Abstract: To address the issues of high computational cost and long cycle time associated with traditional Computational Fluid Dynamics (CFD) in simulating indoor airflow patterns, a dual-branch feed-forward neural network based on classification algorithms was constructed. This network enables the rapid prediction of indoor thermophysical fields and thermal comfort zoning under heating conditions. In view of the distinct physical characteristics of velocity fields and temperature fields, a differentiated classification strategy was proposed: a binary classification (mainstream / low-velocity) was adopted for velocity, while a ternary classification (overcool / comfortable / overheat) was employed for temperature. The research results demonstrate that the average prediction accuracy of the velocity field reaches 95.1%, which can accurately identify mainstream regions and exhibits excellent scale robustness. Meanwhile, the prediction accuracy of the temperature field exceeds 95%, enabling the precise reproduction of the spatial distribution of different temperature zones. Based on this, the prediction of the velocity field by the three-classification was studied. The results show that the cost-benefit ratio of the binary classification is the highest, which can not only meet the design requirements but also avoid the increase of computational costs. While ensuring prediction accuracy, the proposed model can achieve rapid prediction of thermophysical fields, thereby providing a basis for the optimization of air conditioning supply and energy-saving design.

Key words: Dual-branch feed-forward neural network, Classification algorithm, Differentiated classification, Velocity/temperature field, Robustness