Variable taxi time prediction is the core of the Airport Collaborative Decision Making (A-CDM) system. An accurate taxi time prediction contributes to enhancing airport operational efficiency, safety and predictability. The deep dynamic spatio-temporal correlation inherent in airport traffic data is critical for taxi time prediction. However, existing machine learning (deep learning) methods have been unable to thoroughly exploit these correlations. To address this issue, we propose a deep learning-based model called the multi-task dynamic spatio-temporal graph attention network (MT-DSTGAN). Our model also predicts future entire airport traffic flow and taxiing segment traffic flow as auxiliary tasks, with the goal of enhancing the accuracy of aircrafts’ taxi time prediction. The proposed MT-DSTGAN model is implemented and assessed through a case study of Beijing Capital International Airport with a real-world dataset. The advantage of the proposed model, which shows better performance in various evaluation metrics, is demonstrated in a comparative study with other baseline works. In summary, the proposed MT-DSTGAN exhibits promising capabilities in perceiving the dynamic changes in the taxiing process of aircraft and demonstrates the ability to capture complex spatio-temporal correlations in airport traffic data.