Which price to pay? Auto-tuning building MPC controller for optimal economic cost

Jiarui Yu, Jicheng Shi, Wenjie Xu, Colin N. JonesarXiv 2025 · DOI · arXiv

Abstract

Model Predictive Control (MPC) is a premier strategy for energy-efficient building management, yet its practical adoption is hindered by the complexity of hyperparameter tuning. This paper introduces a framework to auto-tune building MPC hyperparameters using a constrained Bayesian optimization algorithm called CONFIG. By optimizing a simplified "Mask MPC" against real-world, complex electricity tariffs, the researchers achieved up to a 26.9% reduction in energy costs while maintaining strict occupant comfort. The results demonstrate that intelligent tuning is as critical to performance as the control strategy itself.

Introduction

While MPC is the "gold standard" for building management due to its predictive capabilities regarding weather and electricity pricing, it remains difficult to implement. Traditionally, engineers tune MPC hyperparameters—such as temperature setpoints or penalty weights—in an "open-loop" fashion. This approach often fails to account for real-world disturbances or the specific nuances of a building's electricity contract, which may include capacity tariffs (charges based on peak usage) and dynamic pricing.

This research addresses the challenge of automatically finding the optimal controller settings to minimize monthly bills without compromising occupant comfort. The proposed solution bridges the gap between theoretical control logic and practical economic performance.

Methodology

The authors propose a three-stage closed-loop design process to automate the tuning of building controllers:

  1. System Identification: A linear AutoRegressive model with eXogenous inputs (ARX) is used to model building dynamics. This choice prioritizes data efficiency and computational practicality for residential applications over more complex machine learning models.
  2. The "Mask MPC": To avoid the high computational overhead of Mixed-Integer Quadratic Programming (MIQP)—typically required to handle heat pump start-up costs—the researchers developed a "Mask MPC." This uses a simple Quadratic Programming (QP) structure but applies a "mask" (lower and upper bounds) to the control signal, allowing it to run efficiently on low-power hardware.
  3. CONFIG Algorithm: The core innovation is the use of the CONFIG (Constrained Efficient Global Optimization) algorithm. This Bayesian Optimization approach treats the building simulation as a "black box," iteratively testing hyperparameter sets to converge on a global optimum that balances electricity costs and comfort levels.
Overview of the closed-loop auto-tuning framework using Bayesian Optimization (CONFIG).

Experimental Setup

The study utilized the BOPTEST (Building Optimization Performance Tests) simulation platform, specifically the bestest_hydronic_heat_pump case representing a residential dwelling in Brussels, Belgium. For economic evaluation, the framework was tested against 12 real-world electricity contracts from the Flanders region (November 2023). These contracts provided a rigorous environment featuring dynamic pricing, day/night tariffs, and capacity tariffs.

Results and Discussion

Significant Economic Savings

The optimized Mask MPC reduced monthly electricity costs by up to 26.90% compared to a standard rule-based controller. Interestingly, the results showed that hyperparameter tuning contributes as much to cost reduction as the predictive ability of the MPC itself; a well-tuned simple controller can outperform an untuned advanced one.

Efficiency and Hardware Viability

The Mask MPC achieved economic performance comparable to a complex MIQP controller while being nearly 40% faster to simulate. In hardware tests, the Mask MPC remained responsive using free solvers, whereas the MIQP version often became unresponsive or required excessive solve times, making the Mask MPC a viable candidate for "edge" devices in typical households.

Contract Sensitivity and Comfort

The framework highlighted that the "best" controller settings are highly dependent on the specific electricity contract. For instance, settings for a flat rate differ significantly from those required for a capacity tariff where peak shaving is the priority. Furthermore, the authors successfully integrated Predicted Mean Vote (PMV)—a complex index of human thermal sensation—as a black-box constraint, ensuring comfort was maintained throughout the optimization process.

Conclusion

This research demonstrates that auto-tuning via Bayesian Optimization can significantly enhance the economic viability of building MPC. By simplifying the optimization structure through the "Mask MPC" and utilizing the CONFIG algorithm, advanced control becomes accessible for residential use. Future work will focus on online optimization, allowing controllers to self-tune in real-time based on live building data, and expanding the framework to handle multi-zone buildings and integrated renewable energy sources.

Summary updated

Abstract

Demand-side management (DSM) programs introduce complex pricing, requiring advanced control for cost minimization. Model Predictive Control (MPC) offers a solution but its performance hinges on appropriate hyperparameter tuning. We propose using Constrained Bayesian Optimization (CONFIG) to automate this process. In a case study, our optimized MPC reduced electricity costs by 26.90% compared to a rule-based controller and by 17.46% versus an manually tuned MPC. Analysis of real contracts further showed that optimal DSM program selection can lower monthly bills by up to 20.18%, demonstrating a data-driven path to significant consumer savings.
BibTeX
@misc{yu2025which,
  author = {Jiarui Yu and Jicheng Shi and Wenjie Xu and Colin N. Jones},
  title = {{Which price to pay? Auto-tuning building MPC controller for optimal economic cost}},
  year = {2025},
  eprint = {2501.10859},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2501.10859},
}
Figure