The rapid expansion of high-rise buildings globally presents notable challenges for firefighting, as traditional methods are often ineffective at that altitude. To resolve this issue, unmanned aerial vehicles (UAVs) offer a promising solution, with tactics centered on glass curtain wall demolition to inject fire suppressants. However, this action drastically alters interior ventilation, potentially triggering flashover and a rapid transition to full-room fire involvement. In this study, the mechanisms and influencing factors of flashover, specifically those induced by glass curtain-wall demolition, are investigated through a series of meticulously designed full-scale experiments.

Experiments were conducted in a 3.0 m (length) × 7.0 m (width) × 2.5 m (height) steel compartment, thereby simulating a standard office or residential room. Additionally, the target glass curtain wall for demolition was simulated using controllable gypsum board opening. Further, fire loads were created using fir wood cribs (moisture content: ~12%), with quantities varying between 6, 12, 15, and 18 cribs. Additionally, a square n-heptane pool fire served as the ignition source. Demolition timing systematically varied between 630, 750, and 870 s, post-ignition, thereby creating six distinct test scenarios. A comprehensive data acquisition system comprising the following components was deployed: The strategically positioned thermocouple arrays (e.g., R1-R5) inside the compartment were used to capture the evolution of the three-dimensional temperature field, especially vertical thermal stratification; an oxygen sensor that monitored volume fraction changes at breathing height (1.5 m); thermal imaging cameras that recorded flame and smoke dynamics; and a high-precision balance that tracked combustible mass loss.

The findings revealed distinct fire development patterns after demolition in the ventilation-controlled regime. Particularly, the temperature rise rate of the hot smoke layer exhibited a characteristic dual-peak trend: "initial peak → decayed oscillation → secondary peak." The combustible mass-loss process was segmented into five stages: initial pyrolysis, accelerated pyrolysis, fluctuating stability, sudden increase, and decay. Furthermore, the indoor oxygen concentration demonstrated a complex seven-stage dynamic evolution: "rapid decrease → slow decrease → accelerated decrease → fluctuating decrease → local recovery → fluctuating increase → stable recovery." This evolution was governed by the interplay of combustion intensity and ventilation. A key finding pertains to the influence of fire load: increasing the load (from 12 to 18 cribs) linearly enhanced the maximum post-demolition temperature rise rate (from 18.7 C/s to 56.2 C/s) but nonlinearly shortened the flashover initiation time (~75 s earlier for a 50% load increase). Crucially, a critical load threshold was identified. Beyond this threshold (between 15 and 18 cribs), the sensitivity of flashover initiation time to load diminished (a reduction of only 19 s), indicating a shift to the oxygen-replenishment rate as the dominant control factor of flashover triggering. Demolition timing was equally a critical factor: delaying demolition (from 630 to 870 s) increased the accumulation of unburned pyrolyzates, which rapidly increased the post-demolition temperature (peak rate up from 13.9 C/s to 51.9 C/s, drastically reduced flashover initiation time (from 93 s to 22 s post-demolition), increased the maximum temperatures, and prolonged high-temperature duration. However, beyond a critical demolition time threshold (~750 s in this setup), the ventilation capacity of the opening became the limiting factor. The oxygen supply rate constrained further intensification, stabilizing the flashover time around a fixed value despite additional delay.

To the best of our knowledge, this study represents the first full-scale quantitative analysis of flashover behavior induced by glass curtain-wall demolition. The results definitively establish the profound, nonlinear influences of fire load and demolition timing, thereby identifying critical thresholds that control flashover dynamics. These results provide valuable insights into key parameters for designing realistic fire experiments related to structural demolition. More importantly, the findings offer a crucial scientific basis for refining fire safety strategies for high-rise buildings, optimizing tactical decision-making regarding UAV-driven window demolition operations, and improving risk assessment protocols.

An accurate measurement of the heat release rate (HRR) of a fire source is crucial for thoroughly understanding the fire evolution process. However, the commonly used oxygen consumption method requires expensive equipment, leading to high operational costs. According to the energy conservation principle, heat released per unit of time during material combustion is closely related to the increase of the room temperature. Machine learning methods have demonstrated considerable potential for exploring relationships between several independent and dependent variables, and attempts have been made to predict fire parameters based on temperature. Therefore, based on previous studies, this paper proposes a comprehensive machine learning framework to predict the HRR using temperature data as input. Furthermore, feature selection techniques are innovatively introduced to obtain key location temperatures to maximize HRR prediction accuracy.

First, fire scenarios with different parameters are simulated in an ISO 9705 room using the fire dynamics simulator (FDS) software. Temperature data at different locations are obtained by gridding thermocouples, and a fire database is constructed. Subsequently, feature selection using recursive feature elimination (RFE) algorithms based on least absolute shrinkage and selection operator (Lasso) and random forest (RF) is performed to obtain two different low-dimensional subsets from high-dimensional simulated temperature features. Control groups with the same number of features are established. Finally, the performance of three typical models, namely, linear regression (LR), K-nearest neighbors (KNN), and light gradient boosting machine (LightGBM), for predicting the HRR are compared using different feature subsets.

The results show that using the subset obtained by RFE based onRF, the LightGBM model demonstrates the lowest root mean square error (RMSE) and mean absolute error (MAE) values of 23.89 kW and 15.49 kW, respectively, indicating the least difference between its predicted and observed values. Furthermore, regarding the coefficient of determination, the LightGBM model reaches the highest value of 0.991 6, close to 1, thereby demonstrating its superior fitting capability. This is primarily due to the complex nonlinear relationship between the trained temperature dataset and HRR. Compared to the KNN and LR models, the histogram algorithm and the leaf-wise growth strategy with a depth limit enable the LightGBM model to give full play to its advantages. Tree-based models, such as LightGBM and XGBoost, can also be used for algorithm model-level improvements in the future. Additionally, deep learning models, such as multilayer fully connected and convolutional neural networks, can be utilized for fitting complex mappings. Compared with LightGBM models trained with manual feature subsets, RFE based on RF decreases the prediction errors (RMSE and MAE values decrease by 46.54% and 50.66%, respectively). The coefficient of determination also increases by 2.1%, validating that this feature selection method can considerably improve prediction accuracy over manual feature selection.

This paper proposes a comprehensive machine learning framework to obtain thermocouple temperature through the FDS fire simulation and then combines the feature selection and prediction models for HRR prediction, substantially improving prediction accuracy over manual feature selection. This comprehensive framework has theoretical and practical significance, thereby opening new pathes for HRR prediction. Subsequent research studies will construct additional comprehensive fire databases, increase the number of feature parameters, or explore algorithm combinations to improve HRR prediction.