Data-driven methods for building control — A review and promising future directions

Emilio T. Maddalena, Yingzhao Lian, Colin N. JonesControl 2019 · DOI

TL;DR

This paper reviews data-driven methods for controlling building HVAC systems, highlighting challenges like sensor noise and operational constraints. It proposes "learning model predictive control" (LMPC) as a promising direction, combining the predictive power of MPC with the adaptability of machine learning to optimize energy efficiency and thermal comfort.

What Problem Does This Paper Address?

Modern buildings consume a significant amount of energy, with HVAC systems being major contributors. Traditionally, these systems are controlled by rule-based systems or simple PID controllers, which are often suboptimal and struggle to adapt to dynamic factors like changing electricity prices or occupant behavior. While advanced control techniques like Model Predictive Control (MPC) offer better performance, they often require complex models and significant tuning, making them difficult to implement in real-world scenarios. This paper addresses the need for more intelligent, adaptive, and practical control strategies for HVAC systems that can leverage available data to improve energy efficiency and maintain occupant comfort without requiring extensive system overhauls.

Methods

The paper provides a comprehensive review of various data-driven control methodologies for HVAC systems. It categorizes these methods into those that process information offline and those that learn online. Key techniques discussed include:

  • Rule-Based Control (RBC): Traditional, often implemented with state machines, but lacks continuous adaptation.
  • Model Predictive Control (MPC): Uses a model of the system to predict future behavior and optimize control actions. The paper distinguishes between standard MPC (model updated offline) and Learning MPC (LMPC) where the model is updated online.
  • Reinforcement Learning (RL): Agents learn optimal control policies through trial and error, directly interacting with the environment.
  • Machine Learning for Prediction: Techniques like neural networks and decision trees are used for forecasting thermal loads, occupancy, and weather.

The authors propose four key attributes for an ideal HVAC control solution: Lifetime adaptability (A1), Safety (A2), Incorporation of exogenous signal forecasts and complex objectives (A3), and Scalability (A4). They evaluate existing methods against these criteria.

Datasets

The paper discusses numerous studies that utilize real-world building data for training and validation. Specific datasets are not explicitly detailed as a benchmark dataset for this review paper. However, the authors mention that data is typically collected from sensors within buildings, including temperature probes, occupancy detectors, luminosity sensors, and power meters. Challenges related to data quality, such as low resolution and missing samples, are highlighted.

Key Insights and Ideas

  • The "Ideal" HVAC Controller: The authors define four crucial attributes (adaptability, safety, forecasting integration, scalability) that a truly effective HVAC control system should possess. This framework provides a valuable lens for evaluating and comparing different control strategies.
  • LMPC as the Sweet Spot: Learning Model Predictive Control (LMPC) emerges as a particularly promising approach. It combines the predictive capabilities of MPC with the ability to adapt online, addressing the limitations of static models and the need for continuous improvement.
  • Interfacing with Existing Systems is Crucial: The paper emphasizes that new control strategies must be able to interface with already deployed industrial controllers rather than requiring complete system replacement. This is vital for practical adoption in the industry.
  • Data Quality is a Major Hurdle: Despite the abundance of sensors, poor sensor calibration and low-resolution data are significant challenges that hinder the effectiveness of data-driven methods.
  • Safety Guarantees are Paramount: For online learning methods, especially RL, ensuring safety (i.e., maintaining thermal comfort and avoiding constraint violations) during the learning process is a critical, yet often unaddressed, challenge.

Key Results

This paper is a review and does not present new experimental results. Instead, it synthesizes findings from numerous existing studies. The key "results" are the insights derived from comparing these studies against the proposed ideal attributes. For instance, the review indicates that:

  • RL excels in adaptability (A1) but often lacks safety guarantees (A2) and can suffer from poor scalability (A4) due to long training times.
  • RBC is straightforward to implement and scalable (A4), and can incorporate forecasts (A3), but lacks continuous adaptation (A1) and safety guarantees (A2).
  • MPC can handle complex objectives (A3) and is generally scalable (A4), but standard MPC lacks online adaptability (A1).
  • LMPC shows promise in addressing A1, A3, and A4, with ongoing research focusing on A2.

Limitations and Future Work

The authors acknowledge several limitations and suggest future research directions:

  • Safety in Online Learning: A significant gap exists in guaranteeing safety during the online learning process for methods like Reinforcement Learning. Future work should focus on developing safe RL techniques that can be integrated with control theory.
  • Model Uncertainty Quantification: While LMPC can adapt models, explicitly quantifying and managing model uncertainty is crucial for robust performance and safety.
  • Interfacing with Legacy Systems: Developing methods to effectively interface with existing, proprietary industrial controllers remains a challenge.
  • Experimental Validation: Many promising LMPC approaches require more extensive experimental validation on complex, real-world HVAC systems with multiple zones and thermal storage.
  • Scalability for Complex Systems: While LMPC shows promise, ensuring scalability for very large and complex HVAC systems with numerous actuators and zones is an ongoing research area.

The paper concludes by advocating for LMPC, particularly when combined with semi-parametric or non-parametric regression for identification, as a promising path forward that balances adaptability, predictive power, and the potential for practical implementation in existing buildings.

Summary updated

BibTeX
@inproceedings{maddalena2019data,
  author = {Emilio T. Maddalena and Yingzhao Lian and Colin N. Jones},
  title = {{Data-driven methods for building control — A review and promising future directions}},
  booktitle = {Control},
  year = {2019},
  doi = {10.1016/j.conengprac.2019.104211},
}
Figure