P-Flash - A Machine Learning-based Model for Flashover Prediction using Recovered Temperature Data.

Area of Science:

• Fire Safety Engineering
• Computational Modeling
• Machine Learning

Background:

• Compartment fires pose significant risks to life and property.
• Accurate flashover prediction is crucial for effective firefighting strategies.
• Heat detector limitations, such as failure at high temperatures, complicate real-time fire monitoring.

Purpose of the Study:

• To develop a predictive model for flashover occurrence in compartment fires using Support Vector Regression (SVR).
• To address the challenge of heat detector temperature limitations in real-world fire scenarios.
• To provide firefighters with actionable, real-time information for enhanced situational awareness and safety.

Main Methods:

• Generated synthetic temperature data from 1000 simulation cases (8 million data points) for a multi-room compartment fire.
• Developed the P-Flash (Prediction model for Flashover occurrence) model utilizing heat detector data, including from adjacent spaces.
• Implemented sequence segmentation and learning from fitting techniques to overcome heat detector temperature limitations.

Main Results:

• The P-Flash model demonstrated reliable prediction capabilities for flashover events.
• Model performance achieved approximately 83% for current and 81% for future flashover prediction, accounting for heat detector failure at 150°C.
• The model can predict flashover even when data from all detectors is unavailable.

Conclusions:

• The data-driven P-Flash model offers a promising tool for real-time flashover prediction in compartment fires.
• This approach can significantly enhance firefighter situational awareness, operational effectiveness, and safety.
• The model's ability to handle incomplete data makes it valuable for practical firefighting applications.