Overcoming the Scalability Bottleneck of Physics-Based Control

46 concepts45 papers

Traditional model-based control strategies, such as Model Predictive Control (MPC), rely on explicit mathematical representations of a system's underlying physics. To optimize a building's heating, ventilation, and air conditioning (HVAC) system, for example, an engineer traditionally derives first-principles thermodynamic equations or constructs resistance-capacitance (RC) networks representing the thermal properties of walls, zones, and ambient disturbances.

While highly precise, this approach suffers from a severe scalability bottleneck. As noted by Roy S. Smith et al. in Control of Multicarrier Energy Systems from Buildings to Networks, the unique physical characteristics of individual structures make constructing high-fidelity thermal models an expensive, bespoke process. Ziwei Xiao demonstrates in Automatic generation of RC models using IoT data from commercial buildings that this manual modeling phase often consumes the majority of total engineering effort in advanced control projects.

This overhead is not unique to buildings. Whether calibrating thermodynamic models for aircraft auxiliary power units (Wlamir Olivares Loesch Vianna et al. in Health Monitoring of an Auxiliary Power Unit Using a Classification Tree), predicting battery degradation across heterogeneous vehicle fleets (Erik Frisk et al. in Data-Driven Lead-Acid Battery Prognostics Using Random Survival Forests), or mapping route-level transit energy consumption (Afiya Ayman et al. in Data-Driven Prediction of Route-Level Energy Use for Mixed-Vehicle Transit Fleets), the requirement for site- or unit-specific physical calibration makes widespread deployment logistically prohibitive.

To resolve this, a major research direction focuses on bypassing or automating physical modeling entirely, shifting toward data-driven predictive control, reinforcement learning, and differentiable system identification.


Data Predictive Control: Bypassing the Equations

If a system's historical operational data is rich enough, the data itself implicitly contains the underlying physical laws. Data Predictive Control (DPC) and Data-Enabled Predictive Control (DeePC) leverage this insight by replacing explicit physical equations with non-parametric or statistical models learned directly from telemetry.

Traditional MPC:  [Physical Parameters] ──> [Differential Equations] ──> [Optimization] ──> Control Action
Data-Driven DPC:  [Historical Data] ──────────────────────────────────> [Optimization] ──> Control Action

The intuition is straightforward: rather than mapping physical properties to equations and then to control actions, DPC maps historical input-output trajectories directly to control decisions. In Experimental demonstration of data predictive control for energy optimization and thermal comfort in buildings, Felix Bünning et al. demonstrated that a DPC framework built on Random Forests could cut cooling energy by 24.9% in a real-world apartment, proving that data-driven models can match or exceed the performance of bespoke physical models.

To handle multi-component systems where physical interactions compound in complexity, Marta Zagorowska et al. applied DeePC in Degradation-aware data-enabled predictive control of energy hubs. By optimizing heat pump and battery schedules directly from data, the authors bypassed the need to mathematically formalize both building thermodynamics and complex chemical battery degradation.

When operating without a strict physical model, managing unexpected environmental or occupancy disturbances becomes a challenge. To address this, Jicheng Shi et al. introduced Disturbance-Adaptive Data-Driven Predictive Control: Trading Comfort Violation for Savings in Building Climate Control. Their framework incorporates an adaptive feedback loop that allows operators to mathematically trade a precise, user-defined percentage of comfort violations for energy savings, ensuring robust performance even when the exact nature of external disturbances is unknown.


Reinforcement Learning: Learning Policies from Interaction

While DPC still solves an optimization problem over a rolling horizon, Deep Reinforcement Learning (DRL) shifts the paradigm toward policy optimization. Instead of predicting future states explicitly, a DRL agent learns a direct mapping from environmental states to optimal control actions by maximizing a reward signal (such as minimizing energy cost while maintaining comfort).

As surveyed by Liang Yu in A Review of Deep Reinforcement Learning for Smart Building Energy Management, model-free DRL eliminates the need to construct explicit thermal models for complex, real-world structures. This model-free nature allows controllers to adapt to highly dynamic environments that are notoriously difficult to capture via physics, such as co-optimizing HVAC systems alongside volatile renewable microgrids (Rumia Masburah et al. in Co-designing Intelligent Control of Building HVACs and Microgrids) or integrating real-time occupant voting feedback (Xinyu Liang et al. in Human-in-the-Loop AI for HVAC Management Enhancing Comfort and Energy Efficiency).

However, purely model-free RL can suffer from slow initial convergence and unsafe exploration phases. To mitigate this, researchers often deploy hybrid architectures. Huan Zhao et al. proposed a framework in Hybrid-Model-Based Deep Reinforcement Learning for Heating, Ventilation, and Air-Conditioning Control that uses a simplified, low-fidelity physical model purely as a safety "protector" and pre-training simulator, leaving the fine-grained, data-driven optimization to the DRL agent. This balances the safety guarantees of physical models with the scalability of data-driven control.


Differentiable Models and Structured Learning

Another facet of this research direction involves replacing traditional physics equations with neural networks or statistical structures that can be integrated directly into optimization pipelines.

In Modeling and Optimization of Complex Building Energy Systems with Deep Neural Networks, Yize Chen et al. demonstrated how deep Recurrent Neural Networks (RNNs) can capture complex building dynamics. Once the RNN is trained on historical data, its weights are "frozen," and controllers can perform gradient descent directly through the neural network to find the optimal control inputs.

[Control Inputs] ──> (Frozen RNN / ICNN) ──> [Predicted Energy/Comfort]
       ▲                                              │
       └─────────── [Gradient Descent Update] ────────┘

Standard deep neural networks, however, do not inherently respect mathematical properties like convexity, which can make the downstream optimization problem non-convex and computationally expensive to solve in real time. To bridge this gap, Felix Bünning et al. utilized Input Convex Neural Networks (ICNNs) in Input Convex Neural Networks for Building MPC. By constraining the network architecture such that the output is convex with respect to the control inputs, the authors guaranteed that the resulting optimization landscape remains convex, enabling fast, reliable, and scalable model-based control without bespoke physical modeling.

Key Foundational Work

Health Monitoring of an Auxiliary Power Unit Using a Classification Tree (2011) — contributed 1 concept elements to this direction.
Health Monitoring of a Pneumatic Valve Using a PIT Based Technique (2010) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_model_based_control_complexity_3fe34875,
  title = {model-based control complexity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/model-based-control-complexity/rdr_3fe34875768a4b3cb8db2a703b7808a9},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_model_based_control_complexity_3fe34875_v1,
  title = {model-based control complexity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/model-based-control-complexity/rdr_3fe34875768a4b3cb8db2a703b7808a9/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}

See something wrong? Report a mistake