurements Wang et al.. For deterministic linear systems, Coulson et al. proved that this direct data-driven formulation is mathematically equivalent to traditional model-based predictive control.
Overcoming Noise and Real-World Uncertainty
In an ideal world, the columns of the Hankel matrix perfectly define the system subspace. In reality, raw data is always corrupted by measurement noise. Noise breaks the exact linear relationships, meaning a naive optimization might choose an erratic vector $g$ to fit the noise, leading to poor control performance or instability.
To make DeePC robust, researchers introduce regularization terms into the objective function. The choice of regularization is not just a heuristic; it has deep theoretical justifications:
- Stochastic and Robust Equivalence: Huang et al. proved that adding a quadratic regularization term on $g$ is mathematically equivalent to solving a min-max robust optimization problem. This ensures the controller performs well under worst-case bounded disturbances.
- Distributional Robustness: By framing the problem through Distributionally Robust Optimization (DRO), Coulson et al. demonstrated that regularization implicitly optimizes performance against the worst-case probability distribution of data noise.
- Safety Guarantees: For safety-critical systems, Berberich et al. proposed a constraint tightening approach. By shrinking the operational boundaries based on the noise variance, they guarantee that the physical system never violates its constraints despite relying entirely on imperfect data.
Scaling Beyond Linear Systems
While Willems' Fundamental Lemma strictly applies to linear time-invariant systems, real-world systems are frequently nonlinear. The behavioral framework has been extended to handle these complexities through several key innovations:
- Lifting to Linear Spaces: Xiong et al. extended DeePC to nonlinear systems by leveraging a Koopman Bilinear Realization. By passing raw data through high-dimensional nonlinear lifting functions, the complex dynamics are mapped into a higher-dimensional linear space where behavioral matrix combinations still hold true.
- Kernel Methods for Efficiency: A major drawback of non-parametric control is that the Hankel matrix grows with the amount of data collected, increasing online computational complexity. To solve this, Alsalti et al. introduced a kernel basis representation derived from the null space of the data matrices, compressing the data into a compact form that drastically reduces online optimization times.
Applications in the Wild
Bypassing the need for explicit physics-based modeling makes this paradigm highly effective for complex, time-varying engineering systems where first-principles modeling is too expensive or inaccurate.
The strategy has optimized building energy hubs and doubled battery lifespans by adapting directly to component degradation Zagorowska et al.. In electrical grids, it stabilizes weak power networks and dampens dangerous oscillations across grid-connected power converters Huang et al.. It has even been deployed in decentralized settings, such as coordinating autonomous vehicles to smooth out urban traffic jams without needing a mathematical model of human driving behavior Wang et al..