Predicting the performance of hybrid ventilation in buildings using a multivariate attention-based biLSTM Encoder-Decoder neural network