Channel-Spatial Fusion Attention for Wind Field Prediction in High-Rise Building Fire Scenarios.

Area of Science:

• Computational fluid dynamics (CFD)
• Aerodynamics
• Artificial intelligence in fire safety engineering

Background:

• Traditional wind-field prediction methods for high-rise fires suffer from poor information fusion and feature representation.
• Accurate wind-field prediction is crucial for understanding fire spread and ensuring safety in tall buildings.

Purpose of the Study:

• To develop an advanced attention mechanism for enhancing wind-field prediction accuracy in high-rise building fires.
• To address limitations in current models regarding information fusion and feature representation under fire conditions.

Main Methods:

• Proposed an Adaptive Channel and Multi-Scale Spatial Fusion Attention Mechanism (CSFAM) for improved adaptive focusing and multi-scale integration.
• Utilized CFD-based scenario modeling to generate a dataset of 1050 wind-field distributions.
• Applied the CSFAM-enhanced multi-layer perceptron (MLP) for wind-field prediction.

Main Results:

• CSFAM-enhanced MLP achieved a mean squared error (MSE) of 0.0004 and a mean absolute error (MAE) of 0.0141.
• The model demonstrated a high coefficient of determination (R²) of 0.9766, outperforming existing methods.
• CSFAM significantly improved the capture of aerodynamic features like vortices induced by fires.

Conclusions:

• CSFAM markedly enhances wind-speed prediction accuracy in high-rise building fires.
• The mechanism improves the identification and expression of complex aerodynamic structures, leading to more robust predictions.
• This approach offers a significant advancement for fire safety engineering in tall structures.