Quality control of data pre-processing for improving prediction performance of ANN model based on CFD simulations: Effect of grid resolutions

With the development of artificial intelligence (AI), neural networks (NNs) have been employed to train extensive datasets for indoor airflow and heat transfer, aiming for fast and precise predictions. Yet, as AI models are based on volume-averaged results by reducing cells in computational fluid dynamics (CFD) to minimize training loads, only a few studies explore how discretized grid resolution affects AI model predictions. This study examines the influence of grid resolutions on indoor airflow and temperature distributions predicted by artificial neural network (ANN) models. CFD simulations were performed to establish a dataset for ANN training and testing. To evaluate the impact of data pre-processing on prediction quality, six grid resolutions were compared. Results indicate a significant influence of the grid scheme on prediction quality, revealing the failure of the cube grid to reduce prediction quality. However, the results show that the relative error increases as the number of cuboid grids increases, indicating a challenge for ANN prediction with an increased grid number. These findings underscore the importance of grid resolutions in improving prediction quality.