🤖 AI

Thermal transmittance prediction based on the application of artificial neural networks on heat flux method results

This paper investigates the use of various Artificial Neural Network architectures to predict heat flux from air temperature data, aiming to significantly reduce the measurement duration required for in-situ thermal transmittance (U-value) assessment of building envelopes under the ISO 9869-1 standard.

Original authors: Sanjin Gumbarević, Bojan Milovanović, Mergim Gaši, Marina Bagarić

Published 2026-05-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sanjin Gumbarević, Bojan Milovanović, Mergim Gaši, Marina Bagarić

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 you own an old house and want to make it energy-efficient. To do that, you need to know how well your walls keep heat inside. This "heat-leaking" ability is called the U-value.

Usually, to measure this, you have to stick a special sensor (a heat-flux sensor) on the wall and leave it there for days, maybe even two weeks. It's like waiting for a slow-cooking stew to be ready; you can't rush it, or the results are wrong. This takes a long time, costs money, and is annoying if you want to test many different walls in a building.

This paper asks a simple question: Can we use a "smart computer brain" (Artificial Neural Networks) to guess what the sensor would have measured, so we don't have to leave it on the wall for so long?

Here is how they tried to solve it, using some everyday analogies:

The Setup: The Weather and the Wall

The researchers put a sensor on a wall to measure three things:

How hot it is inside the room.
How cold it is outside.
How much heat is actually flowing through the wall.

They collected data every 10 minutes for about 3 days. Think of this as recording a "recipe" of how the wall reacts to the weather.

The "Smart Brain" (The AI)

They trained four different types of "smart brains" (Artificial Neural Networks) to learn this recipe. The goal was to teach the brain: "If the inside is this temperature and the outside is that temperature, how much heat should be flowing through the wall?"

They tested four different "brain" architectures:

The Simple Brain (MLP): A basic, straightforward thinker.
The Memory Brain 1 (LSTM): A thinker with a long-term memory, good at remembering what happened a while ago.
The Memory Brain 2 (GRU): A slightly different type of memory thinker, a bit more streamlined.
The Hybrid Brain (LSTM + GRU): A team where one part has long-term memory and the other has a different kind of memory, working together.

The Experiment: Splitting the Data

To see if these brains were actually learning, they split their data into two piles:

The Study Pile (Training): They showed the brain this data so it could learn the patterns.
The Test Pile (Validation): They hid this data from the brain and asked it to predict the heat flow based only on the temperatures.

They tried different splits: sometimes the brain studied 25% of the data and was tested on the rest; other times it studied 50% or 66%.

The Results: Who Won?

The paper found some interesting things:

More Data Helps the "Memory" Brains: When the brain had more data to study (50% or 66%), the Memory Brains (LSTM and GRU) did a great job. They could predict the heat flow very accurately, almost as if they had the sensor on the wall the whole time.
The Simple Brain Struggled: The Simple Brain (MLP) was okay when it had very little data to study, but it wasn't as reliable when the data got bigger. It's like a student who can memorize a short poem but gets confused by a whole book.
The Best Combo: The Hybrid Brain (LSTM + GRU) got the lowest error rate when it studied 66% of the data. However, the researchers noted that having 66% of the data for training might be "cheating" a little because the test pile was so small.
The Real Winner: When looking at a fair split (50% study, 50% test), the LSTM (Memory Brain 1) was the champion. It predicted the heat flow with an error of only about 5.3% compared to the real sensor.

The Catch: The "Surprise" Factor

The paper also found a limitation. The "smart brains" are like students who memorize the weather patterns they studied. If the weather suddenly changes in a way they've never seen before (a "surprise" temperature spike or drop), the brain gets confused and the prediction becomes unstable.

They showed a picture of another wall where the prediction worked fine until the weather did something unexpected, at which point the "smart brain" lost its way.

The Bottom Line

The paper concludes that using Artificial Intelligence to predict heat flow is a promising idea. If you can train the AI with a few days of data, you might be able to move your sensor to a new wall much sooner than the standard two weeks.

However, the method isn't perfect yet. It works best when the weather behaves predictably, and the researchers say they need to do more testing to make sure the "smart brain" doesn't get confused by sudden weather changes.

In short: They built a digital crystal ball that can guess how much heat a wall loses based on temperature. It works pretty well, but it still needs more practice to handle surprise weather.

Technical Summary: Thermal Transmittance Prediction via Artificial Neural Networks on Heat Flux Method Results

Problem Statement
Deep energy renovation of the existing building stock is a priority for the European Union, yet assessing the thermal transmittance (U-value) of existing building elements remains challenging. Existing structures often lack design documentation, and material degradation or moisture can alter thermal properties over time. While the Heat Flux Method (HFM), standardized by ISO 9869-1, is a common experimental approach for determining in-situ U-values, it suffers from a significant limitation: the duration of measurement. To meet standard criteria and ensure reliability, HFM tests often require up to two weeks, particularly when using dynamic data processing methods to minimize deviation from designed U-values. This extended timeframe hinders the efficient planning of renovation processes. Furthermore, deploying multiple sensors to test different elements simultaneously is often financially and operationally inefficient.

Methodology
The research proposes a method to reduce measurement duration by utilizing Artificial Neural Networks (ANN) to predict heat flux based on interior and exterior air temperatures. The core hypothesis is that once an ANN model is trained on a short segment of HFM data, it can predict the heat flux for the remainder of the period using only temperature inputs, allowing the physical sensor to be relocated to other measurement sites.

The study utilized experimental data collected from a multi-layer wall between February 22 and 26, 2019. The dataset comprised 490 entries recorded every 10 minutes, capturing internal temperature ( $T_i$ ), external temperature ( $T_e$ ), and heat flux ( $q$ ).

Four distinct ANN architectures were evaluated for their ability to predict heat flux:

MLP3: A Multilayer Perceptron with one hidden layer containing three neurons.
LSTM100: A Long Short-Term Memory network with 100 units in the hidden layer.
GRU100: A Gated Recurrent Unit network with 100 units in the hidden layer.
LSTMGRU100: A hybrid architecture combining 50 LSTM units and 50 GRU units across two hidden layers.

Each architecture was tested under three different training/validation data split ratios: 1/4, 1/2, and 2/3. The performance of the models was assessed using Root Mean Squared Error (RMSE), Mean Squared Error (MSE), and Mean Absolute Error (MAE) for heat flux prediction. Additionally, predicted U-values were calculated using the standard average method (ISO 9869-1) applied to the predicted heat flux and compared against the measured U-value (0.586 W/(m²K)).

Key Results
The analysis yielded promising results regarding the prediction of heat flux rates based on temperature inputs:

Architecture Performance: Recurrent Neural Network (RNN) architectures (LSTM and GRU) generally provided more reliable results when larger datasets (1/2 and 2/3 training ratios) were used for training. Conversely, the MLP architecture performed relatively better with smaller training datasets (1/4 ratio).
Best Performance: The most accurate prediction for the 1/2 training/validation ratio was achieved by the LSTM100 model, which yielded an RMSE of 1.68 and a relative difference of 5.31% compared to the measured U-value.
Hybrid Model: The LSTMGRU100 model achieved the lowest RMSE (1.22) at the 2/3 training ratio; however, the authors note this may be attributed to the larger proportion of training data rather than inherent architectural superiority.
Limitations: An additional analysis on a different wall revealed potential instability in predictions when boundary temperatures changed in ways not observed during the training period. This suggests that while the method is effective under stable conditions, it may struggle with unseen thermal boundary variations.

Significance and Claims
The paper claims that applying specific classes of ANNs to HFM results offers a viable pathway to reduce the time required for in-situ U-value measurements. By successfully predicting heat flux based on temperature data, the physical heat flux sensor could theoretically be relocated to other locations after a shorter training period, thereby accelerating the assessment of multiple building elements.

However, the authors maintain a modest stance regarding the method's maturity. They acknowledge that the stability of predictions over the long run, particularly with small datasets or under varying boundary conditions, requires further investigation. The study identifies the influence of unobserved boundary temperature changes as a primary limitation, suggesting that strict measuring conditions or regularization techniques may be necessary for more certain conclusions. Consequently, the paper positions these findings as a direction for further research rather than a finalized solution for immediate widespread deployment.