High-Precision and Fast Prediction Method for Office Ventilation Based on POD and Deep Learning

The real-time optimization of indoor thermal comfort and ventilation efficiency in offices is limited by the high computational cost of traditional Computational Fluid Dynamics (CFD) simulations (single simulation taking hours to days). Furthermore, the strong nonlinearity and coupling characteristics of temperature fields further increase the prediction difficulty of surrogate models. This study proposes a two-stage CFD surrogate model that maps five-dimensional operating condition parameters (supply temperature, supply velocity and position (x, y, z)) to POD modal coefficients through a deep neural network, followed by linear reconstruction of the flow field based on POD theory. The main contributions are: (1) a multi-branch temperature network (weighted fusion architecture of main branch + auxiliary branch + residual connection); (2) a Temperature-Aware Attention Mechanism (generating adaptive attention weights in the 2D temperature modal coefficient space); (3) a combination of hierarchical regularization with an intelligent data augmentation strategy. Experiments based on 510 office CFD scenarios demonstrate that the model achieves a Mean Absolute Error (MAE) of 0.210 K for temperature field prediction (34.4% improvement compared to the baseline with MAE of 0.320 K) and 0.0075 m/s for velocity field prediction (24.2% improvement compared to the baseline with MAE of 0.0099 m/s); the coefficient of determination (R2) reaches 0.98 (temperature) and 0.92 (velocity), respectively. The single prediction time is approximately 0.0008 s, 3–4 orders of magnitude faster than traditional CFD. This model provides an effective approach for temperature field prediction in office ventilation scenarios and provides a practical framework for real-time control, optimization, and digital twin applications.