[Article] Effect of passive energy retrofitting strategies by using Artificial Neural Networks: a case of a hospital building

What if artificial intelligence could evaluate thousands of hospital retrofit combinations in seconds instead of requiring more than a year of building simulations?

Hospitals are among the most energy-intensive buildings. They operate continuously, accommodate diverse medical functions, use large quantities of equipment, and must maintain strict indoor environmental conditions. Improving their performance therefore requires more than adding insulation. Designers must simultaneously consider the building envelope, glazing, solar shading, indoor temperature settings, climate, thermal comfort, and operating patterns.

Our new study, published in Proceedings of the Institution of Civil Engineers - Energy, investigates how passive retrofitting strategies and artificial neural networks can be combined to improve the energy performance of a large hospital building.

The research focused on the 38,000 m² Balikesir University Hospital in Turkey, a six-story, 200-bed healthcare facility composed of seven blocks. We first developed and calibrated an EnergyPlus model using measured natural-gas consumption, electricity consumption, weather data, and indoor temperatures. The model was calibrated according to ASHRAE Guideline 14 before being used to test retrofit strategies.

We then examined five key design variables:

• external-wall insulation type and optimum thickness;

• roof insulation type and optimum thickness;

• glazing thermal performance;

• shading-device type and dimensions; and

• indoor temperature settings for heating and cooling.

Combining five alternatives for each variable created 3,125 possible configurations per season, or 6,250 heating and cooling combinations.

Running every combination through conventional simulation would have required approximately 520 days of computation on the computer used in the study.

This is where artificial intelligence changed the workflow.

Using a Taguchi orthogonal design, only 25 representative simulations were required for each season. Of these, 20 were used to train the artificial neural network and five to test it. Once validated, the ANN predicted the remaining combinations within seconds.

The results demonstrate that AI can serve as a reliable surrogate for computationally demanding building simulations.

• The ANN achieved regression coefficients of R = 0.99 for heating and R = 0.96 to 0.99 for cooling.

• The method reduced the required computational workload by approximately 98%.

• Under the measured indoor-temperature conditions, the optimized combinations produced energy savings of 18.66% during the heating period and 72.48% during the cooling period.

• A high-performance glazing option with a U-value of 0.7 W/m²K appeared in the optimal combinations for both seasons.

• The best heating and cooling solutions were not identical. Solar shading was avoided in the optimal heating configuration, while a 1 m overhang was included in the optimal cooling configuration.

• External-wall and roof insulation strongly influenced heating demand, but their effects on cooling demand were not statistically significant in this case.

• For cooling, glazing, shading, and indoor temperature settings were more influential than additional envelope insulation.

The findings reveal an important message: the best retrofit strategy changes with the season, climate, and operational objective.

More insulation is not automatically the best answer. Fixed shading may reduce cooling demand while limiting useful winter solar gains. A glazing system that performs well in one climate may not be optimal in another. Indoor temperature settings can also outweigh some physical envelope interventions, particularly during the cooling season.

For heating-dominated climates, permanent shading should therefore be assessed carefully. When feasible, movable shading can provide a better balance between winter solar access and summer protection.

The broader methodological lesson is equally important: artificial intelligence should not replace building physics. It should accelerate it.

The reliability of the ANN depended on a carefully constructed and calibrated energy model. Once that physical foundation had been established, the neural network made it possible to explore a much larger design space with far less computational effort.

This combined simulation and ANN workflow can support architects, engineers, facility managers, and public authorities in identifying effective retrofit packages for complex, continuously operated buildings. It also opens opportunities for rapid scenario testing across different building types, climates, glazing ratios, weather conditions, and movable-shading systems.

As pressure grows to decarbonize healthcare facilities, we should no longer ask only: Which retrofit measure saves energy?

We should also ask: Which combination performs best for this building, in this climate, during each season, while maintaining the indoor conditions required by its occupants?

Congratulations to Ismail Caner, Nadir Ilten, and Kadriye Ergün for this collaborative work.

📘 Full article: https://doi.org/10.1680/jener.23.00042

📚 Learn more about our research: https://www.sbd.uliege.be/

💻 Subscribe to my newsletter: https://lnkd.in/diTVT5eq

🅱️ Bilibili: b23.tv/bzjL3bn | 🎬 YouTube: https://lnkd.in/erHrfkNf

🌐 Explore previous posts and resources: https://www.shadyattia.org/

🔗 Follow all my professional links, including WeChat 🟩💬 微信: https://lnkd.in/eN3xZhhZ

#ArtificialIntelligence #ArtificialNeuralNetworks #BuildingEnergySimulation #EnergyRetrofit #PassiveDesign #HospitalBuildings #HealthcareFacilities #EnergyEfficiency #BuildingPerformance #EnergyPlus #DesignBuilder #BuildingScience #ThermalComfort #ClimateResponsiveDesign #Decarbonization #SustainableBuildings #UniversityOfLiege #SBDLab