Performance of neural network for indoor airflow prediction: Sensitivity towards weight initialization

Neural networks (NNs) have been proposed as a promising alternative for fast and accurate prediction of indoor airflow. NN training is great importance acquiring results, which essentially nonconvex optimization process through gradient descent-based algorithms. performance at certain solution dependent on the initial parameter values from random initialization, crucial to reliability evaluation model comparisons hyperparameter tuning. In this study, sensitivity airflow towards weight initialization revealed by clarifying two issues equivalence impact strategy. By reproducing non-isothermal airflows, numerical experiments were conducted various scenarios considering different strategies. For each scenario, following same convergence criteria, was repeated obtain multiple solutions concerning / validation errors temperature velocity predictions. The results indicate that modeling capability all are similar; while significant discrepancies among in generalization demonstrated. with sampling intervals larger than [-1, 1] more sensitive weights smaller intervals.