Spatiotemporal Prediction of Fire-Hazard Parameters in Commercial Buildings Considering Active Fire Protection Systems

When a fire breaks out in a large shopping mall or office building, the danger is not just the flames themselves, but the invisible cloud of superheated gas and thick smoke that follows. For people trying to escape, the most critical information is what the air feels like at their breathing level, roughly six feet off the ground. This is where temperature, the amount of oxygen left, the concentration of toxic gases, and how far one can see become matters of life and death. Modern buildings are equipped with active safety systems designed to fight these hazards: sprinklers that spray water to cool the air and mechanical fans that pull smoke out of the room. However, these systems do more than just suppress a fire; they fundamentally change how heat and smoke move through a space. They create a complex, shifting environment that makes it difficult for standard sensors to tell the true story of what is happening at the level where people are trying to run.

Researchers have long used a clever technology called distributed optical-fiber temperature sensing to monitor these fires. Imagine a long, thin glass thread running along the ceiling and down the walls, capable of measuring the temperature at every single point along its length. This creates a continuous, detailed map of heat that is far superior to having just a few scattered thermometers. The challenge arises when the building's safety systems kick in. The water from sprinklers and the air from exhaust fans cool the ceiling and walls where the sensors are, but they also stir up the air in ways that change the temperature at the breathing level in unpredictable ways. If a computer tries to guess the dangerous conditions at the breathing level based only on the temperature readings from the ceiling, it will often be wrong because the relationship between the ceiling and the floor has been altered by the safety systems.

A team of researchers from several Chinese universities and research institutes set out to solve this specific problem. They wanted to create a method that could take the temperature readings from these fiber-optic sensors, even while sprinklers and fans were running, and accurately predict the dangerous conditions at the breathing level. To do this, they did not rely on real-world fires, which are too dangerous and unpredictable to study in a controlled way. Instead, they built a massive digital library of fire scenarios using advanced computer simulations. They modeled commercial spaces of various sizes, from small shops to large atriums, and simulated fires of different intensities. They then ran these simulations three times for each scenario: once with no safety systems, once with only sprinklers, and once with only smoke exhaust, plus a final set where both systems worked together. This allowed them to see exactly how much the safety systems changed the temperature and the composition of the air.

The researchers found that the safety systems had a profound and measurable effect. When only sprinklers were active, the average temperature at the breathing level dropped by more than 30 degrees Celsius compared to an uncontrolled fire. When only smoke exhaust fans were used, the temperature dropped by about 10 degrees. But when both systems worked together, the temperature at the breathing level plummeted by nearly 60 degrees. This massive cooling effect meant that the old rules for guessing the air quality based on ceiling temperature were no longer valid. If a system tried to use the old rules during an active fire, it would significantly overestimate the danger or miss the true conditions entirely. The data showed that the relationship between the heat on the ceiling and the air at the breathing level had been completely rewritten by the water and the airflow.

To fix this, the team turned to machine learning, a type of computer intelligence that learns patterns from data. They trained several different computer models to act as a translator. The goal was to take the raw temperature data from the fiber-optic sensors on the ceiling and walls, along with information about the size of the room and the intensity of the fire, and output a corrected, accurate temperature for the breathing level. They tested four different types of learning algorithms. One of them, a graph neural network, proved to be the most effective. This specific model was particularly good at understanding the spatial layout of the sensors, recognizing how the temperature at one point on the ceiling relates to the temperature at the next point, and how the water from a sprinkler might cool a specific area differently than the air from a fan.

Once they had the best model for correcting the temperature, they used it to predict the other dangerous factors. Because heat, smoke, and gas are all physically linked, knowing the true temperature allows a computer to estimate the amount of carbon dioxide, the remaining oxygen, the density of the smoke, and the visibility. The results were striking. Before the correction, the computer's guesses for carbon dioxide levels were off by nearly 14 percent. After applying the new temperature correction, the error dropped to less than 5 percent. For visibility, which is crucial for finding an exit, the error rate fell from nearly 8 percent down to just under 3 percent. The model was able to accurately predict the density of the smoke and the amount of oxygen remaining, even in the chaotic environment created by the active safety systems.

The researchers then tested whether this new method would work in buildings they had never seen before. They applied their model to digital simulations of a small retail shop, a tall atrium, and a massive commercial hall, none of which were part of the original training data. The model held up remarkably well across all these different shapes and sizes. It maintained high accuracy in predicting the dangerous conditions, proving that the method is not just a solution for one specific room but a robust tool that can adapt to the diverse architecture of modern commercial buildings. This suggests that in a real emergency, a single fiber-optic cable could provide a complete, real-time picture of the safety conditions at the breathing level, even as sprinklers spray and fans roar.

The implications of this work are significant for how we keep large public spaces safe. Currently, fire safety systems are designed to detect a fire and trigger an alarm, but they do not typically provide a detailed, real-time map of the hazardous conditions at the level where people are evacuating. This new approach offers a way to generate that map using the sensors that are already being installed for temperature monitoring. By correcting the data to account for the active safety systems, building managers and firefighters could potentially see exactly where the air is becoming toxic, where the oxygen is running low, and where the smoke is blocking the view. This information could be used to guide people away from danger and toward safe exits, or to help firefighters understand the environment they are entering. While the study was conducted entirely through computer simulations, the results provide a strong foundation for developing a new generation of fire safety monitoring that is smart enough to understand the very systems designed to save lives.