Crowd flow prediction has significant implications for public safety, transportation resource scheduling, and urban transportation planning. This problem involves forecasting future crowd flow by analyzing spatio-temporal features from historical data. In this paper, we present an attention-based deep learning model called STA for city crowd flow prediction. By utilizing the attention mechanism, STA effectively extracts spatio-temporal features from historical data, with its effectiveness confirmed through the MCTP trend evaluation index. Through contrast learning, we demonstrate STA’s ability to effectively capture spatio-temporal dependencies in historical data. Extensive experiments on two real-world datasets highlight STA’s advancements. It exceeds the state-of-the-art baseline, demonstrating a performance improvement of 6%-12%.