Smarter HVAC Systems with AI

How Decision-Focused Learning is Revolutionizing Energy-Efficient Building Control.

Keywords

; ; ;

Published June 28, 2025 By EngiSphere Research Editors

In Brief

A recent research presents a novel HVAC management system that embeds neural network-based thermal models into optimization using decision-focused learning with stochastic smoothing, enabling smarter, more cost-effective, and comfort-aware building energy control.


In Depth

Controlling the temperature in buildings is no small task—especially when you want to keep everyone comfy while saving money and supporting the electrical grid. Could Artificial Intelligence (AI) help? That's exactly what a team of researchers explored in their latest study, and the results are (literally and figuratively).

Welcome to the future of Heating, Ventilation, and Air Conditioning (HVAC) management, where Neural Networks (NNs) and Decision-Focused Learning (DFL) join forces to cut costs, reduce energy use, and keep buildings cozy. Let’s dive into the tech behind this next-gen HVAC solution and what it means for smart energy systems!

Why HVAC Systems Matter for Energy Grids

Buildings are energy-hungry giants, consuming over 30% of global electricity, with HVAC systems gobbling up the lion’s share—about 65% of residential usage in Europe alone.

Because buildings heat up and cool down slowly, there's a golden opportunity: if we intelligently control HVAC timing, we can shift energy use to off-peak hours. This helps:

  • Reduce electricity bills
  • Avoid grid overload
  • Increase renewable energy usage

But to do this, we need models that can predict how buildings behave thermally—and that's where things get tricky…

Modeling Thermal Behavior: Physics vs. Data

There are two main ways to model how a building heats and cools:

  1. Physics-based models (white-box) – These simulate walls, insulation, windows, etc., but they require tons of details and are often inaccurate due to aging and installation quirks.
  2. Data-driven models (black-box) – These use historical data and machine learning to "learn" the patterns. Neural Networks (NNs) shine here because they can learn complex relationships with less manual effort.

This research uses a data-driven approach: NNs trained on real building data to model how temperature changes over time based on HVAC usage and outdoor conditions.

The Big Innovation: Decision-Focused Learning (DFL)

Usually, NNs are trained to predict temperature accurately. But here's the catch: even a small prediction error can lead to bad decisions, like cranking up heating at the wrong time.

Enter Decision-Focused Learning (DFL)!

Instead of just training the model to be accurate, DFL trains it to make better decisions. The focus is on outcomes—like comfort and cost—not just predictions.

How it works
  • The neural network becomes part of an optimization model that schedules HVAC usage.
  • The whole system is trained using feedback from real-world simulations, like actual energy usage and comfort levels.
  • A smart technique called Stochastic Smoothing (SS) is used to help the model learn, even though HVAC decisions involve tricky math with on/off switches (which are not easy to optimize).
Formulating the HVAC Brain: Optimization + AI

The authors turned HVAC control into a Mixed Integer Quadratic Program (MIQP)—a fancy math problem that can decide:

  • When to heat or cool
  • How much power to use
  • What temperature to target

But the secret sauce is this: they embedded a Neural Network as a constraint in the optimization problem. It acts as a thermal behavior predictor, telling the optimizer how the building will react.

Since ReLU-based NNs (those using the Rectified Linear Unit function) can be exactly written with mixed-integer equations, they can be plugged right into the MIQP model.

Tackling the Hard Stuff with Stochastic Smoothing

Usually, optimization models can't “talk back” to machine learning models because of discontinuities (those on/off binary switches we mentioned).

To fix that, the team used Stochastic Smoothing (SS):

  • They added a bit of randomness to NN parameters
  • This smoothed out the optimization landscape
  • Then, using a technique from Reinforcement Learning (called REINFORCE), they learned which parameters led to better decisions

This meant the neural network could get better at making cost-saving, comfort-friendly decisions—even if the math under the hood was complex.

Real-World Test: A 5-Zone Office in Denver, USA

To prove their method works, the team tested it on a realistic building model:

  • A 5-zone office floor (511 m²)
  • Each zone had its own heat pump
  • Simulated weather data from Denver
  • Simulated electricity prices that vary by time-of-day

They used the EnergyPlus simulator (an industry-standard tool) to test their strategies. Instead of training on all 365 days (which is a lot!), they used clustering to pick the most diverse weather days—smart move!

The Results Are In: SS-DFL Beats the Rest!

They compared several training methods:

MethodTotal Cost (incl. Comfort)Comfort ScoreTraining Time
Traditional ML (MSE only)HighPoorFast
QP Relaxation (easy math)MediumOKMedium
Fixed Binary (no randomness)GoodDecentSlow
Stochastic Smoothing (SS)LowestBestWorth it!

SS-trained models outperformed everything else. They had the:

  • Lowest electricity costs
  • Best occupant comfort
  • Smallest prediction error

Surprisingly, smaller neural networks trained with SS beat larger ones trained with traditional methods. That means you get better results without huge computing costs!

What’s Next? Future Prospects

This research opens up exciting doors for energy efficiency and smart buildings:

  1. Smarter Grid Participation - Buildings could automatically shift their HVAC usage to help stabilize the grid—especially when renewables fluctuate.
  2. Scalable HVAC AI - The method works even with non-differentiable simulations or real buildings. That means real-world deployment is possible without needing perfect models.
  3. Handling Uncertainty - Future versions might include chance constraints to manage uncertainty in weather forecasts and electricity prices.
  4. Faster Training with Parallelization - The authors suggest speeding things up using parallel computing—great for big buildings or cities!
Final Thoughts

This study is a perfect example of what happens when engineering meets AI in the real world. By focusing not just on predictions, but on decisions, this method creates practical, cost-saving, and sustainable building control systems.

Imagine a future where every building learns from its environment, optimizes its energy use, and even supports the grid—all thanks to smart algorithms and data-driven thinking. That future is already here, and it looks bright (and well-air-conditioned).


In Terms

Neural Network (NN) - A computer model that mimics the way our brains learn patterns—used here to predict how building temperatures change over time. - More about this concept in the article "Smarter Grids with Brains | How AI Is Supercharging Renewable Energy Microgrids".

HVAC (Heating, Ventilation, and Air Conditioning) - The system that controls indoor climate—heating in winter, cooling in summer, and keeping air fresh all year round. - More about this concept in the article "Vertical Greening Systems: The Green Revolution in Sustainable Buildings".

Optimization - A smart way to choose the best possible settings (like thermostat schedules) to reach a goal—like minimizing energy bills or maximizing comfort. - More about this concept in the article "Building Smarter, Greener | Optimizing Modular Construction Supply Chains with AI & Multi-Agent Systems".

Mixed-Integer Program (MIP) / Mixed-Integer Linear Program (MILP) / Mixed-Integer Quadratic Program (MIQP) - A math problem used in decision-making that includes both continuous (smooth) and discrete (on/off) choices—think of turning HVAC on/off and setting temperatures. - More about this concept in the article "Charging Ahead | Smarter Storage Systems for Electric Trucks!".

Decision-Focused Learning (DFL) - An AI training method that focuses on improving real-world decisions (like energy savings), not just making accurate predictions.

Stochastic Smoothing (SS) - A trick to help AI learn better by adding randomness during training, making it easier to figure out which decisions work best—even when the math gets messy.

ReLU (Rectified Linear Unit) - A simple mathematical function used in neural networks that says: "output zero if the input is negative, or pass it through if it's positive"—helps models learn faster.

Thermal Dynamics (of Buildings) - The way temperatures change inside a building based on heating, cooling, weather, and insulation—like how your room stays warm even after the heater turns off.

Mean Squared Error (MSE) - A measure of how far off predictions are from the real thing; smaller MSE means more accurate predictions.

Demand Response - A strategy where buildings adjust energy use (like shifting heating or cooling) to support the power grid and save money—especially during peak demand hours.

Simulator (e.g., EnergyPlus) - A virtual lab that mimics how a building behaves in real life—used to safely test HVAC strategies before applying them to actual buildings.


Source

Pietro Favaro, Jean-François Toubeau, François Vallée, Yury Dvorkin. Decision-Focused Learning for Neural Network-Constrained Optimization: Application to HVAC Management System. https://doi.org/10.48550/arXiv.2506.19717

From: University of Mons; Johns Hopkins University; IEEE.

© 2026 EngiSphere.com