A promising potential of Unmanned Aerial Vehicles (UAV) in 5G networks is to act as Aerial Base Stations (ABSs) that dynamically extend terrestrial base stations coverage without overloading the infrastructure. However, coverage extension faces crucial challenges such as user mobility and determining the best coordinates for new base station deployment. In this paper, we address this problem based on the prediction of users' spatial distribution that allows Aerial base stations (ABS) to adjust their position accordingly. We first analyze the performance of two machine learning schemes (Long Short Term Memory (LSTM)-based encoder-decoder and self-attention-based Transformer) for user mobility prediction based on a real DataSet. Then, we use these schemes to enhance the ABS deployment algorithm. Numerical results reveal significant gains when applying the proposed mobility prediction models over traditional deployment algorithms. In four hours of the day, both the Transformer and LSTM based models show, respectively, more than 31% and 22% gain in coverage rates compared to regular deployment schemes.