PREDICTION PERFORMANCE OF AIRPORT TRAFFIC USING BILSTM AND CNN-BI-LSTM MODELS

Abstract

The COVID-19 pandemic has had a significant and enduring impact on the aviation industry, necessitating the accurate prediction of airport traffic. This study compares the predictive accuracy of biLSTM (Bidirectional Long Short-Term Memory) and CNN-biLSTM (Convolutional Neural Network-Bidirectional Long Short-Term Memory) models using various optimization techniques such as RMSProp, Stochastic Gradient Descent (SGD), Adam, Nadam, and Adamax. The evaluation is based on Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) indices. In the United States, the biLSTM model utilizing the Nadam optimizer achieved an MAPE score of 9.76%. On the other hand, the CNN-biLSTM model utilizing the Nadam optimizer demonstrated a slightly improved MAPE score of 9.62%. For Australia, the biLSTM model using the Nadam optimizer obtained an MAPE score of 31.52%. However, the CNN-biLSTM model employing the RMSprop optimizer had a marginally higher MAPE score of 33.33%. In Chile, the biLSTM model using the Adam optimizer obtained an MAPE score of 44.04%. Conversely, the CNN-biLSTM model using the RMSprop optimizer had a slightly higher MAPE score of 44.09%. Lastly, in Canada, the biLSTM model using the Nadam optimizer achieved a comparatively low MAPE score of 14.99%. Similarly, the CNN-biLSTM model utilizing the Adam optimizer demonstrated a slightly better MAPE score of 14.75%. These results highlight that the choice of optimization technique, model architecture, and balanced dataset can significantly influence the prediction accuracy of airport traffic.