Mobile phone metadata is now widely used to extract socio-economic activity metrics for cities or regions at scale. Properly used, this data can provide unique insights into downstream tasks and business value. We propose a novel scalable method for clustering urban areas based on the spatio-temporal characteristics of mobile traffic data. The development of deep learning techniques makes deep time series clustering feasible. Our approach utilizes deep learning and contrastive learning methods, including contrastive clustering and spatio-temporal model, where the former provides discriminative clusters and the latter provides spatio-temporal correlation between neighboring regions, respectively. Moreover, a novel data augmentation method has been proposed to improve the generalization of the model, the model can be easily transferred to other cities.