A dual-branch feed forward neural network framework for robust indoor airflow and thermal classification prediction

With the increasing demands for indoor thermal comfort and building energy efficiency, rapid prediction of airflow organization in air-conditioned rooms has become a critical research focus. In this paper, a dual-branch feed forward neural network (FFN) combined with classification algorithms is proposed to predict the velocity and temperature distribution of rooms. Handle the characteristics of majority classes through the main branch; extract the features of minority classes through parallel branches. Based on the air conditioned room model, a dataset is constructed using verified data (with an average deviation of 1%) and trained. This paper successively analyzes the classification accuracy of the mainstream area, the configuration quantity and the out-of-distribution classification performance of the model. The results show that when the sample size is sufficient (more than 20 groups), the prediction accuracy can reach up to 96%. Out-of-distribution prediction tests show that the prediction accuracy of the model for out-of-range data decreases by 15.4%. This value can be restored to an average of 95% by supplementing 5 groups of typical samples to optimize the training data distribution. In addition, this paper further explores the prediction of the temperature field by this method. When the ternary classification model is adopted, the prediction accuracy of the temperature field reaches 95%, and it can also achieve fine partitioning. This method provides technical support for the rapid assessment and energy-saving design of air-conditioned rooms.