Scaling Predictive Control via Data-Driven Modeling

This direction is split into other directions: model-based control complexity. Its last version stays readable and citable here.

Version 1 of 190 papers94 conceptsfrozen 2026-10-04

In classical engineering, controlling a complex system—whether a skyscraper’s HVAC or an aircraft’s auxiliary power unit—requires a "first-principles" model. This is a set of mathematical equations, derived from physics, that describes how the system responds to inputs. While highly accurate, these models suffer from a fatal flaw: they do not scale. Every building has a unique layout, every engine has different wear patterns, and every industrial process has its own quirks.

The research direction emerging from these papers focuses on bypassing this "modeling bottleneck" by using data-driven techniques to learn system dynamics directly from sensors, enabling advanced control and monitoring to be deployed at scale.

The Modeling Bottleneck

Traditional Model Predictive Control (MPC) is the gold standard for efficiency, but as Felix Bünning, Benjamin Huber, and Philipp Heer argue in Experimental demonstration of data predictive control for energy optimization and thermal comfort in buildings, the manual effort required to calibrate these models for individual structures is a primary barrier to adoption.

When a model must be custom-built for every single apartment or machine, the engineering cost often outweighs the energy savings. This has led researchers to explore Data Predictive Control (DPC) and Data-enabled Predictive Control (DeePC), which substitute rigid physical equations with flexible machine learning structures like Random Forests or Neural Networks.

Learning Physics from Residuals

Rather than abandoning physics entirely, many researchers use a "hybrid" approach to maintain reliability. In Health Monitoring of an Auxiliary Power Unit Using a Classification Tree, Wlamir Olivares Loesch Vianna and colleagues demonstrate how to use a coarse thermodynamic model to establish a baseline, then use a decision tree to analyze the "residuals"—the difference between the ideal physics and the messy reality of the sensor data.

This intuition—using data to fill the gaps where physics is too complex to model—is echoed by Neal Master and Claire Tomlin in their work on Learning-Based Model Predictive Control (LBMPC). By combining a simplified thermal model with real-time learning, they account for unpredictable factors like solar gain and occupant behavior that first-principles models often miss.

Solving the Data Hunger Problem

If the weakness of physics models is manual labor, the weakness of machine learning is "data hunger." Training a Deep Reinforcement Learning (DRL) agent from scratch for every new building would take months of data collection.

To solve this, researchers like Shichao Xu and Qi Zhu 0002 propose Transfer Learning for Building HVAC Control. The core idea is to decouple the general control strategy (how to save energy) from the building-specific physics (how fast this specific room cools down). Similarly, Di Wu introduces MetaEMS, a meta-learning framework that allows a controller to adapt to a new environment in a fraction of the time by leveraging experience from a diverse fleet of other buildings.

From Buildings to Vehicles

While much of this research is grounded in building energy management, the same principles are being applied to high-stakes mechanical systems.

  • Aviation: Brian M. de Silva and Jared Callaham use Dynamic Mode Decomposition to learn aircraft physics from flight data, automating sensor fault detection without hand-coded derivations.
  • Electric Vehicles: Alexander Winkler and Jakob Andert embed Deep Neural Networks into Moving Horizon Estimation to monitor motor temperatures, providing a faster and more robust solution than traditional lumped-parameter thermal networks.
  • Railways: Jing He uses CNN-BiLSTM models to identify wheel-rail forces, outperforming rigid mechanical simulations that struggle with real-world non-linearities.

Why It Works: The Shift in Perspective

The success of these methods stems from a shift in how we view "truth" in engineering. In the classical view, the truth is the equation. In the data-driven view, the truth is the sensor stream.

By treating the system as a "black box" or "gray box" that can be learned, researchers like João Paulo Pordeus Gomes and Takashi Yoneyama can develop Badness Indicators for valves or engines that adapt to the specific wear and tear of that unit. This move away from "one-size-fits-all" physics toward "unit-specific" data models is what finally allows advanced control and health monitoring to scale across diverse, real-world fleets.

Papers in this version

Showing 60 of 90 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_scaling_predictive_control_via_data_driv_e2202f36,
  title = {Scaling Predictive Control via Data-Driven Modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scaling-predictive-control-via-data-driven-modeling/rdr_e2202f36f8404d9697f6b8c9df252f5f},
  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_scaling_predictive_control_via_data_driv_e2202f36_v1,
  title = {Scaling Predictive Control via Data-Driven Modeling},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/scaling-predictive-control-via-data-driven-modeling/rdr_e2202f36f8404d9697f6b8c9df252f5f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}