Data-driven predictive control skips building a model of the plant. It predicts the future by taking weighted combinations of recorded past trajectories and picks the weights and the next control inputs together by solving an optimization at every step. Recorded data is noisy, so the optimizer can find weights that exploit the noise. The standard fix penalizes the size of the weight vector. That penalty is blind to physics: it treats every direction in trajectory space the same, so the combined prediction can still create mass, produce negative fluid volume, or let a quadcopter's position change in a way its velocity does not allow. We propose a different penalty. It charges the controller only for the part of its prediction that breaks a known conservation identity, such as a mass balance, a species balance, or a kinematic integration rule. We also project the recorded data onto the set of trajectories that satisfy that identity before the controller ever uses it.
The key observation is that many physical plants have nonlinear dynamics but obey linear accounting identities. A tank's outflow follows Torricelli's square-root law, yet the change in stored volume still equals inflow minus outflow. An identity like this holds exactly at operating points the data never visited. It is also exactly the structure that measurement noise destroys inside the data matrix. Conservation-aware regularization therefore targets two documented failures of the method: its sensitivity to noise, and its collapse when asked to operate outside the region covered by the recorded data. It does this without the linear-plant assumption that limits current hybrid physics-plus-data controllers.
Core Claim
The method we build on is Data-Enabled Predictive Control, called DeePC from here on. It rests on Willems' Fundamental Lemma: for a linear time-invariant system with sufficiently rich inputs, every valid future trajectory is a linear combination $g$ of the columns of block-Hankel matrices built from recorded inputs and outputs.
Let $\mathbf{H}$ stack the past and future data blocks. Suppose a known linear identity $A w = 0$ holds for every true trajectory $w$ of length $L$. For mass balance in the tanks, $A$ encodes $V_{t+1} - V_t - \Delta t\,(q^{in}_t - q^{out}_t) = 0$ at each step. The noise-free data columns all satisfy $A\mathbf{H} = 0$. Noisy columns do not. We propose solving
$$\min_{u,y,g,\sigma_y}\; J(u,y) + \lambda_g \|g\|_2^2 + \lambda_y\|\sigma_y\|^2 + \lambda_\phi \|A\mathbf{H}g\|_2^2,$$
where $J$ is the tracking cost, $\sigma_y$ is the usual slack on the initial-condition match, and $\lambda_\phi$ weights the physics residual of the recombined data trajectory. We pair this with a pre-processing step that replaces $\mathbf{H}$ by $P\mathbf{H}$, where $P = I - A^\top(AA^\top)^{-1}A$ is the orthogonal projector onto trajectories that satisfy the identity.
The claim has three parts. First, the penalty is consistent: on noise-free linear data it is identically zero, so it never biases a correct predictor. Second, it corresponds to a distributionally robust formulation in which the ambiguity set contains only physically admissible perturbations. That should be less conservative than the isotropic set. Third, because the identity holds for the true nonlinear plant at unseen operating points, it constrains extrapolation in exactly the directions where Hankel-span predictors fail.
Relation to Prior Work
The regularization literature supplies the starting point and the theory we extend. Berberich et al. and the survey An overview of systems-theoretic guarantees in data-driven model predictive control use an $L_2$ penalty on $g$ to prove practical robust stability. Data-Enabled Predictive Control: In the Shallows of the DeePC uses $L_1$ to select a sparse set of trajectories. Coulson et al. shows that norm regularization of $g$ equals worst-case optimization over a Wasserstein ball, which is a set of distributions within a fixed transport distance of the empirical data. All three penalties are isotropic in trajectory space. We keep their machinery and make one change: the penalty and the ambiguity set become anisotropic, aligned with known physics. Concretely, the Wasserstein ball becomes a ball intersected with the identity's subspace. We aim to prove a counterpart of the Coulson equivalence for this restricted set.
The closest prior work is Semi-Data-Driven Model Predictive Control. It uses a nominal linear physics model and lets a Hankel matrix model only the residual. It reports that this rescues tracking when a two-tank controller trained at a 15 cm level must track 30 cm. Its guarantee requires both the plant and the nominal model to be linear, and it never forces the recombined trajectory to obey conservation. Our approach needs no nominal model of the dynamics, only an accounting identity. That identity is exact for the nonlinear tank. The two ideas are complementary, so we test them both separately and combined.
Deep DeePC learns $g$ with a neural network. It treats $g$ as an unconstrained regression target and handles constraint breaches only through a fallback optimization. We use it as a reference for computation time, not as a competitor on physical consistency.
The physical-inconsistency literature motivates and constrains the design. ACE2 enforces exact dry-air-mass and moisture conservation with a corrector at every step. Deep Conservation enforces conservation inside an online residual minimization and reports conservation violations reduced by more than an order of magnitude on Burgers' equation. Learned Simulators for Turbulence reports that soft constraint losses mitigated drift but did not eliminate it. This last result is the main reason we treat the projection of $\mathbf{H}$ as a hard structural step rather than relying only on the soft $\lambda_\phi$ term. Deep Conservation also shows that exact conservation needs explicit equations. We sidestep that requirement by using only identities that are known structurally, such as balances and kinematics, not full dynamics.
Research Program
The first strand is theory, restricted to linear identities. For linear plants with noisy data we will bound the prediction error of the projected-plus-penalized predictor and compare it with the isotropic $L_2$ bound. Projection removes the noise components that lie in the row space of $A$. Under independent and identically distributed noise of variance $s^2$, the noise energy per column therefore drops by a fraction equal to rank$(A)/(\text{column dimension})$. That reduction follows directly from projecting isotropic noise onto a subspace. We will also state the restricted-ambiguity-set equivalence and show that the resulting stability margin is no worse than the isotropic one.
The second strand is a careful treatment of nonlinear plants. Willems' lemma fails for nonlinear plants whatever regularizer is used. We make no claim to fix that. The claim is narrower: the identity removes one source of extrapolation error, and we measure how much of the out-of-distribution error that source accounts for.
The third strand is implementation. The method is a convex quadratic program with the same variables as standard DeePC. The projection is computed once offline. Per-step cost should match regularized DeePC, which Data‐enabled predictive control for quadcopters reports at roughly 4 to 6 ms on a quadcopter. We tune $\lambda_g, \lambda_y, \lambda_\phi$ with the differentiable tuning of DeePC-Hunt so that no method wins through hand tuning.
Baselines
The naive reference points are unregularized DeePC and sequential identification plus MPC. In the second, a least-squares state-space model is identified and then controlled with the same output constraints and with the conservation identity as a hard constraint. Data-Enabled Predictive Control: In the Shallows of the DeePC found this two-step pipeline produced more constraint violations than DeePC across 30 noise seeds on the quadcopter. It reports those results only as box plots, so no reported number was found on Lacuna for its cost.
The strongest generic competitor is regularized DeePC with $L_2$ or $L_1$ penalties and slack, tuned with DeePC-Hunt under the same tuning budget and data. For the quadcopter we also start from the reported setting $\lambda_g = 500$, $q = 2$. On the reactor and gene-network benchmarks, the only printed accuracy numbers in the file are a gene-network RMSE of 0.1721 for conventional DeePC and 0.1406 for Deep DeePC, both from Deep DeePC. On the gene network we report against both, though that system has weaker conservation structure.
On the two-tank extrapolation test the strongest physics-aware competitor is SD-MPC and its robust variant rSD-MPC. That paper reports these only in figures, so no reported number was found on Lacuna for the 30 cm tracking error. We will reimplement SD-MPC and rSD-MPC with identical data and the nominal model the paper describes.
Two further baselines isolate our contribution. The first is DeePC with the identity imposed as a hard constraint on the predicted outputs only. The second is DeePC on a low-rank, singular-value-truncated Hankel matrix, a common generic denoiser. Our ablations run projection alone, penalty alone, both together, and both combined with SD-MPC.
The proposal beats the field only if it meets two conditions. It must reduce closed-loop tracking RMSE and constraint violations relative to tuned regularized DeePC and the hard-constraint variant across noise levels, with confidence intervals over at least 30 seeds. It must also match or beat rSD-MPC on the 30 cm test without using a nominal dynamics model.
Why not just…?
Why not just add the conservation law as a hard constraint on the predicted outputs? This is the most serious threat. Standard DeePC already accepts output constraints. Our argument is that with noisy $\mathbf{H}$, forcing $A[\mathbf{U};\mathbf{Y}]g = 0$ makes the optimizer search for weights whose noise happens to cancel. That is a form of the noise exploitation regularization was meant to stop, and it should appear as larger $\|g\|$ and worse realized-versus-predicted error. If the hard-constraint baseline matches us at every noise level, the projection step is redundant and we will report that. The ablation runs on identical data and decides the question.
Why not just increase $\lambda_g$? A larger isotropic penalty shrinks $g$ in every direction, including the directions that carry the true response. The cost is bias and conservatism. The same tension appears in Coulson et al., where a larger penalty means a larger worst-case ball. We will trace tracking error against $\lambda_g$ and against $\lambda_\phi$. The proposal holds up only if its error-versus-strength curve lies below the isotropic one.
Why not just identify a model and run physics-constrained MPC? The quadcopter evidence in Data-Enabled Predictive Control: In the Shallows of the DeePC shows identification bias causing violations. The two-tank results in Semi-Data-Driven Model Predictive Control show nominal models losing accuracy away from their linearization point. We still include this baseline with the identity hard-coded, because on the tank a well-chosen gray-box model could win, and that outcome would be informative.
Why not just use SD-MPC or low-rank denoising? SD-MPC needs a nominal linear model, and its guarantee assumes linearity. Low-rank truncation removes noise without knowing which directions are physical, and it can discard real but weak dynamics. Both are run as baselines. The combined variant tests whether our penalty adds anything on top of SD-MPC.
Proposed Experiments
There are three benchmarks. The first is the cascaded two-tank system from SD-MPC, with nonlinear Torricelli outflow and the 15 cm-to-30 cm extrapolation test, using a mass-balance identity on volume and flows. The second is the reactor-separator process from Deep DeePC, with species balances whose reaction terms are unknown; we use only balances that close without kinetics, such as total inert mass. The third is the nonlinear quadcopter simulation used in the two quadcopter papers, using the discrete position-velocity integration identity.
For each benchmark we sweep measurement-noise standard deviation and data length. We report closed-loop tracking RMSE, number of constraint violations, the physics residual $\|A w\|$ of both predicted and realized trajectories, the gap between predicted and realized open-loop trajectories, $\|g\|$, and solve time per step.
Two checks decide the outcome. The first is whether conservation-aware regularization closes most of the gap between pure DeePC and rSD-MPC at 30 cm without a nominal model. The second is whether it beats the hard-output-constraint variant at moderate-to-high noise. If both checks fail, physics-aware regularization offers nothing beyond imposing constraints the standard way. If only the second succeeds, the contribution narrows to noise robustness, and we will state it that way.
References
- Data-Driven Model Predictive Control with Stability and Robustness Guarantees
- An Overview of Systems-Theoretic Guarantees in Data-Driven Model Predictive Control
- Data-Enabled Predictive Control: In the Shallows of the DeePC
- Distributionally Robust Chance Constrained Data-Enabled Predictive Control
- Semi-Data-Driven Model Predictive Control: A Physics-Informed Data-Driven Control Approach
- Deep DeePC: Data-enabled predictive control with low or no online optimization using deep learning
- ACE2: accurately learning subseasonal to decadal atmospheric variability and forced responses
- Deep Conservation: A Latent-Dynamics Model for Exact Satisfaction of Physical Conservation Laws
- Learned Simulators for Turbulence
- Data‐enabled predictive control for quadcopters
- DeePC-Hunt: Data-enabled Predictive Control Hyperparameter Tuning via Differentiable Optimization