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

To improve the predictive accuracy of wind-field distributions during fires in high-rise buildings, this study targets the shortcomings of traditional prediction methods, including insufficient information fusion and dispersed feature representations under high-rise fire conditions. An efficient attention mechanism, termed Adaptive Channel and Multi-Scale Spatial Fusion Attention Mechanism (CSFAM), is proposed, which endows the model with enhanced adaptive focusing and multi-scale integration capabilities. CSFAM can account for environmental features across multiple dimensions to enable high-spatial-resolution wind-field reconstruction, thereby improving robustness and prediction accuracy in complex environments. To validate the effectiveness of CSFAM for predicting wind fields under high-rise-fire conditions, CFD-based scenario modeling was employed to generate a dataset of 1050 CFD-derived wind-field distributions across diverse inflow-wind and fire-source scenarios, partitioned into training, testing, and validation sets according to the fire-source size. When applying the CSFAM-enhanced multi-layer perceptron (MLP), the wind-field predictions achieved a mean squared error (MSE) of 0.0004, a mean absolute error (MAE) of 0.0141, and an R² of 0.9766, outperforming state-of-the-art methods. The results demonstrate that CSFAM plays a significant role in markedly improving wind-speed prediction accuracy during high-rise-building fires, and enhances the model's ability to identify and express vortex-like and other key aerodynamic features generated by the fire, thereby improving the capture of the complex nonlinear aerodynamic structures induced by fire.