⚡ electrical engineering

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

This contribution proposes the use of constrained Bayesian optimization for the automatic tuning of hyperparameters in model predictive control for building load management, demonstrating significant electricity cost reductions compared to rule-based and manually tuned controllers while simultaneously highlighting the financial benefits of optimal program selection.

Original authors: Jiarui Yu, Jicheng Shi, Wenjie Xu, Colin N. Jones

Published 2026-05-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jiarui Yu, Jicheng Shi, Wenjie Xu, Colin N. Jones

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine your home's heating system is like a very smart, but slightly stubborn chef, trying to prepare a meal (keeping your house warm) while spending as little as possible on ingredients (electricity).

This article is about teaching this chef to shop and cook much more efficiently, especially when ingredient prices change hourly or daily.

Here is the breakdown of their approach using simple analogies:

1. The Problem: The "Confusing Menu"

In many regions, electricity is not available at a fixed price. It is like a restaurant menu where:

Time-of-use tariffs: Lunch is cheap, but dinner is expensive.
Power charges: If you order too many dishes at once (consume a lot of power in a 15-minute window), you get a massive "peak-hour" surcharge.
Fixed fees: You pay a monthly membership fee, regardless of what you eat.

Most people simply set their thermostat to a fixed temperature (like a rule-based controller). This is like the chef preparing the same amount of food at the same time every day, regardless of whether it is "lunchtime" or "dinnertime." In the end, they pay too much.

2. The Solution: The "Smart Chef" (MPC)

The authors use a system called Model Predictive Control (MPC). Imagine this as a chef who checks the weather forecast and the price menu for the next few days.

If electricity is cheap tonight, the chef preheats the oven (warms the house early).
If electricity is expensive tomorrow afternoon, the chef lets the house cool down a bit (but not too much) so they don't have to run the oven then.

The Catch: Even a smart chef must know how to cook. If the chef sets the temperature limits too tightly, the house gets cold (uncomfortable). If they set them too loosely, they waste money. This is called Hyperparameter Tuning. Normally, people have to guess these settings, which is slow and often leads to an inferior meal.

3. The Innovation: The "Auto-Tuning Robot" (CONFIG)

The authors developed a robot (using an algorithm called Constrained Bayesian Optimization or CONFIG) that automatically finds the perfect settings for the chef.

The Digital Twin: Before the robot touches the real house, it tests millions of different settings in a virtual simulation (a "Digital Twin") of a Belgian house.
The Goal: The robot tries to find the settings that lead to the lowest bill, but it has a strict rule: The house must never become too uncomfortable. It uses a "comfort value" (PMV) to ensure people do not freeze or overheat.
The Magic: Unlike other methods that might break the rules to save money, this robot is very cautious. It explores various settings safely to find the absolute best balance between saving money and having a cozy house.

4. The Hurdle: "Minimum Power"

The researchers encountered a tricky problem with heat pumps (the heating device). Even if you ask the pump to run at "1% power," it consumes a lot of electricity just to start the fan (like a car engine idling).

The Old Way: Trying to model this complex "idling" behavior makes the math so difficult that a supercomputer would be required to solve it in real time.
The New Way: They created a "Simplified MPC." It is like putting a filter on the chef's orders. If the chef tries to order a tiny amount of food (low power), the filter says: "No, that is too inefficient. Order either a full meal or nothing at all." This keeps the math simple enough to run on a normal home computer while still saving money.

5. The Results: The Savings

They tested this system in a real scenario with 12 different electricity contracts (different price menus).

Compared to the "dumb" thermostat: The automatically tuned system saved 26.9% on electricity bills.
Compared to the "human-tuned" smart system: Even compared to a smart system manually adjusted by a human expert, the automatically tuned robot saved an additional 17.46%.
Contract Selection: The system also determined which electricity contract was most suitable. By choosing the right "menu," they saved an additional 20.18% compared to the worst possible contract choice.

Summary

The article shows that you do not need a supercomputer or a PhD in engineering to save money on heating. By using a smart, automated "tuning robot" to adjust the settings of a standard heating controller and by choosing the right electricity contract, a homeowner can significantly lower their bills without ever having to freeze. The system handles the difficult math in the background and ensures the house remains comfortable while the wallet stays full.

Technical Summary: Automatic Tuning of a Building MPC Controller for Optimal Economic Costs

Problem Statement
The building sector is responsible for approximately 34% of global energy consumption. While Demand-Side Management (DSM) programs, such as time-of-use and real-time pricing, offer incentives for grid stability, they lead to complex, nonlinear price structures that challenge conventional rule-based controllers. Although Model Predictive Control (MPC) represents a leading solution for such dynamic environments, its performance under real-world conditions is often suboptimal due to a lack of systematic hyperparameter tuning. Conventional MPC design processes frequently fail to reconcile prediction and control objectives under model inaccuracies and environmental disturbances. Furthermore, existing controller tuning methods often lack the ability to handle black-box conditions (such as long-term thermal comfort metrics) or rely on manual reference models, limiting their scalability and effectiveness in minimizing electricity costs under various billing contracts.

Methodology
The authors propose a performance-oriented framework that automates the offline tuning of MPC hyperparameters via Constrained Bayesian Optimization (CONFIG). The methodology functions as follows:

Digital Twin & Modeling: The approach utilizes a digital twin (specifically the BOPTEST platform with a BESTEST test case for a hydronic heat pump) to simulate building operation. A linear, data-driven ARX model identifies the building's thermal dynamics from historical data, focusing on dominant input-output behaviors rather than creating an exact physical replica.
MPC Formulation & Handling MOP: The study addresses the inherent "Minimum Operating Power" (MOP) problem in heat pumps, where small, non-zero control signals trigger inefficient power spikes. To solve this without the computational burden of Mixed-Integer Quadratic Programming (MIQP), the authors introduce a simplified MPC. This approach applies a "masking" function to the control signal to eliminate small inputs that trigger minimum power operation, while maintaining the problem as a manageable Quadratic Programming (QP) formulation.
Optimization Framework (CONFIG): The tuning problem is formulated as a constrained black-box optimization.
- Objective: Minimization of total monthly electricity costs, comprising time-of-use energy charges, capacity tariffs (based on 15-minute peak power), and fixed fees.
- Constraints: Thermal comfort is ensured via a "Predicted Mean Vote" (PMV) constraint. Instead of embedding this nonlinear metric directly into the MPC, the framework uses a Cumulative Distribution Function (CDF) of PMV values over the simulation period. The constraint requires that 80% of the absolute PMV values during occupancy periods remain below 0.5.
- Algorithm: The CONFIG algorithm uses Gaussian Process surrogates to model the objective and constraint functions. It optimizes lower confidence bounds to balance exploration and exploitation, ensuring global optimality guarantees while handling black-box constraints without manual penalty term tuning.
Tuning Variables: The framework automatically tunes four key hyperparameters for the simplified MPC: lower temperature limits for occupied and unoccupied spaces ( $T_{lb}^{occupy}$ , $T_{lb}^{unoccupy}$ ), the maximum control value ( $u_{max}$ ), and the lower bound of the mask ( $u_{low}$ ).

Main Contributions

Automated Tuning via CONFIG: The article presents the application of the CONFIG algorithm for tuning MPC hyperparameters for DSM-based building control, bridging the gap in handling complex, nonlinear price structures and long-term comfort conditions simultaneously.
Simplified MPC for Residential Deployment: The authors propose a "simplified MPC" that approximates the MOP problem via a masking function. This enables the controller to be solved using free, open-source solvers (e.g., SCIP) on standard hardware, avoiding the computational infeasibility of MIQP formulations that frequently exhaust RAM on personal computers.
Contract Selection Guide: The framework is applied to evaluate 12 different Belgian electricity contracts, demonstrating that optimal contract selection combined with tuned control strategies yields significant savings.

Results
A case study involving a single-zone residential building in Brussels, Belgium, was conducted over four months (November to February) across 12 pricing scenarios.

Cost Reduction: The optimized simplified MPC reduced electricity costs by 26.90% compared to a rule-based baseline controller and by 17.46% compared to a manually tuned QP-MPC.
Performance vs. Complexity: The tuned simplified MPC achieved performance equivalent to, and in some months exceeding, that of a computationally intensive MIQP-MPC controller, yet with significantly lower solution times (e.g., 0.15 s vs. 0.80 s per step on a cluster for QP vs. MIQP with SCIP).
Comfort Compliance: All controllers, including the tuned MPC, met the comfort condition (80% of PMV values < 0.5). The tuned MPC achieved this through precise adjustment of temperature limits and the masking function without relaxing comfort requirements.
Contract Savings: The analysis of the 12 contracts revealed that selecting the optimal contract could reduce monthly bills by up to 20.18% compared to the worst-case contract. The study found that dynamic contracts with capacity tariffs (e.g., "ddd") generally performed best, although fixed-price contracts with low fixed fees (e.g., "sdd") were competitive depending on the month.

Significance and Claims
The article claims that the proposed framework offers a data-driven path to significant consumer savings by bridging the gap between complex DSM price structures and practical controller implementation. The authors emphasize the following:

Hyperparameter Tuning is Critical: The performance gains from automatic tuning (17.46% over manual tuning) are comparable to the gains provided by the predictive capability of MPC itself, underscoring the necessity of systematic tuning.
Scalability: By replacing complex MIQP formulations with a masked simplified MPC, the method enables the deployment of advanced control strategies on cost-effective residential hardware without requiring expensive commercial solvers.
Holistic Optimization: The framework successfully integrates long-term comfort conditions and complex billing structures (including capacity tariffs) into a single optimization run, delivering actionable insights for both control strategy and contract selection.

The authors conclude that while offline optimization effectively addresses current implementation challenges, future work should focus on extending the framework with online adaptation and safety mechanisms to enable continuous real-time improvements.